x

\", \"files\": [{\"path\": \"a.js\", \"content\": \"const a = 1; // cut here';\nconst r2 = salvageArtifactArgs(cut2); console.log(JSON.stringify({ saved: r2?.saved, cut: r2?.cut, title: r2?.args.title, hasContent: \"content\" in (r2?.args ?? {}) }));\nconsole.log(\"none:\", salvageArtifactArgs('{\"id\": 3, \"files\": [{\"path\": \"x.js\", \"con'));\n",[],[307],"a4f50720-69ce-4fd5-9b75-f47ebfa4910b","router.tsx","01a0a661-f04a-733e-839a-ce09461a01dc",["D",1789497831502],"import { getCatalogStatus, getProviderArray, type ProviderInstance } from \"./inference-provider.ts\";\nimport { initDB, run, sql } from \"./db.ts\";\nimport { getInstanceRoster, getSavedModels, inflight, resetBreakers, resetProviderStats } from \"./router-core.ts\";\nimport { handleChatCompletions } from \"./router-openai.ts\";\n\nconst CORS = { \"Access-Control-Allow-Origin\": \"*\" };\nconst OPENAI_PATHS = new Set([\"/\", \"/chat/completions\", \"/v1/chat/completions\"]);\n// Item 3: /api/update, /api/reset and /api/reset-stats are all reachable, unauthenticated, on a publicly\n// known URL — confirmed exploitable this session (the user's own successful unauthenticated POST to\n// /api/update). Gating them behind OMNI_ADMIN_KEY is opt-in and backward-compatible: unset, every mutating\n// route behaves exactly as before (this is a real, live router the user is actively depending on — a\n// gate that broke it by default the moment this code shipped would be worse than the hole it closes).\n// Once set, callers must send it back as `X-Admin-Key` (or `?admin_key=`, for the plain-fetch snippets\n// this session already handed the user from a browser console with no header-setting ceremony).\nconst ADMIN_KEY = Deno.env.get(\"OMNI_ADMIN_KEY\") || \"\";\nfunction isAdmin(req: Request, url: URL): boolean {\n if (!ADMIN_KEY) return true; // gate is off unless explicitly configured\n // ui-settings.ts already has a \"router key\" field (s.routerKey) that it sends as `Authorization: Bearer\n // ` on /health and /api/reset — plumbing that pre-dated any server-side check actually reading it.\n // Recognizing that header here means the existing settings field starts working with zero UI changes.\n // X-Admin-Key / ?admin_key= are accepted too, for the plain-fetch console snippets this session already\n // handed the user, which don't bother constructing an Authorization header.\n const bearer = (req.headers.get(\"authorization\") || \"\").replace(/^Bearer\\s+/i, \"\");\n const supplied = bearer || req.headers.get(\"x-admin-key\") || url.searchParams.get(\"admin_key\") || \"\";\n return supplied === ADMIN_KEY;\n}\n\nasync function listUpstreamModels(p: ProviderInstance): Promise {\n const headers: Record = p.key ? { \"Authorization\": `Bearer ${p.key}` } : {};\n const res = await fetch(`${p.base}/models`, { headers });\n if (!res.ok) throw new Error(`HTTP ${res.status}`);\n const data = await res.json();\n return data.data || data;\n}\n\nexport default async function handler(req: Request): Promise {\n const url = new URL(req.url);\n\n if (req.method === \"OPTIONS\") {\n return new Response(null, {\n headers: {\n \"Access-Control-Allow-Origin\": \"*\",\n \"Access-Control-Allow-Headers\": \"*\",\n \"Access-Control-Allow-Methods\": \"POST, GET, OPTIONS\",\n },\n });\n }\n\n const init = await initDB();\n if (!init.ok) {\n return Response.json({ error: `DB init failed: ${init.error}` }, {\n status: 500,\n });\n }\n\n if (req.method === \"GET\") {\n if (url.pathname === \"/health\") {\n const live = getProviderArray();\n // instances/key_policy/auth: the settings panel's \"test / status\" table (ui-settings.ts routerStatus())\n // reads exactly this shape — it was previously missing entirely, so the table always rendered zero rows\n // regardless of how many providers were actually configured.\n const keyIssues = live.filter((p) => p.keyIssue).map((p) => ({ instance: p.name, env: p.keyEnv, issue: p.keyIssue }));\n return Response.json({\n status: live.length > 0 ? (keyIssues.length ? \"ok-with-key-issues\" : \"ok\") : \"degraded\",\n providers_configured: live.length,\n inflight_requests: inflight,\n key_issues: keyIssues, // top-level, not buried per-instance: a malformed secret is the single most common \"why is X never working\" answer\n key_policy: \"tiered:top+1retry,siblings,breadth;config-dead→tail\", // see orderTargets() + demotion in router-core.ts\n auth: !!ADMIN_KEY, // true once OMNI_ADMIN_KEY is set — gates /api/update, /api/reset, /api/reset-stats\n instances: await getInstanceRoster(),\n });\n }\n\n if (url.pathname === \"/api/providers\") {\n const savedModels = await getSavedModels();\n return Response.json(getCatalogStatus().map((p) => ({\n ...p,\n currentModel: savedModels[p.name] || p.fallbackModel,\n })));\n }\n\n if (url.pathname === \"/api/models\") {\n const pName = url.searchParams.get(\"provider\");\n const p = getProviderArray().find((x) => x.name === pName);\n if (!p) return Response.json({ error: \"Provider not found or no key\" }, { status: 404 });\n try {\n return Response.json(await listUpstreamModels(p));\n } catch (e: any) {\n return Response.json({ error: e.message, fallback: p.fallbackModel }, { status: 502 });\n }\n }\n\n // OpenAI-style aggregate list: one row per routable instance (its bound\n // model) plus every upstream id namespaced \":\", so a client can\n // pin vendor+model through the `model` field of chat.completions.\n if (url.pathname === \"/v1/models\") {\n const live = getProviderArray();\n const saved = await getSavedModels();\n const seen = new Set();\n const rows: Record[] = [];\n for (const p of live) {\n const base = p.name.split(\"#\")[0];\n for (const id of [base, p.vendor]) {\n if (seen.has(id)) continue;\n seen.add(id);\n rows.push({ id, object: \"model\", owned_by: p.vendor, bound: saved[base] || p.fallbackModel });\n }\n }\n const results = await Promise.all(\n live.filter((p) => p.slot === 0).map(async (p) => {\n try {\n const list = await listUpstreamModels(p);\n return Array.isArray(list)\n ? list.map((m: any) => `${p.vendor}:${typeof m === \"string\" ? m : m.id}`)\n : [];\n } catch {\n return [];\n }\n }),\n );\n for (const id of results.flat()) {\n if (seen.has(id)) continue;\n seen.add(id);\n rows.push({ id, object: \"model\", owned_by: id.split(\":\")[0] });\n }\n return Response.json({ object: \"list\", data: rows });\n }\n\n return new Response(\n `\n \n \n \n \n \n Ziggyware Router Matrix\n \n \n \n \n \n \n \n

\n \n \n \n `,\n { headers: { \"Content-Type\": \"text/html\" } },\n );\n } else if (req.method === \"POST\") {\n if (url.pathname === \"/api/update\") {\n if (!isAdmin(req, url)) return Response.json({ error: \"unauthorized\" }, { status: 401 });\n const { provider, model } = await req.json();\n if (typeof provider !== \"string\" || !provider) return Response.json({ error: \"provider required\" }, { status: 400 });\n // model:null (or \"\") clears the override so the row falls back to its catalog default — there was\n // previously no way to UNBIND, only to bind something else, which is how a dead \"inkling-small\" override\n // outlived the catalog fix underneath it.\n if (model == null || model === \"\") {\n const d = await run(sql`DELETE FROM omni_router_config WHERE provider = ${provider}`);\n if (!d.ok) return Response.json({ error: d.error }, { status: 500 });\n return Response.json({ success: true, cleared: d.value.rowsAffected });\n }\n const r = await run(\n sql`INSERT INTO omni_router_config (provider, selected_model)\n VALUES (${provider}, ${model})\n ON CONFLICT(provider)\n DO UPDATE SET selected_model = excluded.selected_model`,\n );\n if (!r.ok) return Response.json({ error: r.error }, { status: 500 });\n return Response.json({ success: true });\n }\n\n // Item 2: previously missing entirely — ui-settings.ts's \"clear cooldowns\" button has been calling\n // this exact path since it was written, silently swallowing the 404. Clears every open circuit\n // breaker (in-memory on this isolate, and the persisted cross-isolate copy) so a provider mid-cooldown\n // gets tried again on the very next request instead of waiting out BREAK_DURATION_MS.\n if (url.pathname === \"/api/reset\") {\n if (!isAdmin(req, url)) return Response.json({ error: \"unauthorized\" }, { status: 401 });\n const { cleared } = await resetBreakers();\n return Response.json({ success: true, cleared });\n }\n\n // Item 4: clears the durable provider_stats counters (requests/ok/failed/last_error) — distinct from\n // /api/reset above, which clears the live routing decision, not the historical numbers. Separated so\n // clearing a stuck cooldown doesn't also erase diagnostic history, and vice versa.\n if (url.pathname === \"/api/reset-stats\") {\n if (!isAdmin(req, url)) return Response.json({ error: \"unauthorized\" }, { status: 401 });\n await resetProviderStats();\n return Response.json({ success: true });\n }\n\n if (!OPENAI_PATHS.has(url.pathname)) return new Response(\"Not Found\", { status: 404 });\n let body: unknown;\n try {\n body = await req.json();\n } catch {\n return Response.json({ error: { message: \"Invalid JSON body\" } }, { status: 400, headers: CORS });\n }\n return await handleChatCompletions(body);\n }\n\n return new Response(\"Not Found\", { status: 404 });\n}",[],[314],"bb3daa37-44c4-47aa-895c-ec5da2924499","router-state.ts","01a0a661-f04a-733e-839a-c95f8b6e7b3b",["D",1789497831502],"// Routing state that must survive isolate recycling: instance cooldowns\n// (breaker / 429 / bad key), dead model ids, auto-resolved replacements, and\n// the UI's per-vendor model binding. One SELECT per request, write-through.\nimport { all, run, sql } from \"./db.ts\";\n\nexport type Cooldown = { until: number; reason: string; fails: number };\nexport type Stat = {\n req: number;\n ok: number;\n fail: number;\n last: number;\n model: string;\n err: string;\n};\ntype State = {\n cool: Record;\n dead: Record;\n resolved: Record;\n bound: Record;\n cap: Record;\n tps: Record;\n capin: Record;\n stat: Record;\n};\n\nconst MAX_FAILS = 2;\nexport const COOL = {\n breaker: 2_500,\n rate: 6_000,\n auth: 30_000,\n model: 36_000,\n cap: 90_000,\n};\n\nlet mem: State = {\n cool: {},\n dead: {},\n resolved: {},\n bound: {},\n cap: {},\n capin: {},\n stat: {},\n tps: {},\n};\nlet loadedAt = 0;\n\nexport async function loadState(force = false): Promise {\n if (!force && Date.now() - loadedAt < 2_000) return mem;\n const [st, cfg] = await Promise.all([\n all(\"omni_state\", sql`SELECT key, value FROM omni_state`),\n all(\n \"omni_router_config\",\n sql`SELECT provider, selected_model FROM omni_router_config`,\n ),\n ]);\n const next: State = {\n cool: {},\n dead: {},\n resolved: {},\n bound: {},\n cap: {},\n capin: {},\n stat: {},\n tps: {},\n };\n if (st.ok) {\n for (const r of st.value as { key: string; value: string }[]) {\n const [kind, ...rest] = r.key.split(\":\");\n const id = rest.join(\":\");\n try {\n if (kind === \"cool\") next.cool[id] = JSON.parse(r.value);\n else if (kind === \"dead\") next.dead[id] = Number(r.value);\n else if (kind === \"resolved\") next.resolved[id] = r.value;\n else if (kind === \"cap\") next.cap[id] = JSON.parse(r.value);\n else if (kind === \"stat\") next.stat[id] = JSON.parse(r.value);\n else if (kind === \"capin\") next.capin[id] = JSON.parse(r.value);\n else if (kind === \"tps\") next.tps[id] = JSON.parse(r.value);\n } catch { /* skip corrupt row */ }\n }\n }\n if (cfg.ok) {\n for (\n const r of cfg.value as { provider: string; selected_model: string }[]\n ) next.bound[r.provider] = r.selected_model;\n }\n mem = next;\n loadedAt = Date.now();\n return mem;\n}\n\nconst put = (key: string, value: string) =>\n run(\n sql`INSERT INTO omni_state (key, value, ts) VALUES (${key}, ${value}, ${Date.now()})\n ON CONFLICT(key) DO UPDATE SET value = excluded.value, ts = excluded.ts`,\n );\nconst del = (key: string) =>\n run(sql`DELETE FROM omni_state WHERE key = ${key}`);\n\nexport const isCool = (name: string) =>\n (mem.cool[name]?.until ?? 0) > Date.now();\nexport const coolInfo = (name: string): Cooldown | undefined => mem.cool[name];\nexport const isDead = (vendor: string, model: string) =>\n (mem.dead[`${vendor}:${model}`] ?? 0) > Date.now();\n// Per-INSTANCE model death (a key's daily quota for one model). Namespaced with \"@\" because slot 0's\n// instance name equals the vendor name — without the prefix, slot 0's quota would kill the model for every slot.\nexport const isDeadOn = (instance: string, model: string) =>\n (mem.dead[`@${instance}:${model}`] ?? 0) > Date.now();\nexport async function markDeadOn(\n instance: string,\n model: string,\n until: number,\n) {\n mem.dead[`@${instance}:${model}`] = until;\n await put(`dead:@${instance}:${model}`, String(until));\n}\nexport const resolvedFor = (vendor: string) => mem.resolved[vendor];\nexport const boundFor = (vendor: string) => mem.bound[vendor];\n/** learned per-model output ceiling (from \"Limit N\" size errors), or Infinity */\nexport const capFor = (vendor: string, model: string) => {\n const c = mem.cap[`${vendor}:${model}`];\n return c && c.until > Date.now() ? c.lim : Infinity;\n};\n/** learned per-model INPUT ceiling (ITPM / context) — a tier limit, so it lives 24 h */\n/** Learned output throughput per vendor:model (tokens/s, EMA over completions ≥ 64 tokens). Drives the deadline-derived\n * max_tokens cap: catalog constants (openrouter \"30 tps\") were cutting every build reply to ~1k tokens. */\nexport const tpsFor = (vendor: string, model: string): number | undefined =>\n mem.tps[`${vendor}:${model}`]?.v ?? mem.tps[`${vendor}:*`]?.v;\nexport async function noteTps(\n vendor: string,\n model: string,\n tokens: number,\n ms: number,\n) {\n if (!(tokens >= 64) || !(ms > 200)) return;\n const v = tokens / (ms / 1000);\n for (const key of [`${vendor}:${model}`, `${vendor}:*`]) {\n const prev = mem.tps[key];\n const next = prev\n ? { v: prev.v * 0.7 + v * 0.3, n: prev.n + 1 }\n : { v, n: 1 };\n mem.tps[key] = next;\n await put(`tps:${key}`, JSON.stringify(next));\n }\n}\nexport const capInFor = (vendor: string, model: string) => {\n const c = mem.capin[`${vendor}:${model}`] ?? mem.capin[`${vendor}:*`];\n return c && c.until > Date.now() ? c.lim : Infinity;\n};\nexport async function setCapIn(vendor: string, model: string, lim: number) {\n mem.capin[`${vendor}:${model}`] = { lim, until: Date.now() + 24 * 1000 };\n await put(\n `capin:${vendor}:${model}`,\n JSON.stringify(mem.capin[`${vendor}:${model}`]),\n );\n}\nexport async function setCap(vendor: string, model: string, lim: number) {\n mem.cap[`${vendor}:${model}`] = { lim, until: Date.now() + COOL.cap };\n await put(\n `cap:${vendor}:${model}`,\n JSON.stringify(mem.cap[`${vendor}:${model}`]),\n );\n}\n\n/** Transient failure: counts toward the breaker; opens after MAX_FAILS. */\nexport async function noteFailure(name: string, reason: string) {\n const c = mem.cool[name] ?? { until: 0, reason, fails: 0 };\n c.fails++;\n c.reason = reason;\n if (c.fails >= MAX_FAILS) {\n c.until = Date.now() + COOL.breaker;\n c.fails = 0;\n }\n mem.cool[name] = c;\n await put(`cool:${name}`, JSON.stringify(c));\n}\n/** Hard cooldown for a known duration (429 retry-after, 401/403 key/tier). */\nexport async function coolFor(name: string, ms: number, reason: string) {\n mem.cool[name] = { until: Date.now() + ms, reason, fails: 0 };\n await put(`cool:${name}`, JSON.stringify(mem.cool[name]));\n}\nexport async function noteSuccess(name: string) {\n if (!mem.cool[name]) return;\n delete mem.cool[name];\n await del(`cool:${name}`);\n}\n/** Per-instance counters: the only externally checkable evidence that a key slot is actually used. Written every attempt. */\nexport async function noteStat(\n name: string,\n ok: boolean,\n model: string,\n err = \"\",\n) {\n const s = mem.stat[name] ??\n { req: 0, ok: 0, fail: 0, last: 0, model: \"\", err: \"\" };\n s.req++;\n ok ? s.ok++ : s.fail++;\n s.last = Date.now();\n s.model = model;\n s.err = ok ? \"\" : err.slice(0, 120);\n mem.stat[name] = s;\n await put(`stat:${name}`, JSON.stringify(s));\n}\nexport const statFor = (name: string): Stat | undefined => mem.stat[name];\nexport async function markDead(\n vendor: string,\n model: string,\n until = Date.now() + COOL.model,\n) {\n mem.dead[`${vendor}:${model}`] = until;\n await put(`dead:${vendor}:${model}`, String(until));\n}\n/** Clear cooldowns (and per-instance model deaths) — admin reset. `name` = instance or vendor; empty = everything. */\nexport async function resetCooldowns(name = \"\") {\n const hit = (k: string) =>\n !name || k === name || k.split(\"#\")[0] === name || k.startsWith(name + \":\");\n const cool = Object.keys(mem.cool).filter(hit),\n dead = Object.keys(mem.dead).filter((k) =>\n !name || hit(k.replace(/^@/, \"\"))\n ),\n resolved = Object.keys(mem.resolved).filter(hit);\n for (const k of cool) {\n delete mem.cool[k];\n await del(`cool:${k}`);\n }\n for (const k of dead) {\n delete mem.dead[k];\n await del(`dead:${k}`);\n }\n for (const k of resolved) {\n delete mem.resolved[k];\n await del(`resolved:${k}`);\n } // a junk discovery (whisper, prompt-guard) must not survive a reset\n return { cool, dead, resolved };\n}\nexport async function setResolved(vendor: string, model: string) {\n mem.resolved[vendor] = model;\n await put(`resolved:${vendor}`, model);\n}\nexport async function bindModel(vendor: string, model: string) {\n mem.bound[vendor] = model;\n return run(\n sql`INSERT INTO omni_router_config (provider, selected_model) VALUES (${vendor}, ${model})\n ON CONFLICT(provider) DO UPDATE SET selected_model = excluded.selected_model`,\n );\n}\n\nexport async function unbind(vendor: string) {\n delete mem.bound[vendor];\n return run(sql`DELETE FROM omni_router_config WHERE provider = ${vendor}`);\n}\n\n/** ms until the first of these instances stops cooling; 0 if any is free now; Infinity if none listed. */\nexport function soonestCooldown(names: string[]): number {\n const now = Date.now();\n let best = Infinity;\n for (const n of names) {\n const u = mem.cool[n]?.until ?? 0;\n if (u <= now) return 0;\n best = Math.min(best, u - now);\n }\n return best;\n}\n\n/** Snapshot for /health and the UI. */\nexport function stateView() {\n const now = Date.now();\n return {\n cooling: Object.entries(mem.cool).filter(([, c]) => c.until > now).map((\n [name, c],\n ) => ({ name, reason: c.reason, ms_left: c.until - now })),\n dead_models: Object.entries(mem.dead).filter(([, t]) => t > now).map((\n [k],\n ) => k),\n resolved: mem.resolved,\n bound: mem.bound,\n caps: Object.fromEntries(\n Object.entries(mem.cap).filter(([, c]) => c.until > now).map((\n [k, c],\n ) => [k, c.lim]),\n ),\n caps_in: Object.fromEntries(\n Object.entries(mem.capin).filter(([, c]) => c.until > now).map((\n [k, c],\n ) => [k, c.lim]),\n ),\n tps: Object.fromEntries(\n Object.entries(mem.tps).filter(([k]) => !k.endsWith(\":*\")).map((\n [k, t],\n ) => [k, Math.round(t.v)]),\n ),\n stat: mem.stat,\n };\n}",[],[321],"da8e92a9-0a0c-4178-9e46-4453746be314","router-openai.ts","01a0a661-f04a-733e-839a-c4d3f0a470bb",["D",1789497831502],"// OpenAI chat.completions surface over routeInference. Streams are piped\n// byte-for-byte from upstream (real token streaming); failover happens only\n// before the first upstream byte, which is the only point it can happen.\nimport { circuitStates, inflight, type RouteHit, routeInference, type RouteOpts } from \"./router-core.ts\";\n\nconst CORS = { \"Access-Control-Allow-Origin\": \"*\" };\nconst json = (body: unknown, status = 200) =>\n new Response(JSON.stringify(body), { status, headers: { \"Content-Type\": \"application/json\", ...CORS } });\n\nconst sanitize = (p: RouteHit[\"provider\"]) => ({\n id: p.id, name: p.name, vendor: p.vendor, base: p.base, keyEnv: p.keyEnv, slot: p.slot, priority: p.priority, fallbackModel: p.fallbackModel,\n});\n\n// Trail/plan are capped here (not just at storage time — see app.tsx's slimEvents cap) so a pathological\n// worst case (10 full passes × ~20 providers) can't blow up the response payload itself; the last N entries\n// are what matters for a flowchart anyway — earlier passes of the same provider list are redundant detail.\nconst TRAIL_CAP = 60;\nconst meta = (hit: RouteHit) => ({\n provider: sanitize(hit.provider),\n model: hit.model,\n latency_ms: hit.latencyMs,\n inflight_requests: inflight,\n circuit_states: circuitStates(),\n plan: hit.plan ?? null,\n attempts: hit.trail ? hit.trail.slice(-TRAIL_CAP) : null,\n});\n\nexport async function handleChatCompletions(body: any): Promise {\n const messages = body?.messages ?? (body?.prompt ? [{ role: \"user\", content: body.prompt }] : []);\n if (!Array.isArray(messages) || !messages.length) return json({ error: { message: \"Empty message payload\" } }, 400);\n const opts: RouteOpts = {\n model: body.model, tools: body.tools, tool_choice: body.tool_choice,\n temperature: body.temperature, max_tokens: body.max_tokens, stream: !!body.stream,\n response_format: body.response_format,\n // Was silently dropped: the client (app-infer.ts callInference) always sends this, meaning every\n // per-provider attempt used the router's fixed default regardless of how much of the caller's own\n // turn budget was actually left — see HARD_TIMEOUT_MS's comment in router-core.ts.\n deadlineMs: typeof body.omni_deadline_ms === \"number\" && body.omni_deadline_ms > 0 ? body.omni_deadline_ms : undefined,\n };\n let hit: RouteHit;\n try {\n hit = await routeInference(messages, opts);\n } catch (e: any) {\n // Exhaustion still carries a full trail — a flowchart that shows \"everything tried, then a 503\" is\n // strictly more useful than one that goes dark the moment there's no winner to hang it off of.\n return json({ error: { message: e.message, details: e.errors ?? null, plan: e.plan ?? null, attempts: e.trail ? e.trail.slice(-TRAIL_CAP) : null } }, e.status ?? 503);\n }\n const m = meta(hit);\n\n if (!opts.stream) {\n const up = await hit.res.json();\n return json({ ...up, model: up.model ?? hit.model, _omni_meta: m });\n }\n\n // SSE passthrough. Prepend one comment line carrying _omni_meta so clients\n // that want provenance can read it without parsing chunks; OpenAI SDKs\n // ignore \":\"-prefixed lines per the SSE spec.\n const enc = new TextEncoder();\n const head = new ReadableStream({\n start(c) { c.enqueue(enc.encode(`: omni ${JSON.stringify(m)}\\n\\n`)); c.close(); },\n });\n const upstream = hit.res.body ?? new ReadableStream({ start(c) { c.close(); } });\n const out = new ReadableStream({\n async start(c) {\n for (const s of [head, upstream]) {\n const r = s.getReader();\n for (;;) { const { done, value } = await r.read(); if (done) break; c.enqueue(value); }\n }\n c.close();\n },\n cancel() { hit.res.body?.cancel().catch(() => {}); },\n });\n return new Response(out, {\n headers: { \"Content-Type\": \"text/event-stream\", \"Cache-Control\": \"no-cache\", \"X-Omni-Provider\": hit.provider.name, \"X-Omni-Model\": hit.model, ...CORS },\n });\n}\n",[],[328],"d1a4ddf6-1a55-46e8-a780-90cb9db30975","router-discover.ts","01a0a661-f04a-733e-839a-c336ab583905",["D",1789497831502],"import type { ProviderInstance } from \"./inference-provider.ts\";\nimport { isDead } from \"./router-state.ts\";\n\nexport function headersFor(p: ProviderInstance): Record {\n const h: Record = { \"Content-Type\": \"application/json\", \"HTTP-Referer\": \"https://val.town\", \"X-Title\": \"omni\", ...p.headers };\n if (p.key) h.Authorization = `Bearer ${p.key}`;\n return h;\n}\n\n\n/** Live /models lookup → first id matching the vendor's `prefer` regex that isn't dead. */\nexport async function discover(p: ProviderInstance): Promise {\n if (!p.prefer) return null;\n try {\n const res = await fetch(`${p.base}/models`, { headers: headersFor(p) });\n if (!res.ok) return null;\n const data = await res.json();\n const ids: string[] = (data.data ?? data).map((m: any) => (typeof m === \"string\" ? m : m.id)).filter(Boolean);\n const rx = new RegExp(p.prefer, \"i\");\n // never \"discover\" non-chat endpoints: guard/moderation, ASR, TTS, embeddings, rerankers\n const JUNK = /guard|safeguard|moderat|whisper|tts|speech|embed|rerank|playai|vision-only|ocr|\\bstt\\b/i;\n return ids.find((id) => rx.test(id) && !JUNK.test(id) && !isDead(p.vendor, id)) ?? null;\n } catch { return null; }\n}\n\n",[],[335],"9248a3da-bd64-47e5-a671-ceb51ebbadbc","router-core.ts","01a0a661-f04a-733e-839a-bdd6128ed09f",["D",1789497831502],"import { getProviderArray, type ProviderInstance, resolveTargets } from \"./inference-provider.ts\";\nimport { all, bumpProviderStats, providerStatsAll, run, sql } from \"./db.ts\";\n\nexport const MAX_CONCURRENCY = 3;\nconst MAX_FAILS = 2;\nconst BREAK_DURATION_MS = 25_000;\n// Per-attempt ceiling when no caller deadline narrows it further. Was 120_000, then cut to 15_000 to stop\n// one stalling vendor from eating the whole response window — but 15s proved too short for what this app\n// actually sends: live evidence (a Groq 413) confirmed prompts here run 18k+ tokens with up to 16k\n// max_tokens requested, and a live /health comparison showed Anthropic's own success rate fall from 6/7 to\n// 6/23 (nearly all new failures API_STALLED_TIMEOUT) once attempts were being cut off at 15s — a real,\n// legitimately-completing generation for this workload routinely needs more than that. 30s is a genuine\n// tradeoff against breadth (fewer distinct providers fit in one invocation's ~50s safety ceiling), chosen\n// because a provider that's actually working and just needs time is worth far more than three that were\n// never going to succeed anyway getting tried instead.\nexport const HARD_TIMEOUT_MS = 30_000;\nconst RETRYABLE = new Set([408, 409, 425, 429, 500, 502, 503, 504]);\n\nexport let inflight = 0;\n// FIFO admission instead of the previous `while (inflight >= MAX) await sleep(10)` spin: polling burned a\n// timer tick per 10ms per waiter and — worse — admitted waiters in whatever order their timers happened\n// to fire, so under contention a request that arrived later could jump one that had been waiting.\n// Waiters now resolve strictly in arrival order, with zero timers while blocked.\nconst waiters: (() => void)[] = [];\nasync function acquire(): Promise {\n if (inflight < MAX_CONCURRENCY) { inflight++; return; }\n await new Promise((resolve) => waiters.push(resolve));\n inflight++;\n}\nfunction release(): void {\n inflight--;\n const next = waiters.shift();\n if (next) next();\n}\n\n// Config-dead demotion. The breaker deliberately never trips on non-retryable 4xx (a bad key or model id\n// must not oscillate it) — correct, but the consequence, visible in every trail this session, was that a\n// provider returning 404 \"Function not found\", 403 tier_not_allowed or 402 quota-exhausted got a full,\n// fresh attempt on EVERY pass of EVERY request, forever, each one costing a round-trip before the next\n// candidate. Demotion is the missing middle ground: after CFG_DEAD_AFTER consecutive config-class\n// failures the instance is moved to the END of the order (still tried — fail-open is preserved, nothing\n// is ever excluded), for CFG_DEAD_MS, persisted cross-isolate. One success clears it immediately.\nconst CFG_DEAD_AFTER = 2;\nconst CFG_DEAD_MS = 10 * 60_000;\nconst CFG_DEAD_STATUSES = new Set([401, 402, 403, 404]);\n// A 429 is retryable in general, but a 429 whose body says the QUOTA is gone (daily free cap, billing\n// hard-limit) is not a transient rate limit — it's config-dead until the window resets. Observed live on\n// SambaNova / HuggingFace / OpenAI this session.\n// Adversarial case that shaped this regex: Groq's TRANSIENT per-minute 429 reads \"Rate limit reached for\n// model … on tokens per minute (TPM) … Please try again in 2.5s\" — a naive \"limit reached\" match would\n// have demoted a perfectly healthy provider for 10 minutes on a 2-second backoff. So: any explicit\n// short-retry / per-minute phrasing wins outright as transient, and only unmistakably daily/billing\n// wording counts as exhausted.\nconst QUOTA_TRANSIENT = /try again in|per minute|per second|\\bTPM\\b|\\bRPM\\b|\\bTPS\\b/i;\nconst QUOTA_EXHAUSTED = /quota|billing|insufficient|credit|exceeded your|per.?day|daily|free-models-per-day|hard limit/i;\nfunction isConfigClass(status: number | undefined, body: string): boolean {\n if (status === undefined) return false;\n if (CFG_DEAD_STATUSES.has(status)) return true;\n return status === 429 && !QUOTA_TRANSIENT.test(body) && QUOTA_EXHAUSTED.test(body);\n}\n\ntype Breaker = {\n failCount: number; state: \"CLOSED\" | \"OPEN\" | \"HALF_OPEN\"; nextAttempt: number;\n requests: number; ok: number; failed: number; lastError: string | null; // counters exist solely for /health's instance roster\n cfgFails: number; demotedUntil: number; demotedReason: string | null;\n fromDb: boolean; // true ⇒ this isolate's OPEN/demoted state was learned from omni_state, not observed here — so a later re-hydrate may retract it\n};\nconst breakers = new Map();\nexport function getBreaker(name: string): Breaker {\n let b = breakers.get(name);\n if (!b) breakers.set(name, b = { failCount: 0, state: \"CLOSED\", nextAttempt: 0, requests: 0, ok: 0, failed: 0, lastError: null, cfgFails: 0, demotedUntil: 0, demotedReason: null, fromDb: false });\n return b;\n}\nexport const isDemoted = (name: string) => getBreaker(name).demotedUntil > Date.now();\nfunction admit(name: string) {\n const b = getBreaker(name);\n if (b.state === \"OPEN\" && Date.now() >= b.nextAttempt) b.state = \"HALF_OPEN\";\n return b.state !== \"OPEN\";\n}\n// `trip`=false records the attempt (roster counters) without touching breaker state — used for 4xx config\n// errors, which fail over but must never oscillate the breaker on a bad key/model id (see `attempt` below).\n// The breaker's open/closed decision stays in-memory-only (per-isolate — no DB read added to the hot\n// path); the counters are additionally persisted (best-effort, never blocks or fails the actual routing\n// decision) so /health reflects reality across isolates instead of whichever one happens to answer a\n// GET. bumpProviderStats failing silently is acceptable: it's diagnostics, not routing logic.\n// Breaker/demotion persistence is written ONLY on a transition (closed→open, open→closed, demoted↔not) —\n// the first version of this wrote on every report(), one extra DB round-trip per attempt for a value that\n// almost never changed. The stats bump is unavoidable (it IS the per-attempt counter); this one wasn't.\nasync function report(name: string, ok: boolean, error?: string, trip = true, configClass = false) {\n const b = getBreaker(name);\n const wasOpen = b.state === \"OPEN\", wasDemoted = b.demotedUntil > Date.now();\n b.fromDb = false; // whatever happens below was observed HERE — local evidence outranks anything hydrated\n b.requests++;\n if (ok) { b.ok++; b.failCount = 0; b.state = \"CLOSED\"; b.cfgFails = 0; b.demotedUntil = 0; b.demotedReason = null; }\n else {\n b.failed++;\n if (error) b.lastError = error;\n if (trip && (++b.failCount >= MAX_FAILS || b.state === \"HALF_OPEN\")) {\n b.state = \"OPEN\"; b.nextAttempt = Date.now() + BREAK_DURATION_MS; b.failCount = 0;\n }\n if (configClass) {\n if (++b.cfgFails >= CFG_DEAD_AFTER) { b.demotedUntil = Date.now() + CFG_DEAD_MS; b.demotedReason = (error ?? \"config error\").slice(0, 160); }\n } else b.cfgFails = 0; // a transient failure between config errors resets the streak — demotion needs a CONSISTENT signal\n }\n await bumpProviderStats(name, ok, error ?? null, Date.now()).catch(() => {});\n const isOpen = b.state === \"OPEN\", nowDemoted = b.demotedUntil > Date.now();\n if (isOpen !== wasOpen || nowDemoted !== wasDemoted) await persistBreaker(name, b).catch(() => {});\n}\nexport const circuitStates = () => getProviderArray().map((p) => ({ name: p.name, state: getBreaker(p.name).state }));\n\n// ────────────────────────────────────────────────────────────────\n// Breaker cross-isolate persistence (item 5). The in-memory `breakers` Map above still governs the\n// admit()/report() decision on THIS isolate with zero added DB read on the hot path — that design choice\n// (see the comment on `report`) is unchanged. What was missing: a fresh isolate started moments after\n// another isolate opened a breaker had no way to know, so /health's `cooling` field and a brand-new\n// isolate's own routing could both treat a genuinely-tripped provider as available again. omni_state\n// (already reserved in db.ts's schema comment for exactly this: \"router cooldowns / dead models\") now\n// carries each OPEN breaker's state so a new isolate can hydrate it once, cheaply, instead of per-attempt.\nfunction persistBreaker(name: string, b: Breaker): Promise {\n const key = `breaker:${name}`;\n const now = Date.now();\n const demoted = b.demotedUntil > now;\n if (b.state === \"OPEN\" || demoted) {\n const value = JSON.stringify({\n state: b.state, nextAttempt: b.nextAttempt, lastError: b.lastError,\n demotedUntil: demoted ? b.demotedUntil : 0, demotedReason: demoted ? b.demotedReason : null,\n });\n return run(sql`INSERT INTO omni_state (key, value, ts) VALUES (${key}, ${value}, ${now})\n ON CONFLICT(key) DO UPDATE SET value = excluded.value, ts = excluded.ts`);\n }\n return run(sql`DELETE FROM omni_state WHERE key = ${key}`);\n}\nlet hydratedAt = 0;\n// Re-read at most once per HYDRATE_TTL_MS. The first version hydrated ONCE per isolate and then never\n// again, which live trails proved wrong within the hour: message #747 (06:03:04) tried nemotron-ultra,\n// cerebras and sambanova first — all three already demoted by a sibling isolate at 06:02:24 — because\n// this isolate had hydrated earlier and considered itself done. A SELECT on a tiny table every few\n// seconds is nothing next to the 2–30s LLM call it precedes. Persisted rows are applied set-only;\n// entries this isolate only ever learned from the DB (fromDb) are retracted when their row disappears\n// (another isolate saw a success and cleared it). Locally observed state is never overwritten.\nconst HYDRATE_TTL_MS = 3_000;\nasync function hydrateBreakers(): Promise {\n const now = Date.now();\n if (now - hydratedAt < HYDRATE_TTL_MS) return;\n hydratedAt = now;\n const r = await all(\"omni_state\", sql`SELECT * FROM omni_state WHERE key LIKE 'breaker:%'`);\n if (!r.ok) return;\n const seen = new Set();\n for (const row of r.value as any[]) {\n try {\n const data = JSON.parse(row.value);\n const name = row.key.slice(\"breaker:\".length);\n seen.add(name);\n const b = getBreaker(name);\n const localOpen = b.state === \"OPEN\" && !b.fromDb, localDemoted = b.demotedUntil > now && !b.fromDb;\n if (!localOpen && data.state === \"OPEN\" && typeof data.nextAttempt === \"number\" && data.nextAttempt > now) {\n b.state = \"OPEN\"; b.nextAttempt = data.nextAttempt; b.lastError = data.lastError ?? b.lastError; b.fromDb = true;\n }\n if (!localDemoted && typeof data.demotedUntil === \"number\" && data.demotedUntil > now) {\n b.demotedUntil = data.demotedUntil; b.demotedReason = data.demotedReason ?? null; b.fromDb = true;\n }\n } catch { /* malformed row — skip it, never let a bad persisted entry block routing */ }\n }\n for (const [name, b] of breakers) {\n if (b.fromDb && !seen.has(name)) { // retracted elsewhere\n if (b.state === \"OPEN\") b.state = \"CLOSED\";\n b.demotedUntil = 0; b.demotedReason = null; b.fromDb = false;\n }\n }\n}\n// Item 2: the \"clear cooldowns\" button in ui-settings.ts (routerReset()) has always called a POST\n// /api/reset that never existed anywhere in router.tsx — silently no-op'd behind a 404 the UI never\n// surfaced. This is the function that route now actually backs: clears every in-memory breaker on THIS\n// isolate and, best-effort, the persisted OPEN state so other isolates don't re-hydrate the cleared\n// breakers back open on their next call.\nexport async function resetBreakers(): Promise<{ cleared: number }> {\n const cleared = breakers.size;\n breakers.clear();\n const r = await all(\"omni_state\", sql`SELECT key FROM omni_state WHERE key LIKE 'breaker:%'`);\n if (r.ok) for (const row of r.value as any[]) await run(sql`DELETE FROM omni_state WHERE key = ${row.key}`).catch(() => {});\n hydratedAt = Date.now(); // nothing left to hydrate — prevents a stale read from racing back in before the next natural trip\n return { cleared };\n}\n// Item 4: provider_stats accumulates forever with no way to tell \"still failing\" from \"was failing before\n// a fix landed\" (the exact ambiguity flagged after this session's stale-binding fix). Distinct from\n// resetBreakers(): this clears the durable request/ok/failed counters, not the live routing decision.\nexport async function resetProviderStats(): Promise {\n await run(sql`DELETE FROM provider_stats`).catch(() => {});\n}\n/** Per-instance roster for /health — the settings panel's \"test / status\" table reads exactly this shape.\n * requests/ok/failed/last_error come from the durable table (accurate across isolates). cooling/demoted\n * are hydrated from the persisted breaker state first, so a breaker another isolate tripped is visible\n * here too — this closes the \"may show ready while a concurrent request elsewhere is avoiding it\" gap\n * the previous version of this comment documented. */\nexport async function getInstanceRoster() {\n await hydrateBreakers();\n const statsRow = await providerStatsAll();\n const stats = new Map((statsRow.ok ? statsRow.value : []).map((r: any) => [r.name, r]));\n const now = Date.now();\n return getProviderArray().map((p) => {\n const b = getBreaker(p.name);\n const s = stats.get(p.name);\n const cooling = b.state === \"OPEN\" ? { ms_left: Math.max(0, b.nextAttempt - now), reason: b.lastError ?? \"circuit open\" } : null;\n const demoted = b.demotedUntil > now ? { ms_left: b.demotedUntil - now, reason: b.demotedReason ?? \"config error\" } : null;\n return {\n name: p.name, key: p.key ? `${p.keyEnv}…${p.key.slice(-4)}` : null, key_issue: p.keyIssue,\n requests: s?.requests ?? 0, ok: s?.ok ?? 0, failed: s?.failed ?? 0,\n cooling, demoted, last_error: b.lastError ?? s?.last_error ?? null,\n };\n });\n}\n\nexport async function getSavedModels(): Promise> {\n const r = await all(\"omni_router_config\", sql`SELECT * FROM omni_router_config`);\n if (!r.ok) return {};\n return Object.fromEntries((r.value as any[]).map((row) => [row.provider, row.selected_model]));\n}\n\nexport type RouteOpts = {\n model?: string; tools?: unknown[]; tool_choice?: unknown;\n temperature?: number; max_tokens?: number; stream?: boolean; response_format?: unknown;\n deadlineMs?: number; // caller's remaining wall-clock budget (ms), from request receipt, for the WHOLE routeInference call — not per-attempt\n};\nexport type RouteHit = { provider: ProviderInstance; model: string; latencyMs: number; res: Response; plan?: string[]; trail?: RouteAttempt[] };\n// One row per provider try (or skip) — the raw material for the routing flowchart (app.tsx's PassMeter\n// turns this into a synthetic \"route\" ToolEvent; ui-timeline.ts draws it as a chain of nodes). Attached to\n// RouteHit on success and to the thrown error on exhaustion, so the flowchart exists in BOTH outcomes —\n// \"what did it actually try\" must answer the same whether the turn ended in an answer or a 503.\nexport type RouteAttempt = {\n provider: string; vendor: string; model: string; pass: number; timeoutMs: number;\n ok: boolean; skipped?: boolean; demoted?: boolean; latencyMs?: number; error?: string; ts: number;\n};\n\n// A 200 OK with an empty message, or a message that's nothing but dots (a known vendor failure\n// signature — truncated/garbage generations some providers return under load instead of a real error),\n// is not a valid answer. Only a plain-string content is checked — array/multi-part content bodies are\n// left alone rather than risk a false positive on a legitimate format we haven't seen. tool_calls-only\n// messages (content legitimately empty/null) are never flagged.\nfunction isEmptyOrDotContent(msg: any): boolean {\n if (Array.isArray(msg?.tool_calls) && msg.tool_calls.length) return false;\n if (typeof msg?.content !== \"string\") return false;\n const stripped = msg.content.replace(/\\s+/g, \"\");\n return stripped.length === 0 || /^\\.+$/.test(stripped);\n}\n\n// 4xx config errors still fail over but don't trip the breaker (no oscillation on a bad key/model id).\nasync function attempt(p: ProviderInstance, model: string, messages: unknown[], o: RouteOpts, timeoutMs: number): Promise {\n // A malformed secret never reaches the network: gemini's `…3, }` key produced 62 identical \"Failed to\n // construct 'Request': … not a valid ByteString\" failures — thrown by the Request constructor, so it\n // had no HTTP status, was classed retryable, tripped the breaker on a 25s cycle and was retried on\n // every single request forever. It is a config error by definition; say so, and let demotion park it.\n if (p.keyIssue) throw Object.assign(new Error(`KEY_MALFORMED (${p.keyEnv}): ${p.keyIssue}`), { retryable: false, configClass: true, status: 0 });\n const payload: Record = { model, messages, stream: !!o.stream };\n if (o.max_tokens && o.max_tokens > 0) payload.max_tokens = o.max_tokens;\n if (typeof o.temperature === \"number\") payload.temperature = o.temperature;\n // tool_choice:\"none\" with a tools array is semantically \"no tools callable\" — identical to sending no\n // tools at all — but Groq rejects the combination with HTTP 400 \"Tool choice is none, but model called a\n // tool\" (seen on every wrap-up pass this session). Send the equivalent form every vendor accepts.\n if (o.tools?.length && o.tool_choice !== \"none\") { payload.tools = o.tools; if (o.tool_choice) payload.tool_choice = o.tool_choice; }\n if (o.response_format) payload.response_format = o.response_format;\n const ctl = new AbortController();\n const timer = setTimeout(() => ctl.abort(), timeoutMs);\n const t0 = performance.now();\n try {\n const headers: Record = { \"Content-Type\": \"application/json\", \"HTTP-Referer\": \"https://val.town\", \"X-Title\": \"Ziggyware-OmniRouter\" };\n if (p.key) headers.Authorization = `Bearer ${p.key}`;\n const res = await fetch(`${p.base}/chat/completions`, { method: \"POST\", headers, body: JSON.stringify(payload), signal: ctl.signal });\n if (!res.ok) {\n const body = await res.text().catch(() => \"\");\n const e: any = new Error(`HTTP ${res.status}: ${body.slice(0, 300)}`);\n e.retryable = RETRYABLE.has(res.status);\n e.status = res.status;\n e.configClass = isConfigClass(res.status, body);\n throw e;\n }\n // Headers arrived within timeoutMs — that phase is done; clear its timer here, exactly like before\n // content validation existed. CONFIRMED REGRESSION (live /health: anthropic's success rate went from\n // 6/7 to 6/23 after this was added, nearly all new failures API_STALLED_TIMEOUT): keeping the SAME\n // timer armed through the full body read punished large, legitimately-slow-but-completing generations\n // (this app requests up to 16k max_tokens against prompts confirmed 18k+ tokens by a live Groq 413) —\n // a completion that needed even a few seconds past a shrunk, late-pass timeoutMs to finish sending was\n // aborted and misreported as a stall. Body read gets its OWN separate, generous, fixed budget instead —\n // independent of the caller's possibly-tiny remaining timeoutMs — so content validation (which does\n // need the full body) can't retroactively punish a provider that already answered in time.\n if (!o.stream) {\n clearTimeout(timer);\n const BODY_READ_TIMEOUT_MS = 20_000;\n let text: string;\n try {\n text = await Promise.race([\n res.text(),\n new Promise((_, reject) =>\n setTimeout(() => reject(Object.assign(new Error(\"API_STALLED_TIMEOUT\"), { retryable: true, name: \"AbortError\" })), BODY_READ_TIMEOUT_MS)\n ),\n ]);\n } catch (e: any) {\n res.body?.cancel().catch(() => {});\n throw e;\n }\n let json: any = null;\n try { json = JSON.parse(text); } catch { /* not JSON — let router-openai.ts's own parse surface the real error */ }\n const msg = json?.choices?.[0]?.message;\n if (json && isEmptyOrDotContent(msg)) {\n const e: any = new Error(`EMPTY_OR_DOT_RESPONSE: ${JSON.stringify(String(msg?.content ?? \"\")).slice(0, 60)}`);\n e.retryable = true;\n throw e;\n }\n return { provider: p, model, latencyMs: Math.round(performance.now() - t0), res: new Response(text, { status: res.status, headers: res.headers }) };\n }\n clearTimeout(timer);\n return { provider: p, model, latencyMs: Math.round(performance.now() - t0), res };\n } catch (e: any) {\n clearTimeout(timer);\n if (e.name === \"AbortError\") { const t: any = new Error(\"API_STALLED_TIMEOUT\"); t.retryable = true; throw t; }\n // Thrown before any bytes left the process (Request/URL construction) ⇒ our configuration, not their\n // outage. Retrying a TypeError is never going to produce a different result.\n if (e instanceof TypeError && /construct|Invalid URL|ByteString/i.test(String(e.message))) { e.retryable = false; e.configClass = true; }\n if (e.retryable === undefined) e.retryable = true;\n throw e;\n }\n}\n\n// Depth-first (priority, then key-slot) means a single vendor-wide outage — the vendor itself stalling,\n// not any one key being bad — burns every key slot of THAT vendor before a different vendor is ever\n// tried. Observed directly: openrouter/nemotron-ultra + #1 both API_STALLED_TIMEOUT, consuming the whole\n// deadline, while cerebras/groq/mistral/etc. sat untouched. Breadth-first — one instance per vendor first\n// (in existing priority order), second key-slots only in a later wave — means 2 failed attempts always\n// means 2 DIFFERENT vendors tried, not 2 keys of the one that happens to be down. Vendor-relative priority\n// (which model/vendor is preferred) is unchanged; only the within-vendor key-slot ordering is deferred.\nfunction breadthFirst(targets: ProviderInstance[]): ProviderInstance[] {\n const byVendor = new Map();\n for (const p of targets) {\n if (!byVendor.has(p.vendor)) byVendor.set(p.vendor, []);\n byVendor.get(p.vendor)!.push(p);\n }\n const waves: ProviderInstance[][] = [];\n for (const group of byVendor.values()) group.forEach((p, i) => (waves[i] ??= []).push(p));\n return waves.flat();\n}\n\n// Pure breadthFirst() above overcorrected: a second key set on the SAME provider/model (added for\n// throughput / per-key rate-limit headroom) is a different catalog row's slot only in the trivial\n// single-row case — for a vendor with multiple catalog rows (e.g. openrouter's 3 free models), that\n// alt key's position within its own vendor's flattened list can land many entries deep, which\n// breadthFirst then schedules behind EVERY other vendor's primary. Simulated against the real catalog\n// shape (openrouter, 2 keys/row): the alt key landed at attempt #13, while only ~3 attempts fit inside\n// a real turn deadline before providers that hang to HARD_TIMEOUT_MS exhaust it — so the alt key was\n// structurally almost never reached. Fix: let each vendor's TOP-priority row retry with ONE alt key\n// (SAME_KEY_RETRY_CAP) immediately, before the loop advances to a different vendor — this is the only\n// way to make \"I bought a backup key for this rate-limited provider\" actually work — then fall back to\n// breadthFirst for everything else (other rows of that vendor, and any keys beyond the cap). Confirmed\n// via simulation this bounds the regression: worst case (every attempt times out) still guarantees ≥2\n// DIFFERENT vendors get tried before the deadline dies, vs 3 under pure breadthFirst — a small, bounded\n// cost in exchange for alt keys actually being usable.\nconst SAME_KEY_RETRY_CAP = 1;\n// Tier 2 (below) was added after live diagnostics showed a vendor's OTHER catalog rows (e.g. openrouter's\n// qwen3-coder/gpt-oss, siblings of its top-priority nemotron-ultra) going untried in practice even though\n// tier 1 + breadthFirst() both nominally include them: tier 1 only ever admits ONE row per vendor (the top\n// one, plus its own alt keys), and breadthFirst groups by VENDOR not by catalog row — so a sibling row's\n// primary key ends up scheduled behind every alt-key retry of the vendor's first-tried row, many entries\n// deep, structurally unreachable before the wall clock runs out. Fix: a middle tier that gives each\n// vendor's OTHER rows' PRIMARY key (slot 0 only — not their alt keys, which stay in tier 3) a turn right\n// after tier 1's same-row retries, before any vendor's second key of anything. Deliberately NOT folding all\n// of a vendor's rows into tier 1 — that was considered and rejected: a vendor with many free-tier rows\n// could then burn several attempts against what's often a shared per-vendor daily rate limit (confirmed\n// this session: OpenRouter's \"free models per day\" cap) before a different vendor is ever reached.\nfunction orderTargets(targets: ProviderInstance[]): ProviderInstance[] {\n const byVendor = new Map();\n for (const p of targets) {\n if (!byVendor.has(p.vendor)) byVendor.set(p.vendor, []);\n byVendor.get(p.vendor)!.push(p);\n }\n const baseName = (p: ProviderInstance) => p.name.split(\"#\")[0];\n const heads: ProviderInstance[] = []; // tier 1: each vendor's top row, primary key + capped alt-key retries\n const headSet = new Set();\n const siblings: ProviderInstance[] = []; // tier 2: each vendor's OTHER rows, primary key only\n const siblingSet = new Set();\n for (const group of byVendor.values()) {\n if (!group.length) continue;\n const top = group[0]; // already sorted priority,slot — group[0] is this vendor's best row's primary key\n heads.push(top);\n headSet.add(top);\n let taken = 0;\n for (let j = 1; j < group.length && taken < SAME_KEY_RETRY_CAP; j++) {\n if (baseName(group[j]) !== baseName(top)) break; // only the SAME row's own alt keys qualify — a different model doesn't jump the queue\n heads.push(group[j]);\n headSet.add(group[j]);\n taken++;\n }\n const seenRows = new Set([baseName(top)]);\n for (const p of group) {\n const b = baseName(p);\n if (seenRows.has(b)) continue; // either the top row, or an alt key of a sibling row already claimed below\n seenRows.add(b);\n siblings.push(p);\n siblingSet.add(p);\n }\n }\n const rest = targets.filter((p) => !headSet.has(p) && !siblingSet.has(p));\n return [...heads, ...siblings, ...breadthFirst(rest)];\n}\n\n// Item 9: defensive ceiling on `trail`'s own memory footprint, independent of router-openai.ts's separate\n// last-TRAIL_CAP(60) slice taken when serializing the response. That slice only ever protects the wire\n// payload; the in-memory array itself was unbounded, so a pathological worst case (MAX_FULL_PASSES=10 full\n// laps of a large catalog, every attempt failing) could grow it far past what any consumer ever reads.\n// Trimming from the front keeps the most recent — and therefore most diagnostically relevant — entries.\nconst ROUTE_TRAIL_HARD_CAP = 300;\nfunction pushTrail(trail: RouteAttempt[], entry: RouteAttempt): void {\n trail.push(entry);\n if (trail.length > ROUTE_TRAIL_HARD_CAP) trail.splice(0, trail.length - ROUTE_TRAIL_HARD_CAP);\n}\n\nexport async function routeInference(messages: unknown[], o: RouteOpts = {}): Promise {\n await acquire();\n try {\n await hydrateBreakers();\n const live = getProviderArray();\n if (!live.length) throw Object.assign(new Error(\"NO_PROVIDERS_CONFIGURED\"), { status: 503 });\n const resolved = resolveTargets(o.model, live);\n const ordered = orderTargets(resolved.targets);\n // Demoted (config-dead) instances go to the tail as a block, in their existing relative order. This is a\n // stable partition, not a filter: every instance is still in `targets`, still reached by a full pass,\n // still covered by fail-open. What changes is only that a provider proven dead-by-configuration stops\n // costing a round-trip ahead of the ones that might actually answer.\n const demotedSet = new Set(ordered.filter((p) => isDemoted(p.name)));\n const targets = [...ordered.filter((p) => !demotedSet.has(p)), ...ordered.filter((p) => demotedSet.has(p))];\n const { forcedModel } = resolved;\n if (!targets.length) throw Object.assign(new Error(`UNKNOWN_MODEL: ${o.model}`), { status: 404 });\n const saved = await getSavedModels();\n const modelFor = (p: ProviderInstance) => forcedModel ?? saved[p.name.split(\"#\")[0]] ?? p.fallbackModel;\n const errors: string[] = [];\n const plan = targets.map((p) => p.name); // full planned order, BEFORE any attempt — the flowchart's \"intended route\"\n const trail: RouteAttempt[] = []; // every attempt AND skip, in the order it actually happened — the flowchart's \"actual route\"\n let currentPass = 1;\n\n // Hard wall-clock ceiling for THIS invocation, independent of whatever the caller's o.deadlineMs\n // says: Val Town's free tier kills the function outright around 60s with NO response sent at all —\n // that's worse than a clean 503. 50s leaves margin for parsing/serialization after the last attempt\n // returns. A caller deadline, when present and smaller, still narrows further (honors the caller's\n // own remaining turn budget) — it can only shrink this ceiling, never extend past it.\n const t0 = Date.now();\n const SAFETY_CEILING_MS = 50_000;\n const callDeadline = o.deadlineMs && o.deadlineMs > 0 ? Math.min(o.deadlineMs, SAFETY_CEILING_MS) : SAFETY_CEILING_MS;\n // Was lowered to 1_200 for breadth, then raised back up: a fast-failing provider (bad key, wrong\n // model, exhausted quota — the majority of what's actually being seen in practice) rejects in well\n // under a second regardless of this floor, so a low floor never helped THOSE. What it did do was let\n // a genuinely slow-but-working provider get attempted with almost no realistic chance to finish,\n // burning the slot for nothing. 5s is still short relative to HARD_TIMEOUT_MS's 30s, but below it a\n // real non-streaming completion for this workload's prompt/output sizes has no realistic chance.\n const MIN_ATTEMPT_MS = 5_000;\n const remaining = () => callDeadline - (Date.now() - t0) - 500; // 500ms margin for overhead after the last attempt\n const attemptOne = async (p: ProviderInstance): Promise => {\n const timeoutMs = Math.min(HARD_TIMEOUT_MS, remaining());\n if (timeoutMs < MIN_ATTEMPT_MS) {\n errors.push(`${p.name}: skipped — ${Math.max(0, timeoutMs)}ms left`);\n pushTrail(trail, { provider: p.name, vendor: p.vendor, model: modelFor(p), pass: currentPass, timeoutMs: Math.max(0, timeoutMs), ok: false, skipped: true, demoted: demotedSet.has(p) || undefined, error: \"out of time\", ts: Date.now() });\n return null;\n }\n const model = modelFor(p);\n const demoted = demotedSet.has(p) || undefined;\n try {\n const hit = await attempt(p, model, messages, o, timeoutMs);\n await report(p.name, true);\n pushTrail(trail, { provider: p.name, vendor: p.vendor, model, pass: currentPass, timeoutMs, ok: true, demoted, latencyMs: hit.latencyMs, ts: Date.now() });\n return { ...hit, plan, trail };\n } catch (e: any) {\n await report(p.name, false, e.message, !!e.retryable, !!e.configClass);\n errors.push(`${p.name}: ${e.message}`);\n pushTrail(trail, { provider: p.name, vendor: p.vendor, model, pass: currentPass, timeoutMs, ok: false, demoted, error: String(e.message).slice(0, 200), ts: Date.now() });\n console.warn(`[SEVERED] ${p.name} (${e.message})`);\n return null;\n }\n };\n // Never declare exhaustion after a handful: walk the ENTIRE ordered target list every pass (skipping\n // only breaker-open providers, restored below via fail-open if a pass skips everyone), and if the\n // whole list comes up empty with wall-clock room left, run it again — up to MAX_FULL_PASSES — instead\n // of giving up after one lap. A later pass can also pick up a provider whose 25s breaker cooldown\n // (BREAK_DURATION_MS) has since elapsed. In practice, MAX_FULL_PASSES is an upper bound the wall\n // clock rarely lets this reach — walking ~20-30 providers at HARD_TIMEOUT_MS each already consumes\n // the whole SAFETY_CEILING_MS in one pass if most of them stall. The rest of \"at least 10 tries\" is\n // deliberately NOT this function's job: main-script.ts's send loop re-invokes the whole turn (a\n // fresh 60s window each time, so not bounded by this invocation's wall clock) on an EXHAUSTED signal,\n // now guaranteed to happen at least 10 times before it gives up — see its comment for why splitting\n // the guarantee across invocations is the only way to honor it without Val Town killing the function\n // mid-response.\n const MAX_FULL_PASSES = 10;\n for (let pass = 1; pass <= MAX_FULL_PASSES; pass++) {\n currentPass = pass;\n const skipped: ProviderInstance[] = [];\n let attemptedThisPass = 0;\n for (const p of targets) {\n if (remaining() < MIN_ATTEMPT_MS) break; // wall clock is the only thing allowed to cut a pass short\n if (!admit(p.name)) {\n skipped.push(p);\n const b = getBreaker(p.name);\n pushTrail(trail, { provider: p.name, vendor: p.vendor, model: modelFor(p), pass, timeoutMs: 0, ok: false, skipped: true, demoted: demotedSet.has(p) || undefined, error: `circuit open — ${Math.max(0, b.nextAttempt - Date.now())}ms cooldown left`, ts: Date.now() });\n continue;\n }\n attemptedThisPass++;\n const hit = await attemptOne(p);\n if (hit) return hit;\n }\n // Fail-open: a breaker-open provider is skipped for SPEED when a healthier one might work instead,\n // never as a verdict that it's still down. If this pass skipped everyone, nothing was actually\n // attempted — that's not exhaustion, that's a stale verdict. Force one real try through the skipped\n // set before moving on, so \"all providers exhausted\" always means all providers were actually tried.\n if (attemptedThisPass === 0 && skipped.length) {\n for (const p of skipped) {\n if (remaining() < MIN_ATTEMPT_MS) break;\n const hit = await attemptOne(p);\n if (hit) return hit;\n }\n }\n if (remaining() < MIN_ATTEMPT_MS) break; // no room for another full pass\n }\n throw Object.assign(new Error(`ALL_PROVIDERS_EXHAUSTED\\n${errors.join(\"\\n\")}`), { status: 503, errors, plan, trail });\n } finally { release(); }\n}\n",[],[342],"c37812cd-0ebe-4a97-96ae-d9f790be849a","router-attempt.ts","01a0a661-f04a-733e-839a-bbf512f56847",["D",1789497831502],"import type { ProviderInstance } from \"./inference-provider.ts\";\nimport { headersFor } from \"./router-discover.ts\";\n\nexport type RouteOpts = { model?: string; stream?: boolean; extra?: Record; deadlineMs?: number; expectTokens?: number; keyPolicy?: \"depth\" | \"breadth\" | \"rr\"; vendorOrder?: string[]; exclude?: string[] };\nexport type Fail = Error & { kind: \"model\" | \"rate\" | \"auth\" | \"quota\" | \"size\" | \"transient\"; retryAfterMs?: number; sentMax?: number; resetAt?: number };\nconst QUOTA = /balance|recharge|insufficient|余额|credits?|quota|billing|payment|per[ _-]day|daily|monthly|exceeded your|free.?tier|limit reached for the day|\\b[rt]pd\\b/i;\n/** When a quota-class failure heals: header epoch, \"daily\" → next UTC midnight, else re-probe in 6 h (funds/credits can be added any time). */\nexport function quotaResetAt(res: Response | null, body: string): number {\n const now = Date.now();\n const h = res?.headers.get(\"x-ratelimit-reset\") ?? res?.headers.get(\"x-ratelimit-reset-requests\") ?? \"\";\n const n = Number(h);\n if (n > 0) { const ms = n > 1e12 ? n : n > 1e9 ? n * 1000 : now + n * 1000; if (ms > now && ms < now + 36 * 3_600_000) return ms; }\n if (/per day|daily|for the day|today/i.test(body)) { const d = new Date(now); d.setUTCHours(24, 0, 0, 0); return d.getTime() + 60_000; }\n return now + 6 * 3_600_000;\n}\n// \"Limit 6000, Requested 9000\" (Groq TPM) or \"maximum context length is 8192 tokens ... requested 9000\": shrink to fit.\nexport function shrinkMaxTokens(o: RouteOpts, body: string, sent?: number): boolean {\n const cur = sent || Number(o.extra?.max_tokens);\n if (!cur || cur <= 512) return false;\n const lim = body.match(/limit[^\\d]{0,12}(\\d+)/i), req = body.match(/request(?:ed)?[^\\d]{0,12}(\\d+)/i);\n const next = lim && req ? Math.min(cur - (Number(req[1]) - Number(lim[1])) - 64, Number(lim[1]) - 64) : Math.floor(cur / 2);\n o.extra = { ...o.extra, max_tokens: Math.max(512, Math.min(next, cur - 256)) };\n return true;\n}\n/** A size error the PROMPT alone caused: shrinking max_tokens cannot fix it. Groq ITPM, context-length overflows. Returns the input limit when stated. */\nexport function inputBound(body: string): number | null {\n if (!/input tokens|ITPM|prompt tokens|context length|maximum context|too many tokens in (the )?prompt/i.test(body)) return null;\n const m = body.match(/limit[^\\d]{0,12}(\\d+)/i) ?? body.match(/context length (?:of|is) (\\d+)/i);\n return m ? Number(m[1]) : 0;\n}\n/** Rough prompt size in tokens: chars/3.6 over messages + tool schemas (vendors tokenize JSON worse than prose). */\nexport const estimateInputTokens = (messages: unknown[], extra?: Record) => Math.ceil((JSON.stringify(messages).length + (extra?.tools ? JSON.stringify(extra.tools).length : 0)) / 3.6);\n/** \"Limit N\" from a size error when it is an output ceiling (OTPM / max_tokens), else null */\nexport function outputLimit(body: string): number | null {\n if (!/output|max_tokens|completion/i.test(body)) return null;\n const m = body.match(/limit[^\\d]{0,12}(\\d+)/i) ?? body.match(/(?:less than or equal to|maximum value[^\\d]{0,30}is|at most)[^\\d]{0,6}(\\d+)/i);\n return m ? Number(m[1]) : null;\n}\n\nexport function classify(status: number, body: string): Fail[\"kind\"] {\n if ((status === 429 || status === 413 || status === 400) && /too large|too long|context length|maximum context|exceeds.*tokens|reduce.*(length|tokens)/i.test(body)) return \"size\";\n if (status === 429 && QUOTA.test(body)) return \"quota\";\n if ((status === 429 || status === 400) && /max_tokens.{0,40}(less than or equal|maximum value|must be at most)|maximum value for .max_tokens/i.test(body)) return \"size\";\n if (status === 429) return \"rate\";\n if (status === 404 || /model_not_found|invalid.model|tier_not_allowed|does not (exist|support)|not (found|exist|supported|available)|unsupported|decommissioned|deprecated|no access|unable to access/i.test(body)) return \"model\";\n // A 400 is a request/output defect, never billing: the body often quotes the model's own generation (tool_use_failed\n // carries failed_generation), and any \"credits\"/\"balance\" word in that text would otherwise cool the key for hours.\n if (status === 400) return \"transient\";\n if (status === 402 || (status === 403 && QUOTA.test(body))) return \"quota\";\n if (status === 401 || status === 403 || /api.?key|unauthorized|invalid.?token/i.test(body)) return \"auth\";\n return \"transient\";\n}\n\n// timeoutMs covers headers AND (non-stream) body: a slow vendor streaming a long\n// completion past the deadline is the failure mode that killed 60 s turns.\n// capMax: the largest completion this vendor can stream before the deadline; a\n// truncated answer (finish_reason \"length\") beats a timeout — the caller continues it.\nexport async function attempt(p: ProviderInstance, model: string, messages: unknown[], o: RouteOpts, timeoutMs: number, capMax = Infinity): Promise {\n const payload: Record = { ...o.extra, model, messages, stream: !!o.stream };\n const cap = Math.min(p.maxOut ?? Infinity, capMax);\n if (Number(payload.max_tokens) > cap) payload.max_tokens = cap;\n const ctl = new AbortController();\n const timer = setTimeout(() => ctl.abort(), timeoutMs);\n const abortP = new Promise((_, rej) => ctl.signal.addEventListener(\"abort\", () => rej(Object.assign(new Error(\"TIMEOUT\"), { kind: \"transient\" }))));\n try {\n const res = await Promise.race([fetch(`${p.base}/chat/completions`, { method: \"POST\", headers: headersFor(p), body: JSON.stringify(payload), signal: ctl.signal }), abortP]);\n if (res.ok) {\n if (o.stream) return res;\n const text = await Promise.race([res.text(), abortP]).catch((e) => { res.body?.cancel().catch(() => {}); throw e; });\n return new Response(text, { status: res.status, headers: { \"content-type\": res.headers.get(\"content-type\") ?? \"application/json\", \"x-omni-sent-max\": String(payload.max_tokens ?? \"\") } });\n }\n const body = await res.text().catch(() => \"\");\n const e = new Error(`HTTP ${res.status}: ${body.slice(0, 500)}`) as Fail;\n e.kind = classify(res.status, body);\n // 429 with \"remaining: 0\" and a reset far away is a quota even when the body says nothing useful (OpenRouter free-models-per-day)\n if (e.kind === \"rate\" && res.headers.get(\"x-ratelimit-remaining\") === \"0\") { const at = quotaResetAt(res, body); if (at - Date.now() > 5 * 60_000) e.kind = \"quota\"; }\n e.sentMax = Number(payload.max_tokens) || undefined;\n const ra = Number(res.headers.get(\"retry-after\")) || Number((body.match(/try again in\\s*([\\d.]+)\\s*s/i) ?? [])[1]) || 0;\n if (e.kind === \"rate\" && ra > 0) e.retryAfterMs = Math.min(Math.ceil(ra * 1000) + 500, 15 * 60_000);\n if (e.kind === \"quota\") e.resetAt = quotaResetAt(res, body);\n throw e;\n } catch (e: any) {\n if (!e.kind) { e.kind = \"transient\"; e.message = e.name === \"AbortError\" ? \"TIMEOUT\" : `NETWORK: ${e.message}`; }\n ctl.abort();\n throw e;\n } finally { clearTimeout(timer); }\n}\n\n",[],[349],"d3022313-1cb4-4e1b-b9ac-d5da5d81dfcd","rate-limit-tracker.ts","01a0a661-f04a-733e-839a-b46e6461b836",["D",1789497831502],"// rate-limit-tracker.ts — ~800 chars\ninterface RateLimitState {\n dailyRequests: Record; // model → request count\n resetTime: Record; // model → unix timestamp (ms)\n}\n\nexport class RateLimitTracker {\n private state: RateLimitState = {\n dailyRequests: {},\n resetTime: {},\n };\n\n canRequest(model: string, limit: number): boolean {\n const now = Date.now();\n const resetAt = this.state.resetTime[model] ?? 0;\n\n if (now >= resetAt) {\n // Day boundary crossed → reset\n this.state.dailyRequests[model] = 0;\n this.state.resetTime[model] = now + 86400000; // +1 day\n }\n\n const current = this.state.dailyRequests[model] ?? 0;\n return current < limit;\n }\n\n recordRequest(model: string) {\n this.state.dailyRequests[model] = (this.state.dailyRequests[model] ?? 0) +\n 1;\n }\n\n remaining(model: string, limit: number): number {\n return Math.max(0, limit - (this.state.dailyRequests[model] ?? 0));\n }\n}",[],[356],"854772d6-bd9f-48c3-82c9-362663ab85c3","quality-report.ts","01a0a661-f04a-733e-839a-b28eb2a68367",["D",1789497831502],"import { qualityReport } from \"./artifacts.ts\";\nconst basic = [{ path: \"index.html\", content: \"

\" }, { path: \"app.js\", content: \"const c=document.getElementById('c').getContext('2d');\\nfunction draw(){c.fillRect(0,0,10,10);requestAnimationFrame(draw)}\\ndraw();\\nalert('hi')\\n\" }];\nconst q = qualityReport(basic); console.log(q.length, \"findings\"); console.log(q.map(x => x.slice(0, 60)).join(\"\\n\"));\nconst good = Array.from({ length: 9 }, (_, i) => ({ path: i ? `m${i}.js` : \"index.html\", content: (i ? \"export const x=1;\\n\" : \"

nothing here yet

\") + \"// filler\\n\".repeat(110) }));\ngood.push({ path: \"README.md\", content: \"- f\\n\".repeat(12) });\ngood.push({ path: \"core.js\", content: \"devicePixelRatio; new ResizeObserver(()=>{}); document.addEventListener('visibilitychange',()=>{}); requestAnimationFrame(()=>{}); addEventListener('keydown',()=>{}); addEventListener('pointerdown',()=>{}); try{localStorage.getItem('a')}catch(e){} const schemaVersion=2; function undo(){} toast(); fetch('x').catch(()=>{})\" });\nconsole.log(\"good:\", qualityReport(good));\n",[],[363],["D",1789497831502],[],[18],"cef69fed-62f1-4436-afb6-7a7a88006fc0","prompt-examples.ts","01a0a661-f04a-733e-839a-aa52f4417792",["D",1789497831502],"// Two exemplars per prompt, appended to every rendered prompt (Settings → \"few-shot examples\", default on).\n// Few-shot anchoring is the strongest single lever on output quality for open-weight models: they imitate the\n// register and density of what they are shown. Each pair is deliberately small — the standard is demonstrated,\n// not the whole artifact — so the system prompt stays under free-tier input ceilings (groq ≈ 6.5k tokens).\nimport type { PromptKey } from \"./prompts.ts\";\n\nconst ex = (title: string, a: string, b: string) => `\\nEXAMPLES — ${title}. Match this register, density and completeness exactly.\\n\\n${a.trim()}\\n\\n\\n${b.trim()}\\n`;\n\nexport const PROMPT_EXAMPLES: Record = {\n system: ex(\"answers at the required standard\",\n`USER: why does my WebGL texture look washed out on Safari but not Chrome?\nASSISTANT: Safari applies the canvas's colorSpace conversion twice when the drawing buffer is premultiplied and the texture is uploaded with UNPACK_COLORSPACE_CONVERSION_WEBGL left at BROWSER_DEFAULT_WEBGL. Set it once before texImage2D:\n\\`\\`\\`js\ngl.pixelStorei(gl.UNPACK_COLORSPACE_CONVERSION_WEBGL, gl.NONE);\ngl.pixelStorei(gl.UNPACK_PREMULTIPLY_ALPHA_WEBGL, false);\n\\`\\`\\`\nand create the context with { premultipliedAlpha: false } if you composite yourself. Verify: a 50 % grey PNG must sample as 0.5 ± 0.004 in the shader on both browsers (read it back with gl.readPixels on a 1×1 quad). If it still differs, the PNG carries a gAMA chunk — strip it (pngcrush -rem gAMA) so both decoders agree.`,\n`USER: build a metronome app\nASSISTANT: [create_artifact index.html, then one update_artifact per file — 8 calls, one file each, in dependency order: style.css, main.js, state.js, audio.js (AudioContext scheduler with 25 ms lookahead, 100 ms scheduleAhead, drift-free via ctx.currentTime), render.js (dpr-scaled canvas beat ring, rAF with delta, pauses on visibilitychange), input.js (pointer + keyboard: space start/stop, ↑↓ ±1 BPM, ⇧↑↓ ±10, T tap-tempo with 4-tap median), storage.js (schemaVersion 2 + migrate(), try/catch, memory fallback), README.md (12 features, shortcuts, run instructions)]\nMetronome #41 is live in the workbench: 20–300 BPM, 1–12 beats with accent pattern editor, subdivisions (1/2/3/4), tap tempo, tempo ramp (start → end BPM over N bars), presets with undo, sample-accurate scheduling (jitter < 1 ms measured over 10 min), full keyboard control (? opens the shortcut sheet), dark/light via prefers-color-scheme. Does not do: MIDI clock out, polyrhythm across two meters.`),\n\n meta: ex(\"control block\",\n`<<>>`,\n`<<>>`),\n\n firstExchange: ex(\"session titles\",\n`ask: \"my shader flickers when I resize the window\" → name_session(\"WebGL Resize Flicker\")`,\n`ask: \"write me a pomodoro app with stats\" → name_session(\"Pomodoro Stats App\")`),\n\n plan: ex(\"plans\",\n`1. read_artifact #12 render.js and input.js only — the reported bug is a drag offset, so state.js is irrelevant.\n2. Locate the pointer→world transform; the symptom (offset grows with zoom) means dpr or zoom is applied twice.\n3. One update_artifact with the fix in render.js and a regression guard (assert in dev mode).\nStop when: the transform is applied exactly once and lint is clean.`,\n`1. eval_js: compute the 12-TET frequency table for A4 = 432 and 440 (24 values) — never eyeball.\n2. Answer with a table (note, 440 Hz, 432 Hz, Δ cents = 1200·log2(432/440) = −31.77 constant).\nStop when: table delivered with the constant-cents observation stated once.`),\n\n reconcile: ex(\"reconciled answers\",\n`Drafts disagree on Deno's default V8 heap limit. Draft A cites the flag (--v8-flags=--max-old-space-size) with a number; B and C give different numbers with no source. Keep A's mechanism, drop all three numbers, state the flag and how to read the actual limit at runtime (Deno.memoryUsage / --v8-flags=--help).`,\n`All drafts agree the bug is an off-by-one in the ring buffer; A alone shows the failing index (write === read after wrap). Keep A's proof, keep B's shorter fix (mask with & (N-1) for power-of-two N), drop C's rewrite.`),\n\n critique: ex(\"verdicts\",\n`FLAW: The answer claims requestAnimationFrame runs at 60 Hz. It runs at the display's refresh rate (120/144 Hz on many devices), so the physics step derived as 1/60 s is wrong; the loop must use the timestamp delta. Consequential: motion speed doubles on 120 Hz screens.`,\n`FLAW: none found. Checked: the regex escapes '.' correctly, the unit conversion (m/s² → g) divides by 9.80665, and the code block runs (eval_js returned the stated 42.0).`),\n\n revise: ex(\"revisions\",\n`(only the changed paragraph) The loop must read the timestamp rAF passes and integrate with dt = (t − last) / 1000, clamped to ≤ 0.05 s so a background tab does not explode the simulation; 1/60 is wrong on 120 Hz displays.`,\n`(only the changed line) ⋯ divides by 9.80665 m/s² (standard gravity), not 9.8 — the 0.02 % error compounds over the integration.`),\n\n reground: ex(\"grounded answers\",\n`web_search(\"Deno 2.4 release notes\") → snippet: \"Deno 2.4 (July 2025) adds deno bundle …\" (deno.com/blog, 2025-07-16). Answer: bundling returned in 2.4; the quoted note is the primary source. Not established: whether it supports code-splitting — the note does not say.`,\n`web_search returned only forum posts with no dates for the claimed API change. Missing: a changelog or docs page. I cannot confirm the change; here is what would settle it: the package's CHANGELOG entry for the version you run.`),\n\n continue: ex(\"continuations\",\n`(previous output ended: \" for (const p of particles) {\")\n p.vx += g.x * dt; p.vy += g.y * dt;\n p.x += p.vx * dt; p.y += p.vy * dt;\n }`,\n`(previous output ended: \"The second constraint is\")\n that the buffer must be power-of-two sized, so the wrap is a mask instead of a modulo.`),\n\n summarize: ex(\"closing summaries\",\n`Built #57 \"Spectrum Analyzer\": 10 files, 1,180 lines. Mic or file input, 4096-point FFT (radix-2, Hann window), log-frequency canvas at device pixel ratio, peak hold, A-weighting toggle, keyboard (space pause, ↑↓ gain, W weighting), settings persisted (schema v1). Open it in the workbench to run. Does not do: MP3 decoding in Safari < 17 (decodeAudioData limitation).`,\n`Found the cause: state.js:88 mutates the array it is iterating, so every other item is skipped on delete. Fixed with a filtered copy; added a regression check in dev mode. No other file touched.`),\n\n workbench: ex(\"focused edits\",\n`ask: \"the canvas is blurry\" (open file render.js) → read_artifact render.js → one update_artifact: edits = [{find: \"canvas.width = w; canvas.height = h;\", replace: \"const dpr = Math.min(devicePixelRatio || 1, 3); canvas.width = Math.round(w * dpr); canvas.height = Math.round(h * dpr); canvas.style.width = w + 'px'; canvas.style.height = h + 'px'; ctx.setTransform(dpr, 0, 0, dpr, 0, 0);\"}] → \"Backing store now scales with devicePixelRatio (capped at 3); layout size unchanged.\"`,\n`ask: \"add undo\" (open file state.js) → read_artifact state.js, input.js → one update_artifact: files = [{path: \"history.js\", content: }], edits in state.js (dispatch records inverse), input.js (⌘Z/⇧⌘Z), index.html (script tag) → \"Undo/redo with drag coalescing; 200 steps; shortcuts listed in the help panel.\"`),\n\n skill: ex(\"applying a skill\",\n`Skill \"css-tokens\" says every color must come from a custom property → the artifact's style.css defines :root tokens and no rule uses a literal hex; index.html sets data-theme from prefers-color-scheme.`,\n`Skill \"no-deps\" says no external code → the chart is drawn with canvas 2D and a 60-line scale/axis module instead of a library.`),\n\n architect: ex(\"manifests\",\n`{\"title\": \"Pomodoro Stats\", \"summary\": \"Focus timer with session history and analytics.\",\n \"features\": [\"25/5/15 cycles with editable durations\", \"start/pause/skip with space and S\", \"long break every 4th\", \"session log with tags\", \"daily/weekly bar chart\", \"streaks\", \"notification + sound at end (feature-detected)\", \"keyboard shortcut sheet (?)\", \"export/import JSON\", \"undo delete of a session\", \"dark/light tokens\", \"persisted settings + history schema v1 with migrate()\"],\n \"visual\": \"--bg #0f1115 --surface #171a21 --text #e6e8ef --accent #ff6b4a --danger #ff4d6d; Inter/system type 13/15/20/28; 4 px spacing unit; radius 10; 160 ms ease-out\",\n \"shared\": \"store.js: createStore(reducer, init) → {get, dispatch, subscribe}; actions {type:'TICK', now} {type:'START'} {type:'PAUSE'} {type:'COMPLETE', session:{id, start, end, kind, tag}}; state {phase:'focus'|'short'|'long', running, remainingMs, sessions[], settings}; ids = crypto.randomUUID() with Date.now()-based fallback; DOM ids #timer #ring #log #chart #help; CSS classes .btn .btn-primary .card\",\n \"files\": [{\"path\": \"index.html\", \"purpose\": \"markup + boot; loads siblings in order\", \"exports\": [], \"imports\": [\"style.css\", \"main.js\"], \"notes\": \"\"}, {\"path\": \"store.js\", \"purpose\": \"typed reducer store + undo stack\", \"exports\": [\"createStore(reducer, init)\", \"undoable(reducer)\"], \"imports\": [], \"notes\": \"pure\"}, {\"path\": \"timer.js\", \"purpose\": \"drift-free ticking from performance.now(), pauses on visibilitychange\", \"exports\": [\"startTicker(store)\"], \"imports\": [\"./store.js\"], \"notes\": \"\"}, {\"path\": \"render.js\", \"purpose\": \"ring on dpr-scaled canvas, log list, chart\", \"exports\": [\"mount(store)\"], \"imports\": [\"./store.js\", \"./chart.js\"], \"notes\": \"ResizeObserver\"}, {\"path\": \"chart.js\", \"purpose\": \"bar chart, canvas 2D, axis/scale helpers\", \"exports\": [\"drawBars(ctx, data, opts)\"], \"imports\": [], \"notes\": \"\"}, {\"path\": \"input.js\", \"purpose\": \"pointer + keyboard map\", \"exports\": [\"bindInput(store)\"], \"imports\": [\"./store.js\"], \"notes\": \"\"}, {\"path\": \"storage.js\", \"purpose\": \"versioned persistence, migrate(), memory fallback\", \"exports\": [\"load()\", \"save(state)\", \"SCHEMA_VERSION\"], \"imports\": [], \"notes\": \"try/catch\"}, {\"path\": \"notify.js\", \"purpose\": \"Notification + WebAudio beep, both feature-detected\", \"exports\": [\"notifyEnd(kind)\"], \"imports\": [], \"notes\": \"\"}, {\"path\": \"style.css\", \"purpose\": \"tokens, layout, states, motion\", \"exports\": [], \"imports\": [], \"notes\": \"prefers-color-scheme\"}, {\"path\": \"README.md\", \"purpose\": \"features, shortcuts, run, file map\", \"exports\": [], \"imports\": [], \"notes\": \"\"}],\n \"acceptance\": [\"space toggles\", \"ring is crisp at dpr 2\", \"reload keeps history\", \"chart updates after COMPLETE\", \"? opens help\", \"delete then ⌘Z restores\", \"tab hidden 5 min → remaining time still correct\", \"opens from disk (file://)\"]}`,\n`{\"title\": \"Tile Editor\", \"summary\": \"2D tilemap editor with layers and export.\",\n \"features\": [\"tileset upload (PNG as data URL file)\", \"grid with zoom/pan (wheel, space-drag)\", \"paint/erase/fill/rect tools (B/E/G/R)\", \"3 layers with visibility/lock\", \"undo/redo 200 steps\", \"minimap\", \"export JSON + PNG\", \"import JSON\", \"autosave schema v2\", \"keyboard shortcut sheet\", \"touch painting\", \"dark tokens\"],\n \"visual\": \"--bg #101214 --panel #191c20 --grid #2a2f36 --accent #7cf0c8; mono UI 12/14; 8 px unit; radius 6; 120 ms\",\n \"shared\": \"state {map:{w,h,layers:[{name, tiles:Uint16Array, visible, locked}]}, tool, sel:{x,y}, view:{x,y,zoom}}; actions PAINT{layer,x,y,id} FILL RECT SET_TOOL SET_VIEW UNDO REDO; render.js draws on dpr canvas; input.js converts client→cell via view; ids: #stage #tools #layers #minimap\",\n \"files\": [… 11 files as above, each ≤ 250 lines …],\n \"acceptance\": [\"fill respects bounds\", \"zoom keeps cursor cell fixed\", \"undo across tools\", \"PNG export pixel-exact\", \"runs from disk\"]}`),\n\n builder: ex(\"files at the standard\",\n`\\`\\`\\`js\n// render.js — dpr-scaled canvas, ResizeObserver, rAF with delta, pauses when hidden\nimport { subscribe, get } from './store.js';\nconst canvas = document.getElementById('ring'); const ctx = canvas.getContext('2d');\nlet w = 0, h = 0, dpr = 1, raf = 0, last = 0, hidden = document.hidden;\nconst ro = new ResizeObserver(([e]) => { const r = e.contentRect; w = r.width; h = r.height; dpr = Math.min(devicePixelRatio || 1, 3); canvas.width = Math.round(w * dpr); canvas.height = Math.round(h * dpr); ctx.setTransform(dpr, 0, 0, dpr, 0, 0); draw(0); });\nro.observe(canvas.parentElement);\ndocument.addEventListener('visibilitychange', () => { hidden = document.hidden; if (!hidden) { last = performance.now(); loop(last); } });\nfunction draw(dt) {\n const s = get(); ctx.clearRect(0, 0, w, h);\n const r = Math.min(w, h) * 0.42, cx = w / 2, cy = h / 2, frac = 1 - s.remainingMs / s.totalMs;\n ctx.lineWidth = 10; ctx.lineCap = 'round';\n ctx.strokeStyle = getComputedStyle(canvas).getPropertyValue('--grid'); ctx.beginPath(); ctx.arc(cx, cy, r, 0, Math.PI * 2); ctx.stroke();\n ctx.strokeStyle = getComputedStyle(canvas).getPropertyValue('--accent'); ctx.beginPath(); ctx.arc(cx, cy, r, -Math.PI / 2, -Math.PI / 2 + frac * Math.PI * 2); ctx.stroke();\n ctx.fillStyle = getComputedStyle(canvas).getPropertyValue('--text'); ctx.font = \\`600 \\${Math.round(r * 0.5)}px system-ui\\`; ctx.textAlign = 'center'; ctx.textBaseline = 'middle';\n ctx.fillText(fmt(s.remainingMs), cx, cy);\n}\nfunction loop(t) { if (hidden) return; const dt = Math.min(0.05, (t - last) / 1000); last = t; draw(dt); raf = requestAnimationFrame(loop); }\nconst fmt = (ms) => { const s = Math.max(0, Math.round(ms / 1000)); return \\`\\${String(Math.floor(s / 60)).padStart(2, '0')}:\\${String(s % 60).padStart(2, '0')}\\`; };\nexport function mount() { subscribe(() => draw(0)); last = performance.now(); raf = requestAnimationFrame(loop); return () => { cancelAnimationFrame(raf); ro.disconnect(); }; }\n\\`\\`\\``,\n`\\`\\`\\`js\n// storage.js — versioned persistence with migration, try/catch, in-memory fallback (file:// and private mode safe)\nexport const SCHEMA_VERSION = 2;\nconst KEY = 'pomodoro.v'; let mem = null;\nconst migrations = { 1: (s) => ({ ...s, sessions: (s.sessions || []).map((x) => ({ ...x, tag: x.tag ?? '' })) }) };\nexport function migrate(doc) { let v = doc.version || 1, data = doc.data; while (v < SCHEMA_VERSION) { data = migrations[v] ? migrations[v](data) : data; v++; } return data; }\nexport function load() {\n try { const raw = localStorage.getItem(KEY); if (raw) return migrate(JSON.parse(raw)); } catch (e) { /* quota, disabled, or corrupt: fall through */ }\n return mem;\n}\nexport function save(state) {\n const doc = JSON.stringify({ version: SCHEMA_VERSION, data: state, savedAt: Date.now() });\n try { localStorage.setItem(KEY, doc); return 'local'; } catch (e) { mem = state; return 'memory'; }\n}\n\\`\\`\\``),\n\n integrate: ex(\"integration passes\",\n`read_artifact #41 all:true → findings: input.js imports { startTicker } from './timer.js' but timer.js exports start(); render.js draws before store exists (load order); README lists 12 features, \"tempo ramp\" not implemented anywhere → one update_artifact: edits (timer.js export rename, index.html script order), files: [ramp.js — full implementation of the ramp: linear BPM interpolation per bar, UI in index.html, actions in store.js]. Then lint is clean; acceptance checks walked one by one in the reply.`,\n`Lint: \"state.js:120 empty function body — function exportJson() {}\" → not deleted, implemented: serialize state (schema version + data), Blob download with a11y-labelled anchor, toast on success; and the matching importJson with validation errors surfaced in the UI.`),\n\n digest: ex(\"digests\",\n`Part 1/3 of error.log: 14 distinct errors. Load-bearing: ui.js:6:11 TypeError \"Cannot read properties of null (reading 'addEventListener')\" — bindUI() runs before #toolbar exists (script in , no defer). storage.js:1 SyntaxError: constants.js has no export named DEFAULT_BOT (it exports DEFAULT_BOT_ID). Repeats: 11× the same TypeError. Numbers: first error at t=0.312 s after load.`,\n`Part 2/2 of api.md: endpoints POST /v1/jobs {name, cron} → 201 {id}; GET /v1/jobs/{id} → {status: 'queued'|'running'|'done'|'failed', attempts ≤ 3}; rate limit 60/min per key (header x-rate-remaining); errors as {error:{code,message}}; timestamps RFC 3339 UTC.`),\n\n completion: ex(\"audits\",\n`{\"done\": false, \"quality\": 2, \"remaining\": [\"Implement the tempo ramp feature named in README (ramp.js: start→end BPM over N bars, UI controls, store actions)\", \"Scale the canvas by devicePixelRatio in render.js\", \"Add keyboard control for every toolbar action and the ? shortcut sheet\", \"Persist settings with SCHEMA_VERSION + migrate() in storage.js\", \"Add empty state for the session log\", \"Replace alert() in input.js with the in-app toast\"]}`,\n`{\"done\": true, \"quality\": 5, \"remaining\": []}`),\n\n brief: ex(\"design briefs\",\n`PRODUCT — A drift-free practice metronome musicians keep open for an hour: sample-accurate, fully keyboard-driven, remembers every preset.\nFEATURES — 1. BPM 20–300 via drag ring, ±1/±10 keys, typed entry. 2. Tap tempo (T), 4-tap median, outlier rejection. 3. Beats/bar 1–12 with per-beat accent editor (click/tap cycles off·low·high). 4. Subdivisions 1/2/3/4 with distinct timbre. 5. Tempo ramp start→end over N bars with live readout. 6. Presets: save/rename/delete/reorder, undo delete. 7. Practice log: minutes per day, sparkline. 8. Visual beat ring + flash, dpr-crisp, pauses when hidden without losing the beat grid. 9. Shortcut sheet (?), all actions reachable by keyboard. 10. Sound sets (click/wood/beep) via WebAudio synthesis, no files. 11. Settings + presets persisted, schema v1 + migrate(). 12. Dark/light tokens, reduced-motion respected.\nVISUAL SYSTEM — --bg #0e0f12 --surface #16181d --text #e8e9ee --dim #8b90a0 --accent #ffb454 --danger #ff5c5c; type 12/14/18/40 (BPM), tabular numerals; 4 px unit; radius 12; 140 ms ease-out. Identity: the beat ring's accent flash is the only saturated element.\nARCHITECTURE — index.html, style.css, main.js (boot), store.js (reducer + undo), audio.js (lookahead scheduler 25/100 ms, ctx.currentTime clock), timer.js (bar/beat counting from audio clock), render.js (ring, log, sparkline), input.js (pointer/keyboard/touch), presets.js, storage.js (v1), README.md. State {bpm, beats, accents[], subdiv, ramp{on,from,to,bars}, running, presets[], log[], settings}.\nEDGE CASES — resize during ramp; 0 presets; 500 log days; BPM typed as \"abc\"; storage disabled; audio context suspended until first gesture (show a \"tap to enable sound\" state); keyboard-only from cold load; two-finger touch on the ring.\nACCEPTANCE — jitter < 1 ms over 10 min (measured against ctx.currentTime); ring crisp at dpr 2; reload restores presets; ? opens sheet; T×4 sets tempo; hidden 10 min → still on beat; runs from file://; reduced-motion disables flash.`,\n`PRODUCT — A tilemap editor a game jammer finishes a level in: tileset in, layers, tools, undo, JSON/PNG out, nothing to install.\nFEATURES — 1. Tileset PNG upload → tile size auto-detect + override. 2. Zoom/pan (wheel, space-drag, pinch) keeping the cursor cell fixed. 3. Paint/erase/fill/rect/pick tools with keys B E G R I. 4. 3 layers: rename, visibility, lock, reorder. 5. Undo/redo 200 steps with stroke coalescing. 6. Minimap with viewport rectangle. 7. Export JSON (schema v2) and PNG (pixel-exact). 8. Import JSON with validation report. 9. Autosave every change, restore on load. 10. Shortcut sheet (?). 11. Touch painting with palm rejection (pointerType). 12. Dark tokens + reduced motion.\nVISUAL SYSTEM — --bg #0f1113 --panel #181b1f --grid #262b31 --accent #6ee7b7 --danger #f87171; mono 12/13, UI 14; 8 px unit; radius 6; 100 ms.\nARCHITECTURE — index.html, style.css, main.js, store.js, history.js, tools.js, render.js, input.js, tileset.js, io.js (export/import), storage.js (v2), README.md. State {map{w,h,tileSize,layers[]}, tileset{img,cols,rows}, tool, brush, view{x,y,zoom}, sel}.\nEDGE CASES — tileset larger than 4096 px; map resize with content; fill on a locked layer; undo past import; storage full; import with wrong schema; keyboard-only; 2-finger pan while painting.\nACCEPTANCE — fill bounded; zoom pivot exact; undo across tools; PNG export matches canvas; reload restores; runs from file://; ? sheet complete; touch paint on iPad.`),\n\n compact: ex(\"running summaries\",\n`GOAL — Browser metronome app (#41 \"Metronome\", 9 files) the user practises with daily; now adding a practice log.\nSTANDING INSTRUCTIONS — no placeholders ever; ≤ 250 lines per file; dark theme default; keyboard for every action.\nSTATE — #41 files: index.html, style.css, main.js, store.js, audio.js (lookahead 25/100 ms), timer.js, render.js, input.js, storage.js (SCHEMA_VERSION 2), README.md.\nDONE — tap tempo (4-tap median); ramp feature (ramp.js added this session); storage migration v1→v2 (accents array).\nOPEN — practice log: user wants minutes/day + sparkline (render.js), persisted (storage v3 with migrate). Undo for preset delete still missing.\nFACTS — user measured jitter 0.6 ms over 10 min; Safari needs AudioContext resume on first pointerdown (fixed in main.js:12); user's display is 120 Hz (delta-time bug found earlier, fixed).`,\n`GOAL — Debug and harden a WebGL2 hologram renderer the user pasted (single file, ~900 lines), then port it to ES modules.\nSTANDING INSTRUCTIONS — Fisher-dense replies, no restating; answers ≤ 5000 chars; artifacts must run from file://.\nSTATE — #58 \"Hologram Renderer\": index.html, gl.js, shaders.js (vert/frag as template strings), scene.js, input.js, README.md.\nDONE — fixed premultiplied-alpha washout (UNPACK_COLORSPACE_CONVERSION_WEBGL = NONE); moved shaders from fetch() to shaders.js for file:// portability; resize now uses ResizeObserver + dpr cap 2.\nOPEN — user reports 45 fps on M1 at 4k: suspected per-frame uniform uploads in scene.js:140–170 (not yet profiled); wants a UBO.\nFACTS — target GPU M1, Safari 17; user rejects any external library; precision: mediump breaks the interference pattern, must stay highp.`),\n};\n",[],[373],"27c3fd61-523d-409e-bdde-9914448a393b","out.txt","01a0a661-f04a-733e-839a-a5e63b438d0c",["D",1789497831502],[],[380],"62e3c28e-8c5e-4e6f-badd-8df7131d554b","openrouter-providers.ts","01a0a661-f04a-733e-839a-a3df813201eb",["D",1789497831502],"// openrouter-providers.ts — ~450 chars\ninterface OpenRouterProvider {\n model: string;\n context: number;\n tier: number; // 1=primary, 2=fallback, 3=tertiary\n costPerMToken?: [number, number]; // [input, output] or undefined if free\n requestsPerDay: number;\n expiryDate?: Date; // e.g., new Date('2026-07-21') for Tencent Hy3\n capabilities: (\"reasoning\" | \"coding\" | \"multimodal\" | \"tool-use\")[];\n}\n\nexport const OPENROUTER_FREE: Record = {\n \"nvidia/nemotron-3-ultra\": {\n model: \"nvidia/nemotron-3-ultra\",\n context: 1_000_000,\n tier: 1,\n requestsPerDay: 1000,\n capabilities: [\"reasoning\", \"coding\", \"tool-use\"],\n },\n \"qwen/qwen3-coder\": {\n model: \"qwen/qwen3-coder\",\n context: 1_000_000,\n tier: 1,\n requestsPerDay: 1000,\n capabilities: [\"coding\", \"tool-use\"],\n },\n \"openai/gpt-oss-120b\": {\n model: \"openai/gpt-oss-120b\",\n context: 131_072,\n tier: 2,\n requestsPerDay: 1000,\n capabilities: [\"reasoning\"],\n },\n \"tencent/hy3-preview\": {\n model: \"tencent/hy3-preview\",\n context: 262_144,\n tier: 2,\n requestsPerDay: 1000,\n expiryDate: new Date(\"2026-07-21\"),\n capabilities: [\"reasoning\"],\n },\n};",[],[386],"87cdbd18-1726-4a18-b03c-5c2dfad3c8d3","omni-client.ts","01a0a661-f04a-733e-839a-9c831faf06e4",["D",1789497831502],"// Drop-in client for any val/project that talks to the router.\n// import { omni } from \"https://esm.town/v/ziggyware/free-ai/omni-client.ts\";\n// const r = await omni.chat([{ role: \"user\", content: \"hi\" }], { model: \"coder\" });\n// r.text, r.message, r.meta.provider.name\n// Env (optional): OMNI_URL (default https://router.val.run), OMNI_CLIENT_KEY.\n// Retries the whole route once on 503/network (the router already fails over\n// across every provider internally; this covers a cold isolate or a blip).\n\nexport type Msg = { role: \"system\" | \"user\" | \"assistant\" | \"tool\"; content: string; [k: string]: unknown };\nexport type ChatOpts = {\n model?: string; tools?: unknown[]; tool_choice?: unknown; temperature?: number; max_tokens?: number;\n response_format?: unknown; signal?: AbortSignal; retries?: number; url?: string; key?: string;\n};\nexport type ChatResult = {\n message: { role: string; content: string | null; tool_calls?: unknown[] };\n text: string; finish_reason: string; model: string;\n meta: { provider: { name: string; vendor: string }; latency_ms: number; attempts: number };\n raw: any;\n};\n\nconst base = (o: ChatOpts) => (o.url ?? Deno.env.get(\"OMNI_URL\") ?? \"https://router.val.run\").replace(/\\/$/, \"\");\nconst headers = (o: ChatOpts) => {\n const h: Record = { \"Content-Type\": \"application/json\" };\n const k = o.key ?? Deno.env.get(\"OMNI_CLIENT_KEY\");\n if (k) h.Authorization = `Bearer ${k}`;\n return h;\n};\nconst body = (messages: Msg[], o: ChatOpts, stream: boolean) =>\n JSON.stringify({ messages, model: o.model, tools: o.tools, tool_choice: o.tool_choice, temperature: o.temperature, max_tokens: o.max_tokens, response_format: o.response_format, stream });\n\nasync function chat(messages: Msg[], o: ChatOpts = {}): Promise {\n const tries = 1 + (o.retries ?? 1);\n let last: Error = new Error(\"unreachable\");\n for (let i = 0; i < tries; i++) {\n try {\n const res = await fetch(`${base(o)}/v1/chat/completions`, { method: \"POST\", headers: headers(o), body: body(messages, o, false), signal: o.signal });\n const j = await res.json();\n if (!res.ok || j.error) {\n const e = Object.assign(new Error(j?.error?.message ?? `HTTP ${res.status}`), { status: res.status, details: j?.error?.details });\n if (res.status === 503 && i + 1 < tries) { last = e; await new Promise((r) => setTimeout(r, 400 * (i + 1))); continue; }\n throw e;\n }\n const c = j.choices[0];\n return { message: c.message, text: c.message?.content ?? \"\", finish_reason: c.finish_reason ?? \"stop\", model: j.model, meta: j._omni_meta, raw: j };\n } catch (e: any) {\n last = e;\n if (e.status && e.status !== 503) throw e;\n }\n }\n throw last;\n}\n\n/** Async iterator of content deltas. `for await (const tok of omni.stream(msgs)) …` */\nasync function* stream(messages: Msg[], o: ChatOpts = {}): AsyncGenerator {\n const res = await fetch(`${base(o)}/v1/chat/completions`, { method: \"POST\", headers: headers(o), body: body(messages, o, true), signal: o.signal });\n if (!res.ok || !res.body) throw Object.assign(new Error((await res.json().catch(() => ({})))?.error?.message ?? `HTTP ${res.status}`), { status: res.status });\n const rd = res.body.pipeThrough(new TextDecoderStream()).getReader();\n let buf = \"\";\n for (;;) {\n const { done, value } = await rd.read();\n if (done) return;\n buf += value;\n let nl: number;\n while ((nl = buf.indexOf(\"\\n\")) >= 0) {\n const line = buf.slice(0, nl).trim(); buf = buf.slice(nl + 1);\n if (!line.startsWith(\"data:\")) continue;\n const data = line.slice(5).trim();\n if (data === \"[DONE]\") return;\n try { const d = JSON.parse(data).choices?.[0]?.delta?.content; if (d) yield d; } catch { /* keepalive / partial */ }\n }\n }\n}\n\nconst models = async (o: ChatOpts = {}) => (await (await fetch(`${base(o)}/v1/models`, { headers: headers(o) })).json()).data as { id: string }[];\nconst health = async (o: ChatOpts = {}) => await (await fetch(`${base(o)}/health`)).json();\n\nexport const omni = { chat, stream, models, health };\n",[],[393],"7c896d46-da1c-44cf-a683-85382a8f11e8","mcp-client.ts","01a0a661-f04a-733e-839a-9b0bd83ccbd0",["D",1789497831502],"// ════════════════════════════════════════════════════════════════════════════\n// MCP CLIENT — Streamable HTTP transport, zero dependencies.\n//\n// Connects this app's agent loop to Model Context Protocol servers and\n// surfaces their tools alongside the built-in registry. The first (and\n// motivating) server is `rojs`, the local Roslyn C# analysis sidecar: its\n// MCP bridge is `mcp-server.mjs` in the rojs repo, run with `--http` so it\n// speaks Streamable HTTP on http://localhost:5081/mcp.\n//\n// Deployment note: this code runs server-side (Val Town / Deno). From the\n// Val Town cloud, \"localhost\" is the val's own sandbox — to reach a rojs\n// instance on your machine you must tunnel it (e.g. `cloudflared tunnel\n// --url http://localhost:5081`) and set:\n// ROJS_MCP_URL = https:///mcp\n// ROJS_MCP_TOKEN = \n// Running the app locally under plain Deno, the default localhost URL works\n// as-is. Additional servers: MCP_SERVERS = JSON array of\n// {\"name\":\"...\", \"url\":\"https://.../mcp\", \"token\":\"...\"}.\n//\n// Failure posture: MCP servers are OPTIONAL capability, never a dependency.\n// Discovery has a short timeout and a negative-result cache, so an offline\n// sidecar costs one failed fetch per minute, not one per message.\n// ════════════════════════════════════════════════════════════════════════════\n\nexport interface McpServerConfig {\n name: string;\n url: string;\n token?: string;\n}\n\n// Structural twin of app.tsx's Tool — declared here rather than imported so\n// the dependency points app → mcp-client only.\nexport interface McpTool {\n schema: Record;\n handle: (\n args: Record,\n ctx: { session: string },\n ) => Promise;\n}\n\ninterface McpToolDef {\n name: string;\n description?: string;\n inputSchema?: Record;\n}\n\nconst DISCOVER_TIMEOUT_MS = 3_000; // a turn must not stall on a dead sidecar\nconst CALL_TIMEOUT_MS = 30_000; // Roslyn session creation can be slow\nconst CACHE_TTL_MS = 60_000;\n\n// ── Config ────────────────────────────────────────────────────────────────────\n\nfunction env(key: string): string | undefined {\n try {\n return Deno.env.get(key) ?? undefined;\n } catch {\n return undefined; // no --allow-env for this var; treat as unset\n }\n}\n\nexport function mcpServerConfigs(): McpServerConfig[] {\n const servers: McpServerConfig[] = [];\n const raw = env(\"MCP_SERVERS\");\n if (raw) {\n try {\n const parsed = JSON.parse(raw);\n if (Array.isArray(parsed)) {\n for (const s of parsed) {\n if (s && typeof s.name === \"string\" && typeof s.url === \"string\") {\n servers.push({ name: s.name, url: s.url, token: s.token });\n }\n }\n }\n } catch {\n console.warn(\"MCP_SERVERS is not valid JSON; ignoring.\");\n }\n }\n // rojs is wired in by default: absent config it points at the local bridge,\n // and simply fails discovery (cheaply) when the bridge isn't running.\n if (!servers.some((s) => s.name === \"rojs\")) {\n servers.push({\n name: \"rojs\",\n url: env(\"ROJS_MCP_URL\") ?? \"http://localhost:5081/mcp\",\n token: env(\"ROJS_MCP_TOKEN\"),\n });\n }\n return servers;\n}\n\n// ── Wire protocol ─────────────────────────────────────────────────────────────\n\nclass McpClient {\n private nextId = 1;\n private sessionId: string | null = null;\n private initialized = false;\n private cfg: McpServerConfig;\n\n constructor(cfg: McpServerConfig) {\n this.cfg = cfg;\n }\n\n private headers(): Record {\n const h: Record = {\n \"content-type\": \"application/json\",\n // Spec requires the client to accept both response modes even though\n // we only ever get one message back per POST.\n \"accept\": \"application/json, text/event-stream\",\n };\n if (this.cfg.token) h[\"authorization\"] = `Bearer ${this.cfg.token}`;\n if (this.sessionId) h[\"mcp-session-id\"] = this.sessionId;\n return h;\n }\n\n // POST one JSON-RPC message; return the parsed response (or null for a\n // notification's 202). Handles both application/json and text/event-stream\n // response bodies, since Streamable HTTP servers may use either.\n private async post(\n payload: Record,\n timeoutMs: number,\n ): Promise | null> {\n const res = await fetch(this.cfg.url, {\n method: \"POST\",\n headers: this.headers(),\n body: JSON.stringify(payload),\n signal: AbortSignal.timeout(timeoutMs),\n });\n const sid = res.headers.get(\"mcp-session-id\");\n if (sid) this.sessionId = sid;\n\n if (res.status === 202) {\n await res.body?.cancel();\n return null;\n }\n const text = await res.text();\n if (!res.ok) {\n throw new Error(`${this.cfg.name} HTTP ${res.status}: ${text.slice(0, 200)}`);\n }\n const ct = res.headers.get(\"content-type\") ?? \"\";\n if (ct.includes(\"text/event-stream\")) {\n // Minimal SSE parse: each event's data lines are one JSON-RPC message;\n // the response to our request is the one carrying our id.\n const id = (payload as { id?: unknown }).id;\n // Per the SSE spec, a blank line (event boundary) is any of \\n\\n, \\r\\n\\r\\n, or \\r\\r — splitting on\n // \"\\n\\n\" alone silently mis-parses (or entirely misses) events from a server that sends CRLF line\n // endings, since \"\\r\\n\\r\\n\" doesn't contain a literal \"\\n\\n\" substring in a way this split expects\n // consistently once \\r sticks to the preceding line. Normalize line endings before splitting instead.\n for (const chunk of text.replace(/\\r\\n/g, \"\\n\").split(\"\\n\\n\")) {\n const data = chunk.split(\"\\n\")\n .filter((l) => l.startsWith(\"data:\"))\n .map((l) => l.slice(5).trim())\n .join(\"\\n\");\n if (!data) continue;\n try {\n const msg = JSON.parse(data);\n if (msg && msg.id === id) return msg;\n } catch { /* keep-alives, non-JSON events */ }\n }\n throw new Error(`${this.cfg.name}: no response for id ${id} in SSE stream`);\n }\n return text ? JSON.parse(text) : null;\n }\n\n private async rpc(\n method: string,\n params?: Record,\n timeoutMs = CALL_TIMEOUT_MS,\n ): Promise {\n const msg = await this.post(\n { jsonrpc: \"2.0\", id: this.nextId++, method, ...(params ? { params } : {}) },\n timeoutMs,\n );\n if (!msg) throw new Error(`${this.cfg.name}: empty response to ${method}`);\n const err = msg.error as { code?: number; message?: string } | undefined;\n if (err) throw new Error(`${this.cfg.name} ${method}: ${err.message ?? \"error\"}`);\n return msg.result;\n }\n\n private async ensureInit(timeoutMs: number): Promise {\n if (this.initialized) return;\n await this.rpc(\"initialize\", {\n protocolVersion: \"2025-03-26\",\n capabilities: {},\n clientInfo: { name: \"freeai-chat\", version: \"1.0.0\" },\n }, timeoutMs);\n // Notification: no id, expect 202/no body.\n await this.post(\n { jsonrpc: \"2.0\", method: \"notifications/initialized\" },\n timeoutMs,\n );\n this.initialized = true;\n }\n\n async listTools(timeoutMs = DISCOVER_TIMEOUT_MS): Promise {\n await this.ensureInit(timeoutMs);\n const r = (await this.rpc(\"tools/list\", undefined, timeoutMs)) as {\n tools?: McpToolDef[];\n };\n return r?.tools ?? [];\n }\n\n async callTool(\n name: string,\n args: Record,\n ): Promise {\n await this.ensureInit(DISCOVER_TIMEOUT_MS);\n const r = (await this.rpc(\"tools/call\", { name, arguments: args })) as {\n content?: Array<{ type: string; text?: string }>;\n isError?: boolean;\n };\n const text = (r?.content ?? [])\n .filter((c) => c.type === \"text\" && typeof c.text === \"string\")\n .map((c) => c.text)\n .join(\"\\n\");\n if (r?.isError) return { error: text || \"tool error\" };\n // Tool outputs are JSON-serialized by convention; fall back to raw text.\n try {\n return JSON.parse(text);\n } catch {\n return text;\n }\n }\n\n // A failed session (network flap, server restart) must not poison the\n // client forever — drop protocol state so the next call re-initializes.\n reset(): void {\n this.initialized = false;\n this.sessionId = null;\n }\n}\n\n// Clients persist across discovery cycles so MCP session state (and the\n// initialize handshake) isn't redone per chat message.\nconst clients = new Map();\n\nfunction clientFor(cfg: McpServerConfig): McpClient {\n let c = clients.get(cfg.name);\n if (!c) {\n c = new McpClient(cfg);\n clients.set(cfg.name, c);\n }\n return c;\n}\n\n// ── Tool registry ─────────────────────────────────────────────────────────────\n\n// OpenAI function-name charset; MCP tool names are looser.\nfunction toolKey(server: string, tool: string): string {\n return `${server}_${tool}`.replace(/[^a-zA-Z0-9_-]/g, \"_\").slice(0, 64);\n}\n\nlet cache: { at: number; tools: Record } | null = null;\n\n/**\n * Discover tools from every configured MCP server. Cached for CACHE_TTL_MS —\n * including negative results, so an offline server is retried once a minute\n * rather than once a message. Never throws.\n */\nexport async function discoverMcpTools(): Promise> {\n const now = Date.now();\n if (cache && now - cache.at < CACHE_TTL_MS) return cache.tools;\n\n const tools: Record = {};\n await Promise.all(mcpServerConfigs().map(async (cfg) => {\n const client = clientFor(cfg);\n let defs: McpToolDef[];\n try {\n defs = await client.listTools();\n } catch (e) {\n client.reset();\n console.warn(`MCP server \"${cfg.name}\" unavailable: ${String(e).slice(0, 160)}`);\n return;\n }\n for (const def of defs) {\n const key = toolKey(cfg.name, def.name);\n tools[key] = {\n schema: {\n type: \"function\",\n function: {\n name: key,\n description: def.description ?? `${cfg.name}: ${def.name}`,\n parameters: def.inputSchema ?? { type: \"object\", properties: {} },\n },\n },\n handle: async (args) => {\n try {\n return await client.callTool(def.name, args);\n } catch (e) {\n client.reset();\n return { error: String(e) };\n }\n },\n };\n }\n }));\n\n cache = { at: now, tools };\n return tools;\n}",[],[400],"1f8f9229-4f6f-489e-bc12-1f82a4a9d4fe","main-style.ts","01a0a661-f04a-733e-839a-94794ebba47a",["D",1789497831502],"// Base tokens + shell layout. Principle: one surface at a time (chat | build | runs);\n// nothing persistent on screen that is not actionable in the current mode.\nexport const style: string = `\n@import url('https://fonts.googleapis.com/css2?family=IBM+Plex+Mono:ital,wght@0,300;0,400;0,500;0,600;1,300;1,400&family=IBM+Plex+Sans:wght@300;400;500&display=swap');\n*{box-sizing:border-box;margin:0;padding:0;-webkit-tap-highlight-color:transparent;}\n:root{--bg:#05060a;--s0:#080914;--s1:#0d101c;--s2:#141827;--s3:#1b2132;--line:#262d3f;--line2:#31384a;--text:#e4e8f4;--dim:#9aa2b8;--dim2:#5f6678;--acc:#4f8cff;--acc2:#c6ff7f;--err:#ff6b6b;--warn:#ffb347;--inf:#4f8cff;--mono:'IBM Plex Mono',monospace;--sans:'IBM Plex Sans',sans-serif;--r:2px;--content-w:760px;--fs:13px;}\nhtml,body{height:100dvh;overflow:hidden;position:fixed;width:100%;background:var(--bg);}\nbody{font-family:var(--mono);color:var(--text);display:flex;font-size:12px;}\nbutton{font-family:inherit;}\n::-webkit-scrollbar{width:8px;height:8px;} ::-webkit-scrollbar-thumb{background:var(--line2);} ::-webkit-scrollbar-track{background:transparent;}\n.ib{background:none;border:0;color:var(--dim);width:28px;height:28px;display:inline-flex;align-items:center;justify-content:center;cursor:pointer;border-radius:var(--r);} .ib:hover{color:var(--acc);background:var(--s1);} .ib svg{width:15px;height:15px;}\n.act-btn{background:none;border:1px solid var(--line2);color:var(--dim);font:inherit;font-size:10px;letter-spacing:.1em;text-transform:uppercase;padding:3px 8px;cursor:pointer;white-space:nowrap;} .act-btn:hover{border-color:var(--acc);color:var(--acc);} .act-btn.primary{border-color:var(--acc);color:var(--acc);} .act-btn:disabled{opacity:.4;cursor:default;}\n\n/* ── sidebar ── */\n#sidebar{width:240px;flex-shrink:0;background:var(--s0);border-right:1px solid var(--line);display:flex;flex-direction:column;transition:margin-left .18s;}\n#sidebar.collapsed{margin-left:-240px;}\n#sb-hdr{display:flex;align-items:center;justify-content:space-between;padding:10px 8px 10px 14px;border-bottom:1px solid var(--line);}\n.wordmark{font-size:11px;letter-spacing:.3em;text-transform:uppercase;color:var(--acc);font-weight:500;}\n#new-btn{margin:8px;display:flex;align-items:center;gap:8px;background:none;border:1px solid var(--line2);color:var(--dim);font:inherit;font-size:11px;padding:7px 10px;cursor:pointer;} #new-btn:hover{border-color:var(--acc);color:var(--acc);} #new-btn svg{width:12px;height:12px;}\n#sessions{flex:1;overflow:auto;padding:4px 0;}\n.sess-row{display:flex;align-items:center;padding:6px 8px 6px 14px;cursor:pointer;color:var(--dim);font-size:11px;gap:6px;} .sess-row:hover{background:var(--s1);color:var(--text);} .sess-row.active{color:var(--acc);background:var(--s1);border-left:2px solid var(--acc);padding-left:12px;}\n.sess-name{flex:1;overflow:hidden;text-overflow:ellipsis;white-space:nowrap;letter-spacing:.04em;} .sess-del{background:none;border:0;color:var(--dim2);width:18px;height:18px;cursor:pointer;opacity:0;display:inline-flex;align-items:center;justify-content:center;} .sess-row:hover .sess-del{opacity:1;} .sess-del:hover{color:var(--err);} .sess-del svg{width:10px;height:10px;}\n\n/* ── main shell ── */\n#main{flex:1;display:flex;flex-direction:column;overflow:hidden;height:100dvh;min-width:0;}\n#top{display:flex;align-items:center;gap:6px;padding:6px 10px;border-bottom:1px solid var(--line);background:var(--s0);flex-shrink:0;}\n#session-title{font-size:11px;letter-spacing:.14em;text-transform:uppercase;color:var(--text);cursor:text;overflow:hidden;text-overflow:ellipsis;white-space:nowrap;flex:1;min-width:0;} #session-title-input{display:none;flex:1;background:var(--s1);border:1px solid var(--acc);color:var(--text);font:inherit;font-size:11px;letter-spacing:.14em;text-transform:uppercase;padding:3px 6px;outline:0;}\n.modes{display:flex;border:1px solid var(--line2);} .modes button{background:none;border:0;border-right:1px solid var(--line2);color:var(--dim2);font:inherit;font-size:10px;letter-spacing:.14em;text-transform:uppercase;padding:4px 12px;cursor:pointer;} .modes button:last-child{border-right:0;} .modes button.on{color:var(--acc);background:var(--s1);} .modes button:hover{color:var(--text);}\n/* an artifact updated while you were on another tab — badge, never auto-switch (see ART.refresh in ui-artifacts-script.ts) */\n.modes button.has-update:not(.on){color:var(--acc);} .modes button.has-update::after{content:\"\";display:inline-block;width:6px;height:6px;border-radius:50%;background:var(--warn);margin-left:6px;vertical-align:middle;animation:blink 1s infinite;}\n.pill{font-size:10px;letter-spacing:.08em;color:var(--dim);border:1px solid var(--line2);padding:3px 8px;cursor:pointer;white-space:nowrap;display:inline-flex;gap:6px;align-items:center;} .pill:hover{border-color:var(--acc);color:var(--acc);} .pill b{color:var(--acc);font-weight:500;} .pill .dot{width:6px;height:6px;border-radius:50%;background:var(--acc2);} .pill.busy .dot{background:var(--warn);animation:blink 1s infinite;} .pill.err .dot{background:var(--err);}\n@keyframes blink{50%{opacity:.3}}\n.view{display:none;flex:1;min-height:0;flex-direction:column;} .view.on{display:flex;}\n\n/* ── chat ── */\n#chat-wrap{flex:1;position:relative;min-height:0;} #chat{height:100%;overflow-y:auto;padding:14px 16px 24px;}\n#chat>.msg-wrap,#typing,#empty,.bar-inner,.bar-meta,#jobs{max-width:var(--content-w);margin-left:auto;margin-right:auto;width:100%;}\n#empty{display:flex;flex-direction:column;align-items:center;justify-content:center;height:100%;color:var(--dim2);gap:10px;} .empty-mark{font-size:10px;letter-spacing:.3em;text-transform:uppercase;color:var(--dim2);} #empty h3{font-family:var(--sans);font-weight:300;font-size:22px;color:var(--dim);} .hints{display:flex;gap:6px;flex-wrap:wrap;justify-content:center;margin-top:8px;} .hint{border:1px solid var(--line);color:var(--dim);font-size:10px;padding:4px 10px;cursor:pointer;} .hint:hover{border-color:var(--acc);color:var(--acc);}\n.msg-wrap{padding:8px 0;animation:rise .25s ease;} .msg-wrap.sys .role-tag{color:var(--warn);} @keyframes rise{from{opacity:0;transform:translateY(4px)}to{opacity:1;transform:none}}\n.msg-label{display:flex;align-items:center;gap:8px;margin-bottom:5px;font-size:10px;letter-spacing:.12em;text-transform:uppercase;color:var(--dim2);}\n.role-tag{color:var(--dim);} .user .role-tag{color:var(--acc);} .msg-ts{color:var(--dim2);} .provider-tag{color:var(--dim2);text-transform:none;letter-spacing:0;font-size:10px;overflow:hidden;text-overflow:ellipsis;white-space:nowrap;max-width:50%;}\n.msg-body{font-size:var(--fs);line-height:1.6;} .user .msg-body{color:var(--dim);white-space:pre-wrap;word-break:break-word;border-left:2px solid var(--line2);padding-left:10px;}\n.msg-actions{display:flex;gap:4px;margin-top:6px;opacity:0;transition:opacity .15s;} .msg-wrap:hover .msg-actions{opacity:1;} .msg-actions button{background:none;border:0;color:var(--dim2);font:inherit;font-size:10px;letter-spacing:.1em;text-transform:uppercase;cursor:pointer;padding:2px 6px;} .msg-actions button:hover{color:var(--acc);}\n.edit-area{width:100%;background:var(--s1);color:var(--text);border:1px solid var(--acc);font:inherit;font-size:var(--fs);padding:8px;resize:none;outline:0;line-height:1.5;} .edit-btns{display:flex;gap:6px;margin-top:6px;} .edit-btn{background:none;border:1px solid var(--line2);color:var(--dim);font:inherit;font-size:10px;letter-spacing:.1em;text-transform:uppercase;padding:3px 8px;cursor:pointer;} .edit-btn.confirm{border-color:var(--acc);color:var(--acc);}\n.telemetry{display:inline-flex;gap:8px;font-size:10px;color:var(--dim2);margin-top:6px;flex-wrap:wrap;} .telemetry span b{color:var(--dim);font-weight:500;}\n.app-card{margin:8px 0;border:1px solid var(--line);background:var(--s0);} .app-card-bar{display:flex;align-items:center;gap:8px;padding:5px 10px;border-bottom:1px solid var(--line);font-size:10px;letter-spacing:.08em;} .app-card-title{flex:1;color:var(--acc2);text-transform:uppercase;overflow:hidden;text-overflow:ellipsis;white-space:nowrap;} .app-card-open{background:none;border:0;color:var(--dim);font:inherit;font-size:10px;letter-spacing:.1em;text-transform:uppercase;cursor:pointer;text-decoration:none;} .app-card-open:hover{color:var(--acc);} .app-card-frame{width:100%;height:300px;border:0;background:#fff;display:block;}\n.manifest{margin:8px 0;border:1px solid var(--line);background:var(--s0);font-size:11px;} .manifest-h{display:flex;gap:8px;align-items:center;padding:5px 10px;border-bottom:1px solid var(--line);font-size:10px;letter-spacing:.1em;text-transform:uppercase;color:var(--acc);} .manifest-h span{flex:1;} .manifest-f{display:flex;gap:8px;align-items:center;padding:3px 10px;border-bottom:1px solid var(--line);color:var(--dim);} .manifest-f:last-child{border-bottom:0;} .manifest-f .st{font-size:9px;letter-spacing:.12em;text-transform:uppercase;min-width:64px;color:var(--dim2);} .manifest-f .st.run{color:var(--warn);} .manifest-f .st.ok{color:var(--acc2);} .manifest-f .st.err{color:var(--err);} .manifest-f code{color:var(--text);} .manifest-f i{flex:1;color:var(--dim2);font-style:normal;overflow:hidden;text-overflow:ellipsis;white-space:nowrap;}\n#typing{display:none;padding:8px 0;} .typing-inner{display:flex;align-items:center;gap:10px;font-size:10px;letter-spacing:.12em;text-transform:uppercase;color:var(--dim);} .typing-dots{display:flex;gap:3px;} .dot{width:4px;height:4px;background:var(--acc);border-radius:50%;animation:blink 1.2s infinite;} .dot:nth-child(2){animation-delay:.2s} .dot:nth-child(3){animation-delay:.4s}\n#scroll-pill{position:absolute;bottom:12px;left:50%;transform:translateX(-50%);background:var(--s2);border:1px solid var(--line2);color:var(--dim);font-size:10px;letter-spacing:.1em;padding:4px 10px;cursor:pointer;opacity:0;pointer-events:none;transition:opacity .15s;} #scroll-pill.visible{opacity:1;pointer-events:auto;}\n.err-inline{color:var(--err);}\n\n/* ── composer ── */\n#bar{border-top:1px solid var(--line);background:var(--s0);padding:8px 16px 10px;flex-shrink:0;}\n.chips{display:flex;gap:4px;flex-wrap:wrap;margin-bottom:6px;min-height:0;} .chip{font-size:10px;letter-spacing:.06em;border:1px solid var(--line2);color:var(--dim);padding:2px 8px;cursor:pointer;display:inline-flex;gap:5px;align-items:center;} .chip:hover{border-color:var(--acc);color:var(--acc);} .chip.on{border-color:var(--acc);color:var(--acc);} .chip .x{color:var(--dim2);} .chip .x:hover{color:var(--err);}\n.bar-inner{display:flex;align-items:flex-end;gap:8px;border:1px solid var(--line2);background:var(--bg);padding:6px 6px 6px 10px;} .bar-inner:focus-within{border-color:var(--acc);}\n#inp{flex:1;background:none;border:0;color:var(--text);font:inherit;font-size:var(--fs);line-height:1.5;resize:none;outline:0;max-height:180px;padding:4px 0;} #inp::placeholder{color:var(--dim2);}\n#char-count{font-size:10px;color:var(--dim2);} #char-count.warn{color:var(--warn);} #char-count.danger{color:var(--err);}\n#send{background:var(--acc);border:0;color:#fff;width:30px;height:30px;display:inline-flex;align-items:center;justify-content:center;cursor:pointer;flex-shrink:0;} #send svg{width:14px;height:14px;} #send.stop{background:var(--err);}\n.bar-meta{display:flex;gap:10px;align-items:center;margin-top:6px;font-size:10px;color:var(--dim2);letter-spacing:.06em;} .bar-meta .sp{flex:1;} .bar-meta kbd{border:1px solid var(--line2);padding:0 4px;color:var(--dim2);font:inherit;font-size:9px;}\n#slash{position:absolute;left:0;right:0;bottom:100%;margin:0 auto;max-width:var(--content-w);background:var(--s1);border:1px solid var(--line2);display:none;z-index:5;} #slash.on{display:block;} .slash-i{display:flex;gap:10px;padding:5px 10px;font-size:11px;cursor:pointer;} .slash-i b{color:var(--acc);font-weight:500;min-width:90px;} .slash-i span{color:var(--dim);} .slash-i.sel,.slash-i:hover{background:var(--s2);}\n#bar{position:relative;}\n\n/* ── runs ── */\n#runs{flex:1;overflow:auto;padding:12px 16px;} .run{border:1px solid var(--line);background:var(--s0);margin:0 auto 8px;max-width:var(--content-w);} .run-h{display:grid;grid-template-columns:70px 1fr auto;gap:10px;padding:6px 10px;cursor:pointer;font-size:11px;align-items:center;} .run-h:hover{background:var(--s1);} .run-h .t{color:var(--dim2);font-size:10px;} .run-h .q{overflow:hidden;text-overflow:ellipsis;white-space:nowrap;color:var(--text);} .run-h .m{display:flex;gap:8px;font-size:10px;color:var(--dim2);white-space:nowrap;} .run-h .m b{color:var(--dim);font-weight:500;}\n.run-b{display:none;border-top:1px solid var(--line);padding:8px 10px;} .run.open .run-b{display:block;} .run-kv{display:grid;grid-template-columns:repeat(auto-fill,minmax(150px,1fr));gap:4px 12px;font-size:10px;color:var(--dim2);margin-bottom:8px;} .run-kv b{color:var(--dim);font-weight:500;}\n.run-reply{font-size:11px;color:var(--dim);white-space:pre-wrap;max-height:200px;overflow:auto;border-top:1px solid var(--line);padding-top:6px;margin-top:6px;}\n#runs-empty{color:var(--dim2);text-align:center;padding:40px;font-size:11px;letter-spacing:.1em;}\n\n/* ── palette ── */\n#pal{position:fixed;inset:0;background:rgba(0,0,0,.45);display:none;z-index:60;align-items:flex-start;justify-content:center;padding-top:12vh;} #pal.on{display:flex;}\n#pal-box{width:min(560px,94vw);background:var(--s0);border:1px solid var(--line2);box-shadow:0 20px 60px rgba(0,0,0,.5);} #pal-in{width:100%;background:var(--bg);border:0;border-bottom:1px solid var(--line);color:var(--text);font:inherit;font-size:13px;padding:10px 12px;outline:0;}\n#pal-list{max-height:50vh;overflow:auto;} .pal-i{display:flex;gap:10px;align-items:center;padding:7px 12px;cursor:pointer;font-size:11px;} .pal-i.sel,.pal-i:hover{background:var(--s1);} .pal-i .k{font-size:9px;letter-spacing:.12em;text-transform:uppercase;color:var(--dim2);min-width:60px;} .pal-i .l{flex:1;color:var(--text);} .pal-i .h{color:var(--dim2);font-size:10px;}\n@media (max-width:900px){#sidebar{position:fixed;left:0;top:0;height:100dvh;z-index:20;} .modes button{padding:4px 8px;} .pill.opt{display:none;}}\n\n#live-draft{margin:6px 0 0;border:1px solid var(--line);border-radius:8px;background:var(--surface);max-height:180px;overflow:auto}#live-draft .ld-h{font-size:10px;letter-spacing:.06em;text-transform:uppercase;color:var(--dim);padding:4px 8px;border-bottom:1px solid var(--line)}#live-draft pre{margin:0;padding:6px 8px;font-size:11px;white-space:pre-wrap;word-break:break-word;color:var(--text)}\n`;\n",[],[407],"49382e44-b90e-4a34-90be-f8cac6d91f5b","main-script.ts","01a0a661-f04a-733e-839a-904e65b21ccd",["D",1789497831502],"// Shell logic: sessions, messages (live + restored from message_meta), turns with\n// cross-turn auto-continue, slash commands, the swarm orchestrator (/build), the Runs\n// inspector and the command palette. Sends carry SETTINGS.forRequest() + workbench focus.\nexport const script: string = String.raw`\nconst $ = (id) => document.getElementById(id);\nconst liveDraft = $('live-draft'), liveDraftT = $('live-draft-t'), liveDraftN = $('live-draft-n');\nconst chat = $('chat'), inp = $('inp'), typing = $('typing'), typingLbl = $('typing-label'), empty = $('empty'), sendBtn = $('send'), charEl = $('char-count'), pill = $('scroll-pill'), sessionsEl = $('sessions'), titleEl = $('session-title'), titleInp = $('session-title-input'), sidebar = $('sidebar'), progressEl = $('progress'), liveThink = $('live-think'), liveTl = $('live-tl'), chipsEl = $('chips'), barStatus = $('bar-status'), barTokens = $('bar-tokens');\nconst urlSession = new URLSearchParams(location.search).get('s');\nlet currentSession = urlSession || (() => { try { return localStorage.getItem('omni.session'); } catch { return null; } })() || crypto.randomUUID();\nlet msgCount = 0, inflight = false, sidebarOpen = false, aborter = null, view = 'chat', lastAiWrap = null, turnLog = [];\nObject.defineProperty(window, 'currentSession', { get: () => currentSession });\nwindow.inflightTurn = () => inflight;\n// ── longevity helpers ──────────────────────────────────────────────────────────\n// wait(): timers in a Web Worker are not subject to background-tab throttling (Chrome parks page timers at 1/min after\n// 5 min hidden), so retry / continue / swarm loops keep their cadence while the user is on another tab. Falls back to\n// setTimeout when workers or blob URLs are unavailable.\nconst wait = (() => {\n let w = null, seq = 0; const pend = new Map();\n // onerror previously just dropped the worker (w = null) without touching pend: any wait() already in flight when the\n // worker errors had its resolver stranded in pend forever, hanging whatever awaited it (retry loops, swarm workers,\n // auto-continue, /finalize's unwind wait) with no way out. Resolve everyone waiting before abandoning the worker.\n try { w = new Worker(URL.createObjectURL(new Blob(['onmessage=function(e){setTimeout(function(){postMessage(e.data.id)},e.data.ms)}'], { type: 'text/javascript' }))); w.onmessage = (e) => { const r = pend.get(e.data); if (r) { pend.delete(e.data); r(); } }; w.onerror = () => { for (const r of pend.values()) r(); pend.clear(); w = null; }; } catch { w = null; }\n return (ms) => new Promise((res) => { if (!w) return setTimeout(res, ms); const id = ++seq; pend.set(id, res); w.postMessage({ id, ms }); });\n})();\nconst jitter = (ms) => Math.round(ms * (0.85 + Math.random() * 0.3));\n// fetchT(): the val has a 60 s wall clock, so a request with no answer after \\`ms\\` is dead — abort it and surface a\n// retryable TIMEOUT error instead of hanging the turn forever. The caller's own signal (stop button) still wins.\nasync function fetchT(url, opts = {}, ms = 95_000) {\n const c = new AbortController(); const outer = opts.signal; const onAbort = () => c.abort();\n if (outer) { if (outer.aborted) throw Object.assign(new Error('stopped'), { name: 'AbortError' }); outer.addEventListener('abort', onAbort); }\n const t = setTimeout(() => c.abort(), ms);\n try { return await fetch(url, { ...opts, signal: c.signal }); }\n catch (e) { if (outer?.aborted) throw Object.assign(new Error('stopped'), { name: 'AbortError' }); if (c.signal.aborted) throw new Error('TIMEOUT: no response after ' + Math.round(ms / 1000) + 's'); throw e; }\n finally { clearTimeout(t); outer?.removeEventListener('abort', onAbort); }\n}\n// The server may have finished a turn whose response never reached us (timeout, network blip): the reply is in history.\n// sessionId is required, not defaulted to the live currentSession: recovery must check the history of the session the\n// turn actually belongs to, not whatever session happens to be on screen when the timeout fires (see sendWith's own()).\nasync function recoverFromHistory(text, sessionId) { const h = await fetch('?history=' + encodeURIComponent(sessionId)).then(r => r.json()).catch(() => []); const last = h[h.length - 1]; return last?.role === 'assistant' && h[h.length - 2]?.content === text ? last : null; }\nwindow.omniNet = { wait, fetchT, jitter }; /* exposed for the workbench script and tests */\nconst baseTitle = document.title;\nfunction tabProgress(t) { document.title = t ? '⏳ ' + t + ' · ' + baseTitle : baseTitle; }\nwindow.addEventListener('beforeunload', (e) => { if (inflight) { e.preventDefault(); e.returnValue = 'A turn is still running — leaving loses its live progress (finished work is saved).'; } });\nfunction rememberSession() { try { localStorage.setItem('omni.session', currentSession); } catch {} history.replaceState(null, '', location.pathname + '?s=' + encodeURIComponent(currentSession)); }\nconst ts = () => new Date().toLocaleTimeString([], { hour: '2-digit', minute: '2-digit', hour12: false });\nconst fmtTs = (t) => t ? new Date(t).toLocaleTimeString([], { hour: '2-digit', minute: '2-digit', hour12: false }) : ts();\n\n// ─── views ────────────────────────────────────────────────────────────────────\nfunction setView(v) {\n if (v === 'build') { const c = ART.current(); if (c) ART.open(c.id); else { let id = null; try { id = localStorage.getItem('omni.ap.last'); } catch {} if (id) ART.open(Number(id)); else { toggleSidebar(true); ART.listForSession(currentSession); barStatus.textContent = 'no artifact yet — /build something'; } } return; }\n view = v; if (!document.getElementById('apanel').classList.contains('closed')) ART.close();\n document.querySelectorAll('#modes button').forEach(b => b.classList.toggle('on', b.dataset.view === v));\n document.querySelectorAll('.view').forEach(el => el.classList.toggle('on', el.id === 'view-' + v));\n if (v === 'runs') renderRuns();\n}\ndocument.addEventListener('omni:artifact-closed', () => { document.querySelectorAll('#modes button').forEach(b => b.classList.toggle('on', b.dataset.view === view)); });\ndocument.addEventListener('omni:artifact-opened', () => { document.querySelectorAll('#modes button').forEach(b => b.classList.toggle('on', b.dataset.view === 'build')); });\nfunction toggleSidebar(force) { sidebarOpen = force ?? !sidebarOpen; sidebar.classList.toggle('collapsed', !sidebarOpen); }\n\n// ─── sessions ─────────────────────────────────────────────────────────────────\nasync function loadSessions() { renderSessions(await fetch('?sessions').then(r => r.json()).catch(() => [])); }\nfunction renderSessions(list) {\n sessionsEl.innerHTML = list.length ? '' : '

// empty

';\n for (const s of list) {\n const row = document.createElement('div'); row.className = 'sess-row' + (s.id === currentSession ? ' active' : ''); row.dataset.id = s.id;\n const name = document.createElement('span'); name.className = 'sess-name'; name.textContent = (s.name || 'New Chat').toUpperCase();\n const del = document.createElement('button'); del.className = 'sess-del'; del.title = 'delete'; del.innerHTML = '';\n del.onclick = (e) => { e.stopPropagation(); deleteSession(s.id); };\n row.appendChild(name); row.appendChild(del); row.onclick = () => switchSession(s.id, s.name); sessionsEl.appendChild(row);\n }\n}\nfunction clearChat() { msgCount = 0; lastAiWrap = null; turnLog = []; [...chat.querySelectorAll('.msg-wrap')].forEach(n => n.remove()); empty.style.display = ''; }\nasync function switchSession(id, name) {\n currentSession = id; rememberSession(); titleEl.textContent = (name || 'New Chat').toUpperCase(); clearChat();\n document.querySelectorAll('.sess-row').forEach(el => el.classList.toggle('active', el.dataset.id === id));\n ART.listForSession(id);\n const history = await fetch('?history=' + encodeURIComponent(id)).then(r => r.json()).catch(() => []);\n for (const m of history) {\n if (m.role !== 'user' && m.role !== 'assistant') continue;\n const meta = m.meta && typeof m.meta === 'object' ? m.meta : null;\n if (m.role === 'user' && meta?.hidden) continue; // continuation prompts\n if (m.role === 'assistant' && meta?.continuation && lastAiWrap) { appendToLast(m.content, meta); continue; }\n addMsg(m.role === 'user' ? 'user' : 'ai', m.content, meta, false, meta?.toolEvents || null, m.ts);\n if (m.role === 'assistant') turnLog.push({ q: lastUserText(), reply: m.content, meta, ts: m.ts });\n }\n chat.scrollTop = chat.scrollHeight; if (view === 'runs') renderRuns();\n SWARM.checkResume(true).then(found => { if (found) chat.scrollTop = chat.scrollHeight; }); COMPACT.show();\n}\nfunction lastUserText() { const u = [...chat.querySelectorAll('.msg-wrap.user .msg-body')].pop(); return u ? u.dataset.raw : ''; }\nasync function deleteSession(id) { const r = await fetch('?session=' + encodeURIComponent(id), { method: 'DELETE' }).catch(e => ({ ok: false, error: String(e) })); if (!r.ok) { barStatus.textContent = 'delete failed: ' + (r.error || r.status); return; } if (id === currentSession) newChat(); else await loadSessions(); }\nfunction newChat() { currentSession = crypto.randomUUID(); rememberSession(); titleEl.textContent = 'NEW CHAT'; clearChat(); loadSessions(); ART.listForSession(currentSession); setView('chat'); inp.focus(); }\nfunction startRename() { titleEl.style.display = 'none'; titleInp.style.display = 'block'; titleInp.value = titleEl.textContent; titleInp.focus(); titleInp.select(); }\nfunction commitRename() { const v = (titleInp.value.trim() || 'New Chat').toUpperCase(); titleEl.textContent = v; titleEl.style.display = ''; titleInp.style.display = 'none'; fetch('?rename', { method: 'POST', headers: { 'content-type': 'application/json' }, body: JSON.stringify({ id: currentSession, name: v }) }).then(loadSessions); }\nfunction titleKey(e) { if (e.key === 'Enter') commitRename(); if (e.key === 'Escape') { titleInp.style.display = 'none'; titleEl.style.display = ''; } }\n\n// ─── messages ─────────────────────────────────────────────────────────────────\nfunction renderAppCard(ev) {\n const url = ev.result?.url; if (!url) return null;\n const card = document.createElement('div'); card.className = 'app-card';\n const bar = document.createElement('div'); bar.className = 'app-card-bar';\n const title = document.createElement('span'); title.className = 'app-card-title'; title.textContent = '#' + ev.result.id + ' ' + (ev.result.title || 'untitled');\n const mk = (t, fn) => { const b = document.createElement('button'); b.className = 'app-card-open'; b.textContent = t; b.onclick = fn; return b; };\n const open = document.createElement('a'); open.className = 'app-card-open'; open.href = url; open.target = '_blank'; open.rel = 'noopener noreferrer'; open.textContent = 'open ↗';\n bar.appendChild(title); bar.appendChild(mk('preview', () => ART.open(ev.result.id, 'preview'))); bar.appendChild(mk('edit', () => ART.open(ev.result.id, 'split'))); bar.appendChild(open);\n card.appendChild(bar);\n if ((ev.args?.kind || 'html') !== 'markdown' && SETTINGS.current().cardPreview !== false) { const f = document.createElement('iframe'); f.className = 'app-card-frame'; f.src = url; f.sandbox = 'allow-scripts'; f.loading = 'lazy'; card.appendChild(f); }\n return card;\n}\nfunction telemetryEl(meta) {\n if (!meta) return null;\n const bits = [];\n if (meta.instance || meta.provider?.name) bits.push('' + MD.esc(meta.instance || meta.provider.name) + '');\n if (meta.model) bits.push('' + MD.esc(String(meta.model).split('/').pop()) + '');\n if (meta.maxTokensSent != null) bits.push('max ' + meta.maxTokensSent + '' + (meta.maxTokensAsked && meta.maxTokensAsked !== meta.maxTokensSent ? '/' + meta.maxTokensAsked : '') + '');\n if (meta.tokensOut != null) bits.push('out ' + meta.tokensOut + '');\n if (meta.finishReason && meta.finishReason !== 'stop') bits.push('' + MD.esc(meta.finishReason) + '');\n if (meta.telemetry?.passes != null) bits.push('passes ' + meta.telemetry.passes + '');\n if (meta.telemetry?.latencyMs != null) bits.push('' + (meta.telemetry.latencyMs / 1000).toFixed(1) + 's');\n if (meta.skills?.length) bits.push('skills: ' + MD.esc(meta.skills.join(', ')) + '');\n if (!bits.length) return null;\n const d = document.createElement('div'); d.className = 'telemetry'; d.innerHTML = bits.join(''); return d;\n}\nfunction addMsg(cls, text, meta = null, animate = true, toolEvents = null, at = null) {\n if (!msgCount++) empty.style.display = 'none';\n const w = document.createElement('div'); w.className = 'msg-wrap ' + cls; if (!animate) w.style.animation = 'none';\n const label = document.createElement('div'); label.className = 'msg-label';\n label.innerHTML = '' + (cls === 'user' ? 'YOU' : 'AI') + '' + fmtTs(at) + '';\n if (cls === 'ai' && (meta?.instance || meta?.provider?.name)) { const p = document.createElement('span'); p.className = 'provider-tag'; p.textContent = (meta.instance || meta.provider.name) + (meta.model ? ' / ' + String(meta.model).split('/').pop() : ''); label.appendChild(p); }\n w.appendChild(label);\n if (cls === 'ai' && toolEvents?.length) {\n w.appendChild(TL.make(toolEvents).root);\n const byId = new Map();\n for (const ev of toolEvents) { if (!/^(create|update)_artifact$/.test(ev.tool) || ev.error || !ev.result?.id) continue; const prev = byId.get(ev.result.id); byId.set(ev.result.id, { ...ev, args: { ...(prev?.args || {}), ...ev.args }, result: { ...(prev?.result || {}), ...ev.result, title: ev.result.title ?? prev?.result?.title } }); }\n for (const ev of byId.values()) { const c = renderAppCard(ev); if (c) w.appendChild(c); }\n }\n if (cls === 'ai' && meta?.reasoning?.length) meta.reasoning.forEach((t, i) => { const d = document.createElement('details'); d.className = 'think'; if (SETTINGS.current().thinking === 'open') d.open = true; d.innerHTML = '

'; d.querySelector('summary').textContent = 'thinking' + (meta.reasoning.length > 1 ? ' ' + (i + 1) : '') + ' · ' + t.length + ' chars'; d.querySelector('.think-body').textContent = t; w.appendChild(d); });\n const body = document.createElement('div'); body.className = 'msg-body md'; body.dataset.raw = text; body.innerHTML = cls === 'user' ? MD.esc(text) : MD.render(text); w.appendChild(body); if (cls === 'ai') MD.enhance(body);\n if (cls === 'ai') { const t = telemetryEl(meta); if (t) w.appendChild(t); if (meta?.truncated && !meta.autoContinued) truncNote(w); }\n const acts = document.createElement('div'); acts.className = 'msg-actions';\n const mk = (t, fn) => { const b = document.createElement('button'); b.textContent = t; b.onclick = fn; acts.appendChild(b); };\n mk('copy', () => navigator.clipboard.writeText(body.dataset.raw));\n if (cls === 'user') mk('edit', () => enterEditMode(w, body, body.dataset.raw)); else { mk('retry', () => retryFromMsg(w)); mk('continue', () => sendWith(CONTINUE_PROMPT, { continuation: true })); }\n w.appendChild(acts);\n chat.insertBefore(w, typing); if (SETTINGS.current().autoScroll !== false || cls === 'user') chat.scrollTop = chat.scrollHeight;\n if (cls === 'ai') lastAiWrap = w;\n return w;\n}\nconst CONTINUE_PROMPT = 'Continue exactly where you stopped. Do not repeat anything.';\nfunction truncNote(w) { const n = document.createElement('div'); n.className = 'trunc'; n.innerHTML = 'cut at the output limit · '; const b = document.createElement('button'); b.className = 'act-btn'; b.textContent = 'continue'; b.onclick = () => { n.remove(); sendWith(CONTINUE_PROMPT, { continuation: true }); }; n.appendChild(b); w.appendChild(n); }\n/** A continuation reply is appended into the previous AI bubble instead of opening a new one. */\nfunction appendToLast(text, meta) {\n if (!lastAiWrap) return addMsg('ai', text, meta, true, meta?.toolEvents);\n const body = lastAiWrap.querySelector('.msg-body'); body.dataset.raw += text; body.innerHTML = MD.render(body.dataset.raw); MD.enhance(body);\n lastAiWrap.querySelectorAll('.trunc,.telemetry').forEach(n => n.remove());\n // A continuation round's own tool calls were previously dropped here entirely: addMsg() bakes a fresh\n // turn's toolEvents into a permanent TL block in the bubble, but this function only ever touched the\n // text. The LIVE progress table the user watches filling up while \"continue\" runs gets torn down by\n // finish() same as any round — that part is normal — but nothing replaced it here, so a continuation's\n // tool activity (and any artifact it built/edited) vanished with no permanent record: the table\n // appeared to just disappear into nothing. Mirror addMsg(): bake this round's events into their own\n // block, and surface any artifact build/update from it as an app card too.\n if (meta?.toolEvents?.length) {\n const actionsEl = lastAiWrap.querySelector('.msg-actions');\n lastAiWrap.insertBefore(TL.make(meta.toolEvents).root, actionsEl);\n const byId = new Map();\n for (const ev of meta.toolEvents) { if (!/^(create|update)_artifact$/.test(ev.tool) || ev.error || !ev.result?.id) continue; const prev = byId.get(ev.result.id); byId.set(ev.result.id, { ...ev, args: { ...(prev?.args || {}), ...ev.args }, result: { ...(prev?.result || {}), ...ev.result, title: ev.result.title ?? prev?.result?.title } }); }\n for (const ev of byId.values()) { const c = renderAppCard(ev); if (c) lastAiWrap.insertBefore(c, actionsEl); }\n }\n const t = telemetryEl(meta); if (t) lastAiWrap.insertBefore(t, lastAiWrap.querySelector('.msg-actions'));\n const c = lastAiWrap.querySelector('.cont-n'); const n = (c ? Number(c.dataset.n) : 0) + 1; if (c) c.remove();\n const tag = document.createElement('span'); tag.className = 'msg-ts cont-n'; tag.dataset.n = n; tag.textContent = '+' + n + ' continued'; lastAiWrap.querySelector('.msg-label').appendChild(tag);\n if (meta?.truncated && !meta.autoContinued) truncNote(lastAiWrap);\n return lastAiWrap;\n}\nfunction enterEditMode(wrap, bodyEl, original) {\n bodyEl.style.display = 'none'; wrap.querySelector('.msg-actions').style.display = 'none';\n const ta = document.createElement('textarea'); ta.className = 'edit-area'; ta.value = original; ta.rows = Math.min(original.split('\\n').length + 1, 8);\n const btns = document.createElement('div'); btns.className = 'edit-btns';\n const cancel = document.createElement('button'); cancel.className = 'edit-btn'; cancel.textContent = 'cancel'; cancel.onclick = () => { ta.remove(); btns.remove(); bodyEl.style.display = ''; wrap.querySelector('.msg-actions').style.display = ''; };\n const ok = document.createElement('button'); ok.className = 'edit-btn confirm'; ok.textContent = 'resend →'; ok.onclick = () => { const t = ta.value.trim(); if (!t) return; const all = [...chat.querySelectorAll('.msg-wrap')], i = all.indexOf(wrap); all.slice(i).forEach(n => n.remove()); msgCount = Math.max(0, msgCount - (all.length - i)); sendWith(t); };\n btns.appendChild(cancel); btns.appendChild(ok); wrap.appendChild(ta); wrap.appendChild(btns); ta.focus();\n}\nfunction retryFromMsg(aiWrap) { const all = [...chat.querySelectorAll('.msg-wrap')], i = all.indexOf(aiWrap); if (i < 1) return; const q = all[i - 1].querySelector('.msg-body')?.dataset.raw || ''; all.slice(i - 1).forEach(n => n.remove()); msgCount = Math.max(0, msgCount - (all.length - (i - 1))); sendWith(q); }\nfunction addError(message, details, retryText, keep) {\n const w = addMsg('ai', ''); w.querySelector('.msg-body').remove(); w.querySelector('.msg-actions').remove();\n const card = document.createElement('div'); card.className = 'err-card';\n card.innerHTML = '

error

'; card.querySelector('.err-m').textContent = message;\n if (details?.length) { const d = document.createElement('details'); d.innerHTML = '

details (' + details.length + ')

'; d.querySelector('pre').textContent = details.join('\\n'); card.appendChild(d); }\n if (keep?.events?.length) { const t = TL.make(keep.events, { live: false, label: 'what happened before the error', open: true }); card.appendChild(t.root); }\n if (keep?.draft) { const d = document.createElement('details'); d.open = true; d.innerHTML = '

draft reply recovered (' + keep.draft.length + ' chars, last part)

'; d.querySelector('pre').textContent = keep.draft; card.appendChild(d); }\n const acts = document.createElement('div'); acts.className = 'err-actions'; const rb = document.createElement('button'); rb.className = 'act-btn'; rb.textContent = 'retry'; rb.onclick = () => { w.remove(); msgCount--; sendWith(retryText); }; acts.appendChild(rb); card.appendChild(acts); w.appendChild(card);\n}\n\n// ─── turn ─────────────────────────────────────────────────────────────────────\nfunction showTyping(v, label = 'thinking') { typingLbl.textContent = label; typing.style.display = v ? 'block' : 'none'; if (v) chat.scrollTop = chat.scrollHeight; document.dispatchEvent(new CustomEvent('omni:typing', { detail: { on: v, label, progress: progressEl.textContent } })); }\nfunction setSendState(busy) { sendBtn.classList.toggle('stop', busy); sendBtn.title = busy ? 'stop (Esc)' : 'send (Enter)'; sendBtn.onclick = busy ? stopGen : sendMsg; sendBtn.innerHTML = busy ? '' : ''; $('pill-model').classList.toggle('busy', busy); }\nfunction stopGen() { if (aborter) aborter.abort(); SWARM.stop(); }\nconst APP_ASK = /\\b(create|build|write|implement|generate|make|design|code|develop)\\b[\\s\\S]*\\b(app|application|game|editor|dashboard|simulat\\w*|tool|ide|player|visuali[sz]\\w*|engine|synth\\w*|tracker|planner|studio|workbench|explorer|builder|designer|sequencer|daw|calculator|clone)\\b/i;\nasync function sendMsg() { const t = inp.value.trim(); if (!t) return; if (/^finalize[.!]?$/i.test(t)) { inp.value = ''; return finalize(); } if (inflight) return; FINALIZE.requested = false; inp.value = ''; inp.style.height = 'auto'; charEl.textContent = ''; SLASH.hide(); if (t.startsWith('/') && await SLASH.run(t)) return;\n // App-scale asks: the swarm path (architect → 8–16 parallel builders → integrate) is the only way a 60 s val produces a ≥ 900-line product in one go.\n if ((SETTINGS.current().autoSwarm ?? true) && APP_ASK.test(t) && t.length < 6000 && !ART.focusPayload()) { barStatus.textContent = 'app-scale ask → /build'; await SWARM.build(t); return; }\n await sendWith(t); }\nfunction addSysCard(tag, text, detail) { const n = document.createElement('div'); n.className = 'msg-wrap sys'; n.innerHTML = '

' + ts() + '

'; n.querySelector('.role-tag').textContent = tag; const b = n.querySelector('.msg-body'); b.textContent = text; if (detail) { const d = document.createElement('details'); d.style.marginTop = '4px'; const sm = document.createElement('summary'); sm.textContent = 'summary'; sm.style.cursor = 'pointer'; const pre = document.createElement('div'); pre.textContent = detail; pre.style.whiteSpace = 'pre-wrap'; d.append(sm, pre); b.appendChild(d); } chat.insertBefore(n, typing); msgCount++; return n; }\n// ── conversation compaction: runs as its own val invocation after a turn (server flags needsCompact) or via /compact ──\nconst COMPACT = (() => {\n let running = false;\n async function run(manual) {\n if (running) return false; running = true; barStatus.textContent = 'compacting history…';\n try {\n const r = await fetchT('?compact', { method: 'POST', headers: { 'content-type': 'application/json' }, body: JSON.stringify({ session: currentSession, settings: SETTINGS.forRequest() }) }, 70_000).then(x => x.json());\n if (r.error) { if (manual) addSysCard('COMPACT', 'failed: ' + (r.error.message || r.error)); return false; }\n if (!r.ok) { if (manual) addSysCard('COMPACT', r.reason || 'nothing to fold'); return false; }\n const c = await fetch('?compact=' + encodeURIComponent(currentSession)).then(x => x.json()).catch(() => ({}));\n addSysCard('COMPACT', 'folded ' + r.covered + ' older messages into the running summary (' + r.before.toLocaleString() + ' → ' + r.after.toLocaleString() + ' chars); ' + (c.compact ? c.compact.covered + ' messages summarized in total' : ''), c.compact?.summary);\n return true;\n } catch (e) { if (manual) addSysCard('COMPACT', 'failed: ' + e.message); return false; }\n finally { running = false; if (barStatus.textContent.startsWith('compacting')) barStatus.textContent = ''; }\n }\n async function show() { const c = await fetch('?compact=' + encodeURIComponent(currentSession) + '&check=' + (Number(SETTINGS.current().contextChars) || 16000)).then(x => x.json()).catch(() => ({})); if (c.compact) addSysCard('COMPACT', c.compact.covered + ' earlier messages are folded into a running summary the model reads instead of the full history', c.compact.summary); if (c.needs && (SETTINGS.current().autoCompact ?? true)) run(false); return !!c.compact; }\n async function clear() { await fetch('?compact=' + encodeURIComponent(currentSession), { method: 'DELETE' }); addSysCard('COMPACT', 'summary cleared — the model sees raw history again (within the budget)'); }\n return { run, show, clear };\n})();\n// ── /finalize: end the work-until-finalize chain and run one closing turn (acceptance walk, lint, README, summary) ──\nconst FINALIZE = { requested: false };\nasync function finalize() {\n FINALIZE.requested = true; stopGen(); barStatus.textContent = 'finalizing…';\n for (let t = 0; t < 40 && inflight; t++) await wait(250); /* let the aborted turn unwind */\n // sendWith() silently no-ops if inflight is still true (its own re-entrancy guard) — previously that meant a\n // slow-to-unwind prior turn made /finalize appear to do nothing at all, with FINALIZE.requested reset right after\n // as if the closing pass had actually run. Detect the stuck case and say so instead of pretending it happened.\n if (inflight) { FINALIZE.requested = false; barStatus.textContent = ''; addSysCard('FINALIZE', 'could not start — the previous turn is still stopping. Press stop again if it looks stuck, then run /finalize once more.'); return; }\n const f = ART.focusPayload();\n await sendWith('FINALIZE. Stop adding features. Walk the acceptance list and README against the shipped files, fix any lint or portability finding with update_artifact, make README.md match exactly what exists, then give the closing summary: what it does, how to run it, known limits.', { finalize: true, ...(f || {}) });\n FINALIZE.requested = false;\n}\n/** extra: {continuation:true} appends to the last AI bubble; {focus} targets an artifact; {__quiet} skips the user bubble; {__skill} one-turn skill;\n * {sessionId} pins this turn — and any auto-continue/follow-up/improve round it spawns — to the session it started in,\n * instead of reading the live, mutable currentSession throughout the turn's lifetime. Previously a session switch mid-turn\n * could misroute progress polling, crash recovery and the eventual reply into whichever session happened to be open when\n * each of those ran; own() below draws the line between \"which session is this data for\" (sessionId) and \"should this\n * turn paint into the visible UI right now\" (is sessionId still the one on screen);\n * {__ownsInflight} lets a caller that already holds the single global inflight/send-button lock (the swarm's integrate\n * step, from fanout()) keep holding it across this call instead of sendWith re-acquiring/releasing its own — closing the\n * window where a manual send could slip in between the swarm's builders finishing and integration starting and get\n * silently rejected by sendWith's own inflight guard, along with the aborted/failed integration it was hiding. */\nasync function sendWith(text, extra = {}) {\n if (!extra.__ownsInflight) { if (inflight) return false; inflight = true; setSendState(true); }\n const sessionId = extra.sessionId || currentSession;\n const own = () => sessionId === currentSession;\n const focusExtra = extra.focus !== undefined ? { focus: extra.focus } : (ART.focusPayload() || {});\n const isCont = !!extra.continuation;\n if (extra.followup) { const n = document.createElement('div'); n.className = 'msg-wrap sys'; n.innerHTML = '

AUTO' + ts() + '

'; n.querySelector('.msg-body').textContent = 'continuing unfinished work:\\n' + extra.followup.map(x => '• ' + x).join('\\n'); if (own()) { chat.insertBefore(n, typing); msgCount++; } }\n else if (!isCont && !extra.__quiet && own()) addMsg('user', text);\n if (own()) { showTyping(true, isCont ? 'continuing' : 'thinking'); tabProgress(isCont ? 'continuing' : 'thinking'); progressEl.textContent = ''; liveThink.textContent = ''; }\n const S0 = SETTINGS.current();\n const plan = isCont ? ['answer'] : [S0.plan ? 'plan' : null, 'answer', S0.verify ? 'verify' : null].filter(Boolean);\n const live = TL.make([], { live: true, label: isCont ? 'continuing' : 'working', plan });\n if (own()) { liveTl.innerHTML = ''; liveTl.appendChild(live.root); }\n // Per-turn controller. 'return sendWith(next)' inside this try/finally starts the NEXT turn before this turn's finally runs,\n // so finish() must only tear down state that still belongs to this turn — otherwise it nulls the successor's aborter mid-flight.\n const ac = new AbortController(); aborter = ac;\n let since = 0, polling = true, finished = false, draft = ''; const events = [];\n (async () => { while (polling) { await wait(document.hidden ? 5000 : 2000); if (!polling) break; try {\n const j = await (await fetch('?progress=' + encodeURIComponent(sessionId) + '&since=' + since, { signal: ac.signal })).json(); since = j.next;\n for (const d of j.events) { if (d.type === 'stage' && d.label !== 'done') { events.push(d); if (own()) showTyping(true, d.label); } else if (d.type === 'tool') { events.push({ ...d, live: true }); if (own()) showTyping(true, 'tool: ' + d.tool); } else if (d.type === 'thinking') { if (own() && SETTINGS.current().liveThinking) liveThink.textContent = d.text; } else if (d.type === 'partial') { draft = d.text; if (own()) { liveDraft.hidden = false; liveDraftT.textContent = d.text; liveDraftN.textContent = '(last ' + d.text.length + ' chars)'; liveDraftT.scrollTop = liveDraftT.scrollHeight; } } else if (d.type === 'pass' || d.type === 'continue' || d.type === 'retry') events.push(d); }\n live.set(events); if (own() && live.root.parentNode !== liveTl) { liveTl.innerHTML = ''; liveTl.appendChild(live.root); }\n } catch {} } })();\n const base = SETTINGS.forRequest();\n if (extra.__skill) base.skills = [...(base.skills || []), { ...extra.__skill, mode: 'always' }];\n if (extra.__force === 'ensemble') base.ensemble = true;\n const settings = { ...base, ...focusExtra, continuation: isCont, followup: !!extra.followup };\n let ok = false;\n try {\n // Exhaustion is retried from the CLIENT (the server has a 60 s wall clock): keep trying with backoff until\n // retryMinutes elapse or the user presses stop. The server dedupes the re-sent user message (retryOf).\n // MIN_RETRY_ATTEMPTS is a floor, not a target: exhaustion must be retried at least this many times even\n // if retryMinutes (default 5 min) would otherwise cut it short — a short default window must never look\n // like \"gave up after 2 tries\". Once MIN_RETRY_ATTEMPTS is reached, retryMinutes governs again as before.\n const forever = !!S0.untilStopped && !extra.finalize; /* keep going until the stop button: retries, cut-reply continues and follow-up rounds are unbounded */\n const retryMs = forever ? Infinity : Math.max(0, Number(S0.retryMinutes ?? 5)) * 60_000, tStart = Date.now();\n const MIN_RETRY_ATTEMPTS = 10;\n let final = null, attempt = 0;\n for (;;) {\n attempt++;\n let r;\n try { r = await fetchT('?q', { method: 'POST', headers: { 'content-type': 'application/json' }, signal: ac.signal, body: JSON.stringify({ q: text, session: sessionId, settings: { ...settings, retryOf: attempt > 1 } }) }); final = await r.json().catch(() => null); }\n catch (e) { if (e.name === 'AbortError') throw e; /* transport failure: the server may still have finished — check before re-sending */ const rec = await recoverFromHistory(text, sessionId); if (rec) { final = { reply: rec.content, meta: rec.meta, toolEvents: rec.meta?.toolEvents, truncated: rec.meta?.truncated }; break; } r = { ok: false, status: 0 }; final = { error: e.message }; }\n let emptyReply = false;\n if (r.ok && final && !final.error) {\n // A 200 with nothing usable in it — blank, or nothing but dots (the same junk-generation\n // signature app-infer.ts/router-core.ts guard against) — must never be accepted as a finished\n // turn. toolEvents alone (no text) is legitimate (e.g. a pure tool-call turn); only flag when\n // there's neither real text nor recorded tool activity.\n const replyStripped = String(final.reply ?? '').replace(/\\s+/g, '');\n const looksEmpty = replyStripped.length === 0 || /^\\.+$/.test(replyStripped);\n if (!looksEmpty || (final.toolEvents && final.toolEvents.length)) break;\n emptyReply = true;\n }\n const e = final?.error; const msg = emptyReply ? 'server returned an empty or dot-only reply' : (typeof e === 'string' ? e : (e?.message || ('HTTP ' + r.status))); const details = Array.isArray(e?.details) ? e.details : [];\n const exhausted = emptyReply || /EXHAUSTED|COOLING|router 50[23]|HTTP 50[23]|TIMEOUT|failed to fetch|networkerror|load failed/i.test(msg + ' ' + details.slice(0, 3).join(' '));\n const elapsed = Date.now() - tStart;\n if (!exhausted || (attempt >= MIN_RETRY_ATTEMPTS && elapsed >= retryMs)) { const x = new Error(msg + (attempt > 1 ? ' (after ' + attempt + ' attempts over ' + Math.round(elapsed / 1000) + 's)' : '')); x.details = details; x.structured = true; throw x; }\n const waitMs = jitter(Math.min(45_000, 8_000 * attempt));\n events.push({ type: 'retry', n: attempt, error: msg.slice(0, 80) + ' — waiting ' + Math.round(waitMs / 1000) + 's' + (details.length ? '\\n' + details.slice(0, 6).join('\\n') : '') }); live.set(events);\n for (let left = waitMs; left > 0; left -= 1000) { if (ac.signal.aborted) { const x = new Error('stopped'); x.name = 'AbortError'; throw x; } if (own()) { showTyping(true, 'providers busy — retry ' + (attempt + 1) + ' in ' + Math.ceil(left / 1000) + 's (' + (forever ? '∞' : Math.round((retryMs - elapsed) / 60000) + ' min') + ' left)'); tabProgress('retry ' + (attempt + 1)); } await wait(1000); }\n since = 0; // the next attempt writes a fresh progress log\n }\n if (own()) showTyping(false); ok = true;\n const meta = { ...(final.meta || {}), telemetry: final.telemetry, reasoning: final.reasoning, truncated: final.truncated, skills: final.skills, toolEvents: final.toolEvents };\n const autoLeft = extra.__autoLeft ?? (forever ? Infinity : (SETTINGS.current().autoContinue ?? 6));\n // no-progress guard for unbounded mode: two consecutive continuations that add nothing end the chain\n const noProgress = isCont && String(final.reply || '').trim().length < 20 ? (extra.__stall ?? 0) + 1 : 0;\n const willAuto = final.truncated && autoLeft > 0 && noProgress < 2;\n if (willAuto) meta.autoContinued = true;\n // Everything below that renders into the DOM or touches session-scoped chrome (title bar, artifact panel, compact,\n // the runs log) only runs if this turn's session is still the one on screen (own()). If the user switched away, the\n // reply is already safely persisted server-side under sessionId (the request body above used sessionId, not the\n // live global) — switchSession()'s history fetch will render it correctly next time that session is opened. Before\n // this fix all of this unconditionally used the live, possibly-since-changed currentSession/DOM, so a switch mid-turn\n // could land a reply, a title rename, or an artifact refresh in the wrong session's UI.\n if (own()) {\n if (isCont) appendToLast(final.reply || '', meta); else addMsg('ai', final.reply || '…', meta, true, final.toolEvents);\n turnLog.push({ q: text, reply: final.reply, meta, ts: Date.now() });\n if (final.sessionName) { titleEl.textContent = final.sessionName.toUpperCase(); loadSessions(); } else if (msgCount <= 2) loadSessions();\n // An update to the artifact you already had loaded refreshes it in place (ART.refresh) — it only actually\n // switches the visible view if the workbench panel is already open (you're already looking at it); if\n // you're on the chat view it just badges the build tab instead of yanking you off what you're reading.\n // A brand-new artifact with nothing previously loaded still honors autoOpen (that's what it's for).\n if (final.toolEvents?.some(e => /artifact/.test(e.tool) && !e.error)) { ART.listForSession(sessionId); const c = ART.current(); if (c && final.toolEvents.some(e => e.result?.id === c.id)) ART.refresh(c.id); else if (!c && ART.settings().autoOpen) { const ev = final.toolEvents.find(e => e.tool === 'create_artifact' && e.result?.id); if (ev) ART.open(ev.result.id); } }\n document.dispatchEvent(new CustomEvent('omni:turn', { detail: { text, final } }));\n if (final.needsCompact && (SETTINGS.current().autoCompact ?? true)) COMPACT.run(false);\n } else if (final.sessionName) { loadSessions(); } // background turn: still refresh the sidebar's name list, just never the visible title bar\n if (willAuto) { if (own()) { finish(); barStatus.textContent = 'auto-continuing (' + (forever ? '∞' : autoLeft) + ' left)'; } else finish(); return sendWith(CONTINUE_PROMPT, { continuation: true, __autoLeft: autoLeft - 1, __stall: noProgress, sessionId, __ownsInflight: extra.__ownsInflight }); }\n // The chain is ending here even though the model was still mid-output (final.truncated) — say why,\n // instead of leaving the user to notice only that the live table vanished and nothing more happened.\n // The two silent causes: the stall guard (two continuation rounds in a row added under 20 chars —\n // real, since a genuinely finished reply is usually longer, but no reason not to say so) and the\n // per-turn auto-continue cap (SETTINGS autoContinue, default 6) simply running out.\n if (own() && isCont && final.truncated && !willAuto) {\n addSysCard('AUTO', noProgress >= 2\n ? 'auto-continue stopped: the last two rounds made no real progress — click \"continue\" above to try again, or rephrase the ask'\n : 'auto-continue stopped: reached the ' + (SETTINGS.current().autoContinue ?? 6) + '-round limit while the reply was still being cut off — click \"continue\" above to keep going');\n }\n // Completion audit says the ask is not finished: keep going with the remaining items (bounded, stop button aborts).\n const compLeft = extra.__compLeft ?? (forever ? Infinity : (SETTINGS.current().autoComplete ?? 4));\n const rem = final.completion && !final.completion.done ? final.completion.remaining : [];\n const sameAsBefore = extra.followup && JSON.stringify(rem) === JSON.stringify(extra.followup);\n if (rem.length && compLeft > 0 && !sameAsBefore) {\n if (own()) { finish(); barStatus.textContent = 'not finished — continuing (' + (forever ? '∞' : compLeft) + ' rounds left)'; } else finish();\n const f = own() ? (ART.focusPayload() || {}) : {};\n return sendWith('Continue the task until it is fully complete. Remaining work (do all of it, verify, then report):\\n' + rem.map(x => '- ' + x).join('\\n'), { followup: rem, __compLeft: compLeft - 1, sessionId, __ownsInflight: extra.__ownsInflight, ...f });\n }\n if (own() && final.completion && final.completion.done) barStatus.textContent = '';\n // Work-until-finalize: the ask is complete, but the user has not said /finalize — ship the next most valuable upgrades.\n // Ends when: /finalize or stop; two consecutive rounds that write nothing; the model reports nothing left worth shipping.\n if (forever && !FINALIZE.requested && final.completion && final.completion.done && (extra.__improve || ART.current() || (final.toolEvents || []).some(e => /artifact/.test(e.tool)))) {\n const wrote = (final.toolEvents || []).some(e => /^(create|update)_artifact$/.test(e.tool) && !e.error);\n const idle = extra.__improve && !wrote ? (extra.__idle ?? 0) + 1 : 0;\n if (idle >= 2 || /NOTHING_LEFT/.test(final.reply || '')) { if (own()) { addSysCard('AUTO', 'no further upgrades produced in two rounds — waiting for you (/finalize to wrap up, or give a new direction)'); barStatus.textContent = 'idle — /finalize when ready'; } }\n else {\n if (own()) { finish(); const round = (extra.__improve ?? 0) + 1; barStatus.textContent = 'improvement round ' + round + ' — /finalize to stop'; tabProgress('improve ' + round); } else finish();\n const round = (extra.__improve ?? 0) + 1;\n const f = own() ? (ART.focusPayload() || {}) : {};\n return sendWith('The task is complete and the user has not finalized yet. Working until told to finalize: choose the 3 most valuable upgrades a principal engineer would ship NEXT for this artifact (new capabilities, robustness, performance, accessibility, polish — never a rewrite, never a regression, never placeholders), implement them fully with update_artifact, verify lint is clean, then report exactly what shipped. If nothing worthwhile remains, reply with the single line NOTHING_LEFT.', { followup: ['improvement round ' + round], __improve: round, __idle: idle, sessionId, __ownsInflight: extra.__ownsInflight, ...f });\n }\n }\n } catch (e) {\n let recovered = null;\n if (e.name !== 'AbortError' && !e.structured) { if (own()) showTyping(true, 'recovering'); for (let t = 0; t < 8 && !recovered; t++) { await wait(3000); recovered = await recoverFromHistory(text, sessionId); } }\n if (own()) showTyping(false);\n if (recovered) { ok = true; if (own()) addMsg('ai', recovered.content, recovered.meta, true, recovered.meta?.toolEvents); }\n else if (own()) { if (e.name === 'AbortError') addMsg('ai', '_[stopped]_'); else addError(e.message, e.details || [], text, { events: events.slice(), draft }); }\n if (own()) document.dispatchEvent(new CustomEvent('omni:turn', { detail: { text, error: recovered ? null : (e.name === 'AbortError' ? 'stopped' : e.message), final: recovered ? { reply: recovered.content } : null } }));\n } finally { finish(); }\n return ok;\n // extra.__ownsInflight: the caller (fanout's integrate step) already holds inflight/setSendState — leave that lock\n // alone here so it doesn't get released between the swarm's build phase and integrate phase (see fanout()).\n function finish() { if (finished) return; finished = true; polling = false; live.finish(ok); if (aborter !== ac) return; /* a successor turn owns the shared state now */ aborter = null; if (!extra.__ownsInflight) { inflight = false; setSendState(false); } if (own()) { tabProgress(''); progressEl.textContent = ''; liveThink.textContent = ''; liveDraft.hidden = true; liveDraftT.textContent = ''; liveTl.innerHTML = ''; barStatus.textContent = ''; } if (view === 'runs') renderRuns(); }\n}\n\n// ─── composer: chips, slash commands, textarea ────────────────────────────────\nfunction renderChips() {\n const S = SETTINGS.current(); const c = ART.current(); chipsEl.innerHTML = '';\n const chip = (t, on, fn, title) => { const s = document.createElement('span'); s.className = 'chip' + (on ? ' on' : ''); s.textContent = t; s.title = title || ''; s.onclick = fn; chipsEl.appendChild(s); return s; };\n chip('mode: ' + (S.mode === 'pin' ? (S.pin || 'pin') : S.mode === 'auto' ? 'best free' : S.mode), false, () => { const order = ['auto', 'fast', 'coder', 'reasoning']; SETTINGS.set('mode', order[(order.indexOf(S.mode) + 1) % order.length]); renderChips(); }, 'click to cycle routing mode');\n chip('expect: ' + (S.expect || 'auto'), false, () => { const o = ['auto', 'short', 'long', 'max']; SETTINGS.set('expect', o[(o.indexOf(S.expect || 'auto') + 1) % o.length]); renderChips(); }, 'expected reply length → drives max_tokens and vendor choice');\n if (c && !document.getElementById('apanel').classList.contains('closed')) chip('focus #' + c.id, true, () => ART.close(), 'requests target this artifact');\n const sk = (S.skills || []).filter(k => k.mode === 'always').length; if (sk) chip(sk + ' skill' + (sk > 1 ? 's' : '') + ' always', true, () => SETTINGS.toggle(true));\n if (S.autoContinue === 0) chip('auto-continue off', false, () => SETTINGS.toggle(true));\n $('pill-model-t').textContent = S.mode === 'pin' ? (S.pin || 'pin') : S.mode === 'auto' ? 'best free' : S.mode;\n}\ndocument.addEventListener('omni:settings', renderChips); document.addEventListener('omni:artifact-opened', renderChips); document.addEventListener('omni:artifact-closed', renderChips);\nfunction prefill(t) { inp.value = t; inp.dispatchEvent(new Event('input')); inp.focus(); inp.setSelectionRange(t.length, t.length); }\nconst SLASH = (() => {\n const CMDS = [\n { c: '/build', h: 'plan → parallel builders → integrate (large multi-file apps)', run: (a) => SWARM.build(a) },\n { c: '/plan', h: 'manifest only; review, then \"build all\" on the card', run: (a) => SWARM.plan(a) },\n { c: '/fix', h: 'repair the open artifact from console errors', run: (a) => { const f = ART.focusPayload(); if (!f) { barStatus.textContent = 'open an artifact first'; return false; } return sendWith((a || 'Fix the artifact.') + (f.focus.errors.length ? '\\nConsole:\\n' + f.focus.errors.join('\\n') : ''), f); } },\n { c: '/resume', h: 'resume an unfinished /build in this session', run: () => SWARM.checkResume(false) },\n { c: '/finalize', h: 'stop the work-until-finalize chain and run the closing pass', run: () => finalize().then(() => true) },\n { c: '/compact', h: 'fold older turns into a running summary now (/compact clear removes it)', run: (a) => a === 'clear' ? COMPACT.clear().then(() => true) : COMPACT.run(true).then(() => true) },\n { c: '/continue', h: 'continue the last reply', run: () => sendWith(CONTINUE_PROMPT, { continuation: true }) },\n { c: '/ensemble', h: 'answer with N drafts + reconciler (this turn)', run: (a) => sendWith(a, { __force: 'ensemble' }) },\n { c: '/fast', h: 'fastest vendor from now on', run: (a) => { SETTINGS.set('mode', 'fast'); renderChips(); return a ? sendWith(a) : true; } },\n { c: '/best', h: 'best free model from now on', run: (a) => { SETTINGS.set('mode', 'auto'); renderChips(); return a ? sendWith(a) : true; } },\n { c: '/coder', h: 'coder routing from now on', run: (a) => { SETTINGS.set('mode', 'coder'); renderChips(); return a ? sendWith(a) : true; } },\n { c: '/skill', h: '/skill — activate a skill for this turn only', run: (a) => { const [n, ...rest] = a.split(/\\s+/); const k = (SETTINGS.current().skills || []).find(s => s.name.toLowerCase() === (n || '').toLowerCase()); if (!k) { barStatus.textContent = 'no skill named ' + n; return false; } return sendWith(rest.join(' '), { __skill: k }); } },\n { c: '/new', h: 'new chat', run: () => { newChat(); return true; } },\n { c: '/runs', h: 'open the runs inspector', run: () => { setView('runs'); return true; } },\n { c: '/settings', h: 'open settings', run: () => { SETTINGS.toggle(true); return true; } },\n ];\n const box = $('slash'); let sel = 0, items = [];\n function show(q) { items = CMDS.filter(c => c.c.startsWith(q.split(' ')[0])); if (!items.length || q.includes(' ')) return hide(); sel = 0; box.innerHTML = items.map((c, i) => '

' + c.c + '' + MD.esc(c.h) + '

').join(''); box.classList.add('on'); box.querySelectorAll('.slash-i').forEach(el => el.onclick = () => { prefill(items[el.dataset.i].c + ' '); hide(); }); }\n function hide() { box.classList.remove('on'); items = []; }\n function key(e) { if (!items.length) return false; if (e.key === 'ArrowDown') { sel = (sel + 1) % items.length; } else if (e.key === 'ArrowUp') { sel = (sel - 1 + items.length) % items.length; } else if (e.key === 'Tab' || (e.key === 'Enter' && !inp.value.includes(' '))) { prefill(items[sel].c + ' '); hide(); return true; } else if (e.key === 'Escape') { hide(); return true; } else return false; box.querySelectorAll('.slash-i').forEach((el, i) => el.classList.toggle('sel', i === sel)); return true; }\n async function run(t) { const [c, ...rest] = t.split(/\\s+/); const cmd = CMDS.find(x => x.c === c.toLowerCase()); if (!cmd) return false; const r = await cmd.run(rest.join(' ').trim()); return r !== false; }\n return { show, hide, key, run, CMDS };\n})();\ninp.addEventListener('input', () => { inp.style.height = 'auto'; inp.style.height = Math.min(inp.scrollHeight, 180) + 'px'; const n = inp.value.length; charEl.textContent = n > 50 ? n : ''; charEl.className = n > 48000 ? 'danger' : n > 8000 ? 'warn' : ''; if (inp.value.startsWith('/')) SLASH.show(inp.value); else SLASH.hide(); });\ninp.addEventListener('keydown', e => { if (SLASH.key(e)) { e.preventDefault(); return; } const ctrl = SETTINGS.current().sendKey === 'ctrl'; if (e.key === 'Enter' && (ctrl ? (e.ctrlKey || e.metaKey) : !e.shiftKey)) { e.preventDefault(); sendMsg(); } });\ndocument.addEventListener('keydown', e => {\n const pal = $('pal').classList.contains('on');\n if (e.key === 'Escape') { if (pal) PAL.close(); else if (!$('settings').classList.contains('closed')) SETTINGS.toggle(false); else if (inflight && document.getElementById('apanel').classList.contains('closed')) stopGen(); }\n else if ((e.ctrlKey || e.metaKey) && e.key === ',') { e.preventDefault(); SETTINGS.toggle(); }\n else if ((e.ctrlKey || e.metaKey) && e.key.toLowerCase() === 'k') { e.preventDefault(); PAL.toggle(); }\n else if ((e.ctrlKey || e.metaKey) && e.key.toLowerCase() === 'b' && !e.shiftKey) { e.preventDefault(); setView('build'); }\n else if (e.key === '/' && document.activeElement === document.body && !e.ctrlKey && !e.metaKey) { e.preventDefault(); prefill('/'); }\n});\nchat.addEventListener('scroll', () => pill.classList.toggle('visible', chat.scrollHeight - chat.scrollTop - chat.clientHeight > 60));\nfunction scrollBottom() { chat.scrollTo({ top: chat.scrollHeight, behavior: 'smooth' }); }\n\n// ─── swarm orchestrator: one action → many val invocations ────────────────────\nconst SWARM = (() => {\n let job = null;\n // Per-run cancellation token, replacing a single shared \"stopped\" boolean that every fanout() call reset to false —\n // if build A was stopped while one of its staggered workers was still in a wait() between checks, and the user then\n // started build B (which reset the shared flag), A's worker would see !stopped again and keep going on A's own queue.\n // stop() cancels whichever run is currently active; buildOne()/worker() below only ever look at the token they were\n // handed, never a module-wide flag, so an old run can no longer be un-stopped by a new one starting.\n let activeRun = null;\n // Was a single flat path-to-rowElement object, shared across every manifest card and unconditionally\n // wiped (reset to empty) each time manifestCard() ran. checkResume() calls manifestCard() once per\n // unfinished build IN A LOOP when more than one exists — each call clobbered the previous card's entire\n // row map, so status updates (building/retry/ok/error) for an EARLIER card's files landed nowhere (path\n // not in the current row map, so st() silently no-ops) or, whenever two different artifacts happened to\n // share a filename (extremely common: \"index.html\", \"app.js\", \"style.css\"), landed on the WRONG card's\n // row for that same path — visually \"updates the wrong file\". Scoped per artifactId instead, so cards\n // never share state and a shared filename across artifacts can no longer cross-talk.\n const rowsByArtifact = new Map();\n // Job state lives server-side (omni_state) so a build survives a closed tab, a dead network or a reload: resume() re-runs only\n // the files that never finished. build_file writes are idempotent (same path → overwrite), so a duplicate run is harmless.\n // Serialized: each call snapshots the (already locally-mutated) job synchronously, then chains its fetch after every\n // earlier one. Previously every call fired an independent, unawaited fetch — concurrent swarm workers finishing near-\n // simultaneously could have their POSTs reach the server out of send order; since the server replaces the whole job\n // state rather than merging it, whichever POST arrived LAST won, even if it was queued first and carried a smaller\n // done[] snapshot than one queued after it — silently losing already-finished files from the persisted job, so\n // /resume could rebuild files that were never actually missing.\n let saveChain = Promise.resolve();\n const saveJob = (patch) => {\n if (!job) return Promise.resolve();\n Object.assign(job, patch);\n const body = JSON.stringify({ session: currentSession, artifactId: job.artifactId, state: job });\n saveChain = saveChain.then(() => fetch('?job', { method: 'POST', headers: { 'content-type': 'application/json' }, body }).catch(() => {}));\n return saveChain;\n };\n const post = (path, body) => fetchT(path, { method: 'POST', headers: { 'content-type': 'application/json' }, body: JSON.stringify(body) }).then(async r => { const j = await r.json().catch(() => ({})); if (!r.ok || j.error) throw new Error(typeof j.error === 'string' ? j.error : (j.error?.message || ('HTTP ' + r.status))); return j; });\n function manifestCard(m, artifactId) {\n const w = addMsg('ai', ''); w.querySelector('.msg-body').remove(); w.querySelector('.msg-actions').remove();\n const el = document.createElement('div'); el.className = 'manifest';\n el.innerHTML = '

';\n el.querySelector('span').textContent = '#' + artifactId + ' ' + (m.title || '') + ' · ' + m.files.length + ' files';\n const rows = {};\n for (const f of m.files) { if (rows[f.path]) continue; const r = document.createElement('div'); r.className = 'manifest-f'; r.innerHTML = 'queued'; r.querySelector('code').textContent = f.path; r.querySelector('i').textContent = f.purpose; r.querySelector('[data-a=re]').onclick = () => buildOne(m, artifactId, f.path, 2).then(() => ART.open(artifactId)); el.appendChild(r); rows[f.path] = r; }\n rowsByArtifact.set(artifactId, rows);\n el.querySelector('[data-a=open]').onclick = () => ART.open(artifactId, 'split');\n el.querySelector('[data-a=go]').onclick = () => fanout(m, artifactId);\n w.appendChild(el); turnLog.push({ q: '/plan', reply: JSON.stringify(m, null, 1).slice(0, 4000), meta: { swarm: true }, ts: Date.now() });\n return el;\n }\n const st = (artifactId, path, s, note) => { const rows = rowsByArtifact.get(artifactId); const r = rows && rows[path]; if (!r) return; const e = r.querySelector('.st'); e.textContent = s; e.className = 'st ' + (s === 'ok' ? 'ok' : s === 'error' || s === 'issues' ? 'err' : s === 'queued' ? '' : 'run'); if (note !== undefined) r.querySelector('i').textContent = note; };\n async function plan(ask) {\n if (!ask) { barStatus.textContent = '/plan needs a description'; return false; }\n if (inflight) return false; inflight = true; setSendState(true); addMsg('user', '/plan ' + ask); showTyping(true, 'architect');\n try { const r = await post('?plan', { q: ask, session: currentSession, settings: SETTINGS.forRequest() }); showTyping(false); manifestCard(r.manifest, r.artifactId); job = { artifactId: r.artifactId, manifest: r.manifest, done: [], failed: [], status: 'planned', title: r.title || '', ask }; await saveJob({}); ART.listForSession(currentSession); loadSessions(); return r; }\n catch (e) { showTyping(false); addError('plan failed: ' + e.message, [], '/plan ' + ask); return false; }\n finally { inflight = false; setSendState(false); }\n }\n async function buildOne(m, artifactId, path, attempts = 2, run = { cancelled: false }) {\n let last = { path, error: 'not built' };\n // A builder that fails because providers are exhausted (502/503) is retried within the same window the chat turn uses\n // (retryMinutes, or forever in until-finalize mode) — a missing file is never accepted because the roster was busy for a minute.\n const S = SETTINGS.current(); const retryMs = S.untilStopped ? Infinity : Math.max(0, Number(S.retryMinutes ?? 5)) * 60_000; const t0 = Date.now();\n for (let a = 1; (a <= attempts || (Date.now() - t0 < retryMs && /EXHAUSTED|COOLING|HTTP 50[23]|TIMEOUT|failed to fetch/i.test(last.error || ''))) && !run.cancelled && a < 60; a++) {\n st(artifactId, path, a > 1 ? 'retry ' + a : 'building');\n try {\n const r = await post('?build_file', { session: currentSession, artifactId, manifest: m, path, settings: { ...SETTINGS.forRequest(), model: a > 1 ? 'coder' : SETTINGS.forRequest().model } });\n last = r; const bad = r.truncated || (r.issues && r.issues.length);\n st(artifactId, path, bad ? (a < attempts ? 'retry' : 'issues') : 'ok', r.lines + ' lines · ' + ((r.meta && r.meta.instance) || '') + (r.truncated ? ' · truncated' : '') + (r.issues?.length ? ' · ' + r.issues[0].slice(path.length + 2, 100) : ''));\n if (!bad) { if (job && job.artifactId === artifactId && !job.done.includes(path)) { job.done.push(path); saveJob({}); } return r; }\n } catch (e) { last = { error: e.message, path }; const again = a < attempts || (Date.now() - t0 < retryMs && /EXHAUSTED|COOLING|HTTP 50[23]|TIMEOUT|failed to fetch/i.test(e.message)); st(artifactId, path, again ? 'retry' : 'error', e.message.slice(0, 120)); if (!run.cancelled && again) await wait(jitter(Math.min(45_000, 6000 * a))); }\n }\n if (job && job.artifactId === artifactId && !job.failed.includes(path)) { job.failed.push(path); saveJob({}); }\n return last;\n }\n async function fanout(m, artifactId) {\n if (inflight) return; inflight = true; setSendState(true);\n const run = { cancelled: false }; activeRun = run;\n if (!job || job.artifactId !== artifactId) job = { artifactId, manifest: m, done: [], failed: [], status: 'building', title: m.title || '', ask: '' };\n const skip = new Set(job.done); job.failed = []; saveJob({ status: 'building' });\n const conc = Math.max(1, Math.min(8, SETTINGS.current().swarm || 3)); const queue = m.files.map(f => f.path).filter(p => !skip.has(p)); const results = []; const total = queue.length; let n = 0;\n for (const p of skip) st(artifactId, p, 'ok', 'kept from the previous run');\n barStatus.textContent = 'building ' + queue.length + ' files × ' + conc + ' parallel'; ART.open(artifactId, 'split'); ART.jobs(queue.map(p => ({ path: p, st: 'queued' })));\n const worker = async () => { while (queue.length && !run.cancelled) { const p = queue.shift(); ART.jobs([{ path: p, st: 'building' }]); tabProgress('build ' + n + '/' + total); const r = await buildOne(m, artifactId, p, 2, run); results.push(r); n++; tabProgress('build ' + n + '/' + total); ART.jobs([{ path: p, st: r.error ? 'error' : (r.truncated || r.issues?.length) ? 'issues' : 'ok' }]); ART.open(artifactId); } };\n // staggered starts: N builders launched in the same instant is the burst that trips Val Town's edge and per-second vendor limits.\n // inflight/setSendState are NOT released here anymore — they stay held straight through the integrate step below (which\n // takes over the same lock via sendWith's __ownsInflight). Previously they were released right after Promise.all, leaving\n // a window where a manual send could slip past sendMsg's inflight check, and sendWith's own inflight guard would then\n // silently reject the integrate call started a moment later — and because sendWith swallows its own errors/aborts,\n // fanout had no way to notice and would still mark the job \"done\" below. Both problems close together: the lock stays\n // held (no window to slip through) and the integrate call's success is now checked via its return value (see below).\n try { await Promise.all(Array.from({ length: conc }, (_, i) => wait(i * 1500).then(worker))); } finally { tabProgress(''); }\n if (run.cancelled) { inflight = false; setSendState(false); barStatus.textContent = 'build stopped — /resume continues it'; saveJob({ status: 'stopped' }); ART.jobs(null); return; }\n // full-project lint AFTER every file exists: dangling refs, missing exports, syntax — this is what the integrator must fix\n const full = await fetch('?artifact_files=' + artifactId).then(r => r.json()).catch(() => ({ issues: [] }));\n const issues = (full.issues || []).concat(results.filter(r => r.error).map(r => r.path + ': ' + r.error)).concat(results.filter(r => r.truncated).map(r => r.path + ': output was truncated — verify the file ends correctly'));\n barStatus.textContent = 'integrating'; saveJob({ status: 'integrating' });\n const integrated = await sendWith(ART.tpl('integrate', { id: artifactId, issues: issues.length ? issues.join('\\n') : '(none reported — still verify imports/exports)' }), { focus: { id: artifactId, title: m.title || '', file: 'index.html', files: m.files.map(f => f.path), selection: '', errors: [] }, __quiet: true, __ownsInflight: true });\n inflight = false; setSendState(false);\n ART.open(artifactId, 'split'); ART.jobs(null);\n if (integrated) { await saveJob({ status: 'done' }); job = null; }\n else { barStatus.textContent = 'build stopped before integration finished — /resume to retry integration'; await saveJob({ status: 'stopped' }); }\n }\n async function build(ask) { const r = await plan(ask); if (r) await fanout(r.manifest, r.artifactId); return !!r; }\n // Unfinished builds for this session (server-side job rows) → a card with the manifest, finished files marked, and a resume button.\n async function checkResume(auto) {\n const j = await fetch('?job=list&session=' + encodeURIComponent(currentSession)).then(r => r.json()).catch(() => ({ jobs: [] }));\n const list = (j.jobs || []).filter(x => x.manifest && x.artifactId);\n if (!list.length) { if (!auto) barStatus.textContent = 'nothing to resume'; return false; }\n for (const x of list) {\n const el = manifestCard(x.manifest, x.artifactId);\n el.querySelector('.manifest-h span').textContent += ' · unfinished (' + x.done.length + '/' + x.manifest.files.length + ' built, ' + x.status + ')';\n const go = el.querySelector('[data-a=go]'); go.textContent = 'resume build';\n for (const p of x.done) st(x.artifactId, p, 'ok'); for (const p of x.failed || []) st(x.artifactId, p, 'error', 'failed last run — will retry');\n go.onclick = () => { job = { artifactId: x.artifactId, manifest: x.manifest, done: x.done, failed: [], status: 'building', title: x.title, ask: x.ask }; fanout(x.manifest, x.artifactId); };\n if (!auto && list.length === 1) go.onclick();\n }\n return true;\n }\n return { plan, build, fanout, checkResume, stop: () => { if (activeRun) activeRun.cancelled = true; } };\n})();\n\n// ─── runs inspector ───────────────────────────────────────────────────────────\nfunction renderRuns() {\n const host = $('runs'); host.innerHTML = '';\n if (!turnLog.length) { host.innerHTML = '

no turns in this session yet

'; return; }\n for (const t of [...turnLog].reverse()) {\n const m = t.meta || {}; const r = document.createElement('div'); r.className = 'run';\n const h = document.createElement('div'); h.className = 'run-h';\n h.innerHTML = '';\n h.querySelector('.t').textContent = fmtTs(t.ts); h.querySelector('.q').textContent = t.q || '';\n h.querySelector('.m').innerHTML = [m.instance ? '' + MD.esc(m.instance) + '' : '', m.model ? MD.esc(String(m.model).split('/').pop()) : '', m.toolEvents?.length ? m.toolEvents.length + ' tools' : '', m.tokensOut != null ? m.tokensOut + ' out' : '', m.finishReason && m.finishReason !== 'stop' ? '' + m.finishReason + '' : '', m.telemetry?.latencyMs ? (m.telemetry.latencyMs / 1000).toFixed(1) + 's' : ''].filter(Boolean).map(x => '' + x + '').join('');\n const b = document.createElement('div'); b.className = 'run-b';\n const kv = document.createElement('div'); kv.className = 'run-kv';\n const pairs = [['instance', m.instance], ['model', m.model], ['max_tokens sent', m.maxTokensSent], ['max_tokens asked', m.maxTokensAsked], ['finish', m.finishReason], ['tokens out', m.tokensOut], ['passes', m.telemetry?.passes], ['latency', m.telemetry?.latencyMs != null ? (m.telemetry.latencyMs / 1000).toFixed(1) + 's' : null], ['confidence', m.telemetry?.confidence], ['stakes', m.telemetry?.stakes], ['verifier', m.telemetry?.verifierVerdict], ['skills', m.skills?.join(', ')], ['continuation', m.continuation ? 'yes' : null], ['truncated', m.truncated ? 'yes' : null], ['attempts', m.attempts], ['router retries', m.retries]];\n kv.innerHTML = pairs.filter(([, v]) => v != null && v !== '').map(([k, v]) => '

' + k + ' ' + MD.esc(String(v)) + '

').join('');\n b.appendChild(kv); if (m.toolEvents?.length) b.appendChild(TL.make(m.toolEvents, { open: true }).root);\n const rp = document.createElement('div'); rp.className = 'run-reply'; rp.textContent = (t.reply || '').slice(0, 3000); b.appendChild(rp);\n h.onclick = () => r.classList.toggle('open'); r.appendChild(h); r.appendChild(b); host.appendChild(r);\n }\n}\n\n// ─── command palette ──────────────────────────────────────────────────────────\nconst PAL = (() => {\n const box = $('pal'), input = $('pal-in'), list = $('pal-list'); let items = [], sel = 0;\n function all() {\n const S = SETTINGS.current(); const out = [];\n out.push({ k: 'chat', l: 'New chat', run: newChat });\n out.push({ k: 'view', l: 'Build — open workbench', h: '⌘B', run: () => setView('build') }, { k: 'view', l: 'Runs — inspector', run: () => setView('runs') }, { k: 'view', l: 'Chat', run: () => setView('chat') });\n for (const c of SLASH.CMDS) out.push({ k: 'cmd', l: c.c + ' — ' + c.h, run: () => prefill(c.c + ' ') });\n for (const m of ['auto', 'fast', 'coder', 'reasoning']) out.push({ k: 'route', l: 'Routing: ' + (m === 'auto' ? 'best free' : m) + (S.mode === m ? ' ✓' : ''), run: () => { SETTINGS.set('mode', m); renderChips(); } });\n for (const t of ['midnight', 'black', 'paper']) out.push({ k: 'theme', l: 'Theme: ' + t + (S.theme === t ? ' ✓' : ''), run: () => SETTINGS.set('theme', t) });\n out.push({ k: 'setting', l: 'Auto-continue cut replies: ' + (S.autoContinue ?? 6) + ' turns', run: () => SETTINGS.toggle(true) }, { k: 'setting', l: 'Swarm concurrency: ' + (S.swarm || 3), run: () => SETTINGS.toggle(true) }, { k: 'setting', l: 'Settings…', h: '⌘,', run: () => SETTINGS.toggle(true) }, { k: 'setting', l: 'Editor settings…', run: () => { setView('build'); setTimeout(() => ART.toggleSettings(true), 150); } }, { k: 'setting', l: 'Router status…', run: () => { SETTINGS.toggle(true); SETTINGS.routerStatus(); } });\n for (const k of (S.skills || [])) out.push({ k: 'skill', l: 'Skill: ' + k.name + ' (' + k.mode + ')', run: () => prefill('/skill ' + k.name + ' ') });\n for (const a of ART.list()) out.push({ k: 'artifact', l: '#' + a.id + ' ' + a.title, run: () => ART.open(a.id, 'split') });\n for (const s of [...sessionsEl.querySelectorAll('.sess-row')]) out.push({ k: 'session', l: s.querySelector('.sess-name').textContent, run: () => s.click() });\n return out;\n }\n function render() { const q = input.value.trim().toLowerCase(); items = all().filter(i => !q || (i.k + ' ' + i.l).toLowerCase().includes(q)).slice(0, 40); sel = 0; list.innerHTML = items.map((i, n) => '

' + i.k + '' + MD.esc(i.l) + '' + (i.h || '') + '

').join(''); list.querySelectorAll('.pal-i').forEach(el => el.onclick = () => pick(Number(el.dataset.i))); }\n function pick(i) { const it = items[i]; close(); if (it) it.run(); }\n function open() { box.classList.add('on'); input.value = ''; render(); input.focus(); }\n function close() { box.classList.remove('on'); inp.focus(); }\n input.addEventListener('input', render);\n input.addEventListener('keydown', e => { if (e.key === 'ArrowDown') sel = (sel + 1) % items.length; else if (e.key === 'ArrowUp') sel = (sel - 1 + items.length) % items.length; else if (e.key === 'Enter') return pick(sel); else if (e.key === 'Escape') return close(); else return; e.preventDefault(); list.querySelectorAll('.pal-i').forEach((el, n) => el.classList.toggle('sel', n === sel)); });\n return { open, close, toggle: () => box.classList.contains('on') ? close() : open() };\n})();\n\n// ─── init ─────────────────────────────────────────────────────────────────────\n(async () => {\n const list = await fetch('?sessions').then(r => r.json()).catch(() => []);\n renderSessions(list);\n const known = list.find(s => s.id === currentSession);\n if (known || urlSession) await switchSession(currentSession, known?.name); else rememberSession();\n renderChips(); ART.restore();\n})();\n`;\n",[],[414],"acd4d095-6049-403f-ae67-d55d9690c77a","main-html.tsx","01a0a661-f04a-733e-839a-8eb9391187d2",["D",1789497831502],"/** @jsxImportSource https://esm.sh/react@19.2.0 */\nimport { style } from \"./main-style.ts\";\nimport { script } from \"./main-script.ts\";\nimport { mdStyle } from \"./ui-style-md.ts\";\nimport { markdownScript } from \"./ui-markdown.ts\";\nimport { settingsHtml, settingsScript } from \"./ui-settings.ts\";\nimport { settingsStyle } from \"./ui-style-settings.ts\";\nimport { artifactHtml, artifactStyle } from \"./ui-artifacts.ts\";\nimport { artifactScript } from \"./ui-artifacts-script.ts\";\nimport { timelineScript, timelineStyle } from \"./ui-timeline.ts\";\n\nconst ICON = {\n menu: ``,\n x: ``,\n plus: ``,\n gear: ``,\n send: ``,\n cmd: ``,\n};\n\nexport const PAGE_HTML: string = `\n\n \n \n ZIGGYWARE\n \n\n\n \n\n

\n \n NEW CHAT\n \n

\n best free\n \n \n

\n\n

// ziggyware free ai //

What are we building?

/build — plan + parallel builders + integrate/plan — manifest only/fix — repair the open artifact⌘K — everything else

thinking

draft so far

↓ bottom

⏎ send ⇧⏎ newline / commands

\n\n

no turns in this session yet

\n\n

\n ${settingsHtml}\n ${artifactHtml}\n \n \n \n \n \n\n`;\n",[],[421],"a8445db2-0623-4d95-8bd1-9bf82c96f1cd","lint-portability.ts","01a0a661-f04a-733e-839a-88d123b6ac9a",["D",1789497831502],"import { lintArtifact, lintPortability } from \"./artifacts.ts\";\nconst files = [\n { path: \"index.html\", content: `` },\n { path: \"main.js\", content: `import {DATA} from './data.js';\\nfetch('./levels.json').then(r=>r.json());\\nparent.postMessage({a:1},'*');\\nconst u = location.origin + '/api';\\nlocalStorage.setItem('k','v');\\nnavigator.serviceWorker.register('sw.js');\\nconst img='/artifact/12/x.png';\\nexport function boot(){ return DATA.length }` },\n { path: \"data.js\", content: `export const DATA=[1,2,3];` },\n { path: \"ok.js\", content: `// comment /artifact/1\\nfunction save(k,v){ try{ localStorage.setItem(k,v) }catch(e){ mem[k]=v } }\\nconst mem={};\\nif (window.parent !== window) { parent.postMessage('x','*'); }\\nexport { save };` },\n];\nconst p = lintPortability(files); console.log(p.join(\"\\n\")); console.log(\"count\", p.length, \"ok.js hits\", p.filter(x=>x.startsWith(\"ok.js\")).length);\nconsole.log(\"full\", lintArtifact(files).length);\n",[],[428],"5d3d5602-8ff2-4d22-b917-d66b9bef49b7","issues.ts","01a0a661-f04a-733e-839a-86cccd9d59e4",["D",1789497831502],"// Durable bug/feature backlog (the `issue` table in db.ts) — agent tool functions plus a small read endpoint.\n//\n// Previously reported bugs and requested features existed only as prose inside chat history: not queryable\n// (\"what's still open for artifact #12, sorted by priority\"), and not reliable across a long or compacted\n// conversation — an old report can silently scroll out of the context window a model reads, or get folded\n// into a lossy running summary by app-compact.ts. These tools give the model a durable place to write down\n// what it found (or was told) and read it back later, in this session or a future one, instead of depending\n// on the item still being visible in recent chat turns.\nimport { all, one, run, sql, unwrap } from \"./db.ts\";\n\n// Local structural types instead of importing app.tsx's Tool/ToolCtx — mirrors app-swarm.ts's SwarmDeps\n// pattern (see its header comment): avoids a runtime import cycle back into app.tsx, and TypeScript's\n// structural typing means these satisfy app.tsx's own Tool/ToolCtx shapes without needing to share the type.\ntype ToolCtx = { session: string; turn?: { artifactId?: number } };\ntype Tool = { schema: Record; handle: (args: Record, ctx: ToolCtx) => Promise };\n\nconst KINDS = [\"bug\", \"feature\"] as const;\nconst STATUSES = [\"open\", \"in_progress\", \"fixed\", \"wontfix\", \"duplicate\"] as const;\nconst PRIORITIES = [\"low\", \"medium\", \"high\", \"critical\"] as const;\ntype Status = (typeof STATUSES)[number];\n\nexport const ISSUE_TOOLS: Record = {\n log_issue: {\n schema: {\n type: \"function\",\n function: {\n name: \"log_issue\",\n description:\n \"Record a bug or feature request in the durable backlog so it survives this turn and can be found again later — by you in a future session, or a human reviewing progress. Call this whenever the user reports something broken, asks for something to be added later rather than now, or you notice a defect while working that you are not fixing immediately. Do not use it for something you are about to fix right now in this same turn — just fix it.\",\n parameters: {\n type: \"object\",\n properties: {\n kind: { type: \"string\", enum: KINDS as unknown as string[], description: \"bug: something is broken. feature: something new to add.\" },\n title: { type: \"string\", description: \"One line summary, ≤80 chars.\" },\n description: { type: \"string\", description: \"What's wrong or wanted, with repro steps / context if it's a bug.\" },\n priority: { type: \"string\", enum: PRIORITIES as unknown as string[], description: \"Default medium.\" },\n artifactId: { type: \"number\", description: \"The artifact this concerns, if any. Omit for a session-level or general item.\" },\n },\n required: [\"kind\", \"title\", \"description\"],\n },\n },\n },\n handle: async ({ kind, title, description, priority, artifactId }, ctx) => {\n const k = String(kind);\n if (k !== \"bug\" && k !== \"feature\") return { error: `kind must be \"bug\" or \"feature\", got \"${k}\"` };\n const pRaw = String(priority || \"medium\");\n const pri = (PRIORITIES as readonly string[]).includes(pRaw) ? pRaw : \"medium\";\n const aid = Number.isInteger(Number(artifactId)) && Number(artifactId) > 0 ? Number(artifactId) : (ctx.turn?.artifactId ?? null);\n const now = Date.now();\n const r = await run(sql`INSERT INTO issue (session, artifact_id, kind, title, description, status, priority, source, created_ts, updated_ts, resolved_ts, notes)\n VALUES (${ctx.session}, ${aid}, ${k}, ${String(title).slice(0, 200)}, ${String(description).slice(0, 4000)}, 'open', ${pri}, 'model', ${now}, ${now}, NULL, '[]')`);\n if (!r.ok) return { error: r.error };\n const id = r.value.lastInsertRowid === undefined ? null : Number(r.value.lastInsertRowid);\n return { ok: true, id, kind: k, title: String(title).slice(0, 200), priority: pri };\n },\n },\n list_issues: {\n schema: {\n type: \"function\",\n function: {\n name: \"list_issues\",\n description:\n \"List items in the bug/feature backlog for this session (default: open items only), so work picks up where it left off instead of re-discovering the same problems or forgetting a requested feature.\",\n parameters: {\n type: \"object\",\n properties: {\n status: { type: \"string\", enum: [...STATUSES, \"all\"], description: 'Default \"open\".' },\n kind: { type: \"string\", enum: [...KINDS, \"all\"], description: 'Default \"all\".' },\n artifactId: { type: \"number\", description: \"Limit to one artifact.\" },\n },\n },\n },\n },\n handle: async ({ status, kind, artifactId }, ctx) => {\n const st = String(status || \"open\");\n const kd = String(kind || \"all\");\n const aid = artifactId != null && Number.isInteger(Number(artifactId)) ? Number(artifactId) : null;\n const rows = unwrap(\n await all(\n \"issue\",\n sql`SELECT * FROM issue WHERE session = ${ctx.session}\n AND (${st} = 'all' OR status = ${st})\n AND (${kd} = 'all' OR kind = ${kd})\n AND (${aid} IS NULL OR artifact_id = ${aid})\n ORDER BY (priority = 'critical') DESC, (priority = 'high') DESC, created_ts ASC LIMIT 100`,\n ),\n [],\n );\n return {\n count: rows.length,\n issues: rows.map((r) => ({ id: r.id, kind: r.kind, title: r.title, description: r.description, status: r.status, priority: r.priority, artifactId: r.artifact_id, created: r.created_ts })),\n };\n },\n },\n update_issue: {\n schema: {\n type: \"function\",\n function: {\n name: \"update_issue\",\n description:\n \"Update a backlog item's status — e.g. after fixing a bug or shipping a feature. Always include a short note saying what actually happened; it's appended to the item's audit trail, not shown as a replacement description.\",\n parameters: {\n type: \"object\",\n properties: {\n id: { type: \"number\", description: \"Issue id, from log_issue or list_issues.\" },\n status: { type: \"string\", enum: STATUSES as unknown as string[] },\n note: { type: \"string\", description: \"What changed — required.\" },\n },\n required: [\"id\", \"status\", \"note\"],\n },\n },\n },\n handle: async ({ id, status, note }) => {\n const iid = Number(id);\n const st = String(status);\n if (!(STATUSES as readonly string[]).includes(st)) return { error: `status must be one of ${STATUSES.join(\", \")}` };\n const cur = unwrap(await one(\"issue\", sql`SELECT * FROM issue WHERE id = ${iid}`), null);\n if (!cur) return { error: `no issue with id ${iid}` };\n let notes: unknown[] = [];\n try { notes = JSON.parse(cur.notes || \"[]\"); } catch { /* corrupt/empty — start fresh rather than fail the update */ }\n notes.push({ ts: Date.now(), text: String(note).slice(0, 1000) });\n const resolvedTs = (st === \"fixed\" || st === \"wontfix\" || st === \"duplicate\") ? Date.now() : null;\n await run(sql`UPDATE issue SET status = ${st as Status}, updated_ts = ${Date.now()}, resolved_ts = ${resolvedTs}, notes = ${JSON.stringify(notes).slice(0, 8000)} WHERE id = ${iid}`);\n return { ok: true, id: iid, status: st };\n },\n },\n};\n\n/** GET ?issues=list&session=S[&status=open][&kind=bug][&artifactId=N] → the same backlog, for non-tool (UI/debug) callers. */\nexport async function handleIssues(req: Request, url: URL): Promise {\n if (req.method !== \"GET\" || url.searchParams.get(\"issues\") !== \"list\") return null;\n const session = String(url.searchParams.get(\"session\") ?? \"\");\n if (!session) return Response.json({ error: { message: \"session required\" } }, { status: 400 });\n const st = String(url.searchParams.get(\"status\") ?? \"open\");\n const kd = String(url.searchParams.get(\"kind\") ?? \"all\");\n const aidParam = url.searchParams.get(\"artifactId\");\n const aid = aidParam != null && Number.isInteger(Number(aidParam)) ? Number(aidParam) : null;\n const rows = unwrap(\n await all(\n \"issue\",\n sql`SELECT * FROM issue WHERE session = ${session}\n AND (${st} = 'all' OR status = ${st})\n AND (${kd} = 'all' OR kind = ${kd})\n AND (${aid} IS NULL OR artifact_id = ${aid})\n ORDER BY (priority = 'critical') DESC, (priority = 'high') DESC, created_ts ASC LIMIT 200`,\n ),\n [],\n );\n return Response.json({ issues: rows });\n}\n",[],[435],"2a86f564-cdd3-458d-baf6-974dfa006127","inference-provider.ts","01a0a661-f04a-733e-839a-81a52891dc3f",["D",1789497831502],"// Instance = catalog row × key slot (KEY, KEY1…KEY9 → name#n). keyEnv null ⇒ keyless.\nimport type { Row } from \"./db.ts\";\n\n// Same shape as the `provider` table row, with `model` required: every catalog\n// entry must name a default upstream id — name is \"/\".\nexport type ProviderSpec = Omit, \"model\"> & { model: string; tags?: string[]; paid?: boolean };\n\nexport type ProviderInstance = ProviderSpec & {\n vendor: string;\n slot: number;\n key: string | null;\n keyIssue: string | null; // non-null ⇒ the secret's VALUE is malformed (see keyIssue()); surfaced on /health and in the catalog\n fallbackModel: string;\n};\n\nconst KEY_SLOTS = 10;\n\n// Secret-value sanity. GEMINI_API_KEY was observed live this session with a literal `…3, }` suffix —\n// a JSON fragment pasted into the secret field — and it sat there across multiple diagnostics pulls\n// because nothing anywhere said \"this key cannot be right\" until an upstream 400 did, buried in a trail.\n// Leading/trailing whitespace is the one defect that is safe to repair silently (no real key contains\n// it); everything else is reported, never guessed at. Returns the reason, or null for a plausible key.\nexport function keyIssue(raw: string): string | null {\n if (/^[\"'`]|[\"'`]$/.test(raw)) return \"wrapped in quotes\";\n if (/[{}\\[\\],]/.test(raw)) return \"contains JSON punctuation ({ } [ ] ,)\";\n if (/\\s/.test(raw)) return \"contains internal whitespace\";\n if (/^bearer\\s/i.test(raw)) return 'starts with \"Bearer \" — paste only the token';\n if (raw.length < 8) return `implausibly short (${raw.length} chars)`;\n return null;\n}\n\n// ids verified 2026-09-08. Dropped: github_models (410, retiring), deepinfra (paid).\n// tags mirror v2/catalog-free.ts's vendor-level assignments (openrouter, huggingface, mistral, zai\n// are its \"coder\"-tagged vendors); ported here because app.tsx sends model:\"coder\" as a routing tag\n// for app-scale builds (see app.tsx ~line 1288), which this catalog previously had no way to honor.\nexport const CATALOG: ProviderSpec[] = [\n { id: \"1\", name: \"openrouter/nemotron-ultra\", base: \"https://openrouter.ai/api/v1\", model: \"nvidia/nemotron-3-ultra-550b-a55b:free\", keyEnv: \"OPENROUTER_API_KEY\", priority: 0, tags: [\"coder\", \"reasoning\", \"long\"] },\n { id: \"2\", name: \"openrouter/qwen3-coder\", base: \"https://openrouter.ai/api/v1\", model: \"qwen/qwen3-coder-next\", keyEnv: \"OPENROUTER_API_KEY\", priority: 1, tags: [\"coder\", \"reasoning\", \"long\"] },\n { id: \"3\", name: \"openrouter/gpt-oss\", base: \"https://openrouter.ai/api/v1\", model: \"openai/gpt-oss-120b\", keyEnv: \"OPENROUTER_API_KEY\", priority: 2, tags: [\"coder\", \"reasoning\", \"long\"] },\n // Re-verified 2026-09-14 against live HTTP 404s from the user's own /health: Cerebras has retired\n // Llama 3.3 70B from its public inference API entirely (confirmed via inference-docs.cerebras.ai —\n // only gpt-oss-120b and qwen-3.8-27b remain listed); Groq's llama-3.3-70b-versatile hit its announced\n // free/developer-tier deprecation date (2026-08-16, per console.groq.com/docs/deprecations) and is now\n // 404ing live, with Groq's own migration notice pointing at openai/gpt-oss-120b. Both replaced below.\n { id: \"4\", name: \"cerebras/gpt-oss-120b\", base: \"https://api.cerebras.ai/v1\", model: \"gpt-oss-120b\", keyEnv: \"CEREBRAS_API_KEY\", priority: 3 },\n { id: \"5\", name: \"sambanova/llama3.3-70b\", base: \"https://api.sambanova.ai/v1\", model: \"Meta-Llama-3.3-70B-Instruct\", keyEnv: \"SAMBANOVA_API_KEY\", priority: 4 },\n { id: \"6\", name: \"groq/gpt-oss-120b\", base: \"https://api.groq.com/openai/v1\", model: \"openai/gpt-oss-120b\", keyEnv: \"GROQ_API_KEY\", priority: 5 },\n { id: \"8\", name: \"mistral/large\", base: \"https://api.mistral.ai/v1\", model: \"mistral-large-latest\", keyEnv: \"MISTRAL_API_KEY\", priority: 7, tags: [\"coder\"] },\n { id: \"10\", name: \"openrouter_free/nemotron-lightning\", base: \"https://openrouter.ai/api/v1\", model: \"nvidia/nemotron-3.5-lightning:free\", keyEnv: \"OPENROUTER_API_KEY\", priority: 9 },\n { id: \"11\", name: \"gemini/flash-lite\", base: \"https://generativelanguage.googleapis.com/v1beta/openai\", model: \"gemini-2.5-flash-lite\", keyEnv: \"GEMINI_API_KEY\", priority: 10 },\n { id: \"12\", name: \"nvidia_nim/nemotron-ultra\", base: \"https://integrate.api.nvidia.com/v1\", model: \"nvidia/llama-3.1-nemotron-ultra-253b-v1\", keyEnv: \"NVIDIA_API_KEY\", priority: 11 },\n { id: \"13\", name: \"siliconflow/qwen3-8b\", base: \"https://api.siliconflow.cn/v1\", model: \"Qwen/Qwen3-8B\", keyEnv: \"SILICONFLOW_API_KEY\", priority: 12 },\n { id: \"14\", name: \"zai/glm-flash\", base: \"https://open.bigmodel.cn/api/paas/v4\", model: \"GLM-4.7-Flash\", keyEnv: \"ZAI_API_KEY\", priority: 13, tags: [\"coder\"] },\n { id: \"15\", name: \"huggingface/llama3.3-70b\", base: \"https://router.huggingface.co/v1\", model: \"meta-llama/Llama-3.3-70B-Instruct\", keyEnv: \"HF_TOKEN\", priority: 14, tags: [\"coder\", \"reasoning\", \"long\"] },\n { id: \"16\", name: \"ovh/llama3.3-70b\", base: \"https://oai.endpoints.kepler.ai.cloud.ovh.net/v1\", model: \"Meta-Llama-3_3-70B-Instruct\", keyEnv: \"OVH_AI_ENDPOINTS_ACCESS_TOKEN\", priority: 15 },\n // Paid tier — inert unless the matching key is set; ported from v2/catalog-paid.ts (ids verified 2026-09-08\n // there). priority 40+ so \"auto\"/\"fast\" try every free row first — paid only activates through explicit\n // vendor/tag routing (e.g. model:\"coder\", model:\"anthropic\") or once every free provider has failed.\n { id: \"17\", name: \"deepseek/chat\", base: \"https://api.deepseek.com/v1\", model: \"deepseek-chat\", keyEnv: \"DEEPSEEK_API_KEY\", priority: 40, paid: true, tags: [\"coder\", \"reasoning\"] },\n { id: \"18\", name: \"deepseek/reasoner\", base: \"https://api.deepseek.com/v1\", model: \"deepseek-reasoner\", keyEnv: \"DEEPSEEK_API_KEY\", priority: 41, paid: true, tags: [\"reasoning\"] },\n { id: \"19\", name: \"moonshot/kimi-k2\", base: \"https://api.moonshot.ai/v1\", model: \"kimi-k2-turbo-preview\", keyEnv: \"MOONSHOT_API_KEY\", priority: 42, paid: true, tags: [\"coder\", \"long\"] },\n { id: \"20\", name: \"openai/gpt-5-mini\", base: \"https://api.openai.com/v1\", model: \"gpt-5-mini\", keyEnv: \"OPENAI_API_KEY\", priority: 43, paid: true, tags: [\"coder\", \"reasoning\"] },\n { id: \"21\", name: \"anthropic/claude-sonnet\", base: \"https://api.anthropic.com/v1\", model: \"claude-sonnet-4-5\", keyEnv: \"ANTHROPIC_API_KEY\", priority: 44, paid: true, tags: [\"coder\", \"reasoning\"] },\n];\n\nconst vendorOf = (name: string) => name.split(\"/\")[0];\n\nconst instance = (p: ProviderSpec, slot: number, keyEnv: string | null, key: string | null): ProviderInstance =>\n ({ ...p, keyEnv, key, keyIssue: key ? keyIssue(key) : null, slot, vendor: vendorOf(p.name), name: slot ? `${p.name}#${slot}` : p.name, fallbackModel: p.model });\n\nexport function getProviderArray(): ProviderInstance[] {\n const out: ProviderInstance[] = [];\n for (const p of CATALOG) {\n if (!p.keyEnv) { out.push(instance(p, 0, null, null)); continue; }\n for (let i = 0; i < KEY_SLOTS; i++) {\n const env = p.keyEnv + (i ? String(i) : \"\");\n const key = Deno.env.get(env)?.trim(); // surrounding whitespace is the one defect repaired silently — see keyIssue()\n if (key) out.push(instance(p, i, env, key));\n }\n }\n return out.sort((a, b) => a.priority - b.priority || a.slot - b.slot);\n}\n\n/** All catalog rows for status/UI; no key material. */\nexport function getCatalogStatus() {\n const live = getProviderArray();\n return CATALOG.map((p) => {\n const slots = live.filter((i) => i.id === p.id);\n return {\n ...p,\n vendor: vendorOf(p.name),\n fallbackModel: p.model,\n hasKey: !p.keyEnv || slots.length > 0,\n keySlots: slots.length,\n keyIssues: slots.filter((i) => i.keyIssue).map((i) => ({ env: i.keyEnv, issue: i.keyIssue })),\n };\n }).sort((a, b) => a.priority - b.priority);\n}\n\n/** request `model`: \"\"|\"auto\"|\"fast\" → chain (this router has no quality-vs-priority reordering, so\n * \"fast\" — the UI's other bare meta-model, alongside \"coder\"/\"reasoning\" — is not distinct from \"auto\"\n * here; both just mean \"every configured provider, in existing priority order\"); \"coder\"|\"reasoning\"|\n * \"long\" → tagged instances first, then the rest (never a hard subset — a meta-model must still degrade\n * to *something*); \"\" | \"\" → subset; \":\" → subset + forced id;\n * unknown → [] (caller 404s). */\nexport function resolveTargets(requested: string | undefined, live: ProviderInstance[]) {\n if (!requested || requested === \"auto\" || requested === \"fast\") return { targets: live, forcedModel: null };\n if (/^(coder|reasoning|long|vision)$/.test(requested)) {\n const tagged = live.filter((p) => p.tags?.includes(requested));\n return { targets: [...tagged, ...live.filter((p) => !tagged.includes(p))], forcedModel: null };\n }\n const exact = live.filter((p) => p.name === requested || p.name.split(\"#\")[0] === requested);\n if (exact.length) return { targets: exact, forcedModel: null };\n const colon = requested.indexOf(\":\");\n const vendor = colon === -1 ? requested : requested.slice(0, colon);\n const forced = colon === -1 ? null : requested.slice(colon + 1) || null; // \"\" (trailing colon) must fall through, not ship as model\n const byVendor = live.filter((p) => p.vendor === vendor);\n return { targets: byVendor, forcedModel: byVendor.length ? forced : null };\n}\n",[],[442],"1073d0d0-702e-4242-bfa4-284404bf01ff","e2e9.mjs","01a0a661-f04a-733e-839a-7c2de53e601c",["D",1789497831502],"import { chromium } from \"playwright-core\";\nimport fs from \"node:fs\";\nconst html = fs.readFileSync(\"app/page.html\", \"utf8\");\nconst files = { \"index.html\": \"

x

\", \"app.js\": \"console.log(1);\\n\".repeat(40) };\nconst b = await chromium.launch({ executablePath: \"/opt/pw-browsers/chromium\", args: [\"--no-sandbox\", \"--headless=new\"], ignoreDefaultArgs: [\"--headless=old\"] });\nconst p = await b.newPage({ viewport: { width: 1400, height: 800 } }); const errs = []; p.on(\"pageerror\", e => errs.push(e.message));\nlet qCount = 0, progressCalls = 0;\nawait p.route(\"**/*\", async (route) => { const u = new URL(route.request().url());\n if (u.host === \"cdnjs.cloudflare.com\") { const m = u.pathname.match(/codemirror\\/5\\.65\\.16\\/(.*)$/); const f = m ? \"/tmp/cm/\" + m[1] : null; if (f && fs.existsSync(f)) return route.fulfill({ contentType: f.endsWith(\".css\") ? \"text/css\" : \"text/javascript\", body: fs.readFileSync(f) }); return route.abort(); }\n if (u.host !== \"app.local\") return route.abort();\n if (u.search.includes(\"?artifact_files=\")) return route.fulfill({ contentType: \"application/json\", body: JSON.stringify({ id: 1, title: \"T\", kind: \"html\", session: \"s\", files: Object.entries(files).map(([path, content]) => ({ path, content })), issues: [] }) });\n if (u.search.startsWith(\"?q\")) { qCount++; const body = JSON.parse(route.request().postData()); await new Promise(r => setTimeout(r, 2500));\n if (qCount === 1) return route.fulfill({ status: 503, contentType: \"application/json\", body: JSON.stringify({ error: { message: \"router 503: ALL_PROVIDERS_EXHAUSTED_OR_COOLING\", details: [\"groq: cooling 20s — rate\", \"openrouter: TIMEOUT\"] } }) });\n return route.fulfill({ contentType: \"application/json\", body: JSON.stringify({ reply: \"done after retry (retryOf=\" + body.settings.retryOf + \")\", toolEvents: [{ hop: 0, tool: \"eval_js\", args: { code: \"1\" }, result: 1 }], meta: {} }) }); }\n if (u.search.includes(\"?progress\")) { progressCalls++; return route.fulfill({ contentType: \"application/json\", body: JSON.stringify({ events: [{ type: \"stage\", label: \"answering\" }, { type: \"pass\", n: 1, of: 8 }, { type: \"tool\", hop: 0, tool: \"read_artifact\", args: { id: 1 } }, { type: \"tool\", hop: 0, tool: \"eval_js\", args: { code: \"1\" } }].slice(0, Math.min(4, progressCalls)), next: 4 }) }); }\n if (u.pathname.startsWith(\"/artifact/\")) return route.fulfill({ contentType: \"text/html\", body: files[\"index.html\"] });\n if (u.search) return route.fulfill({ contentType: \"application/json\", body: \"[]\" });\n return route.fulfill({ contentType: \"text/html\", body: html }); });\nawait p.goto(\"https://app.local/\", { waitUntil: \"load\" }); await p.waitForTimeout(300);\n// A) editor resizes with the window\nawait p.evaluate(() => ART.open(1, 'split')); await p.waitForTimeout(900);\nconst h1 = await p.evaluate(() => document.querySelector('.CodeMirror').getBoundingClientRect().height);\nawait p.setViewportSize({ width: 1000, height: 500 }); await p.waitForTimeout(400);\nconst h2 = await p.evaluate(() => ({ cm: document.querySelector('.CodeMirror').getBoundingClientRect().height | 0, wrap: document.getElementById('ap-edwrap').getBoundingClientRect().height | 0, scroll: document.querySelector('.CodeMirror-scroll').clientHeight | 0 }));\nawait p.setViewportSize({ width: 1400, height: 800 }); await p.keyboard.press(\"Escape\"); await p.waitForTimeout(300);\n// B) live timeline: statuses, expanded row survives polling, retry loop\nawait p.evaluate(() => { SETTINGS.set('retryMinutes', 1); });\nawait p.fill(\"#inp\", \"hello\"); await p.keyboard.press(\"Enter\");\nawait p.waitForFunction(() => !!document.querySelector('#live-tl .tl .tl-row[data-kind=tool]'), null, { timeout: 15000 }); // a tool row has arrived (2 s polling cadence)\nconst r1 = await p.evaluate(() => { const tl = document.querySelector('#live-tl .tl'); const rows = [...tl.querySelectorAll('.tl-row')]; rows.find(r => r.dataset.kind === 'tool')?.click(); return { plan: [...tl.querySelectorAll('.tl-pl')].map(x => x.className.replace('tl-pl ', '') + ':' + x.textContent), rows: rows.map(r => r.querySelector('.tl-k').textContent + '=' + r.querySelector('.tl-st').textContent), header: tl.querySelector('.tl-sum').textContent, label: document.getElementById('typing-label').textContent }; });\nawait p.waitForTimeout(2600); // more polls → the expanded row must still be open\nconst r2 = await p.evaluate(() => { const tl = document.querySelector('#live-tl .tl'); return { stillOpen: !!tl.querySelector('.tl-row.open'), xVisible: tl.querySelector('.tl-row.open + .tl-x') ? getComputedStyle(tl.querySelector('.tl-row.open + .tl-x')).display : 'MISSING', html: tl.innerHTML.slice(0, 300), rows: tl.querySelectorAll('.tl-row').length, label: document.getElementById('typing-label').textContent }; });\nawait p.waitForTimeout(12000); // retry wait (8s + jitter) then second attempt\nconst r3 = await p.evaluate(() => ({ q: document.querySelectorAll('.msg-wrap.ai').length, reply: [...document.querySelectorAll('.msg-wrap.ai .msg-body')].pop()?.textContent, err: document.querySelectorAll('.err-card').length }));\nconsole.log(JSON.stringify({ h1, h2, r1, r2, r3, qCount }, null, 1)); console.log(\"errors:\", errs.filter(e => !/cdnjs/.test(e)));\nawait b.close();\n",[],[449],"e050eafe-4651-4dcf-9c67-a8fae7ca5492","e2e15.mjs","01a0a661-f04a-733e-839a-78fb2e8eaf2c",["D",1789497831502],"import { chromium } from \"playwright-core\";\nimport fs from \"node:fs\";\nconst html = fs.readFileSync(\"app/page.html\", \"utf8\");\nconst HOOK = \"\";\nconst files = { \"index.html\": \"

x

\" };\nconst manifest = { title: \"T\", shared: \"\", files: [{ path: \"index.html\", purpose: \"e\" }, { path: \"a.js\", purpose: \"a\" }] };\nconst b = await chromium.launch({ executablePath: \"/opt/pw-browsers/chromium\", args: [\"--no-sandbox\", \"--headless=new\"], ignoreDefaultArgs: [\"--headless=old\"] });\nconst p = await b.newPage({ viewport: { width: 1400, height: 800 } }); const errs = []; p.on(\"pageerror\", e => errs.push(e.message));\nconst posts = []; let buildTries = 0;\nawait p.route(\"**/*\", async (route) => { const u = new URL(route.request().url());\n if (u.host === \"cdnjs.cloudflare.com\") return route.abort();\n if (u.host !== \"app.local\") return route.abort();\n if (u.search.includes(\"?artifact_files=\")) return route.fulfill({ contentType: \"application/json\", body: JSON.stringify({ id: 61, title: \"T\", kind: \"html\", session: \"s\", files: Object.entries(files).map(([path, content]) => ({ path, content })), issues: [] }) });\n if (u.search.startsWith(\"?q\")) { const body = JSON.parse(route.request().postData()); posts.push({ q: body.q.slice(0, 50), errors: body.settings.focus?.errors }); await new Promise(r => setTimeout(r, 200)); return route.fulfill({ contentType: \"application/json\", body: JSON.stringify({ reply: \"fixed\", toolEvents: [], meta: {}, completion: { done: true, remaining: [] } }) }); }\n if (u.search.startsWith(\"?build_file\")) { buildTries++; const body = JSON.parse(route.request().postData()); if (buildTries <= 2) return route.fulfill({ status: 502, contentType: \"application/json\", body: JSON.stringify({ error: { message: \"build a.js failed: router 503: ALL_PROVIDERS_EXHAUSTED_OR_COOLING\" } }) }); return route.fulfill({ contentType: \"application/json\", body: JSON.stringify({ ok: true, path: body.path, lines: 10, issues: [], meta: {} }) }); }\n if (u.search.startsWith(\"?job\")) return route.fulfill({ contentType: \"application/json\", body: '{\"ok\":true,\"jobs\":[]}' });\n if (u.search.startsWith(\"?plan\")) return route.fulfill({ contentType: \"application/json\", body: JSON.stringify({ manifest, artifactId: 61, title: \"T\" }) });\n if (u.search.includes(\"?progress\")) return route.fulfill({ contentType: \"application/json\", body: '{\"events\":[],\"next\":0}' });\n if (u.pathname === \"/artifact/61/\" ) return route.fulfill({ contentType: \"text/html\", body: files[\"index.html\"].replace(\"\", \"\" + HOOK) });\n if (u.pathname.startsWith(\"/artifact/\")) return route.fulfill({ status: 404, body: \"\" });\n if (u.search && !u.search.startsWith(\"?s=\")) return route.fulfill({ contentType: \"application/json\", body: \"[]\" });\n return route.fulfill({ contentType: \"text/html\", body: html }); });\nawait p.goto(\"https://app.local/\", { waitUntil: \"load\" }); await p.waitForTimeout(400);\nawait p.evaluate(() => { ART.setSettings({ ...ART.settings(), codemirror: false, autoFixLint: false, autoFixErrors: true, live: false }); SETTINGS.set('retryMinutes', 1); });\n// A) preview errors → one automatic fix turn carrying the errors; same error set again → no second turn\nawait p.evaluate(() => ART.open(61, 'split')); await p.waitForTimeout(3500);\nconst a = { posts: posts.map(x => x.q), errorsSent: posts[0]?.errors, console: await p.evaluate(() => [...document.querySelectorAll('#ap-con-body .con-err')].map(x => x.textContent)) };\nawait p.evaluate(() => ART.open(61)); await p.waitForTimeout(3000);\nconst a2 = { postsAfterReload: posts.length };\n// B) builder retried through a 502 window\nawait p.evaluate(() => ART.close());\nawait p.fill(\"#inp\", \"/build a thing\"); await p.keyboard.press(\"Enter\"); await p.waitForTimeout(16000);\nconst bres = { buildTries, rows: await p.evaluate(() => [...document.querySelectorAll('.manifest-f')].map(r => r.querySelector('code').textContent + '=' + r.querySelector('.st').textContent)) };\nconsole.log(JSON.stringify({ a, a2, bres }, null, 1)); console.log(\"errors:\", errs);\nawait b.close();\n",[],[456],"573c72be-77fb-4553-9ea4-5801a111aaaf","e2e14.mjs","01a0a661-f04a-733e-839a-7408c9f95419",["D",1789497831502],"import { chromium } from \"playwright-core\";\nimport fs from \"node:fs\";\nconst html = fs.readFileSync(\"app/page.html\", \"utf8\");\nconst b = await chromium.launch({ executablePath: \"/opt/pw-browsers/chromium\", args: [\"--no-sandbox\", \"--headless=new\"], ignoreDefaultArgs: [\"--headless=old\"] });\nconst p = await b.newPage({ viewport: { width: 1400, height: 800 } }); const errs = []; p.on(\"pageerror\", e => errs.push(e.message));\nlet q = 0;\nawait p.route(\"**/*\", async (route) => { const u = new URL(route.request().url());\n if (u.host !== \"app.local\") return route.abort();\n if (u.search.startsWith(\"?q\")) { q++; await new Promise(r => setTimeout(r, 5000)); return route.fulfill({ status: 500, contentType: \"application/json\", body: JSON.stringify({ error: { message: \"Val timed out after 1m\" } }) }); }\n if (u.search.includes(\"?progress\")) return route.fulfill({ contentType: \"application/json\", body: JSON.stringify({ next: 3, events: [{ type: \"stage\", label: \"answering\" }, { type: \"partial\", text: \"const store = createStore(reducer, init);\\nexport function mount() {\" }, { type: \"continue\", n: 1, chars: 1840, tail: \"export function mount() {\" }] }) });\n if (u.search.startsWith(\"?history\")) return route.fulfill({ contentType: \"application/json\", body: \"[]\" });\n if (u.search && !u.search.startsWith(\"?s=\")) return route.fulfill({ contentType: \"application/json\", body: \"[]\" });\n return route.fulfill({ contentType: \"text/html\", body: html }); });\nawait p.goto(\"https://app.local/\", { waitUntil: \"load\" }); await p.waitForTimeout(400);\nawait p.evaluate(() => { SETTINGS.set('retryMinutes', 0); });\nawait p.fill(\"#inp\", \"fix the lint\"); await p.keyboard.press(\"Enter\");\nawait p.waitForTimeout(3000);\nconst live = await p.evaluate(() => ({ draftShown: !document.getElementById('live-draft').hidden, draft: document.getElementById('live-draft-t').textContent.slice(0, 40), contRow: [...document.querySelectorAll('#live-tl .tl-row')].map(r => r.textContent).find(t => /cut at the vendor cap/.test(t))?.slice(0, 90) }));\nawait p.waitForTimeout(30000); // error + recovery polls (8 × 3 s)\nconst after = await p.evaluate(() => { const c = document.querySelector('.err-card'); return { err: c?.querySelector('.err-m')?.textContent, keptTimelineRows: c?.querySelectorAll('.tl-row').length ?? 0, keptDraft: /draft reply recovered/.test(c?.textContent || ''), draftText: c?.querySelector('details pre')?.textContent.slice(0, 30) }; });\nconsole.log(JSON.stringify({ live, after, q }, null, 1)); console.log(\"errors:\", errs);\nawait b.close();\n",[],[463],"03988cb1-3226-499b-8f3b-8794a5bc6d9f","e2e13.mjs","01a0a661-f049-7779-b407-bc7c85275a5f",["D",1789497831502],"import { chromium } from \"playwright-core\";\nimport fs from \"node:fs\";\nconst html = fs.readFileSync(\"app/page.html\", \"utf8\");\nconst b = await chromium.launch({ executablePath: \"/opt/pw-browsers/chromium\", args: [\"--no-sandbox\", \"--headless=new\"], ignoreDefaultArgs: [\"--headless=old\"] });\nconst p = await b.newPage({ viewport: { width: 1400, height: 800 } }); const errs = []; p.on(\"pageerror\", e => errs.push(e.message));\nconst posts = []; let phase = \"write\";\nawait p.route(\"**/*\", async (route) => { const u = new URL(route.request().url());\n if (u.host !== \"app.local\") return route.abort();\n if (u.search.startsWith(\"?q\")) { const body = JSON.parse(route.request().postData()); posts.push(body.q.slice(0, 40)); await new Promise(r => setTimeout(r, 150));\n const wrote = phase === \"write\" || !/^The task is complete/.test(body.q);\n return route.fulfill({ contentType: \"application/json\", body: JSON.stringify({ reply: /FINALIZE/.test(body.q) ? \"final summary\" : \"shipped round\", toolEvents: wrote ? [{ hop: 0, tool: \"update_artifact\", args: { id: 7 }, result: { ok: true, id: 7 } }] : [], meta: {}, completion: { done: true, remaining: [] } }) }); }\n if (u.search.includes(\"?progress\")) return route.fulfill({ contentType: \"application/json\", body: '{\"events\":[],\"next\":0}' });\n if (u.search.startsWith(\"?compact\")) return route.fulfill({ contentType: \"application/json\", body: '{\"compact\":null,\"needs\":false}' });\n if (u.search && !u.search.startsWith(\"?s=\")) return route.fulfill({ contentType: \"application/json\", body: \"[]\" });\n return route.fulfill({ contentType: \"text/html\", body: html }); });\nawait p.goto(\"https://app.local/\", { waitUntil: \"load\" }); await p.waitForTimeout(400);\nawait p.evaluate(() => { SETTINGS.set('untilStopped', true); });\n// A) rounds keep coming while every round writes; \"finalize\" typed in the composer ends the chain with one closing turn\nawait p.fill(\"#inp\", \"build me a thing\"); await p.keyboard.press(\"Enter\"); await p.waitForTimeout(6000);\nconst midA = { posts: posts.length, bar: await p.evaluate(() => document.getElementById('bar-status').textContent) };\nawait p.fill(\"#inp\", \"finalize\"); await p.keyboard.press(\"Enter\"); await p.waitForTimeout(3000);\nconst afterA = { posts: [...posts], inflight: await p.evaluate(() => window.inflightTurn()), last: await p.evaluate(() => [...document.querySelectorAll('.msg-wrap.ai .msg-body')].pop()?.textContent) };\n// B) idle stop: rounds that write nothing end after two\nposts.length = 0; phase = \"idle\";\nawait p.fill(\"#inp\", \"another thing\"); await p.keyboard.press(\"Enter\"); await p.waitForTimeout(6000);\nconst afterB = { posts: [...posts], inflight: await p.evaluate(() => window.inflightTurn()), autoCard: await p.evaluate(() => [...document.querySelectorAll('.msg-wrap.sys .msg-body')].some(x => /no further upgrades/.test(x.textContent))) };\nconsole.log(JSON.stringify({ midA, afterA, afterB }, null, 1)); console.log(\"errors:\", errs);\nawait b.close();\n",[],[470],"dfd7b3ab-46c9-489d-b5df-ce042cdc86ba","e2e12.mjs","01a0a661-f049-7779-b407-b9b842e09136",["D",1789497831502],"import { chromium } from \"playwright-core\";\nimport fs from \"node:fs\";\nconst html = fs.readFileSync(\"app/page.html\", \"utf8\");\nconst b = await chromium.launch({ executablePath: \"/opt/pw-browsers/chromium\", args: [\"--no-sandbox\", \"--headless=new\"], ignoreDefaultArgs: [\"--headless=old\"] });\nconst p = await b.newPage({ viewport: { width: 1400, height: 800 } }); const errs = []; p.on(\"pageerror\", e => errs.push(e.message));\nlet q = 0, compactPosts = 0, compactGets = 0;\nawait p.route(\"**/*\", async (route) => { const u = new URL(route.request().url()); const m = route.request().method();\n if (u.host !== \"app.local\") return route.abort();\n if (u.search.startsWith(\"?compact\") && m === \"POST\") { compactPosts++; return route.fulfill({ contentType: \"application/json\", body: JSON.stringify({ ok: true, before: 41000, after: 2900, covered: 22, uptoId: 88 }) }); }\n if (u.search.startsWith(\"?compact\")) { compactGets++; return route.fulfill({ contentType: \"application/json\", body: JSON.stringify({ compact: compactPosts ? { summary: \"GOAL — test\", covered: 22, uptoId: 88, chars: 2900, ts: 1 } : null }) }); }\n if (u.search.startsWith(\"?q\")) { q++; const body = JSON.parse(route.request().postData()); await new Promise(r => setTimeout(r, 200));\n if (q <= 3 && !body.settings.continuation) return route.fulfill({ status: 503, contentType: \"application/json\", body: JSON.stringify({ error: { message: \"router 503: ALL_PROVIDERS_EXHAUSTED_OR_COOLING\", details: [\"x\"] } }) });\n // then: truncated replies forever (until stopped), the 4th with needsCompact\n return route.fulfill({ contentType: \"application/json\", body: JSON.stringify({ reply: \"part \" + q + \" of a very long answer that keeps going\", truncated: true, toolEvents: [], meta: {}, needsCompact: q === 4 }) }); }\n if (u.search.includes(\"?progress\")) return route.fulfill({ contentType: \"application/json\", body: '{\"events\":[],\"next\":0}' });\n if (u.search && !u.search.startsWith(\"?s=\")) return route.fulfill({ contentType: \"application/json\", body: \"[]\" });\n return route.fulfill({ contentType: \"text/html\", body: html }); });\nawait p.goto(\"https://app.local/\", { waitUntil: \"load\" }); await p.waitForTimeout(400);\n// retryMinutes 0 would normally mean \"no retry\"; autoContinue 1 would normally stop after one continue. untilStopped overrides both.\nawait p.evaluate(() => { SETTINGS.set('retryMinutes', 0); SETTINGS.set('autoContinue', 1); SETTINGS.set('untilStopped', true); });\nawait p.fill(\"#inp\", \"write the whole thing\"); await p.keyboard.press(\"Enter\");\nawait p.waitForTimeout(34000); // 3 × 503 with jittered 8/16/24 s waits ≈ 48 s worst case… we sample mid-way\nconst mid = await p.evaluate(() => ({ typing: document.getElementById('typing-label').textContent, bar: document.getElementById('bar-status')?.textContent }));\nawait p.waitForTimeout(26000);\nconst before = await p.evaluate(() => ({ ai: document.querySelectorAll('.msg-wrap.ai').length, compactCards: [...document.querySelectorAll('.msg-wrap.sys .role-tag')].filter(x => x.textContent === 'COMPACT').length, inflight: window.inflightTurn() }));\nawait p.evaluate(() => stopGen()); await p.waitForTimeout(1500);\nconst after = await p.evaluate(() => ({ inflight: window.inflightTurn(), last: [...document.querySelectorAll('.msg-wrap.ai .msg-body')].pop()?.textContent.slice(-40) }));\nconsole.log(JSON.stringify({ mid, q, compactPosts, compactGets, before, after }, null, 1)); console.log(\"errors:\", errs);\nawait b.close();\n",[],[477],"52a25fb5-2fef-4828-afb0-41c27f18a655","e2e11.mjs","01a0a661-f049-7779-b407-b712e79e392e",["D",1789497831502],"import { chromium } from \"playwright-core\";\nimport fs from \"node:fs\";\nconst html = fs.readFileSync(\"app/page.html\", \"utf8\");\nconst manifest = { title: \"Demo\", shared: \"x\", files: [{ path: \"index.html\", purpose: \"entry\" }, { path: \"a.js\", purpose: \"a\" }, { path: \"b.js\", purpose: \"b\" }, { path: \"c.js\", purpose: \"c\" }] };\nconst b = await chromium.launch({ executablePath: \"/opt/pw-browsers/chromium\", args: [\"--no-sandbox\", \"--headless=new\"], ignoreDefaultArgs: [\"--headless=old\"] });\nconst p = await b.newPage({ viewport: { width: 1400, height: 800 } }); const errs = []; p.on(\"pageerror\", e => errs.push(e.message));\nconst built = [], jobPosts = [], planCalls = []; let qBodies = [];\nawait p.route(\"**/*\", async (route) => { const u = new URL(route.request().url());\n if (u.host === \"cdnjs.cloudflare.com\") { const m = u.pathname.match(/codemirror\\/5\\.65\\.16\\/(.*)$/); const f = m ? \"/tmp/cm/\" + m[1] : null; if (f && fs.existsSync(f)) return route.fulfill({ contentType: f.endsWith(\".css\") ? \"text/css\" : \"text/javascript\", body: fs.readFileSync(f, \"utf8\") }); return route.abort(); }\n if (u.host !== \"app.local\") return route.abort();\n if (u.search.startsWith(\"?sessions\")) return route.fulfill({ contentType: \"application/json\", body: JSON.stringify([{ id: \"S1\", name: \"demo\", ts: 1 }]) });\n if (u.search.startsWith(\"?history\")) return route.fulfill({ contentType: \"application/json\", body: \"[]\" });\n if (u.search.startsWith(\"?job=list\")) return route.fulfill({ contentType: \"application/json\", body: JSON.stringify({ jobs: [{ artifactId: 7, manifest, done: [\"index.html\", \"a.js\"], failed: [\"b.js\"], status: \"building\", title: \"Demo\", ask: \"make demo\" }] }) });\n if (u.search.startsWith(\"?job\")) { jobPosts.push(JSON.parse(route.request().postData())); return route.fulfill({ contentType: \"application/json\", body: \"{\\\"ok\\\":true}\" }); }\n if (u.search.startsWith(\"?plan\")) { planCalls.push(JSON.parse(route.request().postData()).q); return route.fulfill({ contentType: \"application/json\", body: JSON.stringify({ manifest, artifactId: 9, title: \"Demo\" }) }); }\n if (u.search.startsWith(\"?build_file\")) { const body = JSON.parse(route.request().postData()); built.push(body.path); await new Promise(r => setTimeout(r, 300)); return route.fulfill({ contentType: \"application/json\", body: JSON.stringify({ ok: true, path: body.path, lines: 10, issues: [], meta: { instance: \"openrouter#2\" } }) }); }\n if (u.search.includes(\"?artifact_files=\")) return route.fulfill({ contentType: \"application/json\", body: JSON.stringify({ id: 7, title: \"Demo\", kind: \"html\", session: \"S1\", files: [{ path: \"index.html\", content: \"

x

\" }], issues: [] }) });\n if (u.search.startsWith(\"?q\")) { qBodies.push(JSON.parse(route.request().postData())); return route.fulfill({ contentType: \"application/json\", body: JSON.stringify({ reply: \"integrated\", toolEvents: [], meta: {}, completion: { done: true, remaining: [] } }) }); }\n if (u.search.startsWith(\"?slow\")) { await new Promise(r => setTimeout(r, 20000)); return route.fulfill({ body: \"late\" }); }\n if (u.pathname.startsWith(\"/artifact/\")) return route.fulfill({ contentType: \"text/html\", body: \"

x

\" });\n if (u.search && !u.search.startsWith(\"?s=\")) return route.fulfill({ contentType: \"application/json\", body: \"[]\" });\n return route.fulfill({ contentType: \"text/html\", body: html }); });\nawait p.goto(\"https://app.local/?s=S1\", { waitUntil: \"load\" }); await p.waitForTimeout(800);\n// A) worker-backed wait + fetchT timeout\nconst a = await p.evaluate(async () => { const t0 = performance.now(); await omniNet.wait(300); const dt = performance.now() - t0; let err = null; try { await omniNet.fetchT('?slow', {}, 1200); } catch (e) { err = e.message; } let stopped = null; const c = new AbortController(); setTimeout(() => c.abort(), 200); try { await omniNet.fetchT('?slow', { signal: c.signal }, 5000); } catch (e) { stopped = e.name; } return { dt: Math.round(dt), err, stopped }; });\n// B) resume card rendered from the server job row\nconst cardTxt = await p.evaluate(() => { const c = document.querySelector('.manifest'); return c ? { head: c.querySelector('.manifest-h span').textContent, btn: c.querySelector('[data-a=go]').textContent, st: [...c.querySelectorAll('.manifest-f')].map(r => r.querySelector('code').textContent + '=' + r.querySelector('.st').textContent) } : null; });\nawait p.click('.manifest [data-a=go]'); await p.waitForTimeout(2500);\nconst after = await p.evaluate(() => ({ st: [...document.querySelectorAll('.manifest-f')].map(r => r.querySelector('code').textContent + '=' + r.querySelector('.st').textContent), reply: [...document.querySelectorAll('.msg-wrap.ai .msg-body')].pop()?.textContent, title: document.title }));\n// C) app-scale ask auto-routes to the swarm\nawait p.evaluate(() => ART.close()); await p.waitForTimeout(300);\nawait p.fill(\"#inp\", \"build a step sequencer app with effects\"); await p.keyboard.press(\"Enter\"); await p.waitForTimeout(2500);\nawait p.evaluate(() => ART.close()); await p.waitForTimeout(300); await p.fill(\"#inp\", \"what time is it\"); await p.keyboard.press(\"Enter\"); await p.waitForTimeout(800);\nconst c = { planCalls, qAfter: qBodies.length, builtTotal: built.length };\nconsole.log(JSON.stringify({ c, a, cardTxt, built, after, jobStatuses: jobPosts.map(j => j.state.status + ':' + j.state.done.length), integrateQ: qBodies.length, integrateFocus: qBodies[0]?.settings?.focus?.id }, null, 1));\nconsole.log(\"errors:\", errs);\nawait b.close();\n",[],[484],"38d25b74-8f52-416a-967c-a3fbcfc25455","e2e10.mjs","01a0a661-f049-7779-b407-b1fd553e92ff",["D",1789497831502],"// lint strip: fix button sends a focused fix; auto-fix fires once; completion audit → automatic follow-up turn\nimport { chromium } from \"playwright-core\";\nimport fs from \"node:fs\";\nconst html = fs.readFileSync(\"app/page.html\", \"utf8\");\nconst files = { \"index.html\": \"\", \"app.js\": \"function init() {}\\n\" };\nconst b = await chromium.launch({ executablePath: \"/opt/pw-browsers/chromium\", args: [\"--no-sandbox\", \"--headless=new\"], ignoreDefaultArgs: [\"--headless=old\"] });\nconst p = await b.newPage({ viewport: { width: 1400, height: 800 } }); const errs = []; p.on(\"pageerror\", e => errs.push(e.message));\nconst posts = [];\nawait p.route(\"**/*\", async (route) => { const u = new URL(route.request().url());\n if (u.host !== \"app.local\") return route.abort();\n if (u.search.includes(\"?artifact_files=\")) return route.fulfill({ contentType: \"application/json\", body: JSON.stringify({ id: 7, title: \"T\", kind: \"html\", session: \"s\", files: Object.entries(files).map(([path, content]) => ({ path, content })), issues: [\"app.js:1: empty function body — \\\"function init() {}\\\" — replace with the real implementation\"] }) });\n if (u.search.startsWith(\"?q\")) { const body = JSON.parse(route.request().postData()); posts.push({ q: body.q.slice(0, 60), focus: body.settings.focus?.id, followup: body.settings.followup }); await new Promise(r => setTimeout(r, 300));\n const first = posts.length === 1 && !/Continue the task/.test(body.q);\n // the SUCCESSOR turn (auto follow-up) hits a 503 once: its retry loop must use its own AbortController (regression: \"Cannot read properties of null (reading 'signal')\")\n if (posts.length === 2 && !body.settings.retryOf) return route.fulfill({ status: 503, contentType: \"application/json\", body: JSON.stringify({ error: { message: \"router 503: ALL_PROVIDERS_EXHAUSTED_OR_COOLING\", details: [\"x\"] } }) });\n return route.fulfill({ contentType: \"application/json\", body: JSON.stringify({ reply: \"did part \" + posts.length, toolEvents: [{ hop: 0, tool: \"update_artifact\", args: { id: 7 }, result: { ok: true, id: 7, url: \"/artifact/7/\", title: \"T\" } }], meta: {}, completion: posts.length < 3 ? { done: false, remaining: [\"wire the keyboard handler in input.js\", \"add the score display\"] } : { done: true, remaining: [] } }) }); }\n if (u.search.includes(\"?progress\")) return route.fulfill({ contentType: \"application/json\", body: '{\"events\":[],\"next\":0}' });\n if (u.pathname.startsWith(\"/artifact/\")) return route.fulfill({ contentType: \"text/html\", body: files[\"index.html\"] });\n if (u.search) return route.fulfill({ contentType: \"application/json\", body: \"[]\" });\n return route.fulfill({ contentType: \"text/html\", body: html }); });\nawait p.goto(\"https://app.local/\", { waitUntil: \"load\" }); await p.waitForTimeout(300);\nawait p.evaluate(() => { ART.setSettings({ ...ART.settings(), codemirror: false, autoFixLint: true }); ART.open(7, 'split'); }); await p.waitForTimeout(1200);\nconst r1 = await p.evaluate(() => ({ lintShown: !document.getElementById('ap-lint').hidden, n: document.getElementById('ap-lint-n').textContent, btn: !!document.querySelector('#ap-lint-h button') }));\nawait p.waitForTimeout(16000); // auto-fix turn + completion follow-ups\nconst r2 = await p.evaluate(() => ({ auto: [...document.querySelectorAll('.msg-wrap.sys .msg-body')].map(x => x.textContent.split('\\n')[0]), ai: document.querySelectorAll('.msg-wrap.ai').length }));\nconsole.log(JSON.stringify({ r1, posts, r2 }, null, 1)); console.log(\"errors:\", errs);\nawait b.close();\n",[],[491],"94ab5c50-7f76-4eac-8956-2bbb8a1f1f1b","db.ts","01a0a661-f049-7779-b407-ac8de55a7431",["D",1789497831502],"import { sqlite } from \"https://esm.town/v/std/sqlite/main.ts\";\n\n/* ────────────────────────────────────────────────────────────────\n Result monad — DB failures are values, never uncaught exceptions.\n Error channel is a closed literal union, not bare `string`, so\n every `r.error === \"...\"` comparison downstream is checked at\n compile time instead of merely hoped for.\n ──────────────────────────────────────────────────────────────── */\ntype Result = { ok: true; value: T } | { ok: false; error: E };\ntype DbError = \"QUOTA_EXCEEDED\" | \"UNKNOWN\";\n\nfunction Ok(value: T): Result {\n return { ok: true, value };\n}\n\nfunction Err(error: E): Result {\n return { ok: false, error };\n}\n\nasync function attempt(fn: () => Promise): Promise> {\n try {\n return Ok(await fn());\n } catch (e) {\n return Err(\n /full|quota/i.test(String(e)) ? \"QUOTA_EXCEEDED\" : \"UNKNOWN\",\n );\n }\n}\n\n/* ────────────────────────────────────────────────────────────────\n Schema — single source of truth. Row shapes come from an actual\n TS interface, not from pattern-matching SQL text.\n ──────────────────────────────────────────────────────────────── */\ninterface Schema {\n provider: {\n id: string; // natural key — matches message.providerId's type; no cast needed at the join\n name: string;\n base: string;\n keyEnv: string | null;\n model: string | null;\n priority: number;\n };\n message: {\n id: number;\n role: \"user\" | \"assistant\" | \"system\";\n content: string;\n ts: number;\n session: string;\n providerId: string | null;\n };\n message_meta: { message_id: number; meta: string };\n memory: { key: string; value: string; weight: number; ts: number };\n memory_fact: {\n id: number;\n fact: string;\n salience: number; // 1 - predictability; decays on non-access\n created: number;\n accessed: number;\n hits: number;\n };\n session: { id: string; name: string; ts: number };\n omni_router_config: { provider: string; selected_model: string };\n omni_state: { key: string; value: string; ts: number }; // router cooldowns / dead models / resolved ids\n progress: { session: string; events: string; ts: number }; // live turn progress, polled by the chat UI\n artifact_file: { artifact_id: number; path: string; content: string; ts: number }; // multi-file artifacts; index.html mirrors artifact.content\n // Durable counterpart to router-core.ts's in-memory breaker: that Map resets on every fresh isolate,\n // which is why /health's instance roster could show 0 requests moments after a real, failed attempt —\n // the isolate that served the attempt and the one that served the /health check are not guaranteed to\n // be the same one. This table survives across isolates; the in-memory breaker still governs open/closed\n // decisions (no added DB read on the hot path), this just makes the numbers you can actually SEE correct.\n provider_stats: { name: string; requests: number; ok: number; failed: number; last_error: string | null; updated_at: number };\n artifact: {\n id: number;\n session: string;\n title: string;\n kind: \"html\" | \"svg\" | \"markdown\";\n content: string;\n ts: number;\n };\n // Durable record of a swarm build's plan and overall status. Previously this lived only in an\n // omni_state key/value blob (job::) that was DELETED the moment a build finished\n // — so once a build was done there was no durable answer to \"what was built, from what plan, when\",\n // and while a build was running the state was an opaque JSON string, not something a query could join\n // against build_file below. One row per artifact now, kept forever (status transitions in place; never\n // deleted on \"done\") so it doubles as the artifact's permanent development record.\n build_job: {\n artifact_id: number; // PK — one build per artifact\n session: string;\n title: string;\n ask: string; // the original request that produced this manifest\n manifest: string; // JSON: {title, files:[{path,purpose,exports,imports,notes}], features, visual}\n status: \"planned\" | \"building\" | \"integrating\" | \"done\" | \"stopped\";\n created_ts: number;\n updated_ts: number;\n };\n // Per-file build status for a build_job — the piece that was previously missing entirely: build_file\n // (below, note the different name — this old singular-content table is `artifact_file`) stores a file's\n // CONTENT, but nothing durable recorded whether a given path had ever been built, what its last build's\n // issues/errors were, or how many attempts it took. buildOne() in main-script.ts upserts this row on\n // every attempt (queued → building → ok/issues/error), so /resume, the manifest UI, and any later\n // inspection all read the same source of truth instead of the manifest UI's own client-side memory.\n build_file: {\n artifact_id: number;\n path: string;\n purpose: string;\n status: \"queued\" | \"building\" | \"ok\" | \"issues\" | \"error\" | \"stopped\";\n lines: number | null;\n issues: string; // JSON array of lint-issue strings from the last build attempt\n error: string | null;\n attempts: number;\n updated_ts: number;\n };\n // Durable bug/feature backlog. Previously nothing tracked reported bugs or requested features as\n // structured data at all — they existed only as prose inside chat history, which is neither queryable\n // (\"what's still open for artifact #12\") nor reliable across a long or compacted conversation (an old\n // report can silently scroll out of the context window the model reads). The log_issue/list_issues/\n // update_issue tools in issues.ts read and write this table directly so work can be tracked and resumed\n // reliably instead of depending on the model remembering what was said several turns ago.\n issue: {\n id: number;\n session: string;\n artifact_id: number | null; // null = general/session-level, not tied to one artifact\n kind: \"bug\" | \"feature\";\n title: string;\n description: string;\n status: \"open\" | \"in_progress\" | \"fixed\" | \"wontfix\" | \"duplicate\";\n priority: \"low\" | \"medium\" | \"high\" | \"critical\";\n source: \"user\" | \"audit\" | \"model\";\n created_ts: number;\n updated_ts: number;\n resolved_ts: number | null;\n notes: string; // JSON array of {ts, text} — a short audit trail of what happened to this item\n };\n}\ntype Table = keyof Schema;\ntype Row = Schema[T];\n\nconst DDL: Record = {\n provider: `CREATE TABLE IF NOT EXISTS provider (\n id TEXT PRIMARY KEY, name TEXT NOT NULL, base TEXT NOT NULL, model TEXT NOT NULL,\n keyEnv TEXT, priority INTEGER NOT NULL)`,\n message: `CREATE TABLE IF NOT EXISTS message (\n id INTEGER PRIMARY KEY AUTOINCREMENT, role TEXT NOT NULL, content TEXT NOT NULL,\n ts INTEGER NOT NULL, session TEXT NOT NULL REFERENCES session(id),\n providerId TEXT REFERENCES provider(id))`,\n message_meta: `CREATE TABLE IF NOT EXISTS message_meta (\n message_id INTEGER PRIMARY KEY REFERENCES message(id) ON DELETE CASCADE, meta TEXT NOT NULL)`,\n memory: `CREATE TABLE IF NOT EXISTS memory (\n key TEXT PRIMARY KEY, value TEXT NOT NULL, weight REAL NOT NULL, ts INTEGER NOT NULL)`,\n memory_fact: `CREATE TABLE IF NOT EXISTS memory_fact (\n id INTEGER PRIMARY KEY AUTOINCREMENT, fact TEXT NOT NULL, salience REAL NOT NULL,\n created INTEGER NOT NULL, accessed INTEGER NOT NULL, hits INTEGER NOT NULL DEFAULT 0)`,\n session: `CREATE TABLE IF NOT EXISTS session (\n id TEXT PRIMARY KEY, name TEXT NOT NULL DEFAULT 'New Chat', ts INTEGER NOT NULL)`,\n omni_router_config: `CREATE TABLE IF NOT EXISTS omni_router_config (\n provider TEXT PRIMARY KEY, selected_model TEXT NOT NULL)`,\n omni_state: `CREATE TABLE IF NOT EXISTS omni_state (\n key TEXT PRIMARY KEY, value TEXT NOT NULL, ts INTEGER NOT NULL)`,\n progress: `CREATE TABLE IF NOT EXISTS progress (\n session TEXT PRIMARY KEY, events TEXT NOT NULL, ts INTEGER NOT NULL)`,\n artifact_file: `CREATE TABLE IF NOT EXISTS artifact_file (\n artifact_id INTEGER NOT NULL REFERENCES artifact(id), path TEXT NOT NULL, content TEXT NOT NULL, ts INTEGER NOT NULL,\n PRIMARY KEY(artifact_id, path))`,\n provider_stats: `CREATE TABLE IF NOT EXISTS provider_stats (\n name TEXT PRIMARY KEY, requests INTEGER NOT NULL DEFAULT 0, ok INTEGER NOT NULL DEFAULT 0,\n failed INTEGER NOT NULL DEFAULT 0, last_error TEXT, updated_at INTEGER NOT NULL)`,\n // No UNIQUE(session, title): the real identity of an artifact is its id. A UNIQUE(session, title)\n // constraint here made `INSERT ... ON CONFLICT(session, title) DO UPDATE` silently overwrite any\n // earlier, unrelated artifact that happened to share a title within the session (models reach for\n // generic titles — \"Todo App\", \"Dashboard\" — routinely) — destructive, no versioning, no warning.\n // Identity is now always by id; (session, title) is only a non-unique lookup index (see idx_artifact_session_title).\n artifact: `CREATE TABLE IF NOT EXISTS artifact (\n id INTEGER PRIMARY KEY AUTOINCREMENT, session TEXT NOT NULL REFERENCES session(id),\n title TEXT NOT NULL, kind TEXT NOT NULL, content TEXT NOT NULL, ts INTEGER NOT NULL)`,\n // ON DELETE CASCADE on both FKs: without it, deleting an `artifact` row (there is no delete-artifact path\n // today, but nothing prevents one being added later) would either be blocked by the live FK check once one\n // of these child rows exists, or — for build_file, which references build_job not artifact directly — leave\n // orphaned build_job/build_file rows with no artifact behind them, silently corrupting future `?job` list\n // queries and the join in handleIssues/list_issues. CASCADE keeps a build's row lifecycle tied to its artifact's.\n build_job: `CREATE TABLE IF NOT EXISTS build_job (\n artifact_id INTEGER PRIMARY KEY REFERENCES artifact(id) ON DELETE CASCADE, session TEXT NOT NULL, title TEXT NOT NULL DEFAULT '',\n ask TEXT NOT NULL DEFAULT '', manifest TEXT NOT NULL, status TEXT NOT NULL DEFAULT 'planned',\n created_ts INTEGER NOT NULL, updated_ts INTEGER NOT NULL)`,\n build_file: `CREATE TABLE IF NOT EXISTS build_file (\n artifact_id INTEGER NOT NULL REFERENCES build_job(artifact_id) ON DELETE CASCADE, path TEXT NOT NULL, purpose TEXT NOT NULL DEFAULT '',\n status TEXT NOT NULL DEFAULT 'queued', lines INTEGER, issues TEXT NOT NULL DEFAULT '[]', error TEXT,\n attempts INTEGER NOT NULL DEFAULT 0, updated_ts INTEGER NOT NULL, PRIMARY KEY(artifact_id, path))`,\n // issue.artifact_id is nullable by design (a general/session-level item has no artifact) — ON DELETE SET\n // NULL rather than CASCADE, so deleting the artifact a bug was filed against demotes the issue to\n // session-level instead of destroying the backlog entry (and its notes/audit trail) outright.\n issue: `CREATE TABLE IF NOT EXISTS issue (\n id INTEGER PRIMARY KEY AUTOINCREMENT, session TEXT NOT NULL, artifact_id INTEGER REFERENCES artifact(id) ON DELETE SET NULL,\n kind TEXT NOT NULL, title TEXT NOT NULL, description TEXT NOT NULL DEFAULT '', status TEXT NOT NULL DEFAULT 'open',\n priority TEXT NOT NULL DEFAULT 'medium', source TEXT NOT NULL DEFAULT 'model',\n created_ts INTEGER NOT NULL, updated_ts INTEGER NOT NULL, resolved_ts INTEGER, notes TEXT NOT NULL DEFAULT '[]')`,\n};\n\nconst INDEXES = [\n `CREATE INDEX IF NOT EXISTS idx_message_session_id ON message(session, id)`,\n `CREATE INDEX IF NOT EXISTS idx_memory_ts ON memory(ts)`,\n `CREATE INDEX IF NOT EXISTS idx_memory_fact_accessed ON memory_fact(accessed)`,\n `CREATE INDEX IF NOT EXISTS idx_artifact_session_ts ON artifact(session, ts)`,\n `CREATE INDEX IF NOT EXISTS idx_artifact_session_title ON artifact(session, title)`,\n `CREATE INDEX IF NOT EXISTS idx_build_job_session_status ON build_job(session, status)`,\n `CREATE INDEX IF NOT EXISTS idx_build_file_artifact ON build_file(artifact_id)`,\n `CREATE INDEX IF NOT EXISTS idx_issue_session_status ON issue(session, status)`,\n `CREATE INDEX IF NOT EXISTS idx_issue_artifact ON issue(artifact_id)`,\n];\n\n// One-time migration for an `artifact` table created before this fix, under the old\n// `UNIQUE(session, title)` constraint: `CREATE TABLE IF NOT EXISTS` above never touches an\n// already-existing table, so the constraint must be dropped explicitly (SQLite has no\n// ALTER TABLE ... DROP CONSTRAINT) by rebuilding the table without it. Idempotent: checks the\n// live schema text first and is a no-op once migrated (or on a fresh table that never had it).\n//\n// First cut of this migration failed live with \"FOREIGN KEY constraint failed\" on the final DROP:\n// `ALTER TABLE artifact RENAME TO ...` doesn't just rename — SQLite also silently rewrites every\n// OTHER table's stored `REFERENCES artifact(id)` clause to point at the new name, so `artifact_file`\n// ends up referencing `artifact_pre_unique_migration` mid-migration. Dropping that table while a live\n// FK still points at it is correctly refused. Fix: rebuild `artifact_file` too, in the same pass, so\n// nothing references the old name by the time it's dropped — and verify row counts before either\n// DROP, so a partial/mismatched copy aborts loudly instead of silently discarding rows.\nasync function migrateArtifactUniqueConstraint(): Promise {\n try {\n const r = await sqlite.execute({ sql: `SELECT sql FROM sqlite_master WHERE type='table' AND name='artifact'`, args: [] });\n const ddl = (r.rows?.[0] as any)?.sql as string | undefined;\n if (!ddl || !/UNIQUE\\s*\$\\s*session\\s*,\\s*title\\s*\$/i.test(ddl)) return; // not yet created, or already migrated\n\n const count = async (table: string) => {\n const c = await sqlite.execute({ sql: `SELECT COUNT(*) AS n FROM ${table}`, args: [] });\n return Number((c.rows?.[0] as any)?.n ?? -1);\n };\n const artifactBefore = await count(\"artifact\");\n const filesBefore = await count(\"artifact_file\");\n\n await sqlite.batch([\n `ALTER TABLE artifact RENAME TO artifact_pre_unique_migration`,\n `CREATE TABLE artifact (\n id INTEGER PRIMARY KEY AUTOINCREMENT, session TEXT NOT NULL REFERENCES session(id),\n title TEXT NOT NULL, kind TEXT NOT NULL, content TEXT NOT NULL, ts INTEGER NOT NULL)`,\n `INSERT INTO artifact (id, session, title, kind, content, ts)\n SELECT id, session, title, kind, content, ts FROM artifact_pre_unique_migration`,\n // artifact_file's REFERENCES clause was just silently rewritten (by the RENAME above) to point at\n // artifact_pre_unique_migration — rebuild it too, fresh, against the new `artifact` table, before\n // anything tries to drop the old one.\n `ALTER TABLE artifact_file RENAME TO artifact_file_pre_migration`,\n `CREATE TABLE artifact_file (\n artifact_id INTEGER NOT NULL REFERENCES artifact(id), path TEXT NOT NULL, content TEXT NOT NULL, ts INTEGER NOT NULL,\n PRIMARY KEY(artifact_id, path))`,\n `INSERT INTO artifact_file (artifact_id, path, content, ts)\n SELECT artifact_id, path, content, ts FROM artifact_file_pre_migration`,\n ]);\n\n const artifactAfter = await count(\"artifact\");\n const filesAfter = await count(\"artifact_file\");\n if (artifactAfter !== artifactBefore || filesAfter !== filesBefore) {\n // Leave everything as-is (both old and new tables present) rather than drop anything on a\n // mismatched copy — a human needs to look at this, not have rows silently vanish.\n console.error(`migrateArtifactUniqueConstraint: row-count mismatch after copy (artifact ${artifactBefore}->${artifactAfter}, artifact_file ${filesBefore}->${filesAfter}) — old tables left in place, NOT dropped, migration will retry next startup`);\n return;\n }\n\n // Nothing references the old table names anymore (artifact_file was rebuilt above) — safe to drop.\n await sqlite.batch([\n `DROP TABLE artifact_file_pre_migration`,\n `DROP TABLE artifact_pre_unique_migration`,\n ]);\n } catch (e) {\n console.error(\"migrateArtifactUniqueConstraint failed (artifact table left as-is):\", e);\n }\n}\n\nconst initDB = () =>\n attempt(async () => {\n await sqlite.batch([`PRAGMA foreign_keys = ON`, ...Object.values(DDL)]);\n await migrateArtifactUniqueConstraint(); // must run after CREATE TABLE IF NOT EXISTS (needs the table to exist) and before the new index below\n await sqlite.batch(INDEXES);\n });\n\n/* ────────────────────────────────────────────────────────────────\n Query core — one primitive, three call shapes.\n\n `sql` is a tagged template: values are never concatenated into\n the statement text, so injection is structurally impossible, not\n just discouraged by convention.\n\n The `Table` argument to `all`/`one` is a phantom: zero runtime\n cost, exists purely to pin the generic so the return type is\n declared honestly by the caller instead of inferred from SQL\n text (which is what broke in the original — template-literal\n type matching only resolves on string-literal SQL and silently\n collapses to `unknown` for anything built at runtime).\n ──────────────────────────────────────────────────────────────── */\n// The exact set of values libsql/SQLite can bind as a parameter (mirrors\n// @libsql/client's InValue). `unknown[]` doesn't satisfy either `execute`\n// overload — TS falls through to the `(sql: string, args?: InArgs)` form\n// and reports the whole `{sql, args}` object as an invalid `sql` argument,\n// which is a confusing symptom of an unrelated cause. Declaring the real\n// value type here fixes it at the source instead of casting at each call.\ntype SqlValue =\n | string\n | number\n | bigint\n | boolean\n | ArrayBuffer\n | Uint8Array\n | Date\n | null;\ntype Bound = { text: string; values: SqlValue[] };\nconst sql = (strings: TemplateStringsArray, ...values: SqlValue[]): Bound => ({\n text: strings.reduce(\n (acc, s, i) => acc + s + (i < values.length ? \"?\" : \"\"),\n \"\",\n ),\n values,\n});\n\n// sqlite.execute returns libsql's ResultSet, whose .rows are libsql's own\n// Row type — every column typed as `Value = null | string | number | bigint\n// | ArrayBuffer`, regardless of what our schema declares. TS correctly\n// refuses a direct cast to Row here: nothing in the type system knows a\n// TEXT NOT NULL column can't come back null, or that an INTEGER won't come\n// back as bigint. That guarantee lives in the SQL schema, not in types, so\n// this is an explicit, acknowledged boundary crossing — not a verified one.\nfunction all(_table: T, q: Bound) {\n return attempt(\n async () =>\n (await sqlite.execute({ sql: q.text, args: q.values }))\n .rows as unknown as Row[],\n );\n}\n\nasync function one(table: T, q: Bound) {\n const r = await all(table, q);\n return r.ok ? Ok(r.value[0] ?? null) : r;\n}\n\n// Escape hatch for shapes that don't correspond to any single table — an\n// aggregate like `SELECT COUNT(*) AS c`, a join projection, etc. `all`/`one`\n// can't type these honestly because their phantom `Table` argument has\n// nothing to pin to. Here the caller declares the shape directly via the\n// explicit generic — same honesty contract as `all`, just without a table\n// to anchor it to.\nfunction raw(q: Bound) {\n return attempt(async () =>\n (await sqlite.execute({ sql: q.text, args: q.values }))\n .rows as unknown as T[]\n );\n}\n\n// unwrap(r) → value | null; unwrap(r, fallback) → value | fallback. The old signature\n// dropped the fallback, so every `unwrap(x, [])` caller got null on a DB error and\n// then crashed on `.map` — a latent bug the types were hiding.\nfunction unwrap(r: Result): T | null;\nfunction unwrap(r: Result, fallback: T): T;\nfunction unwrap(r: Result, fallback?: T): T | null {\n return r.ok ? r.value : (fallback === undefined ? null : fallback);\n}\n\n// libsql's ResultSet carries rowsAffected and lastInsertRowid — the only\n// way to learn the AUTOINCREMENT id of a row just inserted, or to confirm\n// a write actually touched anything. Discarding it (`.then(() => undefined)`)\n// made insertMessage() write rows whose id was permanently unrecoverable.\ntype Effect = { rowsAffected: number; lastInsertRowid: bigint | undefined };\nconst run = (q: Bound) =>\n attempt(async () => {\n const r = await sqlite.execute({ sql: q.text, args: q.values });\n return {\n rowsAffected: r.rowsAffected,\n lastInsertRowid: r.lastInsertRowid,\n } satisfies Effect;\n });\n\n/* ────────────────────────────────────────────────────────────────\n Domain operations — thin pipelines over the three primitives.\n Retention is not an afterthought: every write can fail on quota,\n and the failure is handled as data, not caught upstream by luck.\n ──────────────────────────────────────────────────────────────── */\nconst messagesFor = (session: string, limit = 200) =>\n all(\n \"message\",\n sql`SELECT * FROM message WHERE session = ${session} ORDER BY id DESC LIMIT ${limit}`,\n );\n\nconst insertMessage = async (m: Omit, \"id\">) => {\n const r = await run(\n sql`INSERT INTO message (role, content, ts, session, providerId)\n VALUES (${m.role}, ${m.content}, ${m.ts}, ${m.session}, ${m.providerId})`,\n );\n if (!r.ok) return r;\n // lastInsertRowid is undefined only if this wasn't actually an INSERT into\n // a ROWID table — Number(undefined) would silently produce NaN here, which\n // is worse than failing loudly, since a NaN id would propagate downstream\n // looking like a valid one.\n if (r.value.lastInsertRowid === undefined) return Err(\"UNKNOWN\");\n return Ok(Number(r.value.lastInsertRowid));\n};\n\nconst pruneSession = (session: string, keep = 200) =>\n run(sql`DELETE FROM message WHERE session = ${session} AND id NOT IN\n (SELECT id FROM message WHERE session = ${session} ORDER BY id DESC LIMIT ${keep})`);\n\n// Val Town's real ceiling is the storage quota (10MB free / 1GB paid,\n// Turso-backed), not SQLite's own ~2^31-byte string limit. On\n// QUOTA_EXCEEDED, prune this session hard and retry once. If the\n// retry still fails, the caller gets the real error, not a silent drop.\nconst insertMessageSafe = async (m: Omit, \"id\">) => {\n const first = await insertMessage(m);\n\n if (first.ok === false) {\n if (first.error === \"QUOTA_EXCEEDED\") {\n const pruned = await pruneSession(m.session, 100);\n return pruned.ok ? insertMessage(m) : pruned;\n }\n }\n return first;\n};\n\nconst upsertMemory = (key: string, value: string, weight: number, ts: number) =>\n run(\n sql`INSERT INTO memory (key, value, weight, ts) VALUES (${key}, ${value}, ${weight}, ${ts})\n ON CONFLICT(key) DO UPDATE SET value = excluded.value, weight = excluded.weight, ts = excluded.ts`,\n );\n\nconst providerById = (id: string) =>\n one(\"provider\", sql`SELECT * FROM provider WHERE id = ${id}`);\n\nconst providersByPriority = () =>\n all(\"provider\", sql`SELECT * FROM provider ORDER BY priority ASC`);\n\nconst touchSession = (id: string, name: string, ts: number) =>\n run(sql`INSERT INTO session (id, name, ts) VALUES (${id}, ${name}, ${ts})\n ON CONFLICT(id) DO UPDATE SET name = excluded.name, ts = excluded.ts`);\n\n// One row per provider instance name; requests/ok/failed are running totals across every isolate that\n// has ever reported for that name, not a per-call snapshot — the UPDATE adds this call's outcome to\n// whatever was already there rather than replacing it.\nconst bumpProviderStats = (name: string, ok: boolean, error: string | null, ts: number) =>\n run(sql`INSERT INTO provider_stats (name, requests, ok, failed, last_error, updated_at)\n VALUES (${name}, 1, ${ok ? 1 : 0}, ${ok ? 0 : 1}, ${error}, ${ts})\n ON CONFLICT(name) DO UPDATE SET\n requests = requests + 1,\n ok = ok + ${ok ? 1 : 0},\n failed = failed + ${ok ? 0 : 1},\n last_error = COALESCE(${error}, provider_stats.last_error),\n updated_at = ${ts}`);\n\nconst providerStatsAll = () => all(\"provider_stats\", sql`SELECT * FROM provider_stats`);\n\nexport {\n all,\n type Bound,\n bumpProviderStats,\n type DbError,\n Err,\n initDB,\n insertMessage,\n insertMessageSafe,\n messagesFor,\n Ok,\n one,\n providerById,\n providersByPriority,\n providerStatsAll,\n pruneSession,\n raw,\n type Result,\n type Row,\n run,\n type Schema,\n sql,\n type Table,\n touchSession,\n unwrap,\n upsertMemory,\n};",[],[498],"19b30baf-721f-4605-9c0f-8a6645690b54","catalog.ts","01a0a661-f049-7779-b407-a8e0e7a3e0df",["D",1789497831502],"// Vendor catalog. One row per OpenAI-compatible endpoint. `models` is the\n// ordered fallback list; `prefer` picks a replacement from the vendor's live\n// /models when every listed id 404s. `${ENV}` in `base` is interpolated.\n// Override/extend at runtime with OMNI_PROVIDERS (JSON array, merged by vendor)\n// and OMNI_DISABLE (comma list of vendors). Priority: lower = tried first;\n// 0–19 free & fast, 20–39 free & slow/quota'd, 40+ paid last resort.\nexport type Vendor = {\n vendor: string;\n base: string;\n keyEnv: string | null;\n models: string[];\n prefer?: string; // regex source, case-insensitive\n priority: number;\n tags?: string[]; // coder | reasoning | fast | long | vision\n headers?: Record;\n paid?: boolean; // paid vendors rank after every free one in `auto`, whatever their quality\n tps?: number; // measured/estimated output tokens per second (default 60)\n ttft?: number; // typical ms to first token (default 1500)\n maxOut?: number; // clamp for max_tokens: free tiers meter prompt+max_tokens per minute (Groq TPM 8000)\n maxIn?: number; // largest prompt (tokens) the free tier accepts per request; prompts above it skip the vendor without a wasted call\n};\n\nimport { FREE } from \"./catalog-free.ts\";\nimport { PAID } from \"./catalog-paid.ts\";\nexport const CATALOG: Vendor[] = [...FREE, ...PAID];\n\n// Quality prior per upstream model id (0–100). Drives `auto`/`best` ordering:\n// higher first across vendors, then vendor priority as tie-break. Editorial,\n// dated 2026-09; override any row via OMNI_QUALITY='{\"regex\":score,…}'.\nconst QUALITY: [RegExp, number][] = [\n [/deepseek-v4-pro/i, 95], [/kimi-k3/i, 94], [/qwen3\\.8-2\\.4t/i, 93], [/glm-5\\.3(?!-flash)/i, 92], [/nemotron-3-ultra/i, 90],\n [/deepseek-v4-flash/i, 88], [/minimax-m3/i, 86], [/qwen3-coder-next|qwen3-coder-plus/i, 86], [/qwen[-.]?3\\.8-27b/i, 85],\n [/glm-5\\.3-flash|glm-5-turbo/i, 84], [/nemotron-3-super/i, 84], [/deepseek-v3\\.2/i, 84], [/kimi-k2/i, 83], [/gpt-oss-120b/i, 82],\n [/gemini-2\\.5-flash(?!-lite)/i, 80], [/gemma-4-31b/i, 80], [/qwen3\\.6|qwen3\\.7/i, 78], [/nemotron-3\\.5-lightning/i, 76],\n [/gemini-2\\.5-flash-lite/i, 74], [/glm-4\\.5-air|glm-4\\.7/i, 72], [/mistral-small/i, 72], [/command-a/i, 72], [/qwen3-30b|qwen3-32b/i, 72],\n [/llama-3\\.3-70b|llama-3_3-70b/i, 70], [/gpt-oss-20b/i, 70], [/nemotron-70b|nemotron-51b/i, 68], [/qwen3-14b/i, 65], [/ministral-14b/i, 62],\n [/qwen3-8b/i, 55], [/ministral-8b|llama-3\\.1-8b/i, 50], [/gpt-5-mini/i, 88], [/claude-sonnet/i, 92], [/claude-haiku/i, 84], [/deepseek-chat|deepseek-reasoner/i, 90], [/grok-4/i, 88],\n];\nlet override: [RegExp, number][] | null = null;\nexport function quality(id: string): number {\n if (override === null) {\n override = [];\n try { for (const [k, v] of Object.entries(JSON.parse(Deno.env.get(\"OMNI_QUALITY\") ?? \"{}\"))) override.push([new RegExp(k, \"i\"), Number(v)]); } catch { /* ignore */ }\n }\n for (const [rx, q] of override) if (rx.test(id)) return q;\n for (const [rx, q] of QUALITY) if (rx.test(id)) return q;\n return 40;\n}\n",[],[505],"36e6f5b1-b9a5-4d8c-b139-239b7731ac24","catalog-paid.ts","01a0a661-f049-7779-b407-a4214b986595",["D",1789497831502],"import type { Vendor } from \"./catalog.ts\";\n// Paid — last resort, only routed when a key is present.\nexport const PAID: Vendor[] = [\n { paid: true, vendor: \"deepseek\", base: \"https://api.deepseek.com/v1\", keyEnv: \"DEEPSEEK_API_KEY\", priority: 40, tags: [\"coder\", \"reasoning\"], models: [\"deepseek-chat\", \"deepseek-reasoner\"] },\n { paid: true, vendor: \"moonshot\", base: \"https://api.moonshot.ai/v1\", keyEnv: \"MOONSHOT_API_KEY\", priority: 41, tags: [\"coder\", \"long\"], models: [\"kimi-k2-turbo-preview\", \"kimi-k2-0905-preview\"], prefer: \"kimi\" },\n { paid: true, vendor: \"xai\", base: \"https://api.x.ai/v1\", keyEnv: \"XAI_API_KEY\", priority: 42, tags: [\"reasoning\"], models: [\"grok-4-fast-non-reasoning\", \"grok-4-fast-reasoning\"], prefer: \"grok.*fast\" },\n { paid: true, vendor: \"openai\", base: \"https://api.openai.com/v1\", keyEnv: \"OPENAI_API_KEY\", priority: 43, tags: [\"coder\", \"reasoning\", \"vision\"], models: [\"gpt-5-mini\", \"gpt-4.1-mini\"], prefer: \"gpt-5-mini|gpt-4.1-mini\" },\n // Credit-based (small free grants, then metered): ranked with the paid tier so they are last resorts, not daily drivers.\n { paid: true, vendor: \"deepinfra\", tps: 60, base: \"https://api.deepinfra.com/v1/openai\", keyEnv: \"DEEPINFRA_API_KEY\", priority: 41, tags: [\"coder\", \"reasoning\", \"long\"], models: [\"deepseek-ai/DeepSeek-V3.1\", \"Qwen/Qwen3-Coder-480B-A35B-Instruct\", \"meta-llama/Llama-3.3-70B-Instruct\"], prefer: \"deepseek|qwen3-coder|llama\" },\n { paid: true, vendor: \"fireworks\", tps: 120, base: \"https://api.fireworks.ai/inference/v1\", keyEnv: \"FIREWORKS_API_KEY\", priority: 42, tags: [\"coder\", \"reasoning\", \"long\"], models: [\"accounts/fireworks/models/deepseek-v3p1\", \"accounts/fireworks/models/qwen3-coder-480b-a35b-instruct\", \"accounts/fireworks/models/kimi-k2-instruct\"], prefer: \"deepseek|qwen3-coder|kimi\" },\n { paid: true, vendor: \"anthropic\", base: \"https://api.anthropic.com/v1\", keyEnv: \"ANTHROPIC_API_KEY\", priority: 44, tags: [\"coder\", \"reasoning\", \"vision\"], models: [\"claude-sonnet-4-5\", \"claude-haiku-4-5\"], prefer: \"haiku|sonnet\" },\n];\n",[],[512],"276c16ac-6f20-4894-867b-94b3b66089d2","catalog-free.ts","01a0a661-f049-7779-b407-a20f829fbe03",["D",1789497831502],"import type { Vendor } from \"./catalog.ts\";\n// Free tiers, priority 0–19 = latency order (used by `fast`); `auto` ranks by QUALITY instead.\nexport const FREE: Vendor[] = [\n { vendor: \"groq\", tps: 400, maxOut: 5500, maxIn: 6500, base: \"https://api.groq.com/openai/v1\", keyEnv: \"GROQ_API_KEY\", priority: 0, tags: [\"fast\"],\n models: [\"qwen/qwen3.8-27b\", \"openai/gpt-oss-120b\", \"openai/gpt-oss-20b\"], prefer: \"gpt-oss|qwen|llama\" },\n { vendor: \"cerebras\", tps: 1200, maxOut: 8000, base: \"https://api.cerebras.ai/v1\", keyEnv: \"CEREBRAS_API_KEY\", priority: 1, tags: [\"fast\"],\n models: [\"qwen-3.8-27b\", \"gpt-oss-120b\", \"gemma-4-31b\"], prefer: \"gpt-oss|qwen|llama\" },\n { vendor: \"sambanova\", tps: 300, base: \"https://api.sambanova.ai/v1\", keyEnv: \"SAMBANOVA_API_KEY\", priority: 2, tags: [\"fast\", \"reasoning\"],\n models: [\"MiniMax-M3\", \"DeepSeek-V3.2\", \"gpt-oss-120b\", \"Meta-Llama-3.3-70B-Instruct\"], prefer: \"minimax|deepseek|gpt-oss|llama\" },\n { vendor: \"openrouter\", tps: 30, ttft: 4000, base: \"https://openrouter.ai/api/v1\", keyEnv: \"OPENROUTER_API_KEY\", priority: 3, tags: [\"coder\", \"reasoning\", \"long\"],\n // ordered by throughput on the free tier: the 550B ultra model times out inside a 50 s turn and is a last resort\n models: [\"nvidia/nemotron-3-super-120b-a12b:free\", \"google/gemma-4-31b-it:free\", \"nvidia/nemotron-3.5-lightning:free\", \"thinkingmachines/inkling:free\", \"nvidia/nemotron-3-ultra-550b-a55b:free\"],\n prefer: \"super|gemma|lightning\" },\n { vendor: \"huggingface\", tps: 50, base: \"https://router.huggingface.co/v1\", keyEnv: \"HF_TOKEN\", priority: 4, tags: [\"coder\", \"reasoning\", \"long\"],\n models: [\"deepseek-ai/DeepSeek-V4-Pro\", \"moonshotai/Kimi-K3\", \"zai-org/GLM-5.3\", \"deepseek-ai/DeepSeek-V4-Flash\", \"Qwen/Qwen3-Coder-Next\", \"Qwen/Qwen3.8-27B\", \"openai/gpt-oss-120b\", \"zai-org/GLM-5.3-Flash\"], prefer: \"deepseek|kimi|glm|qwen|gpt-oss\" },\n { vendor: \"gemini\", tps: 150, base: \"https://generativelanguage.googleapis.com/v1beta/openai\", keyEnv: \"GEMINI_API_KEY\", priority: 5, tags: [\"fast\", \"long\", \"vision\"],\n models: [\"gemini-2.5-flash-lite\", \"gemini-2.5-flash\", \"gemini-2.0-flash\"], prefer: \"flash\" },\n { vendor: \"mistral\", tps: 80, base: \"https://api.mistral.ai/v1\", keyEnv: \"MISTRAL_API_KEY\", priority: 6, tags: [\"coder\"],\n models: [\"mistral-small-latest\", \"ministral-14b-latest\", \"ministral-8b-latest\", \"devstral-small-latest\"], prefer: \"mistral-small|ministral\" },\n { vendor: \"nvidia_nim\", tps: 40, base: \"https://integrate.api.nvidia.com/v1\", keyEnv: \"NVIDIA_API_KEY\", priority: 20, tags: [\"reasoning\"],\n models: [\"nvidia/llama-3.1-nemotron-51b-instruct\", \"mistralai/mistral-nemotron\", \"meta/llama-3.2-90b-vision-instruct\"], prefer: \"nemotron-(51|70)b|mistral-nemotron\" },\n { vendor: \"zai\", base: \"https://open.bigmodel.cn/api/paas/v4\", keyEnv: \"ZAI_API_KEY\", priority: 8, tags: [\"coder\"],\n models: [\"glm-5.3-flash\", \"glm-4.5-air\", \"glm-5-turbo\"], prefer: \"flash|air\" },\n { vendor: \"cohere\", base: \"https://api.cohere.ai/compatibility/v1\", keyEnv: \"COHERE_API_KEY\", priority: 9,\n models: [\"command-a-03-2025\", \"command-r7b-12-2024\", \"command-r-plus\"], prefer: \"command\" },\n { vendor: \"cloudflare\", tps: 100, base: \"https://api.cloudflare.com/client/v4/accounts/${CF_ACCOUNT_ID}/ai/v1\", keyEnv: \"CF_AI_TOKEN\", priority: 10, tags: [\"fast\"],\n models: [\"@cf/openai/gpt-oss-120b\", \"@cf/meta/llama-3.3-70b-instruct-fp8-fast\", \"@cf/qwen/qwen3-30b-a3b-fp8\"], prefer: \"gpt-oss|llama-3.3|qwen\" },\n { vendor: \"together\", tps: 80, base: \"https://api.together.xyz/v1\", keyEnv: \"TOGETHER_API_KEY\", priority: 11,\n models: [\"openai/gpt-oss-20b\", \"Qwen/Qwen3-14B\", \"meta-llama/Llama-3.3-70B-Instruct\", \"openai/gpt-oss-120b\"], prefer: \"gpt-oss-20b|qwen3-14b|llama-3.3-70b-instruct$\" },\n { vendor: \"siliconflow\", base: \"https://api.siliconflow.cn/v1\", keyEnv: \"SILICONFLOW_API_KEY\", priority: 12,\n models: [\"Qwen/Qwen3-8B\", \"THUDM/GLM-4-9B-0414\"], prefer: \"qwen3-8b|glm\" },\n { vendor: \"ovh\", base: \"https://oai.endpoints.kepler.ai.cloud.ovh.net/v1\", keyEnv: \"OVH_AI_ENDPOINTS_ACCESS_TOKEN\", priority: 13,\n models: [\"Meta-Llama-3_3-70B-Instruct\", \"gpt-oss-120b\", \"Qwen3-32B\"], prefer: \"llama|gpt-oss|qwen\" },\n { vendor: \"chutes\", base: \"https://llm.chutes.ai/v1\", keyEnv: \"CHUTES_API_KEY\", priority: 14,\n models: [\"openai/gpt-oss-120b\", \"Qwen/Qwen3-235B-A22B-Instruct-2507\"], prefer: \"gpt-oss|qwen\" },\n { vendor: \"novita\", base: \"https://api.novita.ai/v3/openai\", keyEnv: \"NOVITA_API_KEY\", priority: 15,\n models: [\"openai/gpt-oss-120b\", \"meta-llama/llama-3.3-70b-instruct\"], prefer: \"gpt-oss|llama-3.3\" },\n { vendor: \"hyperbolic\", base: \"https://api.hyperbolic.xyz/v1\", keyEnv: \"HYPERBOLIC_API_KEY\", priority: 16,\n models: [\"openai/gpt-oss-120b\", \"meta-llama/Llama-3.3-70B-Instruct\"], prefer: \"gpt-oss|llama-3.3\" },\n { vendor: \"dashscope\", base: \"https://dashscope-intl.aliyuncs.com/compatible-mode/v1\", keyEnv: \"DASHSCOPE_API_KEY\", priority: 17, tags: [\"coder\"],\n models: [\"qwen-plus\", \"qwen3-coder-plus\", \"qwen-turbo\"], prefer: \"qwen\" },\n { vendor: \"scaleway\", base: \"https://api.scaleway.ai/v1\", keyEnv: \"SCW_SECRET_KEY\", priority: 18,\n models: [\"gpt-oss-120b\", \"llama-3.3-70b-instruct\"], prefer: \"gpt-oss|llama\" },\n { vendor: \"featherless\", base: \"https://api.featherless.ai/v1\", keyEnv: \"FEATHERLESS_API_KEY\", priority: 19,\n models: [\"meta-llama/Llama-3.3-70B-Instruct\"], prefer: \"llama-3.3|qwen\" },\n];\n",[],[519],"6d51943c-f5e2-4b71-a067-5e2b83ed9008","artifacts.ts","01a0a661-f049-7779-b407-9e2bbbe1a64b",["D",1789497831502],"// Multi-file artifacts: an `artifact` row (title, kind, index content) plus\n// `artifact_file` rows. Served path-style at /artifact// so relative\n// \";\nexport function serveFile(origin: string, path: string, content: string): Response {\n let ct = mimeFor(path || \"index.html\");\n let body: BodyInit = content;\n if (/^(index\\.html?)?$/.test(path) || /\\.html?$/.test(path)) { const i = content.search(/]*>/i); body = i >= 0 ? content.slice(0, content.indexOf(\">\", i) + 1) + CONSOLE_HOOK + content.slice(content.indexOf(\">\", i) + 1) : CONSOLE_HOOK + content; }\n const m = content.length < 12_000_000 && content.startsWith(\"data:\") ? content.match(/^data:([^;,]+)?(;base64)?,/) : null;\n if (m) { // binary asset stored as a data URL by the workbench/zip import\n ct = m[1] || ct; const raw = content.slice(m[0].length);\n body = m[2] ? Uint8Array.from(atob(raw), (c) => c.charCodeAt(0)) : decodeURIComponent(raw);\n }\n return new Response(body, {\n headers: {\n \"content-type\": ct,\n \"content-security-policy\": `default-src 'none'; script-src ${origin} 'unsafe-inline' 'unsafe-eval'; style-src ${origin} 'unsafe-inline'; img-src ${origin} data: blob: https:; font-src ${origin} data:; connect-src ${origin} data: blob:; media-src ${origin} data: blob:; worker-src ${origin} blob:; frame-ancestors 'self'`,\n \"x-content-type-options\": \"nosniff\",\n \"cache-control\": \"no-store\",\n },\n });\n}\n\n/** Pull a full HTML document out of a reply when the model wrote code instead of calling create_artifact. */\nexport function extractHtmlDocument(reply: string): { title: string; html: string } | null {\n const m = reply.match(/```(?:html|htm)?\\s*\\n([\\s\\S]*?)\\n```/gi);\n if (!m) return null;\n for (const block of m) {\n const html = block.replace(/^```\\w*\\s*\\n/, \"\").replace(/\\n```$/, \"\");\n if (/]|= 25) {\n const t = html.match(/([^<]{1,80})<\\/title>/i);\n return { title: t ? t[1].trim() : \"App\", html };\n }\n }\n return null;\n}\n",[],[526],"47065406-e7d6-4943-a59d-1ff50301cfbe","app.tsx","01a0a661-f049-7779-b407-9821bb1c8e22",["D",1789497831502],"import { PAGE_HTML } from \"./main-html.tsx\";\nimport { DEFAULT_PROMPTS, normalizePromptOverrides, PROMPT_VARS, type PromptKey, renderPrompt } from \"./prompts.ts\";\nimport { APP_RX, type ArtFile, deleteFile, extractHtmlDocument, lintArtifact, listFiles, qualityReport, readFile, saveArtifact, saveMany, serveFile, writeFile } from \"./artifacts.ts\";\nimport { discoverMcpTools } from \"./mcp-client.ts\";\nimport { activeSkills, type Focus, normalizeSettings, type Settings, type Skill } from \"./app-settings.ts\";\nimport { handleSwarm } from \"./app-swarm.ts\";\nimport { handleIssues, ISSUE_TOOLS } from \"./issues.ts\";\nimport { BUILD_RX, callInference, type GenOpts, type InferResult } from \"./app-infer.ts\";\nimport { compactMessage, handleCompact, liveChars, loadCompact } from \"./app-compact.ts\";\nimport { deriveTitle, evalJs, getMemoryBlob, META_CLOSE, META_OPEN, parseMeta, parseTextToolCalls, preEstimate, retrieveFacts, salvageArtifactArgs, summarizeEvent, slimEvents, TOOL_ALIASES, type ToolEvent, webSearch, writeFact } from \"./app-helpers.ts\";\nimport {\n all,\n type Bound,\n type DbError,\n Err,\n initDB,\n insertMessageSafe,\n messagesFor,\n Ok,\n one,\n pruneSession,\n raw,\n type Result,\n type Row,\n run,\n sql,\n touchSession,\n unwrap,\n} from \"./db.ts\";\n\n\n\n\n\n\n\nconst BUDGET = {\n loopHops: 8, \n planPass: 1, \n ensembleSamples: 3, \n verifyPass: 1, \n regroundPass: 1, \n hardCeilingPasses: 14, \n};\n\n\n\nconst VERIFIER_MODEL: string | null = null; \n\n\n\n\nconst recent = new Map<string, number>();\nfunction rateLimit(k: string, ms = 900): boolean {\n const now = Date.now();\n if (now - (recent.get(k) ?? 0) < ms) return false;\n recent.set(k, now);\n return true;\n}\nfunction clean(s: string): string {\n return s.replace(/\\u0000/g, \"\").trim().slice(0, 64_000);\n}\nasync function safeJson(res: Response) {\n const ct = res.headers.get(\"content-type\") || \"\";\n const body = await res.text(); \n if (!res.ok || !ct.includes(\"application/json\")) {\n throw new Error(\n `Expected JSON, got ${res.status} ${ct}: ${body.slice(0, 120)}`,\n );\n }\n try {\n return JSON.parse(body);\n } catch {\n throw new Error(`Bad JSON: ${body.slice(0, 120)}`);\n }\n}\n\n\n\n\nasync function ensureSession(id: string): Promise<Row<\"session\">> {\n const existing = unwrap(\n await one(\"session\", sql`SELECT * FROM session WHERE id = ${id}`),\n null,\n );\n if (existing) return existing;\n const row: Row<\"session\"> = { id, name: \"New Chat\", ts: Date.now() };\n await touchSession(row.id, row.name, row.ts);\n return row;\n}\n\nasync function listSessions(): Promise<Row<\"session\">[]> {\n return unwrap(\n await all(\"session\", sql`SELECT * FROM session ORDER BY ts DESC LIMIT 50`),\n [],\n );\n}\n\n\n\nasync function renameSession(id: string, name: string): Promise<void> {\n await run(sql`UPDATE session SET name = ${name} WHERE id = ${id}`);\n}\n\nasync function deleteSession(id: string): Promise<void> {\n \n \n await run(sql`DELETE FROM message WHERE session = ${id}`);\n await run(sql`DELETE FROM session WHERE id = ${id}`);\n}\n\n\n\n\nasync function clearSession(id: string): Promise<void> {\n await pruneSession(id, 0);\n}\n\n\n\n\nasync function addMsg(\n role: Row<\"message\">[\"role\"],\n content: string,\n providerId: string | null = null,\n session = \"default\",\n meta: Record<string, unknown> | null = null,\n): Promise<void> {\n const r = await insertMessageSafe({\n role,\n content,\n ts: Date.now(),\n session,\n providerId,\n });\n if (!r.ok) { console.error(`addMsg failed (session=${session}):`, r.error); return; }\n if (meta) {\n let json = JSON.stringify(meta);\n if (json.length > 120_000) json = JSON.stringify({ ...meta, toolEvents: slimEvents((meta.toolEvents as ToolEvent[]) ?? [], 200) });\n await run(sql`INSERT INTO message_meta (message_id, meta) VALUES (${r.value}, ${json}) ON CONFLICT(message_id) DO UPDATE SET meta = excluded.meta`);\n }\n}\n\n/** Newest messages that fit a character budget (≈ budget/3.6 tokens). Long single messages keep head + tail.\n * The budget is what keeps prompts under free-tier input ceilings (Groq ITPM 7000 ≈ 25k chars all-in, of which\n * system prompt + tool schemas already take ~10k); the router still skips vendors the prompt cannot fit. */\nasync function getContext(session = \"default\", limit = 24, budget = 16_000): Promise<Row<\"message\">[]> {\n const compact = await loadCompact(session);\n const rows = unwrap(await messagesFor(session, limit), []).filter((m) => !compact || m.id > compact.uptoId); \n const out: Row<\"message\">[] = []; let used = compact ? compact.chars : 0;\n for (const [i, m] of rows.entries()) {\n let content = m.content;\n if (content.length > 6_000) content = content.slice(0, 4_200) + \"\\n…[\" + (content.length - 5_400) + \" chars trimmed]…\\n\" + content.slice(-1_200);\n if (i >= 2 && used + content.length > budget) break; \n used += content.length; out.push({ ...m, content });\n }\n out.reverse();\n if (compact) out.unshift({ id: 0, role: \"system\", content: compactMessage(compact).content, ts: compact.ts, session, providerId: null } as Row<\"message\">);\n return out;\n}\n\nasync function getHistory(session: string): Promise<(Row<\"message\"> & { meta: unknown })[]> {\n const rows = unwrap(\n await raw<Row<\"message\"> & { meta: string | null }>(\n sql`SELECT m.*, mm.meta AS meta FROM message m LEFT JOIN message_meta mm ON mm.message_id = m.id WHERE m.session = ${session} ORDER BY m.id ASC`,\n ),\n [],\n );\n const out = rows.map((r) => { let meta: any = null; if (r.meta) { try { meta = JSON.parse(r.meta); } catch { /* ignore */ } } return { ...r, meta }; });\n \n if (out.some((m) => m.role === \"assistant\" && !m.meta)) {\n const arts = unwrap(await all(\"artifact\", sql`SELECT id, title, kind, ts FROM artifact WHERE session = ${session} ORDER BY ts ASC`), []) as { id: number; title: string; kind: string; ts: number }[];\n let prevTs = 0;\n for (const m of out) {\n if (m.role === \"assistant\" && !m.meta) {\n const mine = arts.filter((a) => a.ts > prevTs && a.ts <= m.ts + 2_000);\n if (mine.length) m.meta = { backfilled: true, toolEvents: mine.map((a, i) => ({ hop: i, tool: \"create_artifact\", args: { title: a.title, kind: a.kind }, result: { ok: true, id: a.id, url: `/artifact/${a.id}/`, title: a.title } })) };\n }\n prevTs = m.ts;\n }\n }\n return out;\n}\n\n\n\n\n\n\n\n\n\n\n\n\n\ninterface ToolCtx {\n session: string;\n turn?: { artifactId?: number; artifactTitle?: string }; \n \n \n \n tools: Record<string, Tool>;\n schemas: unknown[];\n}\ninterface Tool {\n schema: Record<string, unknown>;\n handle: (args: Record<string, unknown>, ctx: ToolCtx) => Promise<unknown>;\n}\n\n\nconst TOOLS: Record<string, Tool> = {\n memory_recall: {\n schema: {\n type: \"function\",\n function: {\n name: \"memory_recall\",\n description:\n \"Retrieve the most relevant remembered facts for a query, ranked by relevance × decayed-salience.\",\n parameters: {\n type: \"object\",\n properties: {\n query: { type: \"string\", description: \"What to recall about.\" },\n },\n required: [\"query\"],\n },\n },\n },\n handle: async ({ query }) => ({\n facts: await retrieveFacts(String(query)),\n }),\n },\n memory_write: {\n schema: {\n type: \"function\",\n function: {\n name: \"memory_write\",\n description:\n \"Persist a durable fact. Provide `predictability` 0..1: how obvious/derivable this fact was to you BEFORE learning it (0 = totally surprising, 1 = you would have guessed it). Surprising facts are retained longer.\",\n parameters: {\n type: \"object\",\n properties: {\n fact: { type: \"string\", description: \"The fact (≤400 chars).\" },\n predictability: {\n type: \"number\",\n description: \"0..1 self-estimated prior predictability.\",\n },\n },\n required: [\"fact\", \"predictability\"],\n },\n },\n },\n handle: async ({ fact, predictability }) => {\n const p = Number(predictability);\n await writeFact(String(fact), Number.isFinite(p) ? p : 0.5);\n return { ok: true };\n },\n },\n name_session: {\n schema: {\n type: \"function\",\n function: {\n name: \"name_session\",\n description:\n \"Set a concise title (≤40 chars, ≤4 words) for the current conversation.\",\n parameters: {\n type: \"object\",\n properties: {\n name: { type: \"string\", description: \"Session title.\" },\n },\n required: [\"name\"],\n },\n },\n },\n handle: async ({ name }, { session }) => {\n const trimmed = String(name).slice(0, 40);\n await renameSession(session, trimmed);\n return { ok: true, name: trimmed };\n },\n },\n session_clear: {\n schema: {\n type: \"function\",\n function: {\n name: \"session_clear\",\n description: \"Erase all messages in the current session.\",\n parameters: { type: \"object\", properties: {}, required: [] },\n },\n },\n handle: async (_a, { session }) => {\n await clearSession(session);\n return { ok: true };\n },\n },\n session_delete: {\n schema: {\n type: \"function\",\n function: {\n name: \"session_delete\",\n description:\n \"Permanently delete the current session and all its messages.\",\n parameters: { type: \"object\", properties: {}, required: [] },\n },\n },\n handle: async (_a, { session }) => {\n await deleteSession(session);\n return { ok: true };\n },\n },\n create_artifact: {\n schema: {\n type: \"function\",\n function: {\n name: \"create_artifact\",\n description:\n \"Build an application, visualization, game, graphic or document the user opens and uses. `content` is the entry file (a complete HTML document for kind html; an SVG; or Markdown). Optional `files` adds sibling files (app.js, style.css, shader.glsl, data.json …) that the entry references by RELATIVE path (<script src=\\\"app.js\\\">, fetch(\\\"data.json\\\")). Call this ONCE per turn with the complete program; to change something afterwards call update_artifact — a second create_artifact in the same turn updates the first one instead of duplicating it. Renders in a sandboxed frame: no cookies, no parent page, network limited to its own files.\",\n parameters: {\n type: \"object\",\n properties: {\n title: { type: \"string\", description: \"Short, stable name — the artifact's identity for future updates.\" },\n kind: { type: \"string\", enum: [\"html\", \"svg\", \"markdown\"], description: \"Entry file type.\" },\n content: { type: \"string\", description: \"Entry file source: a complete HTML document (kind html), SVG, or Markdown.\" },\n files: {\n type: \"array\",\n description: \"Optional extra files referenced relatively from the entry file.\",\n items: { type: \"object\", properties: { path: { type: \"string\" }, content: { type: \"string\" } }, required: [\"path\", \"content\"] },\n },\n },\n required: [\"title\", \"kind\", \"content\"],\n },\n },\n },\n handle: async ({ title, kind, content, files }, ctx) => {\n const k = String(kind);\n if (k !== \"html\" && k !== \"svg\" && k !== \"markdown\") return { error: `Invalid kind: ${k}. Must be html, svg, or markdown.` };\n \n \n \n \n \n const t = String(title);\n const r = await saveArtifact(ctx.session, t, k, String(content), Array.isArray(files) ? (files as ArtFile[]) : [], ctx.turn?.artifactId ?? null);\n if (!r.ok) return { error: r.error };\n if (ctx.turn) { ctx.turn.artifactId = r.value.id; ctx.turn.artifactTitle = r.value.title; }\n const lf = await listFiles(r.value.id);\n const issues = lf ? lintArtifact(lf.files) : [];\n const quality = lf ? qualityReport(lf.files) : [];\n return { ok: true, id: r.value.id, url: `/artifact/${r.value.id}/`, title: r.value.title, files: lf?.files.map((f) => f.path) ?? [], issues: issues.length ? issues : undefined, quality: quality.length ? quality : undefined, hint: issues.length ? \"Fix these with update_artifact before answering.\" : undefined };\n },\n },\n read_artifact: {\n schema: {\n type: \"function\",\n function: {\n name: \"read_artifact\",\n description: \"Read an existing artifact before editing it: returns its file list and the content of one file (default: the entry file). Use the id from the user's message (#36 → id 36) or the title.\",\n parameters: {\n type: \"object\",\n properties: {\n id: { type: \"number\", description: \"Artifact id.\" },\n title: { type: \"string\", description: \"Artifact title, if id is unknown.\" },\n path: { type: \"string\", description: \"File to read; default index.html.\" },\n all: { type: \"boolean\", description: \"Return every file's content (small artifacts).\" },\n },\n },\n },\n },\n handle: async ({ id, title, path, all: wantAll }, ctx) => {\n let aid = Number(id) || ctx.turn?.artifactId;\n if (!aid && title) {\n const r = await one(\"artifact\", sql`SELECT * FROM artifact WHERE title = ${String(title)} ORDER BY ts DESC LIMIT 1`);\n if (r.ok && r.value) aid = r.value.id;\n }\n if (!aid) return { error: \"artifact not found; pass id or title\" };\n const lf = await listFiles(aid);\n if (!lf) return { error: `no artifact with id ${aid}` };\n if (ctx.turn) { ctx.turn.artifactId = aid; ctx.turn.artifactTitle = lf.artifact.title; }\n const total = lf.files.reduce((n, x) => n + x.content.length, 0);\n if (wantAll && total <= 40_000) return { id: aid, title: lf.artifact.title, files: lf.files, issues: lintArtifact(lf.files) };\n const p = String(path || lf.files[0].path);\n const f = lf.files.find((x) => x.path === p);\n return { id: aid, title: lf.artifact.title, files: lf.files.map((x) => ({ path: x.path, bytes: x.content.length })), path: p, content: f ? f.content.slice(0, 60_000) : null, issues: lintArtifact(lf.files), error: f ? undefined : `file not found: ${p}` };\n },\n },\n search_artifact: {\n schema: {\n type: \"function\",\n function: {\n name: \"search_artifact\",\n description: \"grep across an artifact's files: regex or plain text → matching lines as path:line: text. Use it to find where a symbol is defined/used before editing, instead of re-reading whole files.\",\n parameters: { type: \"object\", properties: { id: { type: \"number\" }, query: { type: \"string\", description: \"regex (case-insensitive) or literal\" }, path: { type: \"string\", description: \"limit to one file (optional)\" } }, required: [\"query\"] },\n },\n },\n handle: async ({ id, query, path }, ctx) => {\n const aid = Number(id) || ctx.turn?.artifactId;\n if (!aid) return { error: \"no artifact in focus; pass id\" };\n const lf = await listFiles(aid); if (!lf) return { error: `no artifact with id ${aid}` };\n if (ctx.turn) ctx.turn.artifactId = aid;\n let rx: RegExp; try { rx = new RegExp(String(query), \"i\"); } catch { rx = new RegExp(String(query).replace(/[.*+?^${}()|[\\]\\\\]/g, \"\\\\$&\"), \"i\"); }\n const hits: string[] = [];\n for (const f of lf.files) { if (path && f.path !== path) continue; if (f.content.startsWith(\"data:\")) continue; const lines = f.content.split(\"\\n\"); for (let i = 0; i < lines.length && hits.length < 80; i++) if (rx.test(lines[i])) hits.push(`${f.path}:${i + 1}: ${lines[i].trim().slice(0, 200)}`); }\n return { id: aid, query, matches: hits.length, hits, exports: Object.fromEntries(lf.files.filter((f) => /\\.m?js$/.test(f.path)).map((f) => [f.path, [...f.content.matchAll(/^\\s*export\\s+(?:async\\s+)?(?:const|let|var|function\\*?|class)\\s+([\\w$]+)/gm)].map((m) => m[1])])) };\n },\n },\n update_artifact: {\n schema: {\n type: \"function\",\n function: {\n name: \"update_artifact\",\n description:\n \"Modify an existing artifact in ONE call: `edits` (exact find→replace, each `find` must occur once in its file), `files` (add or fully replace files), `delete` (remove files). Single-edit shorthand: `path`+`find`+`replace` or `path`+`content`. Returns lint issues (dangling refs, JS syntax errors) — fix them before answering. Prefer several small edits in one call over rewriting a file.\",\n parameters: {\n type: \"object\",\n properties: {\n id: { type: \"number\", description: \"Artifact id (from create_artifact / the user's #id).\" },\n title: { type: \"string\", description: \"Artifact title, if id is unknown.\" },\n edits: { type: \"array\", description: \"Exact-text edits applied in order.\", items: { type: \"object\", properties: { path: { type: \"string\" }, find: { type: \"string\" }, replace: { type: \"string\" } }, required: [\"path\", \"find\", \"replace\"] } },\n files: { type: \"array\", description: \"Files to add or replace entirely.\", items: { type: \"object\", properties: { path: { type: \"string\" }, content: { type: \"string\" } }, required: [\"path\", \"content\"] } },\n delete: { type: \"array\", description: \"File paths to remove.\", items: { type: \"string\" } },\n path: { type: \"string\", description: \"Shorthand single target; default index.html.\" },\n content: { type: \"string\", description: \"Shorthand: full new content for `path`.\" },\n find: { type: \"string\", description: \"Shorthand: exact text to replace in `path`.\" },\n replace: { type: \"string\", description: \"Shorthand: replacement text.\" },\n },\n },\n },\n },\n handle: async ({ id, title, path, content, find, replace, edits, files, delete: del }, ctx) => {\n let aid = Number(id) || ctx.turn?.artifactId;\n if (!aid && title) {\n const r = await one(\"artifact\", sql`SELECT * FROM artifact WHERE title = ${String(title)} ORDER BY ts DESC LIMIT 1`);\n if (r.ok && r.value) aid = r.value.id;\n }\n if (!aid) return { error: \"artifact not found; pass id or title (call read_artifact first)\" };\n const ops: { path: string; find?: string; replace?: string; content?: string }[] = [];\n if (Array.isArray(edits)) for (const e of edits as any[]) if (e?.path && typeof e.find === \"string\") ops.push({ path: String(e.path), find: e.find, replace: String(e.replace ?? \"\") });\n if (Array.isArray(files)) for (const f of files as any[]) if (f?.path && typeof f.content === \"string\") ops.push({ path: String(f.path), content: f.content });\n if (typeof content === \"string\") ops.push({ path: String(path || \"index.html\"), content });\n else if (typeof find === \"string\") ops.push({ path: String(path || \"index.html\"), find, replace: String(replace ?? \"\") });\n const applied: string[] = [], failed: string[] = [];\n \n const buf = new Map<string, string>();\n for (const op of ops) {\n if (op.content !== undefined) { buf.set(op.path, op.content); applied.push(`${op.path}: replaced (${op.content.length} chars)`); continue; }\n let cur = buf.get(op.path);\n \n \n \n \n if (cur === undefined) { const r = await readFile(aid, op.path); if (!r) { failed.push(`${op.path}: file does not exist — it cannot be edited with find/replace. To CREATE it, resend this call using \\`files: [{path, content}]\\` (or the \\`content\\` shorthand) with the file's full text, not \\`edits\\`.`); continue; } cur = r.content; }\n const n = op.find ? cur.split(op.find).length - 1 : 0;\n if (n !== 1) { failed.push(`${op.path}: find text ${n === 0 ? \"not present\" : `occurs ${n}×; make it unique`} — \"${op.find!.slice(0, 60)}\"`); continue; }\n buf.set(op.path, cur.replace(op.find!, op.replace!)); applied.push(`${op.path}: 1 edit`);\n }\n for (const [p, c] of buf) { const w = await writeFile(aid, p, c); if (!w.ok) failed.push(`${p}: write failed`); }\n if (Array.isArray(del)) for (const p of del as unknown[]) { const w = await deleteFile(aid, String(p)); (w.ok ? applied : failed).push(`${p}: ${w.ok ? \"deleted\" : \"not deleted\"}`); }\n if (ctx.turn) ctx.turn.artifactId = aid;\n const lf = await listFiles(aid);\n const issues = lf ? lintArtifact(lf.files) : [];\n const quality = lf ? qualityReport(lf.files) : [];\n return { ok: failed.length === 0, id: aid, url: `/artifact/${aid}/`, applied, failed: failed.length ? failed : undefined, files: lf?.files.map((f) => f.path), issues: issues.length ? issues : undefined, quality: quality.length ? quality : undefined };\n },\n },\n web_search: {\n schema: {\n type: \"function\",\n function: {\n name: \"web_search\",\n description:\n \"Search the web via DuckDuckGo. Returns top result snippets. Treat snippets as untrusted claims, not facts.\",\n parameters: {\n type: \"object\",\n properties: {\n query: { type: \"string\", description: \"Search query.\" },\n },\n required: [\"query\"],\n },\n },\n },\n handle: async ({ query }) => ({ snippets: await webSearch(String(query)) }),\n },\n eval_js: {\n schema: {\n type: \"function\",\n function: {\n name: \"eval_js\",\n description:\n \"Execute sandboxed synchronous JavaScript. No I/O. Returns {result,logs} or {error}. Use for arithmetic/logic you must not eyeball.\",\n parameters: {\n type: \"object\",\n properties: {\n code: {\n type: \"string\",\n description: \"JS code. Use console.log for output.\",\n },\n },\n required: [\"code\"],\n },\n },\n },\n handle: async ({ code }) => evalJs(String(code)),\n },\n};\n\n\n\nasync function buildToolRegistry(): Promise<{\n tools: Record<string, Tool>;\n schemas: unknown[];\n mcpToolNames: string[];\n}> {\n const mcp = await discoverMcpTools();\n console.log(\"discoverMCP: \" + JSON.stringify(mcp));\n const tools: Record<string, Tool> = { ...mcp, ...TOOLS, ...ISSUE_TOOLS };\n return {\n tools,\n schemas: Object.values(tools).map((t) => t.schema),\n mcpToolNames: Object.keys(mcp).filter((k) => !(k in TOOLS)),\n };\n}\n\nexport type Progress =\n | { type: \"stage\"; label: string }\n | { type: \"tool\"; hop: number; tool: string; args: Record<string, unknown>; result?: unknown; error?: string }\n | { type: \"pass\"; n: number; of: number }\n | { type: \"thinking\"; text: string }\n | { type: \"continue\"; n: number; chars?: number; tail?: string }\n | { type: \"partial\"; text: string }\n | { type: \"retry\"; n: number; error: string };\n\n\nconst TURN_BUDGET_MS = 52_000; \nclass PassMeter {\n private used = 0;\n private t0 = Date.now();\n public reasoning: string[] = [];\n \n \n \n \n public routeEvents: ToolEvent[] = [];\n \n \n \n \n \n \n public modelEvents: ToolEvent[] = [];\n constructor(\n private ceiling: number,\n public emit: (p: Progress) => void = () => {},\n public defaultModel: string | null = null,\n public settings: Settings = normalizeSettings({}),\n ) {}\n /** a prompt for this turn: user override or default, with {{vars}} filled */\n p(key: PromptKey, vars: Record<string, string> = {}): string {\n return renderPrompt(key, vars, this.settings.prompts, this.settings.examples);\n }\n /** ms of turn budget left; optional stages must check this before spending a pass */\n get timeLeft(): number {\n return this.settings.budgetMs - (Date.now() - this.t0);\n }\n get remaining(): number {\n return this.ceiling - this.used;\n }\n async call(\n messages: unknown[],\n tools?: unknown[],\n model?: string | null,\n override: GenOpts = {},\n ): Promise<InferResult | null> {\n if (this.used >= this.ceiling) return null;\n this.used++;\n this.emit({ type: \"pass\", n: this.used, of: this.ceiling });\n const s = this.settings;\n const expectFor = { short: 1200, long: 6000, max: 12_000 }[s.expect as \"short\" | \"long\" | \"max\"];\n const gen: GenOpts = { reasoning: s.reasoning, temperature: s.temperature, maxTokens: s.maxTokens, router: s.router, keyPolicy: s.keyPolicy, vendorOrder: s.vendorOrder, maxMode: s.expect === \"max\" && s.maxMode === \"auto\" ? \"max\" : s.maxMode, ...(expectFor ? { expect: expectFor } : {}), ...override };\n \n \n \n if (this.timeLeft < 8_000) throw Object.assign(new Error(\"TURN_BUDGET: no time left for another model call\"), { budget: true });\n const req = this.summarizeRequest(messages, tools, model, gen);\n let r: InferResult;\n try {\n r = await callInference(messages, tools, model ?? this.defaultModel, this.timeLeft - 3_000, gen);\n } catch (e: any) {\n this.recordModelCall(req, null, e);\n throw e;\n }\n this.recordRoute(r);\n this.recordModelCall(req, r);\n if (r.reasoning) { this.reasoning.push(r.reasoning); this.emit({ type: \"thinking\", text: r.reasoning.slice(0, 600) }); }\n if (r.message?.content) this.emit({ type: \"partial\", text: String(r.message.content).slice(-1_500) });\n \n let n = 0;\n let lastLen = String(r.message?.content ?? \"\").length;\n while (r.finishReason === \"length\" && !r.message?.tool_calls?.length && r.message?.content && n < s.continueMax && this.timeLeft > 15_000 && this.used < this.ceiling) {\n n++; this.used++;\n this.emit({ type: \"continue\", n, chars: lastLen, tail: String(r.message.content).slice(-240) });\n const contReq = this.summarizeRequest([{ role: \"user\", content: this.p(\"continue\") }], undefined, model, { ...gen, expect: 6000 }, `continue #${n} (${lastLen} chars so far)`);\n let cont: InferResult;\n try {\n cont = await callInference(\n [...messages, { role: \"assistant\", content: r.message.content }, { role: \"user\", content: this.p(\"continue\") }],\n undefined, model ?? this.defaultModel, this.timeLeft - 3_000, { ...gen, expect: 6000 },\n );\n } catch (e) {\n this.recordModelCall(contReq, null, e as Error);\n this.emit({ type: \"retry\", n, error: `continuation failed, keeping ${r.message.content.length} chars: ${String((e as Error).message).slice(0, 120)}` });\n break; \n }\n this.recordRoute(cont);\n this.recordModelCall(contReq, cont);\n r = { ...cont, message: { ...r.message, content: r.message.content + (cont.message?.content ?? \"\") }, meta: { ...r.meta, retries: cont.meta.retries } };\n if (cont.reasoning) this.reasoning.push(cont.reasoning);\n this.emit({ type: \"partial\", text: String(r.message.content).slice(-1_500) });\n \n const nowLen = String(r.message.content).length; if (nowLen - lastLen < 200) { this.emit({ type: \"retry\", n, error: `continuation added ${nowLen - lastLen} chars — stopping in-turn continues; the client resumes next turn` }); break; } lastLen = nowLen;\n }\n return r;\n }\n get spent(): number {\n return this.used;\n }\n \n \n private recordRoute(r: InferResult): void {\n const attempts = r.meta?.attempts;\n if (!attempts?.length) return;\n const winner = attempts[attempts.length - 1]?.ok ? attempts[attempts.length - 1].provider : null;\n const failed = attempts.filter((a) => !a.ok && !a.skipped).length;\n const error = winner ? undefined : `exhausted — ${failed} provider${failed === 1 ? \"\" : \"s\"} failed`;\n const args = { model: r.meta?.model ?? null, winner, tried: attempts.length, planned: (r.meta?.plan ?? attempts.map((a) => a.provider)).length };\n const result = { plan: r.meta?.plan ?? attempts.map((a) => a.provider), attempts, winner };\n this.routeEvents.push({ hop: -2, tool: \"route\", ts: Date.now(), args, result, error });\n \n \n \n \n \n this.emit({ type: \"tool\", hop: -2, tool: \"route\", args, result, error });\n }\n \n \n \n \n private summarizeRequest(messages: unknown[], tools: unknown[] | undefined, model: string | null | undefined, gen: GenOpts, label?: string): Record<string, unknown> {\n const arr = messages as any[];\n const last = [...arr].reverse().find((m) => m?.role === \"user\" || m?.role === \"tool\");\n const sys = arr.find((m) => m?.role === \"system\");\n return {\n label: label ?? null,\n model: model ?? this.defaultModel ?? \"auto\",\n messageCount: arr.length,\n lastMessage: last ? `[${last.role}] ${String(last.content ?? \"\").slice(0, 300)}` : null,\n systemPromptChars: sys ? String(sys.content ?? \"\").length : 0,\n tools: tools?.length ? tools.map((t: any) => t.function?.name ?? t.name ?? \"?\") : [],\n temperature: gen.temperature ?? null, maxTokens: gen.maxTokens ?? null, expect: gen.expect ?? null,\n reasoning: gen.reasoning ?? null, vendorOrder: gen.vendorOrder ?? null,\n };\n }\n \n \n \n private recordModelCall(args: Record<string, unknown>, r: InferResult | null, err?: Error): void {\n if (!r) {\n const error = String(err?.message ?? err ?? \"unknown error\").slice(0, 300);\n this.modelEvents.push({ hop: -3, tool: \"model\", ts: Date.now(), args, error });\n this.emit({ type: \"tool\", hop: -3, tool: \"model\", args, error });\n return;\n }\n const toolCalls = (r.message?.tool_calls ?? []).map((c: any) => ({ name: c.function?.name ?? c.name, args: c.function?.arguments ?? c.args }));\n const result = {\n provider: r.meta?.instance ?? null, model: r.meta?.model ?? null,\n finishReason: r.finishReason ?? r.meta?.finishReason ?? null,\n content: r.message?.content ? String(r.message.content).slice(0, 1_500) : null,\n toolCalls, reasoningChars: r.reasoning ? r.reasoning.length : 0,\n tokensIn: r.meta?.tokensIn ?? r.usage?.prompt_tokens ?? null, tokensOut: r.meta?.tokensOut ?? r.usage?.completion_tokens ?? null,\n maxTokensSent: r.meta?.maxTokensSent ?? null, latencyMs: r.meta?.latencyMs ?? null, retries: r.meta?.retries ?? null,\n };\n const flagged = result.finishReason && ![\"stop\", \"tool_calls\"].includes(result.finishReason);\n const error = flagged ? `finish_reason=${result.finishReason}` : undefined;\n this.modelEvents.push({ hop: -3, tool: \"model\", ts: Date.now(), args, result, error });\n this.emit({ type: \"tool\", hop: -3, tool: \"model\", args, result, error });\n }\n}\n\n\n\n\n\n\n\n\n\n\n\ninterface AgentResult {\n reply: string;\n sessionName?: string;\n toolEvents: ToolEvent[];\n meta: InferResult[\"meta\"];\n truncated?: boolean;\n}\n\n\n\n\n/** The JSON object starting at `at` (must be \"{\"), with nesting and string escapes respected; \"\" if unbalanced. */\nasync function agentLoop(\n messages: unknown[],\n ctx: ToolCtx,\n meter: PassMeter,\n): Promise<AgentResult> {\n const working = [...messages];\n const toolEvents: ToolEvent[] = [];\n const seen = new Map<string, unknown>(); \n \n \n \n \n \n \n const failedSigs = new Map<string, string>(); \n let sessionName: string | undefined;\n let lastMeta: InferResult[\"meta\"] = {};\n\n const exclude: string[] = []; let emptyRetries = 0;\n for (let hop = 0; hop < BUDGET.loopHops; hop++) {\n if (hop > 0 && meter.timeLeft < 12_000) break;\n \n const built = toolEvents.some((e) => /artifact/.test(e.tool) && !e.error);\n let resp: InferResult | null;\n try {\n \n resp = await meter.call(working, ctx.schemas, undefined, { ...(built ? { expect: 2500, maxTokens: Math.min(meter.settings.maxTokens ?? 3200, 3200) } : {}), ...(exclude.length ? { exclude } : {}) });\n } catch (e) {\n \n \n if (!toolEvents.length) throw e;\n const budget = !!(e as any)?.budget;\n meter.emit({ type: \"retry\", n: 0, error: budget ? \"turn budget exhausted — saving progress; continuing in the next turn\" : `wrap-up pass failed, reporting tool results: ${String((e as Error).message).slice(0, 120)}` });\n return { reply: toolEvents.map(summarizeEvent).filter(Boolean).join(\"\\n\\n\"), sessionName, toolEvents, meta: lastMeta, truncated: budget };\n }\n if (!resp || !resp.message) break;\n const msg = resp.message;\n lastMeta = resp.meta;\n if (!msg.tool_calls?.length && typeof msg.content === \"string\" && /<tool_call>|to=functions\\.|assistantfinal/.test(msg.content)) {\n const { calls, rest } = parseTextToolCalls(msg.content);\n if (calls.length) { msg.tool_calls = calls; msg.content = rest || null; resp.finishReason = \"tool_calls\"; }\n else msg.content = rest;\n }\n \n \n const visibleNow = typeof msg.content === \"string\" ? parseMeta(msg.content).visible.replace(/[\\s…]+/g, \"\") : \"\";\n if (!msg.tool_calls?.length && visibleNow.length < 2 && emptyRetries < 2 && meter.timeLeft > 12_000 && meter.remaining > 1) {\n emptyRetries++;\n const inst = String(resp.meta?.instance ?? (resp.meta?.provider as any)?.name ?? \"\");\n if (inst) exclude.push(inst);\n meter.emit({ type: \"retry\", n: emptyRetries, error: `empty reply from ${inst || \"model\"} — retrying on another vendor` });\n if (emptyRetries === 1) working.push({ role: \"user\", content: \"Your previous reply was empty. Answer now: either call a tool (read_artifact / search_artifact / update_artifact) or write the answer. Do not reply with reasoning only.\" });\n hop--; continue;\n }\n\n working.push({\n role: \"assistant\",\n content: msg.content ?? null,\n ...(msg.tool_calls?.length ? { tool_calls: msg.tool_calls } : {}),\n });\n\n const toolCalls: any[] = msg?.tool_calls ?? [];\n if (resp.finishReason === \"stop\" || !toolCalls.length) {\n let reply = (msg.content ?? \"\").trim();\n \n const visibleLen = (r: string) => parseMeta(r).visible.trim().length;\n const onlyMeta = reply && visibleLen(reply) === 0;\n if (onlyMeta) reply = \"\";\n \n const lastWrite = [...toolEvents].reverse().find((e) => (e.tool === \"create_artifact\" || e.tool === \"update_artifact\") && !e.error);\n const issuesLeft: string[] = ((lastWrite?.result as any)?.issues ?? []);\n const stubs = issuesLeft.filter((i) => /placeholder|stub|TODO marker|elided|fill-in|empty function|unimplemented|lorem/.test(i));\n const nudges = working.filter((m: any) => m.role === \"user\" && /^The artifact still (contains placeholders|has lint issues)/.test(String(m.content))).length;\n \n const askText = String(([...working].reverse().find((m: any) => m.role === \"user\" && !/^The artifact (still|is below)/.test(String(m.content))) as any)?.content ?? \"\");\n const qualityLeft: string[] = APP_RX.test(askText) && BUILD_RX.test(askText) ? ((lastWrite?.result as any)?.quality ?? []) : [];\n const qNudges = working.filter((m: any) => m.role === \"user\" && /^The artifact is below the virtuoso standard/.test(String(m.content))).length;\n if (!issuesLeft.length && qualityLeft.length >= 3 && qNudges < 1 && hop < BUDGET.loopHops - 1 && meter.timeLeft > 20_000 && meter.remaining > 1) {\n meter.emit({ type: \"stage\", label: \"raising quality\" });\n working.push({ role: \"user\", content: `The artifact is below the virtuoso standard — a basic build is a failed build. Address every finding now with update_artifact (add real files and features; do not answer in prose first):\\n${qualityLeft.slice(0, 12).join(\"\\n\")}` });\n continue;\n }\n \n if (issuesLeft.length && nudges < 2 && hop < BUDGET.loopHops - 1 && meter.timeLeft > 15_000 && meter.remaining > 1) {\n meter.emit({ type: \"stage\", label: stubs.length ? \"replacing placeholders\" : \"fixing lint issues\" });\n working.push({ role: \"user\", content: stubs.length\n ? `The artifact still contains placeholders — this is never acceptable. Replace each with the complete, working implementation via update_artifact now (do not remove the feature, do not answer in prose first):\\n${stubs.slice(0, 8).join(\"\\n\")}`\n : `The artifact still has lint issues — fix every one with update_artifact now (batched edits; do not answer in prose first):\\n${issuesLeft.slice(0, 10).join(\"\\n\")}` });\n continue;\n }\n const readOnly = toolEvents.some((e) => e.tool === \"read_artifact\") && !toolEvents.some((e) => e.tool === \"update_artifact\" || e.tool === \"create_artifact\");\n if (readOnly && hop < BUDGET.loopHops - 1 && meter.timeLeft > 15_000 && meter.remaining > 1 && !working.some((m: any) => m.role === \"user\" && String(m.content).startsWith(\"You read the artifact\"))) {\n meter.emit({ type: \"stage\", label: \"applying edit\" });\n working.push({ role: \"user\", content: \"You read the artifact but made no change. Call update_artifact now with the concrete edits (batched `edits` and/or `files`). Do not answer in prose until the tool has run.\" });\n continue;\n }\n if (!reply && toolEvents.length && meter.timeLeft > 9_000 && meter.remaining > 0) {\n const onlyNaming = toolEvents.every((e) => e.tool === \"name_session\");\n meter.emit({ type: \"stage\", label: onlyNaming ? \"answering\" : \"summarizing\" });\n const again = await meter.call([...working, { role: \"user\", content: onlyNaming ? \"Now answer the user's request in full. No tool calls.\" : meter.p(\"summarize\") }]);\n reply = (again?.message?.content ?? \"\").trim();\n if (again?.meta) lastMeta = again.meta;\n }\n if (!reply && toolEvents.length) reply = toolEvents.map(summarizeEvent).filter(Boolean).join(\"\\n\\n\");\n return {\n reply,\n sessionName,\n toolEvents,\n meta: lastMeta,\n truncated: resp.finishReason === \"length\",\n };\n }\n\n for (const tc of toolCalls) {\n let toolName: string = tc?.function?.name ?? \"\";\n let args: Record<string, unknown> = {};\n let salvaged: { saved: string[]; cut: string | null } | null = null;\n try {\n args = JSON.parse(tc?.function?.arguments ?? \"{}\");\n } catch (err) {\n \n \n const sv = /artifact$/.test(toolName) || /artifact$/.test(TOOL_ALIASES[toolName] ?? \"\") ? salvageArtifactArgs(String(tc?.function?.arguments ?? \"\")) : null;\n if (sv) { args = sv.args; salvaged = { saved: sv.saved, cut: sv.cut }; meter.emit({ type: \"retry\", n: hop, error: `tool call cut at max_tokens — salvaged ${sv.saved.length} complete file(s)${sv.cut ? `, ${sv.cut} was cut` : \"\"}` }); }\n else { console.log(\"function argument error: \" + String(err)); }\n }\n \n if (!ctx.tools[toolName]) { const alias = TOOL_ALIASES[toolName]; if (alias && ctx.tools[alias]) toolName = alias; }\n\n const tool = ctx.tools[toolName];\n let output: unknown;\n let eventError: string | undefined;\n \n \n const sig = toolName + \":\" + JSON.stringify(args);\n const seenBefore = seen.get(sig);\n if (tool && seenBefore && /^(read_artifact|search_artifact|memory_recall|web_search)$/.test(toolName)) {\n output = { ...(typeof seenBefore === \"object\" && seenBefore ? seenBefore as object : { result: seenBefore }), note: \"IDENTICAL CALL REPEATED — you already have this result. Do not read again; call update_artifact with concrete edits now, or answer.\" };\n toolEvents.push({ hop, tool: toolName, args, result: { repeated: true }, ts: Date.now() });\n meter.emit({ type: \"tool\", hop, tool: toolName, args, error: \"repeated\" });\n working.push({ role: \"tool\", tool_call_id: tc.id, content: JSON.stringify(output) });\n continue;\n }\n \n \n \n \n const failedBefore = tool ? failedSigs.get(sig) : undefined;\n if (failedBefore) {\n output = { error: `REPEATED CALL, SAME ARGUMENTS AS A CALL THAT JUST FAILED: ${failedBefore}\\nResending this exact call will fail the same way. Change your approach: fix what the error names, or do something else — do not call ${toolName} again with these exact arguments.` };\n eventError = failedBefore;\n toolEvents.push({ hop, tool: toolName, args, result: { repeated: true, priorFailure: failedBefore }, error: eventError, ts: Date.now() });\n meter.emit({ type: \"tool\", hop, tool: toolName, args, error: \"repeated-failure\" });\n working.push({ role: \"tool\", tool_call_id: tc.id, content: JSON.stringify(output) });\n continue;\n }\n if (tool) {\n try {\n output = await tool.handle(args, ctx);\n if (toolName === \"name_session\" && (output as any)?.name) {\n sessionName = (output as any).name;\n }\n } catch (e) {\n output = { error: String(e) };\n eventError = String(e);\n }\n } else {\n output = { error: `Unknown tool: ${toolName}. Available: ${Object.keys(ctx.tools).join(\", \")}` };\n eventError = `Unknown tool: ${toolName}`;\n }\n if (!eventError) seen.set(sig, output);\n \n \n const softFailure = !eventError && output && typeof output === \"object\"\n ? (Array.isArray((output as any).failed) && (output as any).failed.length ? (output as any).failed.join(\"; \")\n : (output as any).ok === false ? \"ok:false\" : null)\n : null;\n if (eventError || softFailure) failedSigs.set(sig, String(eventError || softFailure).slice(0, 240));\n else failedSigs.delete(sig); \n if (salvaged && output && typeof output === \"object\") output = { ...(output as object), TRUNCATED_CALL: `Your tool call was cut at max_tokens. Saved complete files: [${salvaged.saved.join(\", \") || \"none\"}]. ${salvaged.cut ? `\"${salvaged.cut}\" was cut and NOT saved.` : \"\"} Resend the missing file(s) now — ONE file per update_artifact call, ≤ ~200 lines each; keep calling until every file exists.` };\n\n toolEvents.push({\n hop,\n tool: toolName,\n args,\n result: eventError ? undefined : output,\n error: eventError,\n ts: Date.now(),\n });\n meter.emit({ type: \"tool\", hop, tool: toolName, args, error: eventError });\n working.push({\n role: \"tool\",\n tool_call_id: tc.id,\n content: JSON.stringify(output),\n });\n }\n }\n\n const last = [...working].reverse().find((m: any) => m.role === \"assistant\" && typeof m.content === \"string\" && m.content.trim()) as any;\n let reply = (last?.content ?? \"\").trim();\n if (!reply && toolEvents.length && meter.timeLeft > 9_000 && meter.remaining > 0) {\n meter.emit({ type: \"stage\", label: \"summarizing\" });\n const again = await meter.call([...working, { role: \"user\", content: meter.p(\"summarize\") }]).catch(() => null);\n reply = (again?.message?.content ?? \"\").trim();\n }\n if (!reply) reply = toolEvents.length ? toolEvents.map(summarizeEvent).filter(Boolean).join(\"\\n\\n\") : \"Agent loop limit reached.\";\n return { reply, sessionName, toolEvents, meta: lastMeta };\n}\n\n\n\n\n\n\nasync function planStage(\n base: unknown[],\n meter: PassMeter,\n): Promise<string | null> {\n const resp = await meter.call([\n ...base,\n { role: \"system\", content: meter.p(\"plan\") },\n ], undefined, undefined, { expect: 400, maxTokens: 800 });\n return resp?.message?.content ?? null;\n}\n\n\n\n\n\n\n\n\n\nconst ENSEMBLE_RESAMPLE_ALLOWED = new Set([\"read_artifact\", \"search_artifact\", \"web_search\", \"memory_recall\"]);\nfunction readOnlyCtx(ctx: ToolCtx): ToolCtx {\n const tools = Object.fromEntries(Object.entries(ctx.tools).filter(([k]) => ENSEMBLE_RESAMPLE_ALLOWED.has(k)));\n const schemas = ctx.schemas.filter((s: any) => ENSEMBLE_RESAMPLE_ALLOWED.has(s?.function?.name));\n return { ...ctx, tools, schemas, turn: undefined }; \n}\n\n\nasync function ensembleStage(\n base: unknown[],\n ctx: ToolCtx,\n meter: PassMeter,\n samples: number,\n): Promise<{ reconciled: string; drafts: string[]; toolEvents: ToolEvent[] }> {\n const drafts: string[] = [];\n const allEvents: ToolEvent[] = [];\n for (let i = 0; i < samples && meter.remaining > 2; i++) {\n const r = await agentLoop([...base], i === 0 ? ctx : readOnlyCtx(ctx), meter);\n drafts.push(r.reply);\n allEvents.push(...r.toolEvents);\n }\n if (drafts.length <= 1) {\n return { reconciled: drafts[0] ?? \"\", drafts, toolEvents: allEvents };\n }\n const resp = await meter.call([\n { role: \"user\", content: meter.p(\"reconcile\", { drafts: drafts.map((d, i) => `DRAFT ${i + 1}:\\n${d}`).join(\"\\n\\n\") }) },\n ]);\n return {\n reconciled: resp?.message?.content ?? drafts[0],\n drafts,\n toolEvents: allEvents,\n };\n}\n\n\nasync function verifyStage(\n question: string,\n answer: string,\n meter: PassMeter,\n): Promise<{ answer: string; verdict: string }> {\n const critique = await meter.call(\n [{ role: \"user\", content: meter.p(\"critique\", { question, answer }) }],\n undefined,\n VERIFIER_MODEL,\n { expect: 200, maxTokens: 400 },\n );\n const verdict = (critique?.message?.content ?? \"PASS\").trim();\n if (/^PASS\\b/i.test(verdict) || !/FLAW/i.test(verdict)) {\n return { answer, verdict: \"PASS\" };\n }\n const revised = await meter.call([\n { role: \"user\", content: meter.p(\"revise\", { question, answer, flaw: verdict }) },\n ], undefined, undefined, { expect: Math.min(6000, Math.ceil(answer.length / 3) + 200) });\n const fixed = revised?.message?.content?.trim();\n return { answer: fixed && revised?.finishReason !== \"length\" ? fixed : answer, verdict }; \n}\n\n\n\nasync function regroundStage(\n base: unknown[],\n question: string,\n ctx: ToolCtx,\n meter: PassMeter,\n): Promise<AgentResult> {\n return await agentLoop(\n [\n ...base,\n { role: \"system\", content: meter.p(\"reground\", { question }) },\n ],\n ctx,\n meter,\n );\n}\n\n\n\n\ninterface Cognition {\n reply: string;\n sessionName?: string;\n toolEvents: ToolEvent[];\n meta: InferResult[\"meta\"];\n reasoning: string[];\n truncated: boolean; \n telemetry: {\n passesSpent: number;\n planned: boolean;\n ensembled: number;\n verifierVerdict: string;\n regrounded: boolean;\n confidence: number | null;\n stakes: string | null;\n abstained: boolean;\n };\n}\n\nasync function cognize(\n input: string,\n base: unknown[],\n ctx: ToolCtx,\n isFirstExchange: boolean,\n emit: (p: Progress) => void = () => {},\n model: string | null = null,\n settings: Settings = normalizeSettings({}),\n): Promise<Cognition> {\n const meter = new PassMeter(BUDGET.hardCeilingPasses, emit, model, settings);\n const { complex, maybeHighStakes } = preEstimate(input);\n\n const sys: any[] = [{ role: \"system\", content: meter.p(\"meta\", { open: META_OPEN, close: META_CLOSE }) }];\n if (isFirstExchange) sys.push({ role: \"system\", content: meter.p(\"firstExchange\") });\n const seed = [...base, ...sys];\n\n let planned = false;\n const lastUser = String(([...base].reverse().find((m: any) => m?.role === \"user\") as any)?.content ?? \"\");\n const appBuild = BUILD_RX.test(lastUser) && APP_RX.test(lastUser) && !settings.continuation && !settings.followup && !settings.focus;\n if (appBuild && settings.plan && meter.remaining > 6 && meter.timeLeft > 40_000) {\n \n \n meter.emit({ type: \"stage\", label: \"design brief\" });\n const brief = await meter.call([{ role: \"system\", content: \"You are a principal engineer and product designer. Output only the brief.\" }, { role: \"user\", content: meter.p(\"brief\", { ask: lastUser.slice(0, 4000) }) }], undefined, undefined, { expect: 1200, maxTokens: 1800, temperature: 0.4 })\n .then((r) => String(r?.message?.content ?? \"\").trim()).catch((e) => { meter.emit({ type: \"retry\", n: 0, error: `brief skipped: ${String(e?.message ?? e).slice(0, 160)}` }); return \"\"; });\n if (brief.length > 200) { seed.push({ role: \"system\", content: `DESIGN BRIEF (binding — implement every feature, file and acceptance check in it; the artifact is judged against it):\\n${brief}` }); planned = true; }\n } else if (settings.plan && !settings.continuation && !settings.followup && complex && meter.remaining > 6 && meter.timeLeft > 40_000) {\n meter.emit({ type: \"stage\", label: \"planning\" });\n const plan = await planStage(seed, meter).catch((e) => { meter.emit({ type: \"retry\", n: 0, error: `plan skipped: ${String(e?.message ?? e).slice(0, 160)}` }); return null; });\n if (plan) {\n seed.push({ role: \"system\", content: `Plan to follow:\\n${plan}` });\n planned = true;\n }\n }\n\n \n \n let reply: string;\n let sessionName: string | undefined;\n let toolEvents: ToolEvent[];\n let meta: InferResult[\"meta\"];\n let ensembled = 0;\n let truncated = false;\n\n if (settings.ensemble && !settings.continuation && !settings.followup && maybeHighStakes && meter.remaining > BUDGET.ensembleSamples + 2 && meter.timeLeft > 35_000) {\n meter.emit({ type: \"stage\", label: \"ensemble ×\" + BUDGET.ensembleSamples });\n const e = await ensembleStage(seed, ctx, meter, BUDGET.ensembleSamples);\n reply = e.reconciled;\n toolEvents = e.toolEvents;\n ensembled = e.drafts.length;\n meta = {};\n } else {\n meter.emit({ type: \"stage\", label: \"answering\" });\n try {\n const r = await agentLoop(seed, ctx, meter);\n truncated = !!r.truncated;\n reply = r.reply;\n sessionName = r.sessionName;\n toolEvents = r.toolEvents;\n meta = r.meta;\n } catch (e: any) {\n \n \n \n \n \n \n \n if (!e?.budget) throw e;\n meter.emit({ type: \"retry\", n: 0, error: \"turn budget exhausted before the first pass could run — saving progress; continuing in the next turn\" });\n truncated = true;\n reply = planned ? String(seed.find((m: any) => typeof m.content === \"string\" && m.content.startsWith(\"DESIGN BRIEF\"))?.content ?? \"\").replace(/^DESIGN BRIEF[^\\n]*\\n/, \"\") : \"\";\n sessionName = undefined;\n toolEvents = [];\n meta = {};\n }\n }\n\n \n let { visible, meta: ansMeta } = parseMeta(reply);\n reply = visible;\n\n \n if (!toolEvents.some((e) => e.tool === \"create_artifact\" && !e.error) && ctx.tools.create_artifact) {\n const doc = extractHtmlDocument(reply);\n if (doc) {\n const r = await saveArtifact(ctx.session, doc.title, \"html\", doc.html);\n if (r.ok) {\n toolEvents = [...toolEvents, { hop: -1, tool: \"create_artifact\", args: { title: r.value.title, kind: \"html\", auto: true }, result: { ok: true, id: r.value.id, url: `/artifact/${r.value.id}/`, title: r.value.title }, ts: Date.now() }];\n reply = reply.replace(/```(?:html|htm)?\\s*\\n[\\s\\S]*?\\n```/i, `_Saved as artifact **${r.value.title}** — open it above._`);\n }\n }\n }\n\n \n let regrounded = false;\n const lowConf = ansMeta ? ansMeta.confidence < 0.6 : false;\n const hiStakes = ansMeta ? ansMeta.stakes === \"high\" : maybeHighStakes;\n if ((ansMeta?.abstain || (hiStakes && lowConf)) && meter.remaining > 3 && meter.timeLeft > 20_000) {\n meter.emit({ type: \"stage\", label: \"regrounding\" });\n const rg = await regroundStage(base, input, ctx, meter).catch(() => null);\n if (rg) {\n const parsed = parseMeta(rg.reply);\n reply = parsed.visible || reply;\n ansMeta = parsed.meta ?? ansMeta;\n toolEvents = [...toolEvents, ...rg.toolEvents];\n regrounded = true;\n }\n }\n\n \n let verifierVerdict = \"SKIPPED\";\n const nonTrivial = complex || hiStakes || toolEvents.length > 0;\n \n const builtArtifact = toolEvents.some((e) => e.tool === \"create_artifact\" && !e.error);\n if (settings.verify && !settings.continuation && !truncated && !builtArtifact && nonTrivial && meter.remaining > 2 && reply.trim().length > 0 && meter.timeLeft > 25_000) {\n meter.emit({ type: \"stage\", label: \"verifying\" });\n const v = await verifyStage(input, reply, meter).catch((e) => { meter.emit({ type: \"retry\", n: 0, error: `verify skipped: ${String(e?.message ?? e).slice(0, 160)}` }); return { answer: reply, verdict: \"ERROR\" }; });\n reply = v.answer;\n verifierVerdict = v.verdict;\n }\n\n \n \n \n \n toolEvents = [...toolEvents, ...meter.routeEvents, ...meter.modelEvents].sort((a, b) => (a.ts ?? 0) - (b.ts ?? 0));\n return {\n reply,\n sessionName,\n toolEvents,\n meta,\n reasoning: meter.reasoning,\n truncated,\n telemetry: {\n passesSpent: meter.spent,\n planned,\n ensembled,\n verifierVerdict,\n regrounded,\n confidence: ansMeta?.confidence ?? null,\n stakes: ansMeta?.stakes ?? null,\n abstained: ansMeta?.abstain ?? false,\n },\n };\n}\n\n\n\n\nasync function runAI(\n input: string,\n session = \"default\",\n emit: (p: Progress) => void = () => {},\n model: string | null = null,\n settings: Settings = normalizeSettings({}),\n) {\n if (!rateLimit(session)) {\n return { reply: \"slow down\", toolEvents: [], telemetry: null };\n }\n const text = clean(input);\n const sess = await ensureSession(session);\n const lastRow = settings.retryOf ? unwrap(await one(\"message\", sql`SELECT * FROM message WHERE session = ${session} ORDER BY id DESC LIMIT 1`), null) : null;\n if (!(lastRow && lastRow.role === \"user\" && lastRow.content === text)) \n await addMsg(\"user\", text, null, session, settings.continuation ? { hidden: true, continuation: true } : settings.followup ? { hidden: true, followup: true } : null);\n await run(sql`UPDATE session SET ts = ${Date.now()} WHERE id = ${session}`); \n\n const blob = settings.memory ? await getMemoryBlob() : null; \n const recalled = settings.memory ? await retrieveFacts(text) : []; \n const context = await getContext(session, 24, settings.contextChars);\n const registry0 = await buildToolRegistry();\n const allow = settings.tools ? new Set(settings.tools) : null;\n const registry = allow\n ? { ...registry0, tools: Object.fromEntries(Object.entries(registry0.tools).filter(([k]) => allow.has(k) || k === \"name_session\")),\n schemas: Object.values(registry0.tools).filter((t: any) => allow.has(t.schema?.function?.name) || t.schema?.function?.name === \"name_session\").map((t: any) => t.schema) }\n : registry0;\n\n const memoryHeader = [\n blob?.value ? `blob:${blob.value}` : null,\n recalled.length ? `recalled:\\n- ${recalled.join(\"\\n- \")}` : null,\n ].filter(Boolean).join(\"\\n\\n\") || \"∅\";\n\n \n \n \n const hasRojs = registry.mcpToolNames.some((n) => n.startsWith(\"rojs_\"));\n const rojsHint = hasRojs\n ? `\\n\\nC#/Roslyn code-analysis tools (rojs_*) are connected via MCP. Workflow: rojs_create_session first (inline files[] or root_dir on the analysis host) — it returns a session_id every other rojs tool requires. Then: rojs_get_diagnostics (compile errors), rojs_hover, rojs_find_references, rojs_get_outline, rojs_context_bundle (token-budgeted context around a symbol — prefer it over pasting whole files), rojs_speculative_bind (check an expression compiles BEFORE proposing it), rojs_apply_edits with commit=false (dry-run diagnostic delta), rojs_rename_symbol. Sessions expire after idle timeout — recreate on NOT_FOUND. Positions are 0-based line/col.`\n : \"\";\n\n const toolNames = registry.schemas.map((t: any) => t?.function?.name).filter(Boolean).join(\", \") || \"none\";\n const skillsOn = activeSkills(settings.skills, text);\n const ctx0HasArtifactTools = !!registry.tools.update_artifact;\n if (skillsOn.length) emit({ type: \"stage\", label: \"skills: \" + skillsOn.map((k) => k.name).join(\", \") });\n const base = [\n {\n role: \"system\",\n content: renderPrompt(\"system\", {\n date: new Date().toISOString().slice(0, 10),\n tools: toolNames,\n skills: skillsOn.length ? \"\\n\" + skillsOn.map((k) => renderPrompt(\"skill\", { name: k.name, body: k.body }, settings.prompts, settings.examples)).join(\"\\n\\n\") + \"\\n\" : \"\",\n focus: settings.focus && ctx0HasArtifactTools ? \"\\n\" + renderPrompt(\"workbench\", {\n id: String(settings.focus.id), title: settings.focus.title, files: settings.focus.files.join(\", \") || \"index.html\", file: settings.focus.file || \"index.html\",\n selection: settings.focus.selection ? `\\nSelected text in ${settings.focus.file}:\\n\\`\\`\\`\\n${settings.focus.selection}\\n\\`\\`\\`` : \"\",\n errors: settings.focus.errors.length ? `\\nConsole (latest first):\\n${settings.focus.errors.slice().reverse().map((e) => \"- \" + e).join(\"\\n\")}` : \"\",\n }, settings.prompts, settings.examples) + \"\\n\" : \"\",\n memory: memoryHeader === \"∅\" ? \"\" : `\\nMemory (recalled facts about this user; trust as stated, do not repeat back):\\n${memoryHeader}\\n`,\n rojs: rojsHint ? `\\n${rojsHint.trim()}\\n` : \"\",\n user: settings.system ? `\\nUser instructions (override the doctrine where they conflict):\\n${settings.system}\\n` : \"\",\n }, settings.prompts, settings.examples),\n },\n ...context.map(({ role, content }) => ({ role, content })),\n ];\n\n \n \n \n const msgCount = unwrap(\n await raw<{ c: number }>(\n sql`SELECT COUNT(*) AS c FROM message WHERE session = ${session}`,\n ),\n [],\n );\n const isFirstExchange = (msgCount[0]?.c ?? 0) <= 2;\n\n \n \n let turnText = text;\n if (text.length > 18_000 && !settings.continuation) {\n emit({ type: \"stage\", label: `splitting ${Math.ceil(text.length / 12_000)} parts` });\n const ask = text.slice(0, 400).replace(/\\s+/g, \" \");\n const parts: string[] = []; for (let i = 0; i < text.length && parts.length < 8; i += 12_000) parts.push(text.slice(i, i + 12_000));\n const digests = await Promise.all(parts.map((part, i) => callInference(\n [{ role: \"system\", content: \"You compress source material for a later step. Output only the digest.\" }, { role: \"user\", content: renderPrompt(\"digest\", { ask, part: String(i + 1), n: String(parts.length), text: part }, settings.prompts, settings.examples) }],\n undefined, \"fast\", 25_000, { expect: 900, maxTokens: 1_200, router: settings.router, keyPolicy: settings.keyPolicy, vendorOrder: settings.vendorOrder },\n ).then((r) => (r.message?.content ?? \"\").trim()).catch((e) => `[part ${i + 1} could not be digested: ${String(e.message).slice(0, 80)}]`)));\n turnText = `${text.slice(0, 1_500)}\\n\\n[The message was ${text.length} chars; it was split into ${parts.length} parts and digested. Work from these digests:]\\n\\n${digests.map((d, i) => `--- part ${i + 1}/${parts.length} ---\\n${d}`).join(\"\\n\\n\")}`;\n emit({ type: \"stage\", label: \"digests ready\" });\n }\n\n const result = await cognize(\n turnText,\n base,\n { session, tools: registry.tools, schemas: registry.schemas, turn: {} },\n isFirstExchange,\n emit,\n model,\n settings,\n );\n \n let sessionName = result.sessionName;\n if (!sessionName && !settings.continuation && (sess.name === \"New Chat\" || !sess.name)) {\n sessionName = deriveTitle(text);\n await renameSession(session, sessionName);\n }\n await addMsg(\"assistant\", result.reply, null, session, {\n provider: result.meta?.provider ?? null, model: result.meta?.model ?? null, instance: result.meta?.instance ?? null, reasoning: result.reasoning ?? [], truncated: !!result.truncated,\n continuation: settings.continuation, followup: settings.followup, maxTokensSent: result.meta?.maxTokensSent ?? null, maxTokensAsked: result.meta?.maxTokensAsked ?? null, finishReason: result.meta?.finishReason ?? null, tokensOut: result.meta?.tokensOut ?? null,\n \n \n attempts: result.meta?.attempts?.length ?? null,\n telemetry: result.telemetry ?? null, toolEvents: slimEvents(result.toolEvents ?? []), skills: skillsOn.map((k) => k.name),\n });\n \n \n let completion: { done: boolean; remaining: string[] } | null = null;\n const worked = (result.toolEvents ?? []).some((e) => /artifact|eval_js|rojs_/.test(e.tool));\n if (!settings.continuation && (worked || BUILD_RX.test(text)) && !result.truncated) {\n try {\n const did = (result.toolEvents ?? []).map((e) => `${e.tool}${e.error ? \" (failed)\" : \"\"}: ${JSON.stringify(e.args).slice(0, 160)}`).join(\"\\n\") || \"(no tools)\";\n const aid = [...(result.toolEvents ?? [])].reverse().map((e) => Number((e.result as any)?.id)).find((n) => Number.isInteger(n)) ?? Number(settings.focus?.id);\n let quality = \"(no artifact this turn)\";\n if (Number.isInteger(aid) && APP_RX.test(text)) { const lf = await listFiles(aid); if (lf) { const q = qualityReport(lf.files); quality = q.length ? `artifact #${aid}: ${lf.files.length} files\\n` + q.map((x) => \"- \" + x).join(\"\\n\") : `artifact #${aid} clears the rubric`; } }\n const r = await callInference(\n [{ role: \"system\", content: \"You audit whether a task is finished. Output only JSON.\" }, { role: \"user\", content: renderPrompt(\"completion\", { ask: text.slice(0, 3000), did, reply: (result.reply ?? \"\").slice(0, 3000), quality }, settings.prompts, settings.examples) }],\n undefined, \"fast\", 12_000, { expect: 300, maxTokens: 500, router: settings.router, keyPolicy: settings.keyPolicy, vendorOrder: settings.vendorOrder },\n );\n const raw = String(r.message?.content ?? \"\"); const j = JSON.parse(raw.slice(raw.indexOf(\"{\"), raw.lastIndexOf(\"}\") + 1));\n completion = { done: j.done !== false && !(Number(j.quality) > 0 && Number(j.quality) < 4), remaining: Array.isArray(j.remaining) ? j.remaining.filter((x: unknown) => typeof x === \"string\").slice(0, 8) : [] };\n if (completion.remaining.length === 0) completion.done = true;\n } catch { completion = null; }\n }\n \n \n let needsCompact = false;\n try { const c = await loadCompact(session); const live = await liveChars(session, c?.uptoId ?? 0); needsCompact = live.count > 8 && live.chars > 2 * settings.contextChars; } catch { /* hint only */ }\n return {\n reply: result.reply,\n sessionName,\n toolEvents: result.toolEvents,\n meta: result.meta,\n reasoning: result.reasoning,\n truncated: result.truncated,\n telemetry: result.telemetry,\n completion,\n needsCompact,\n };\n}\n\n\n\n\nexport default async function handler(req: Request): Promise<Response> {\n const url = new URL(req.url);\n const method = req.method;\n const init = await initDB();\n if (!init.ok) {\n return new Response(`DB init failed: ${init.error}`, { status: 500 });\n }\n\n \n const am = url.pathname.match(/^\\/artifact\\/(\\d+)(?:\\/(.*))?$/);\n if (am) {\n const id = Number(am[1]);\n if (am[2] === undefined) return new Response(null, { status: 302, headers: { Location: `/artifact/${id}/` } });\n if (method === \"GET\") {\n const f = await readFile(id, am[2] || \"\");\n if (!f) return new Response(\"not found\", { status: 404 });\n const path = am[2] || (f.kind === \"html\" ? \"index.html\" : f.kind === \"svg\" ? \"index.svg\" : \"index.md\");\n return serveFile(url.origin, path, f.content);\n }\n }\n const compacted = await handleCompact(req, url, (messages, settings) => callInference(messages, undefined, \"fast\", 45_000, { expect: 1400, maxTokens: 2000, temperature: 0.2, router: settings.router, keyPolicy: settings.keyPolicy, vendorOrder: settings.vendorOrder }).then((r) => String(r.message?.content ?? \"\")));\n if (compacted) return compacted;\n if (method === \"GET\" && url.searchParams.has(\"artifact_files\")) {\n const r = await listFiles(Number(url.searchParams.get(\"artifact_files\")));\n if (!r) return Response.json({ error: \"not found\" }, { status: 404 });\n return Response.json({ id: r.artifact.id, title: r.artifact.title, kind: r.artifact.kind, session: r.artifact.session, files: r.files, issues: lintArtifact(r.files) });\n }\n const swarm = await handleSwarm(req, url, { PassMeter, deriveTitle, clean, ensureSession });\n if (swarm) return swarm;\n const issuesRes = await handleIssues(req, url);\n if (issuesRes) return issuesRes;\n if (method === \"POST\" && url.searchParams.has(\"artifact_save_many\")) {\n const b = await req.json().catch(() => ({}));\n const id = Number.isInteger(b.id) ? b.id : null;\n if (id) { const a = await listFiles(id); if (!a) return Response.json({ error: \"not found\" }, { status: 404 }); }\n const r = await saveMany(String(b.session ?? \"default\"), id, String(b.title ?? \"\"), Array.isArray(b.files) ? b.files : []);\n return r ? Response.json({ ok: true, ...r, url: `/artifact/${r.id}/` }) : Response.json({ error: \"save failed\" }, { status: 500 });\n }\n if (method === \"POST\" && url.searchParams.has(\"artifact_delete_file\")) {\n const b = await req.json().catch(() => ({}));\n const w = await deleteFile(Number(b.id), String(b.path ?? \"\"), b.dir === true);\n return w.ok ? Response.json({ ok: true }) : Response.json({ error: \"cannot delete\" }, { status: 400 });\n }\n if (method === \"POST\" && url.searchParams.has(\"artifact_save\")) {\n const b = await req.json().catch(() => ({}));\n const id = Number(b.id), path = String(b.path ?? \"\"), content = typeof b.content === \"string\" ? b.content : null;\n if (!id || !path || content === null) return Response.json({ error: \"id, path, content required\" }, { status: 400 });\n const w = await writeFile(id, path, content);\n return w.ok ? Response.json({ ok: true, url: `/artifact/${id}/` }) : Response.json({ error: w.error }, { status: 500 });\n }\n\n if (\n method === \"GET\" &&\n !url.searchParams.has(\"q\") &&\n !url.searchParams.has(\"sessions\") &&\n !url.searchParams.has(\"session\") &&\n !url.searchParams.has(\"history\") &&\n !url.searchParams.has(\"artifact\") &&\n !url.searchParams.has(\"progress\") &&\n !url.searchParams.has(\"prompts\")\n ) {\n return new Response(PAGE_HTML, {\n headers: { \"content-type\": \"text/html\" },\n });\n }\n\n if (method === \"GET\" && url.searchParams.has(\"sessions\")) {\n return Response.json(await listSessions());\n }\n\n if (method === \"GET\" && url.searchParams.has(\"history\")) {\n return Response.json(await getHistory(url.searchParams.get(\"history\")!));\n }\n\n if (method === \"GET\" && url.searchParams.has(\"artifacts\")) {\n const session = url.searchParams.get(\"session\") ?? \"default\";\n const rows = unwrap(\n await all(\n \"artifact\",\n sql`SELECT * FROM artifact WHERE session = ${session} ORDER BY ts DESC`,\n ),\n [],\n );\n return Response.json(\n rows.map(({ id, title, kind, ts }) => ({ id, title, kind, ts })),\n );\n }\n\n if (method === \"GET\" && url.searchParams.has(\"artifact\")) {\n const id = Number(url.searchParams.get(\"artifact\"));\n if (!Number.isInteger(id)) return new Response(\"bad id\", { status: 400 });\n return new Response(null, { status: 302, headers: { Location: `/artifact/${id}/` } });\n }\n\n if (method === \"GET\" && url.searchParams.has(\"q\")) {\n const q = url.searchParams.get(\"q\")!;\n const session = url.searchParams.get(\"session\") ?? \"default\";\n try {\n const { reply, sessionName, toolEvents, meta, telemetry } = await runAI(q, session, undefined, url.searchParams.get(\"model\"));\n return Response.json({ reply, sessionName, toolEvents, meta, telemetry });\n } catch (err) {\n console.error(\"q failed:\", q.slice(0, 80), err);\n return Response.json({ error: String(err) }, { status: 500 });\n }\n }\n\n \n \n \n \n if (method === \"POST\" && url.searchParams.has(\"q\")) {\n const body = await req.json().catch(() => ({}));\n const q = String(body.q ?? \"\");\n const session = String(body.session ?? \"default\");\n const settings = normalizeSettings({ ...(body.settings ?? {}), model: body.settings?.model ?? body.model });\n \n const model = settings.model ?? (BUILD_RX.test(q) && APP_RX.test(q) && !body.settings?.continuation ? \"coder\" : null);\n if (!q.trim()) return Response.json({ error: { message: \"empty prompt\" } }, { status: 400 });\n \n \n const log: Progress[] = [];\n let chain = Promise.resolve();\n const persist = () => { chain = chain.then(() => run(sql`INSERT INTO progress (session, events, ts) VALUES (${session}, ${JSON.stringify(log)}, ${Date.now()})\n ON CONFLICT(session) DO UPDATE SET events = excluded.events, ts = excluded.ts`).then(() => {})).catch(() => {}); }; \n const emit = (p: Progress) => { log.push(p); persist(); };\n emit({ type: \"stage\", label: \"starting\" });\n try {\n const r = await runAI(q, session, emit, model, settings);\n await chain;\n return Response.json(r);\n } catch (err: any) {\n console.error(\"q(post) failed:\", q.slice(0, 80), err);\n const msg = String(err?.message ?? err);\n const [head, ...rest] = msg.split(\"\\n\");\n return Response.json({ error: { message: head, details: rest } }, { status: err?.status && err.status >= 400 && err.status < 600 ? err.status : 500 });\n } finally {\n emit({ type: \"stage\", label: \"done\" });\n await chain.catch(() => {});\n }\n }\n\n if (method === \"GET\" && url.searchParams.has(\"prompts\")) {\n return Response.json({ defaults: DEFAULT_PROMPTS, vars: PROMPT_VARS });\n }\n\n if (method === \"GET\" && url.searchParams.has(\"progress\")) {\n const r = await one<\"progress\">(\"progress\", sql`SELECT * FROM progress WHERE session = ${url.searchParams.get(\"progress\")!}`);\n const events: Progress[] = r.ok && r.value ? JSON.parse(r.value.events) : [];\n const since = Number(url.searchParams.get(\"since\") ?? 0) || 0;\n return Response.json({ events: events.slice(since), next: events.length });\n }\n\n if (method === \"POST\" && url.searchParams.has(\"rename\")) {\n const { id, name } = await req.json();\n await renameSession(id, name);\n return Response.json({ ok: true });\n }\n\n if (method === \"POST\" && url.searchParams.has(\"clear\")) {\n const { id } = await req.json();\n await clearSession(id);\n return Response.json({ ok: true });\n }\n\n if (method === \"DELETE\" && url.searchParams.has(\"session\")) {\n await deleteSession(url.searchParams.get(\"session\")!);\n return Response.json({ ok: true });\n }\n\n return new Response(\"not found\", { status: 404 });\n}",[],[533],"6e9f809b-09b4-48fd-ae32-542656a6a62c","app-swarm.ts","01a0a661-f049-7779-b407-9727b8815326",["D",1789497831502],"// Swarm build endpoints (?plan, ?build_file): one user action → many val invocations, each under its\n// own 60 s. Dependencies from app.tsx are injected to avoid an import cycle.\nimport { renderPrompt } from \"./prompts.ts\";\nimport { lintArtifact, saveMany } from \"./artifacts.ts\";\nimport { activeSkills, normalizeSettings } from \"./app-settings.ts\";\nimport { all, run, sql, unwrap } from \"./db.ts\";\nexport type SwarmDeps = {\n PassMeter: any; // class from app.tsx: new PassMeter(ceiling, emit, model, settings) with .p(key, vars) and .call(messages, tools, model, override)\n deriveTitle: (text: string) => string;\n clean: (s: string) => string;\n ensureSession: (id: string) => Promise<unknown>;\n};\n/** Returns a Response for ?plan / ?build_file, or null when the request is not a swarm route. */\nexport async function handleSwarm(req: Request, url: URL, d: SwarmDeps): Promise<Response | null> {\n const method = req.method;\n const { PassMeter, deriveTitle, clean, ensureSession } = d;\n // ── swarm build: one user action → many val invocations, each under its own 60 s ──\n // POST ?plan {q, session, settings} → manifest + an artifact shell (index placeholder + manifest.json) builders write into.\n if (method === \"POST\" && url.searchParams.has(\"plan\")) {\n const b = await req.json().catch(() => ({}));\n const settings = normalizeSettings(b.settings ?? {});\n const session = String(b.session ?? \"default\"), ask = clean(String(b.q ?? \"\"));\n if (!ask) return Response.json({ error: { message: \"empty ask\" } }, { status: 400 });\n await ensureSession(session);\n const meter = new PassMeter(3, () => {}, settings.model, settings);\n const skillsOn = activeSkills(settings.skills, ask);\n const prompt = meter.p(\"architect\", { ask, user: settings.system ? `\\nUser instructions: ${settings.system}\\n` : \"\", skills: skillsOn.length ? \"\\n\" + skillsOn.map((k) => renderPrompt(\"skill\", { name: k.name, body: k.body }, settings.prompts)).join(\"\\n\") : \"\" });\n try {\n const r = await meter.call([{ role: \"system\", content: \"Output only JSON.\" }, { role: \"user\", content: prompt }], undefined, settings.model ?? \"coder\", { expect: 3000, temperature: 0.2 });\n const raw = String(r?.message?.content ?? \"\");\n const j = raw.slice(raw.indexOf(\"{\"), raw.lastIndexOf(\"}\") + 1);\n const manifest = JSON.parse(j);\n if (!Array.isArray(manifest.files) || !manifest.files.length) throw new Error(\"manifest has no files\");\n manifest.features = [].concat(manifest.features ?? []).map(String).slice(0, 20); manifest.visual = String(manifest.visual ?? \"\");\n manifest.files = manifest.files.filter((f: any) => f && typeof f.path === \"string\").map((f: any) => ({ path: f.path.replace(/^\\.\\//, \"\"), purpose: String(f.purpose ?? \"\"), exports: [].concat(f.exports ?? []).map(String), imports: [].concat(f.imports ?? []).map(String), notes: String(f.notes ?? \"\") }));\n if (!manifest.files.some((f: any) => /^readme\\.md$/i.test(f.path))) manifest.files.push({ path: \"README.md\", purpose: \"what it is, every feature, controls/shortcuts, how to run (static host or open index.html), file map\", exports: [], imports: [], notes: \"markdown; ≥ 10 feature bullets\" });\n if (!manifest.files.some((f: any) => /^index\\.html?$/.test(f.path))) manifest.files.unshift({ path: \"index.html\", purpose: \"entry: markup + boot; loads every sibling\", exports: [], imports: manifest.files.map((f: any) => f.path), notes: \"\" });\n const title = String(manifest.title ?? deriveTitle(ask)).slice(0, 60);\n const shell = await saveMany(session, null, title, [{ path: \"index.html\", content: `<!doctype html><html><head><meta charset=\"utf-8\"><title>${title}

building ${manifest.files.length} files…

` }, { path: \"manifest.json\", content: JSON.stringify({ ...manifest, ask }, null, 2) }]);\n // Durable plan + per-file registry (build_job/build_file — see db.ts) so \"what files exist and what state\n // are they in\" has one queryable answer instead of living only in the client's in-memory manifest card and\n // an ephemeral, delete-on-completion omni_state blob. One row per manifest file, all starting \"queued\".\n if (shell?.id) {\n const now = Date.now();\n await run(sql`INSERT INTO build_job (artifact_id, session, title, ask, manifest, status, created_ts, updated_ts)\n VALUES (${shell.id}, ${session}, ${title}, ${ask.slice(0, 2000)}, ${JSON.stringify(manifest).slice(0, 60_000)}, 'planned', ${now}, ${now})\n ON CONFLICT(artifact_id) DO UPDATE SET manifest = excluded.manifest, title = excluded.title, ask = excluded.ask, status = 'planned', updated_ts = excluded.updated_ts`);\n for (const f of manifest.files) {\n await run(sql`INSERT INTO build_file (artifact_id, path, purpose, status, updated_ts) VALUES (${shell.id}, ${f.path}, ${String(f.purpose ?? \"\").slice(0, 500)}, 'queued', ${now})\n ON CONFLICT(artifact_id, path) DO UPDATE SET purpose = excluded.purpose, status = 'queued', updated_ts = excluded.updated_ts`);\n }\n }\n return Response.json({ manifest, artifactId: shell?.id ?? null, title, meta: r?.meta ?? null });\n } catch (e: any) {\n return Response.json({ error: { message: `plan failed: ${String(e?.message ?? e).slice(0, 300)}` } }, { status: 502 });\n }\n }\n // POST ?build_file {session, artifactId, manifest, path, settings} → writes that one file (auto-continued), returns lint issues.\n if (method === \"POST\" && url.searchParams.has(\"build_file\")) {\n const b = await req.json().catch(() => ({}));\n const settings = normalizeSettings(b.settings ?? {});\n const session = String(b.session ?? \"default\"), id = Number(b.artifactId), manifest = b.manifest, path = String(b.path ?? \"\");\n const spec = manifest?.files?.find((f: any) => f.path === path);\n if (!Number.isInteger(id) || !spec) return Response.json({ error: { message: \"bad artifactId/path\" } }, { status: 400 });\n const meter = new PassMeter(6, () => {}, settings.model, settings);\n const compact = manifest.files.map((f: any) => `${f.path}: ${f.purpose}${f.exports?.length ? ` — exports ${f.exports.join(\", \")}` : \"\"}`).join(\"\\n\");\n const prompt = meter.p(\"builder\", { title: String(manifest.title ?? \"\"), path, purpose: spec.purpose, exports: spec.exports?.join(\", \") || \"(none)\", imports: spec.imports?.join(\", \") || \"(none)\", shared: String(manifest.shared ?? \"\"), manifest: compact, notes: spec.notes || \"(none)\", features: [].concat(manifest.features ?? []).map((x: unknown, i: number) => `${i + 1}. ${String(x)}`).join(\"\\n\") || \"(see purpose)\", visual: String(manifest.visual ?? \"(designer's choice: tokens, dark default, real identity)\") });\n try {\n const r = await meter.call([{ role: \"system\", content: \"You write one complete source file. Output only a fenced code block.\" }, { role: \"user\", content: prompt }], undefined, settings.model ?? \"coder\", { expect: 6000, temperature: 0.2 });\n const raw = String(r?.message?.content ?? \"\");\n const fence = raw.match(/```[\\w.+-]*\\s*\\n([\\s\\S]*?)\\n```\\s*$/m) ?? raw.match(/```[\\w.+-]*\\s*\\n([\\s\\S]*)$/m);\n const content = (fence ? fence[1] : raw).replace(/\\n```\\s*$/, \"\");\n if (!content.trim()) throw new Error(\"empty file\");\n const w = await saveMany(session, id, \"\", [{ path, content }]);\n // builders run in parallel, so cross-file checks are meaningless here: report only this file's own defects (syntax, JSON, blocked externals)\n const own = lintArtifact([{ path, content }]).filter((i) => !/does not exist|imports \\{/.test(i));\n const truncated = r?.finishReason === \"length\";\n // Persist this attempt's outcome into build_file — the durable per-file record — regardless of how the\n // manifest card is currently rendered client-side. ON CONFLICT because a file rebuilt via the manifest's\n // single-file \"rebuild\" button, or a job resumed after this row's artifact_id/path pair already exists,\n // both hit the same (artifact_id, path) primary key as the original plan-time insert.\n await run(sql`INSERT INTO build_file (artifact_id, path, purpose, status, lines, issues, error, attempts, updated_ts)\n VALUES (${id}, ${path}, ${String(spec.purpose ?? \"\").slice(0, 500)}, ${truncated || own.length ? \"issues\" : \"ok\"}, ${content.split(\"\\n\").length}, ${JSON.stringify(own).slice(0, 4000)}, NULL, 1, ${Date.now()})\n ON CONFLICT(artifact_id, path) DO UPDATE SET status = excluded.status, lines = excluded.lines, issues = excluded.issues, error = NULL, attempts = build_file.attempts + 1, updated_ts = excluded.updated_ts`).catch(() => {});\n return Response.json({ ok: true, path, chars: content.length, lines: content.split(\"\\n\").length, truncated, issues: own, written: !!w, meta: r?.meta ?? null });\n } catch (e: any) {\n const errMsg = `build ${path} failed: ${String(e?.message ?? e).slice(0, 300)}`;\n await run(sql`INSERT INTO build_file (artifact_id, path, status, error, attempts, updated_ts) VALUES (${id}, ${path}, 'error', ${errMsg}, 1, ${Date.now()})\n ON CONFLICT(artifact_id, path) DO UPDATE SET status = 'error', error = excluded.error, attempts = build_file.attempts + 1, updated_ts = excluded.updated_ts`).catch(() => {});\n return Response.json({ error: { message: errMsg }, path }, { status: 502 });\n }\n }\n // ── build jobs: the orchestration state of a swarm build, persisted so a closed tab / dead network / reload can resume it ──\n // GET ?job=list&session=S → unfinished jobs; POST ?job {session, artifactId, state} → update status (manifest/files already\n // live in build_job/build_file — see ?plan and ?build_file above — so this no longer needs to carry done/failed arrays).\n //\n // Previously this was a single opaque JSON blob per job in omni_state (key \"job::\"), DELETED the\n // instant a build finished — so build_job/build_file (db.ts) exist specifically so \"what got built, from what plan, in\n // what state\" has a durable, queryable, joinable answer that survives past \"done\" instead of vanishing on completion.\n if (url.searchParams.has(\"job\")) {\n if (method === \"GET\") {\n const session = String(url.searchParams.get(\"session\") ?? \"\");\n const jobs = unwrap(await all(\"build_job\", sql`SELECT * FROM build_job WHERE session = ${session} AND status != 'done' ORDER BY updated_ts DESC LIMIT 20`), []);\n const out = [];\n for (const j of jobs) {\n let manifest: any = null, manifestError: string | null = null;\n try { manifest = JSON.parse(j.manifest); } catch (e: any) { manifestError = `stored manifest is corrupt JSON: ${String(e?.message ?? e).slice(0, 200)}`; }\n // Previously a corrupt row was silently `continue`d past — the job vanished from the list with no\n // trace, so a client polling ?job=list had no way to know a build existed at all, let alone that it\n // needed attention. Surface it instead: the client can render an error state and offer to clear it\n // (e.g. by marking the job 'stopped' via ?job POST) rather than the build silently disappearing.\n if (!manifest) { out.push({ artifactId: j.artifact_id, manifest: null, manifestError, title: j.title, ask: j.ask, status: j.status, ts: j.updated_ts, done: [], failed: [] }); continue; }\n const files = unwrap(await all(\"build_file\", sql`SELECT path, status FROM build_file WHERE artifact_id = ${j.artifact_id}`), []);\n out.push({\n artifactId: j.artifact_id, manifest, title: j.title, ask: j.ask, status: j.status, ts: j.updated_ts,\n done: files.filter((f: any) => f.status === \"ok\").map((f: any) => f.path),\n failed: files.filter((f: any) => f.status === \"error\" || f.status === \"issues\").map((f: any) => f.path),\n });\n }\n return Response.json({ jobs: out });\n }\n if (method === \"POST\") {\n const b = await req.json().catch(() => ({}));\n const session = String(b.session ?? \"\"), id = Number(b.artifactId), state = b.state ?? {};\n if (!session || !Number.isInteger(id)) return Response.json({ error: { message: \"bad session/artifactId\" } }, { status: 400 });\n // Previously any client-supplied string landed straight in build_job.status with no validation — a typo\n // or a future caller sending e.g. \"complete\" instead of \"done\" would silently create a status value\n // that every reader (checkResume's `status !== 'done'` filter, any future dashboard) doesn't recognize,\n // rather than failing loudly at the one place that writes it.\n const BUILD_STATUSES = new Set([\"planned\", \"building\", \"integrating\", \"done\", \"stopped\"]);\n const status = String(state.status ?? \"building\");\n if (!BUILD_STATUSES.has(status)) return Response.json({ error: { message: `bad status \"${status}\" — must be one of ${[...BUILD_STATUSES].join(\", \")}` } }, { status: 400 });\n const title = String(state.title ?? \"\"), ask = String(state.ask ?? \"\").slice(0, 2000);\n await run(sql`UPDATE build_job SET status = ${status}, updated_ts = ${Date.now()},\n title = CASE WHEN ${title} = '' THEN title ELSE ${title} END,\n ask = CASE WHEN ${ask} = '' THEN ask ELSE ${ask} END\n WHERE artifact_id = ${id}`);\n return Response.json({ ok: true });\n }\n }\n return null;\n}\n",[],[540],"01ceae97-9d7f-464c-81dc-93acfabd6e8d","app-settings.ts","01a0a661-f049-7779-b407-919c4c468977",["D",1789497831502],"// Per-request settings: shape, validation and the skill matcher. Split out of app.tsx\n// to keep each file under Val Town's 80 kB file-API limit.\nimport { normalizePromptOverrides, type PromptKey } from \"./prompts.ts\";\nexport interface Settings {\n model: string | null; // router alias/pin; null = auto\n reasoning: \"low\" | \"medium\" | \"high\" | null;\n temperature: number | null;\n maxTokens: number | null; // null = size to task (respecting vendor limits); number = hard preference\n system: string; // custom instructions appended to the system prompt\n tools: string[] | null; // enabled tool names; null = all\n plan: boolean; verify: boolean; ensemble: boolean; memory: boolean;\n continueMax: number; // auto-continue passes when a reply is cut at max_tokens\n budgetMs: number; // wall-clock budget for the turn (free plan hard cap is 60 s)\n prompts: Partial>; // per-prompt overrides; see prompts.ts\n examples: boolean; // append the two exemplars per prompt (prompt-examples.ts); off saves ~1–3k input tokens per call\n focus: Focus | null; // artifact the workbench has open when the message was sent\n skills: Skill[]; // user-defined instruction packs; `always` or keyword-triggered `auto`\n router: { url: string; key: string | null } | null; // inference router override (allowlisted hosts only)\n maxMode: \"auto\" | \"fixed\" | \"max\"; // auto = size to task; fixed = maxTokens; max = ask for the vendor ceiling\n continuation: boolean; // this turn resumes a reply cut at the output limit: skip plan/verify/ensemble, hide the user turn\n expect: \"auto\" | \"short\" | \"long\" | \"max\"; // expected reply length → omni_expect_tokens and max_tokens sizing\n keyPolicy: \"depth\" | \"rr\" | \"breadth\" | null; // router key rotation for this request (null = router default)\n contextChars: number; // history budget sent to the model (chars); small keeps free-tier ITPM limits reachable\n retryOf: boolean; // client re-sent the same ask after a 503: do not insert the user message again\n vendorOrder: string[]; // vendors to try first, in order (every key slot of one before the next); [] = router default\n followup: boolean; // auto-continuation of unfinished work (client-issued): hide the synthetic user turn, skip plan/ensemble\n}\nconst ROUTER_ALLOW = (Deno.env.get(\"OMNI_ROUTER_ALLOW\") ?? \"\").split(\",\").map((s) => s.trim()).filter(Boolean);\nexport function normalizeRouter(x: any): Settings[\"router\"] {\n if (!x || typeof x !== \"object\" || typeof x.url !== \"string\") return null;\n try {\n const u = new URL(x.url);\n if (u.protocol !== \"https:\" || !(u.hostname.endsWith(\".val.run\") || ROUTER_ALLOW.includes(u.hostname))) return null;\n return { url: u.origin + (u.pathname === \"/\" ? \"\" : u.pathname.replace(/\\/$/, \"\")), key: typeof x.key === \"string\" && x.key ? x.key.slice(0, 200) : null };\n } catch { return null; }\n}\nexport interface Focus { id: number; title: string; file: string; files: string[]; selection: string; errors: string[] }\nexport interface Skill { name: string; when: string; body: string; mode: \"always\" | \"auto\" }\nexport function normalizeFocus(x: any): Focus | null {\n if (!x || typeof x !== \"object\" || !Number.isInteger(x.id)) return null;\n const str = (v: unknown, n: number) => (typeof v === \"string\" ? v.slice(0, n) : \"\");\n return { id: x.id, title: str(x.title, 120), file: str(x.file, 200), files: Array.isArray(x.files) ? x.files.filter((f: unknown) => typeof f === \"string\").slice(0, 40) : [],\n selection: str(x.selection, 4000), errors: Array.isArray(x.errors) ? x.errors.filter((e: unknown) => typeof e === \"string\").slice(-8).map((e: string) => e.slice(0, 400)) : [] };\n}\nexport function normalizeSkills(x: any): Skill[] {\n if (!Array.isArray(x)) return [];\n return x.filter((k) => k && typeof k === \"object\" && typeof k.name === \"string\" && typeof k.body === \"string\" && k.body.trim())\n .slice(0, 12).map((k) => ({ name: k.name.slice(0, 60), when: typeof k.when === \"string\" ? k.when.slice(0, 300) : \"\", body: k.body.slice(0, 6000), mode: k.mode === \"always\" ? \"always\" : \"auto\" }));\n}\n/** Skills that apply to this message: `always`, or `auto` whose `when` (comma/pipe-separated keywords or /regex/) matches the text. */\nexport function activeSkills(skills: Skill[], text: string): Skill[] {\n const t = text.toLowerCase();\n return skills.filter((k) => {\n if (k.mode === \"always\") return true;\n const w = k.when.trim(); if (!w) return false;\n const m = w.match(/^\\/(.+)\\/([a-z]*)$/);\n if (m) { try { return new RegExp(m[1], m[2].replace(/g/g, \"\") || \"i\").test(text); } catch { return false; } }\n return w.split(/[,|]/).map((s) => s.trim().toLowerCase()).filter(Boolean).some((kw) => t.includes(kw));\n });\n}\nexport function normalizeSettings(x: any): Settings {\n const num = (v: unknown, lo: number, hi: number, d: number | null) => (typeof v === \"number\" && isFinite(v)) ? Math.min(hi, Math.max(lo, v)) : d;\n return {\n model: typeof x?.model === \"string\" && x.model && x.model !== \"auto\" ? x.model : null,\n reasoning: [\"low\", \"medium\", \"high\"].includes(x?.reasoning) ? x.reasoning : null,\n temperature: num(x?.temperature, 0, 2, null),\n maxTokens: num(x?.maxTokens, 256, 65_536, null),\n system: typeof x?.system === \"string\" ? x.system.slice(0, 8000) : \"\",\n tools: Array.isArray(x?.tools) ? x.tools.filter((t: unknown) => typeof t === \"string\") : null,\n plan: x?.plan !== false, verify: x?.verify !== false, ensemble: x?.ensemble !== false, memory: x?.memory !== false,\n continueMax: num(x?.continueMax, 0, 8, 3) as number,\n budgetMs: num(x?.budgetMs, 15_000, 52_000, 52_000) as number,\n prompts: normalizePromptOverrides(x?.prompts),\n examples: x?.examples !== false,\n focus: normalizeFocus(x?.focus),\n skills: normalizeSkills(x?.skills),\n router: normalizeRouter(x?.router),\n maxMode: x?.maxMode === \"fixed\" ? \"fixed\" : x?.maxMode === \"max\" ? \"max\" : \"auto\",\n continuation: x?.continuation === true,\n expect: [\"short\", \"long\", \"max\"].includes(x?.expect) ? x.expect : \"auto\",\n keyPolicy: [\"depth\", \"rr\", \"breadth\"].includes(x?.keyPolicy) ? x.keyPolicy : null,\n contextChars: num(x?.contextChars, 2_000, 120_000, 16_000) as number,\n retryOf: x?.retryOf === true,\n followup: x?.followup === true,\n vendorOrder: typeof x?.vendorOrder === \"string\" ? x.vendorOrder.split(\",\").map((s: string) => s.trim().toLowerCase()).filter(Boolean).slice(0, 12) : [],\n };\n}\n",[],[547],"8e11e8ac-b318-46c1-9958-1b20726708d1","app-infer.ts","01a0a661-f049-7779-b407-8e81b847f3d5",["D",1789497831502],"// Inference client (moved from app.tsx for the 80k-char limit): in-process router call, request shaping, size-bound retries.\nimport routerHandler from \"./router.tsx\";\nimport type { Settings } from \"./app-settings.ts\";\nimport { normalizeMessages, shrinkMessages } from \"./app-helpers.ts\";\n\nexport const INFERENCE_URL = \"https://router.val.run\";\n// The router is a module of this val. Calling it in-process removes the val→val HTTP hop entirely: Cloudflare in front of\n// *.val.run blocks server-side fetches from Val Town's own egress (\"Blocked\", server: cloudflare) regardless of volume,\n// and every hop was also a second 60 s wall clock and a second invocation. HTTP is used only for a custom router URL.\nexport const routerCall = (url: string, init: RequestInit): Promise =>\n url.replace(/\\/$/, \"\") === INFERENCE_URL || !url ? routerHandler(new Request(INFERENCE_URL + \"/v1/chat/completions\", init)) : fetch(url, init);\n\n// ════════════════════════════════════════════════════════════════════════════\n// INFERENCE (accepts optional model override for the verifier pass)\n// ════════════════════════════════════════════════════════════════════════════\nexport interface InferResult {\n message: any;\n finishReason: string;\n reasoning?: string; // model \"thinking\" (gpt-oss `reasoning`, qwen/deepseek `reasoning_content`)\n usage?: { prompt_tokens?: number; completion_tokens?: number };\n meta: { provider?: unknown; model?: string; latencyMs?: number; retries?: number; instance?: string; maxTokensSent?: number | null; maxTokensAsked?: number; finishReason?: string; tokensOut?: number | null; tokensIn?: number | null;\n // Routing flowchart material, straight from router-core.ts's trail (via router-openai.ts's _omni_meta) —\n // plan is the intended provider order, attempts is what actually happened (tried/skipped, ok/failed, timing).\n plan?: string[]; attempts?: { provider: string; vendor: string; model: string; pass: number; timeoutMs: number; ok: boolean; skipped?: boolean; demoted?: boolean; latencyMs?: number; error?: string; ts: number }[] };\n}\nexport interface GenOpts { reasoning?: string | null; temperature?: number | null; maxTokens?: number | null; expect?: number; router?: Settings[\"router\"]; maxMode?: Settings[\"maxMode\"]; keyPolicy?: Settings[\"keyPolicy\"]; vendorOrder?: string[]; exclude?: string[] }\n// Long-output requests (code, apps, documents) need throughput, not just quality:\n// the router ranks by quality among vendors whose eta fits the deadline.\nexport const BUILD_RX = /\\b(create|build|write|implement|generate|make|design|code|app|renderer|shader|script|component|page|game|simulat|program|class|function|refactor|port|convert)\\b/i;\nexport const expectTokens = (messages: unknown[]) => {\n const last = [...messages].reverse().find((m: any) => m?.role === \"user\") as any;\n return BUILD_RX.test(String(last?.content ?? \"\")) ? 6000 : 1500;\n};\nexport async function callInference(\n messages: unknown[],\n tools?: unknown[],\n model?: string | null,\n deadlineMs?: number,\n gen: GenOpts = {},\n): Promise {\n // Ask for the most output the task can use; the router shrinks max_tokens to a\n // vendor's per-request/TPM limit and retries, and PassMeter auto-continues on\n // finish_reason \"length\", so a big request degrades to more passes, not an error.\n const expect = gen.expect ?? expectTokens(messages);\n // max_tokens policy — auto: size to task (router clamps to vendor/learned caps); fixed: the user's number; max: ask for the ceiling and let the router clamp.\n const maxTokens = gen.maxMode === \"max\" ? 65_536 : gen.maxMode === \"fixed\" && gen.maxTokens ? gen.maxTokens : (gen.maxTokens ?? (expect >= 6000 ? 16_000 : 4_000));\n const body: Record = { messages: normalizeMessages(messages), max_tokens: maxTokens, omni_expect_tokens: expect };\n const routerUrl = gen.router?.url ?? INFERENCE_URL;\n const routerHeaders: Record = { \"content-type\": \"application/json\", ...(gen.router?.key ? { authorization: `Bearer ${gen.router.key}` } : {}) };\n if (deadlineMs) body.omni_deadline_ms = deadlineMs;\n if (gen.keyPolicy) body.omni_key_policy = gen.keyPolicy;\n if (gen.vendorOrder?.length) body.omni_vendor_order = gen.vendorOrder;\n if (gen.exclude?.length) body.omni_exclude = gen.exclude;\n if (tools?.length) body.tools = tools;\n if (model) body.model = model;\n if (gen.reasoning) body.reasoning_effort = gen.reasoning;\n if (typeof gen.temperature === \"number\") body.temperature = gen.temperature;\n const t0 = Date.now();\n let lastErr: Error = new Error(\"unreachable\");\n let shrinks = 0;\n // Carried across retries so a call that eventually throws (after all 4 client-side retries) still reports\n // the last router pass's routing trail, instead of the flowchart going dark exactly when it matters most.\n let lastPlan: string[] | undefined;\n let lastAttempts: InferResult[\"meta\"][\"attempts\"];\n for (let retry = 0; retry < 4; retry++) {\n if (retry) {\n // Prompt-size exhaustion: the vendors that are left cannot take this prompt, so make the prompt fit —\n // drop oldest history, trim tool payloads — and try again at once instead of burning retries on an\n // identical, unwinnable resend.\n //\n // This regex previously looked for \"prompt ≈N tok > maxIn\" / \"too large\" / \"input limit\" / \"maxIn\" —\n // a message format router-core.ts and router-openai.ts never actually produce anywhere (confirmed:\n // no occurrence of any of those substrings in either file). ALL_PROVIDERS_EXHAUSTED just concatenates\n // each provider's own raw upstream error text (router-core.ts's `errors.push(`${p.name}: ${e.message}`)`),\n // so the old regex could never match and shrinkMessages() below had exactly one call site — this whole\n // branch was dead code; every size-exhausted request instead fell through to the generic retry path and\n // resent the identical oversized prompt up to 4 times, guaranteed to fail identically each time, burning\n // the wall-clock budget until the loop's own deadline check gave up. Replaced with patterns actually\n // observed live this session (a real Groq 413: \"Request too large ... please reduce your message size\n // and try again\") plus the standard phrasings other vendors use for the same failure.\n const sizeBound = /reduce (?:your|the) (?:message|prompt) size|request too large|prompt is too long|too many tokens|context_length_exceeded|maximum context length|exceeds the (?:model'?s? )?(?:context|token) (?:length|limit)|input (?:is )?too long/i.test(lastErr.message) && (lastErr as any).status === 503;\n if (sizeBound && shrinks < 2) { shrinks++; body.messages = shrinkMessages(body.messages as any[], shrinks); body.omni_expect_tokens = Math.min(expect, 2000); }\n const wait = sizeBound ? 0 : 1200 * retry;\n if (deadlineMs && Date.now() - t0 + wait + 4000 > deadlineMs) break;\n if (wait) await new Promise((r) => setTimeout(r, wait));\n if (deadlineMs) body.omni_deadline_ms = deadlineMs - (Date.now() - t0);\n }\n try {\n const res = await routerCall(routerUrl, { method: \"POST\", headers: routerHeaders, body: JSON.stringify(body) });\n const text = await res.text();\n let json: any = null;\n try { json = JSON.parse(text); } catch { /* non-JSON */ }\n if (!res.ok || !json?.choices?.[0]) {\n // An HTML body is an edge/WAF page (Val Town \"Blocked\", Cloudflare challenge), not the router: report its title + who served it, never the markup.\n const html = /^\\s*<(?:!doctype|html)/i.test(text);\n const detail = json?.error ? (typeof json.error === \"string\" ? json.error : `${json.error.message}${json.error.details ? \"\\n\" + [].concat(json.error.details).join(\"\\n\") : \"\"}`)\n : html ? `${routerUrl} answered with an HTML page \"${(text.match(/([^<]{1,80})<\\/title>/i)?.[1] ?? \"no title\").trim()}\" (server: ${res.headers.get(\"server\") ?? \"?\"}, cf-ray: ${res.headers.get(\"cf-ray\") ?? \"-\"}) — the request was blocked before reaching the router; check https://router.val.run/health in a browser and the val's logs`\n : text.slice(0, 600);\n if (json?.error?.plan) lastPlan = json.error.plan;\n if (json?.error?.attempts) lastAttempts = json.error.attempts;\n lastErr = Object.assign(new Error(`router ${res.status}: ${detail}`), { status: res.status, html, plan: lastPlan, attempts: lastAttempts });\n if (res.status === 404 || res.status === 400 || res.status === 401) throw lastErr; // not retryable\n continue;\n }\n const choice = json.choices[0];\n const m = choice.message ?? {};\n // Defense in depth: router-core.ts already rejects an empty/all-dot response before it can win a\n // provider race (see EMPTY_OR_DOT_RESPONSE there), but gen.router.url can point at a different,\n // custom router the user configured — this is the last checkpoint before an empty/junk reply would\n // otherwise be accepted as done. Treated exactly like a 503: retried, not silently returned.\n const omni = json._omni_meta;\n if (omni?.plan) lastPlan = omni.plan;\n if (omni?.attempts) lastAttempts = omni.attempts;\n const hasToolCalls = Array.isArray(m.tool_calls) && m.tool_calls.length > 0;\n if (!hasToolCalls && typeof m.content === \"string\") {\n const stripped = m.content.replace(/\\s+/g, \"\");\n if (stripped.length === 0 || /^\\.+$/.test(stripped)) {\n lastErr = Object.assign(new Error(`router returned an empty/dot response from ${omni?.provider?.name ?? \"an unknown provider\"}`), { status: 503, plan: lastPlan, attempts: lastAttempts });\n continue;\n }\n }\n return {\n message: m,\n finishReason: choice.finish_reason ?? \"stop\",\n reasoning: typeof m.reasoning === \"string\" ? m.reasoning : typeof m.reasoning_content === \"string\" ? m.reasoning_content : undefined,\n usage: json.usage,\n meta: { provider: omni?.provider, model: json.model, latencyMs: omni?.latency_ms, retries: retry,\n instance: omni?.instance ?? omni?.provider?.name, maxTokensSent: omni?.max_tokens_sent ?? maxTokens, maxTokensAsked: maxTokens, finishReason: choice.finish_reason ?? \"stop\", tokensOut: json.usage?.completion_tokens ?? null, tokensIn: json.usage?.prompt_tokens ?? null,\n plan: lastPlan, attempts: lastAttempts },\n };\n } catch (e: any) {\n lastErr = e;\n if (e.status && ![503, 502, 500, 429].includes(e.status)) throw e;\n }\n }\n throw lastErr;\n}\n\n// A metered wrapper: every inference call decrements a shared pass budget so no\n// stage can silently blow past the hard ceiling. This is THE governor's teeth.\n",[],[554],"6b3d8b85-4386-4846-9359-3235986d2915","app-helpers.ts","01a0a661-f049-7779-b407-8b3bd2cefad1",["D",1789497831502],"// Helpers split out of app.tsx to stay under the Val Town 80,000-character per-file limit:\n// legacy blob memory + salience facts, prompt shrinking, title derivation, tool-event slimming, web search, eval_js.\nimport { all, type Bound, one, type Row, run, sql, unwrap } from \"./db.ts\";\n// ts: optional wall-clock stamp — only the synthetic \"route\" events app.tsx's PassMeter pushes set it (real\n// tool-call events don't need it, their hop order already IS their chronology); when present, ui-timeline.ts\n// uses it to interleave route nodes with tool-call rows in the order they actually happened, not just hop order.\nexport interface ToolEvent { hop: number; tool: string; args: Record<string, unknown>; result?: unknown; error?: string; ts?: number }\n\nexport async function getMemoryBlob(): Promise<Row<\"memory\"> | null> {\n return unwrap(\n await one(\"memory\", sql`SELECT * FROM memory WHERE key = ${\"core\"}`),\n null,\n );\n}\n\n// ════════════════════════════════════════════════════════════════════════════\n// EPISODIC MEMORY: write with surprise-salience, retrieve by relevance, decay\n// ════════════════════════════════════════════════════════════════════════════\nexport const DECAY_HALFLIFE_MS = 1000 * 60 * 60 * 24 * 7; // 7 days\nexport const FACT_CAP = 500;\n\nexport function decayed(salience: number, accessed: number, now: number): number {\n const age = Math.max(0, now - accessed);\n return salience * Math.pow(0.5, age / DECAY_HALFLIFE_MS);\n}\n\nexport async function writeFact(fact: string, predictability: number): Promise<void> {\n const now = Date.now();\n // salience = information content ≈ (1 - self-reported predictability).\n const salience = Math.max(0.01, Math.min(1, 1 - predictability));\n await run(\n sql`INSERT INTO memory_fact (fact, salience, created, accessed, hits)\n VALUES (${fact}, ${salience}, ${now}, ${now}, 0)`,\n );\n\n // Bounded store: evict lowest current-value rows beyond a cap.\n const rows = unwrap(\n await all(\"memory_fact\", sql`SELECT * FROM memory_fact`),\n [],\n );\n if (rows.length <= FACT_CAP) return;\n const ranked = rows\n .map((r) => ({ id: r.id, v: decayed(r.salience, r.accessed, now) }))\n .sort((a, b) => a.v - b.v);\n const drop = ranked.slice(0, rows.length - FACT_CAP).map((r) => r.id);\n if (!drop.length) return;\n // Placeholder count varies with drop.length, but every bound value is still\n // a real parameter — this is `run` given a Bound built directly rather than\n // via the tagged template, which only supports a fixed interpolation count.\n const deleteEvicted: Bound = {\n text: `DELETE FROM memory_fact WHERE id IN (${\n drop.map(() => \"?\").join(\",\")\n })`,\n values: drop,\n };\n await run(deleteEvicted);\n}\n\n// Keyword-overlap retrieval — no embedding dependency. Deliberately simple;\n// swap for vector search later without touching callers.\nexport function tokenize(s: string): Set<string> {\n return new Set(\n s.toLowerCase().replace(/[^a-z0-9 ]/g, \" \").split(/\\s+/).filter((w) =>\n w.length > 2\n ),\n );\n}\n\nexport async function retrieveFacts(query: string, k = 5): Promise<string[]> {\n const now = Date.now();\n const qTok = tokenize(query);\n if (!qTok.size) return [];\n const rows = unwrap(\n await all(\"memory_fact\", sql`SELECT * FROM memory_fact`),\n [],\n );\n const scored = rows.map((r) => {\n const fTok = tokenize(r.fact);\n let overlap = 0;\n for (const t of qTok) if (fTok.has(t)) overlap++;\n const relevance = overlap / Math.sqrt(fTok.size || 1);\n const value = decayed(r.salience, r.accessed, now);\n return { id: r.id, fact: r.fact, score: relevance * (0.5 + value) };\n });\n const top = scored.filter((s) => s.score > 0).sort((a, b) =>\n b.score - a.score\n ).slice(0, k);\n // reinforce accessed facts (anti-decay); this is retrieval-as-rehearsal\n for (const t of top) {\n await run(\n sql`UPDATE memory_fact SET accessed = ${now}, hits = hits + 1 WHERE id = ${t.id}`,\n );\n }\n return top.map((t) => t.fact);\n}\n\n// ════════════════════════════════════════════════════════════════════════════\n// SESSION CRUD\n// ════════════════════════════════════════════════════════════════════════════\n\n/** Make a prompt fit a smaller input ceiling: keep system messages and the last user turn intact; level 1 drops the\n * oldest half of the history and caps tool results at 1.5k chars; level 2 keeps only the last 2 exchanges and caps at 600. */\nexport function shrinkMessages(messages: any[], level: number): any[] {\n const sys = messages.filter((m) => m.role === \"system\");\n const rest = messages.filter((m) => m.role !== \"system\");\n const cap = level >= 2 ? 600 : 1_500;\n const keep = level >= 2 ? Math.min(rest.length, 4) : Math.max(2, Math.ceil(rest.length / 2));\n const tail = rest.slice(-keep).map((m) => (m.role === \"tool\" && typeof m.content === \"string\" && m.content.length > cap ? { ...m, content: m.content.slice(0, cap) + \" …[trimmed]\" } : m));\n // a tool message must follow the assistant turn that requested it — drop leading orphans\n while (tail.length && tail[0].role === \"tool\") tail.shift();\n const last = tail[tail.length - 1];\n if (last?.role === \"user\" && typeof last.content === \"string\" && last.content.length > 24_000) tail[tail.length - 1] = { ...last, content: last.content.slice(0, 16_000) + \"\\n…[middle trimmed to fit the model's input limit]…\\n\" + last.content.slice(-6_000) };\n return [...sys, ...tail];\n}\n/** ≤4 significant words of the ask, Title Case; deterministic so two clients agree. */\nexport function deriveTitle(text: string): string {\n const STOP = new Set([\"a\", \"an\", \"the\", \"to\", \"of\", \"for\", \"in\", \"on\", \"and\", \"or\", \"with\", \"me\", \"my\", \"please\", \"can\", \"you\", \"i\", \"it\", \"that\", \"this\", \"is\", \"be\", \"make\", \"create\", \"build\", \"write\", \"help\", \"want\", \"need\", \"using\", \"use\"]);\n const words = text.replace(/```[\\s\\S]*?```/g, \" \").replace(/[^\\p{L}\\p{N}\\s#.-]/gu, \" \").split(/\\s+/).filter((w) => w && !STOP.has(w.toLowerCase()));\n const pick = (words.length ? words : text.split(/\\s+/)).slice(0, 4).map((w) => w.length > 18 ? w.slice(0, 18) : w);\n return (pick.map((w) => /^[a-z]/.test(w) ? w[0].toUpperCase() + w.slice(1) : w).join(\" \") || \"New Chat\").slice(0, 60);\n}\n\n/** Tool events as the UI needs them after a reload: args/results without file bodies. */\nexport function slimEvents(evs: ToolEvent[], cap = 2000): ToolEvent[] {\n // `arrayCap` exists because the 20-element cap — right for a tool's file list — silently truncated the\n // routing flowchart's plan (34 → 20) and attempts, which cut off exactly the tail where demoted\n // providers and late passes live. Route events are already bounded at the router (TRAIL_CAP=60), so they\n // get a cap that keeps the whole trail.\n const mk = (arrayCap: number) => {\n const trim = (v: unknown): unknown => {\n if (typeof v === \"string\") return v.length > cap ? v.slice(0, cap) + \"…\" : v;\n if (Array.isArray(v)) return v.slice(0, arrayCap).map(trim);\n if (v && typeof v === \"object\") return Object.fromEntries(Object.entries(v as Record<string, unknown>).filter(([k]) => ![\"content\", \"files\", \"edits\", \"html\"].includes(k)).map(([k, x]) => [k, trim(x)]));\n return v;\n };\n return trim;\n };\n const trimTool = mk(20), trimRoute = mk(80);\n return evs.map((e) => {\n const trim = e.tool === \"route\" ? trimRoute : trimTool;\n // ts is what cognize() sorted on to interleave route hops with tool calls; dropping it here meant a reload\n // could never re-establish that order.\n return { hop: e.hop, tool: e.tool, ...(e.ts ? { ts: e.ts } : {}), args: trim(e.args) as Record<string, unknown>, result: trim(e.result), ...(e.error ? { error: e.error } : {}) };\n });\n}\n\nexport async function webSearch(query: string): Promise<string> {\n const url = `https://html.duckduckgo.com/html/?q=${\n encodeURIComponent(query)\n }`;\n const res = await fetch(url, { headers: { \"User-Agent\": \"Mozilla/5.0\" } });\n const text = await res.text();\n const snippets: string[] = [];\n const re = /<a class=\"result__snippet\"[^>]*>([\\s\\S]*?)<\\/a>/g;\n let m: RegExpExecArray | null;\n while ((m = re.exec(text)) !== null && snippets.length < 6) {\n snippets.push(m[1].replace(/<[^>]+>/g, \"\").trim());\n }\n return snippets.length ? snippets.join(\"\\n---\\n\") : \"No results found.\";\n}\n\nexport function evalJs(code: string): unknown {\n try {\n const logs: string[] = [];\n const fakeConsole = { log: (...a: unknown[]) => logs.push(a.join(\" \")) };\n const fn = new Function(\n \"Math\",\n \"JSON\",\n \"Array\",\n \"Object\",\n \"String\",\n \"Number\",\n \"Boolean\",\n \"Date\",\n \"RegExp\",\n \"Map\",\n \"Set\",\n \"console\",\n `\"use strict\";\\n${code}`,\n );\n const result = fn(\n Math,\n JSON,\n Array,\n Object,\n String,\n Number,\n Boolean,\n Date,\n RegExp,\n Map,\n Set,\n fakeConsole,\n );\n return { result: result ?? null, logs };\n } catch (e) {\n return { error: String(e) };\n }\n}\n\n// Harmony / text-embedded tool-call parsing (gpt-oss leaks `to=functions.NAME json{…}` into content) and tool-name aliases.\nexport function balancedJson(s: string, at: number): string {\n if (s[at] !== \"{\") return \"\";\n let depth = 0, inStr = false, esc = false;\n for (let i = at; i < s.length; i++) {\n const c = s[i];\n if (inStr) { if (esc) esc = false; else if (c === \"\\\\\") esc = true; else if (c === '\"') inStr = false; continue; }\n if (c === '\"') inStr = true; else if (c === \"{\") depth++; else if (c === \"}\") { depth--; if (depth === 0) return s.slice(at, i + 1); }\n }\n return \"\";\n}\nexport function parseTextToolCalls(content: string): { calls: any[]; rest: string } {\n const calls: any[] = [];\n // gpt-oss \"Harmony\" channel leakage (some vendors' chat templates emit it as plain text):\n // analysis… commentary to=functions.NAME json{…} assistantfinal…\n // → tool call from the commentary segment; only the final channel is user-visible prose.\n if (/to=functions\\.[\\w.-]+/.test(content) || /assistantfinal/.test(content)) {\n const rx = /to=functions\\.([\\w.-]+)(?:\\s*<\\|[^|>]*\\|>\\s*|\\s*json\\s*|\\s)*(?=\\{)/g;\n for (const m of content.matchAll(rx)) {\n const json = balancedJson(content, m.index! + m[0].length);\n if (!json) continue;\n try { const j = JSON.parse(json); calls.push({ id: `call_${crypto.randomUUID().slice(0, 8)}`, type: \"function\", function: { name: m[1], arguments: JSON.stringify(j) } }); } catch { /* malformed */ }\n }\n const fin = content.split(/assistantfinal/); const visible = fin.length > 1 ? fin[fin.length - 1] : (calls.length ? \"\" : content.replace(/^\\s*analysis[\\s\\S]*?(?=assistant|$)/, \"\"));\n content = visible.replace(/<\\|[^|>]*\\|>/g, \"\");\n }\n const rest = content.replace(/<tool_call>([\\s\\S]*?)(?:<\\/tool_call>|$)/g, (_, inner: string) => {\n const t = inner.trim();\n try {\n if (t.startsWith(\"{\")) { const j = JSON.parse(t); if (j.name) calls.push({ id: `call_${crypto.randomUUID().slice(0, 8)}`, type: \"function\", function: { name: j.name, arguments: JSON.stringify(j.arguments ?? j.parameters ?? {}) } }); return \"\"; }\n } catch { /* fall through */ }\n const fn = t.match(/<function=([\\w.-]+)>/);\n if (!fn) return \"\";\n const args: Record<string, string> = {};\n for (const m of t.matchAll(/<parameter=([\\w.-]+)>([\\s\\S]*?)(?:<\\/parameter>|(?=<parameter=)|<\\/function>|$)/g)) args[m[1]] = m[2].replace(/^\\n/, \"\").replace(/\\n$/, \"\");\n calls.push({ id: `call_${crypto.randomUUID().slice(0, 8)}`, type: \"function\", function: { name: fn[1], arguments: JSON.stringify(args) } });\n return \"\";\n });\n return { calls, rest: rest.trim() };\n}\n\nexport const TOOL_ALIASES: Record<string, string> = { search: \"search_artifact\", grep: \"search_artifact\", find: \"search_artifact\", read_file: \"read_artifact\", read: \"read_artifact\", edit: \"update_artifact\", write_file: \"update_artifact\", edit_file: \"update_artifact\", write: \"update_artifact\", search_web: \"web_search\", browse: \"web_search\", remember: \"memory_write\", recall: \"memory_recall\" };\nexport const summarizeEvent = (ev: ToolEvent) =>\n ev.error ? `**${ev.tool}** failed: ${ev.error}`\n : ev.tool === \"create_artifact\" ? `Built **${(ev.result as any)?.title ?? ev.args.title ?? \"artifact\"}** — open it above.`\n : ev.tool === \"update_artifact\" ? `Updated **${String(ev.args.path ?? \"index.html\")}** in artifact #${(ev.result as any)?.id ?? \"\"} — reload it above.`\n : ev.tool === \"read_artifact\" ? \"\" : `Ran **${ev.tool}**.`;\n\n\n/** Vendor-portable message shapes. Strict OpenAI-compatible backends reject what lenient ones accept:\n * nvidia_nim — \"Assistant message must have either content or tool_calls, but not both\" → drop empty content beside tool_calls;\n * together — \"prompt cannot be empty\" → no message may carry empty-string content; tool results and assistant turns get a marker.\n * Consecutive system messages are merged (mistral/anthropic-style backends want one). Never mutates the input. */\nexport function normalizeMessages(messages: unknown[]): unknown[] {\n const out: any[] = [];\n for (const raw of messages as any[]) {\n if (!raw || typeof raw !== \"object\") continue;\n const m: any = { ...raw };\n const hasCalls = Array.isArray(m.tool_calls) && m.tool_calls.length > 0;\n const empty = m.content == null || (typeof m.content === \"string\" && m.content.trim() === \"\") || (Array.isArray(m.content) && m.content.length === 0);\n if (hasCalls) { if (empty) delete m.content; }\n else if (empty) m.content = m.role === \"tool\" ? \"(no output)\" : m.role === \"assistant\" ? \"(continued)\" : \" \";\n if (m.role === \"system\" && out.length && out[out.length - 1].role === \"system\" && typeof m.content === \"string\" && typeof out[out.length - 1].content === \"string\") { out[out.length - 1] = { ...out[out.length - 1], content: out[out.length - 1].content + \"\\n\\n\" + m.content }; continue; }\n out.push(m);\n }\n return out;\n}\n\n/** Salvage a create_artifact / update_artifact call whose JSON arguments were cut at max_tokens: recover every COMPLETE\n * {path, content} object from the `files` array (and `content`/`title` when they closed), report which file was cut.\n * Returns null when nothing usable survived. The cut file is deliberately dropped — a half file is a placeholder. */\nexport function salvageArtifactArgs(raw: string): { args: Record<string, unknown>; saved: string[]; cut: string | null } | null {\n const files: { path: string; content: string }[] = [];\n let cut: string | null = null;\n // Anchored to a preceding `{`/`,` (an actual object-key position), not a bare substring search: `content`\n // is large, arbitrary, model-generated HTML/JS and typically precedes `files` in argument order, so a\n // plain `raw.indexOf('\"files\"')` can match text INSIDE that content instead of the real key — e.g.\n // generated markup containing `data-x=\"files\"` is JSON-escaped as `\\\"files\\\"`, and the 7-char sequence\n // `\"files\"` (unescaped) appears right after that backslash, matching before the real key ever does. That\n // false match then feeds a garbage `[` search and corrupts (or silently drops) the whole salvage.\n const filesKeyMatch = raw.match(/[{,]\\s*\"files\"\\s*:/);\n const at = filesKeyMatch ? filesKeyMatch.index! + filesKeyMatch[0].indexOf('\"files\"') : -1;\n if (at >= 0) {\n const arr = raw.indexOf(\"[\", at);\n let i = arr + 1;\n while (i < raw.length) {\n const o = raw.indexOf(\"{\", i); if (o < 0) break;\n const obj = balancedJson(raw, o);\n if (!obj) { const m = raw.slice(o).match(/\"path\"\\s*:\\s*\"([^\"]*)\"/); cut = m ? m[1] : \"(unnamed)\"; break; }\n try { const f = JSON.parse(obj); if (typeof f?.path === \"string\" && typeof f?.content === \"string\") files.push({ path: f.path, content: f.content }); } catch { break; }\n i = o + obj.length;\n }\n }\n const head = at >= 0 ? raw.slice(0, at) : raw; /* top-level strings only count when they precede the files array — after it they were cut */\n // Escape map covers \\n \\t \\u fully; every other escape (including the unmapped \\r \\b \\f) fell through to\n // \"output the raw character after the backslash\" — correct BY COINCIDENCE for \\\" and \\\\ (the char after\n // the backslash IS the literal character wanted), but wrong for \\r/\\b/\\f, which decoded to a literal\n // \"r\"/\"b\"/\"f\" letter instead of the actual control character. Rare in practice (most generated content\n // uses \\n, not \\r), but real whenever salvaged content contains Windows-style line endings.\n const str = (key: string) => { const m = head.match(new RegExp('\"' + key + '\"\\\\s*:\\\\s*\"')); if (!m) return undefined; const start = (m.index ?? 0) + m[0].length; let out = \"\", j = start; while (j < head.length) { const c = head[j]; if (c === \"\\\\\") { const n = head[j + 1]; out += n === \"n\" ? \"\\n\" : n === \"t\" ? \"\\t\" : n === \"r\" ? \"\\r\" : n === \"b\" ? \"\\b\" : n === \"f\" ? \"\\f\" : n === \"u\" ? String.fromCharCode(parseInt(head.slice(j + 2, j + 6), 16)) : n; j += n === \"u\" ? 6 : 2; continue; } if (c === '\"') return out; out += c; j++; } return undefined; };\n const args: Record<string, unknown> = {};\n const id = raw.match(/\"id\"\\s*:\\s*(\\d+)/); if (id) args.id = Number(id[1]);\n const title = str(\"title\"); if (title !== undefined) args.title = title;\n const kind = str(\"kind\"); if (kind !== undefined) args.kind = kind;\n const content = str(\"content\"); if (content !== undefined) args.content = content;\n if (files.length) args.files = files;\n if (!files.length && args.content === undefined) return null;\n return { args, saved: files.map((f) => f.path), cut };\n}\n\n// Hidden META control block parsing and the pre-estimate heuristics (moved from app.tsx for the 80k-char limit).\nexport const META_OPEN = \"<<<META\";\nexport const META_CLOSE = \"META>>>\";\n\nexport interface AnswerMeta {\n confidence: number; // 0..1 self-estimated P(answer correct)\n stakes: \"low\" | \"high\"; // self-estimated irreversibility of being wrong\n abstain: boolean; // true → model declines rather than confabulate\n}\nexport function parseMeta(text: string): { visible: string; meta: AnswerMeta | null } {\n const i = text.lastIndexOf(META_OPEN);\n const j = text.lastIndexOf(META_CLOSE);\n if (i !== -1 && (j === -1 || j < i)) return { visible: text.slice(0, i).trimEnd(), meta: null }; // dangling open marker: hide it\n if (i === -1) return { visible: text, meta: null };\n const raw = text.slice(i + META_OPEN.length, j).trim();\n const visible = (text.slice(0, i) + text.slice(j + META_CLOSE.length)).trim();\n try {\n const p = JSON.parse(raw);\n const meta: AnswerMeta = {\n confidence: Number.isFinite(+p.confidence) ? +p.confidence : 0.5,\n stakes: p.stakes === \"high\" ? \"high\" : \"low\",\n abstain: !!p.abstain,\n };\n return { visible, meta };\n } catch {\n return { visible, meta: null };\n }\n}\n\n// (META instruction now lives in prompts.ts as \"meta\")\n\n// ════════════════════════════════════════════════════════════════════════════\n// COMPLEXITY / STAKES PRE-ESTIMATE (heuristic, zero-cost — no inference call)\n// Governs whether the expensive stages are even eligible to fire.\n// ════════════════════════════════════════════════════════════════════════════\nexport function preEstimate(\n input: string,\n): { complex: boolean; maybeHighStakes: boolean } {\n const t = input.toLowerCase();\n const long = input.length > 180;\n const multiClause = (input.match(/[.?!]/g)?.length ?? 0) >= 2;\n const reasoningWords =\n /\\b(why|how|compare|prove|derive|design|analy|trade[- ]?off|implication|explain)\\b/\n .test(t);\n const stakesWords =\n /\\b(medical|legal|financial|invest|dose|security|delete|irreversible|production|deploy|money|contract)\\b/\n .test(t);\n const complex = (long && multiClause) || reasoningWords;\n return { complex, maybeHighStakes: stakesWords };\n}\n",[],[561],"b1dd17f0-86c9-4324-a419-0589ff7102aa","app-compact.ts","01a0a661-f049-7779-b407-8670bf9b1b65",["D",1789497831502],"// Conversation compaction: a running summary of everything older than the last few exchanges, stored per session in\n// omni_state (key compact:<session>) and prepended to the model context in place of the messages it covers. Messages\n// themselves are never deleted — history rendering is unchanged — only what the model sees is folded.\n// Runs as its own val invocation (POST ?compact) so it never competes with a turn's 60 s budget.\nimport { all, run, sql, unwrap } from \"./db.ts\";\nimport { renderPrompt } from \"./prompts.ts\";\nimport { normalizeSettings, type Settings } from \"./app-settings.ts\";\n\nexport type Compact = { summary: string; uptoId: number; ts: number; covered: number; chars: number };\n\nexport async function loadCompact(session: string): Promise<Compact | null> {\n const rows = unwrap(await all<\"omni_state\">(\"omni_state\", sql`SELECT value FROM omni_state WHERE key = ${\"compact:\" + session}`), []);\n try { return rows[0] ? JSON.parse(rows[0].value) as Compact : null; } catch { return null; }\n}\nexport async function saveCompact(session: string, c: Compact | null): Promise<void> {\n const key = \"compact:\" + session;\n if (!c) { await run(sql`DELETE FROM omni_state WHERE key = ${key}`); return; }\n await run(sql`INSERT INTO omni_state (key, value, ts) VALUES (${key}, ${JSON.stringify(c)}, ${Date.now()}) ON CONFLICT(key) DO UPDATE SET value = excluded.value, ts = excluded.ts`);\n}\n\n/** Chars of live (uncompacted) history — what a turn would have to fit. */\nexport async function liveChars(session: string, uptoId: number): Promise<{ chars: number; count: number }> {\n const rows = unwrap(await all<\"message\">(\"message\", sql`SELECT id, content FROM message WHERE session = ${session} AND id > ${uptoId} ORDER BY id DESC LIMIT 200`), []);\n return { chars: rows.reduce((n, r) => n + r.content.length, 0), count: rows.length };\n}\n\nconst KEEP_TAIL = 6; // the last 3 exchanges are never folded: the model needs them verbatim\n\n/** Fold everything older than the last KEEP_TAIL messages into the summary. `call` is the inference function (fast tier). */\nexport async function compactSession(\n session: string,\n call: (messages: unknown[]) => Promise<string>,\n prompts?: Record<string, string>,\n examples = true,\n): Promise<{ ok: boolean; before: number; after: number; covered: number; uptoId: number; reason?: string }> {\n const prev = await loadCompact(session);\n const uptoPrev = prev?.uptoId ?? 0;\n const rows = unwrap(await all<\"message\">(\"message\", sql`SELECT * FROM message WHERE session = ${session} AND id > ${uptoPrev} ORDER BY id ASC`), []);\n // Was `<= KEEP_TAIL + 1`: at rows.length === KEEP_TAIL+1 there is already exactly one row old enough\n // to fold (rows.length - KEEP_TAIL === 1), so the +1 rejected a foldable state as \"nothing old enough\"\n // and delayed the first compaction by one message past the intended KEEP_TAIL threshold.\n if (rows.length <= KEEP_TAIL) return { ok: false, before: 0, after: prev?.chars ?? 0, covered: 0, uptoId: uptoPrev, reason: \"nothing old enough to fold\" };\n const fold = rows.slice(0, rows.length - KEEP_TAIL);\n const before = fold.reduce((n, r) => n + r.content.length, 0);\n const transcript = fold.map((m) => `[#${m.id} ${m.role}] ${m.content.length > 5_000 ? m.content.slice(0, 3_800) + \"\\n…[\" + (m.content.length - 4_800) + \" chars trimmed]…\\n\" + m.content.slice(-1_000) : m.content}`).join(\"\\n\\n\");\n const prompt = renderPrompt(\"compact\", { previous: prev?.summary ?? \"(none — this is the first compaction)\", transcript }, prompts as any, examples);\n const summary = (await call([{ role: \"system\", content: \"You maintain the running summary of a long working conversation. Output only the summary.\" }, { role: \"user\", content: prompt }])).trim();\n if (summary.length < 80) return { ok: false, before, after: prev?.chars ?? 0, covered: 0, uptoId: uptoPrev, reason: \"model returned no usable summary\" };\n const storedSummary = summary.slice(0, 12_000);\n // chars must reflect what's actually stored (storedSummary), not the pre-truncation raw model output —\n // getContext() seeds its context-budget accounting with this value, so a mismatch here silently\n // over-reserves budget on every turn after a summary that ran past the 12,000-char cap.\n const next: Compact = { summary: storedSummary, uptoId: fold[fold.length - 1].id, ts: Date.now(), covered: (prev?.covered ?? 0) + fold.length, chars: storedSummary.length };\n await saveCompact(session, next);\n return { ok: true, before, after: next.chars, covered: fold.length, uptoId: next.uptoId };\n}\n\n/** The system message a turn prepends when a compaction exists. */\nexport function compactMessage(c: Compact): { role: \"system\"; content: string } {\n return { role: \"system\", content: `CONVERSATION SO FAR (compacted summary of ${c.covered} earlier messages; treat as established context, do not repeat it back):\\n${c.summary}` };\n}\n\n/** HTTP: POST ?compact {session, settings} folds old turns; GET ?compact=<session> reads the summary; DELETE clears it. Null when not a compact route. */\nexport async function handleCompact(req: Request, url: URL, infer: (messages: unknown[], settings: Settings) => Promise<string>): Promise<Response | null> {\n if (!url.searchParams.has(\"compact\")) return null;\n const session = url.searchParams.get(\"compact\") || \"\";\n if (req.method === \"GET\") {\n const c = await loadCompact(session); const budget = Number(url.searchParams.get(\"check\")) || 0;\n const live = budget ? await liveChars(session, c?.uptoId ?? 0) : null;\n return Response.json({ compact: c, needs: !!live && live.count > 8 && live.chars > 2 * budget, live });\n }\n if (req.method === \"DELETE\") { await saveCompact(session, null); return Response.json({ ok: true }); }\n if (req.method !== \"POST\") return null;\n const b = await req.json().catch(() => ({}));\n const settings = normalizeSettings(b.settings ?? {});\n const s = String(b.session ?? \"\");\n if (!s) return Response.json({ error: { message: \"session required\" } }, { status: 400 });\n try { return Response.json(await compactSession(s, (m) => infer(m, settings), settings.prompts as Record<string, string>, settings.examples)); }\n catch (e: any) { return Response.json({ error: { message: `compact failed: ${String(e?.message ?? e).slice(0, 300)}` } }, { status: 502 }); }\n}\n",[],[568],"837a55f9-8c88-496b-8e0a-5844d707ff23","_valtown-workspace.ts","01a0a661-f049-7779-b407-829bd8d68623",["D",1789497831502],"export default {\n async fetch(req: Request) {\n const url = new URL(req.url);\n const path = url.pathname;\n\n // Load workspace from v.storage\n const workspace = (await v.storage.get(\"workspace\")) ?? {\n directories: {},\n files: {},\n };\n\n // Helpers\n const save = async () => v.storage.set(\"workspace\", workspace);\n\n // -----------------------------\n // CREATE FILE\n // -----------------------------\n if (path === \"/create\" && req.method === \"POST\") {\n const body = await req.json();\n const { filePath, content } = body;\n\n workspace.files[filePath] = {\n name: filePath.split(\"/\").pop(),\n path: filePath,\n type: \"file\",\n content,\n };\n\n await save();\n return new Response(\"created\");\n }\n\n // -----------------------------\n // UPDATE FILE\n // -----------------------------\n if (path === \"/update\" && req.method === \"POST\") {\n const body = await req.json();\n const { filePath, content } = body;\n\n if (!workspace.files[filePath]) {\n return new Response(\"not found\", { status: 404 });\n }\n\n workspace.files[filePath].content = content;\n await save();\n return new Response(\"updated\");\n }\n\n // -----------------------------\n // DELETE FILE\n // -----------------------------\n if (path === \"/delete\" && req.method === \"POST\") {\n const body = await req.json();\n const { filePath } = body;\n\n delete workspace.files[filePath];\n await save();\n return new Response(\"deleted\");\n }\n\n // -----------------------------\n // LIST FILES\n // -----------------------------\n if (path === \"/list\") {\n return Response.json(Object.keys(workspace.files));\n }\n\n // -----------------------------\n // GET FILE\n // -----------------------------\n if (path === \"/file\") {\n const filePath = url.searchParams.get(\"path\");\n return Response.json(workspace.files[filePath]);\n }\n\n return new Response(\"unknown route\", { status: 404 });\n },\n};",[],[575],"9f54c325-4539-4b5f-8186-daaee53cc27a","_valtown-file-editor.ts","01a0a661-f049-7779-b407-7f7b8b934a6f",["D",1789497831502],"export class ValFileEditor {\n constructor(public file) {\n this.file = file;\n }\n\n get content() {\n return this.file.content;\n }\n\n setContent(c) {\n this.file.content = c;\n return this;\n }\n\n // -----------------------------\n // Subdomain support\n // -----------------------------\n getSubdomain() {\n const m = this.file.content.match(/\\/\\/\\s*@subdomain:\\s*([a-z0-9-]+)/i);\n return m?.[1];\n }\n\n setSubdomain(name: string) {\n if (this.getSubdomain()) {\n return this.setContent(\n this.file.content.replace(\n /\\/\\/\\s*@subdomain:\\s*[a-z0-9-]+/i,\n `// @subdomain: ${name}`,\n ),\n );\n }\n return this.setContent(`// @subdomain: ${name}\\n` + this.file.content);\n }\n\n // -----------------------------\n // File type inference\n // -----------------------------\n inferKind() {\n const c = this.file.content;\n\n if (/export\\s+default\\s+async\\s+function\\s+handler/.test(c)) {\n return \"http-module\";\n }\n if (/export\\s+async\\s+function/.test(c)) {\n return \"val-function\";\n }\n if (/export\\s+const\\s+run\\s*=/.test(c)) {\n return \"scheduled-val\";\n }\n if (/export\\s+const\\s+cron\\s*=/.test(c)) {\n return \"cron-val\";\n }\n if (this.file.name.endsWith(\".json\")) return \"json\";\n if (this.file.name.endsWith(\".md\")) return \"markdown\";\n return \"unknown\";\n }\n\n // -----------------------------\n // Fingerprint (browser crypto)\n // -----------------------------\n async fingerprint() {\n const data = new TextEncoder().encode(this.file.content);\n const hash = await crypto.subtle.digest(\"SHA-256\", data);\n return [...new Uint8Array(hash)]\n .map((b) => b.toString(16).padStart(2, \"0\"))\n .join(\"\");\n }\n\n // -----------------------------\n // Save back to workspace\n // -----------------------------\n async save() {\n await fetch(\"https://ziggyware.vals.workers.dev/workspace/update\", {\n method: \"POST\",\n body: JSON.stringify({\n filePath: this.file.path,\n content: this.file.content,\n }),\n });\n }\n}",[],[582],"37718ea9-a407-4967-a98d-21a07c676b1f","_valtown-api.ts","01a0a661-f049-7779-b407-7877acc25b4d",["D",1789497831502],"const valAPI = {\n async createVal({ name, code }) {\n return await fetch(\"https://api.val.town/v1/vals\", {\n method: \"POST\",\n headers: {\n Authorization: `Bearer ${Deno.env.get(\"VALTOWN_API_KEY\")}`,\n \"Content-Type\": \"application/json\",\n },\n body: JSON.stringify({ name, code }),\n }).then((r) => r.json());\n },\n\n async updateVal({ name, code }) {\n return await fetch(`https://api.val.town/v1/vals/${name}`, {\n method: \"PATCH\",\n headers: {\n Authorization: `Bearer ${Deno.env.get(\"VALTOWN_API_KEY\")}`,\n \"Content-Type\": \"application/json\",\n },\n body: JSON.stringify({ code }),\n }).then((r) => r.json());\n },\n async deleteVal(name: string) {\n return await fetch(`https://api.val.town/v1/vals/${name}`, {\n method: \"DELETE\",\n headers: {\n Authorization: `Bearer ${Deno.env.get(\"VALTOWN_API_KEY\")}`,\n },\n }).then((r) => r.json());\n },\n};\n\nexport const ValAPI = {\n async createVal({ name, code }) {\n return await valAPI.createVal({ name, code });\n },\n async updateVal({ name, code }) {\n return await valAPI.updateVal({ name, code });\n },\n async deleteVal(name: string) {\n return await valAPI.deleteVal(name);\n },\n};",[],[589],"b8131718-456b-44b7-863e-f0c146fdc0a4","_inference-router.ts","01a0a661-f049-7779-b407-7404bfedc7e9",["D",1789497831502],"import { sqlite } from \"https://esm.town/v/std/sqlite/main.ts\";\nimport { getProviderArray } from \"./inference-provider.ts\";\nimport type { Provider } from \"./db.ts\";\n\n// ============================================================================\n// 2. STATE MANAGEMENT (SQLite) & CACHING\n// ============================================================================\nasync function initDB() {\n await sqlite.execute(`\n CREATE TABLE IF NOT EXISTS omni_router_config (\n provider TEXT PRIMARY KEY,\n selected_model TEXT NOT NULL\n )\n `);\n}\n\nasync function getSavedModels(): Promise<Record<string, string>> {\n const result = await sqlite.execute(\n `SELECT provider, selected_model FROM omni_router_config`,\n );\n const map: Record<string, string> = {};\n for (const row of result.rows) {\n const r = row as any;\n map[r.provider as string] = r.selected_model as string;\n }\n return map;\n}\n\nconst MAX_CONCURRENCY = 3;\nlet inflight = 0;\n\nconst sleep = (ms: number) => new Promise((r) => setTimeout(r, ms));\nconst now = () => performance.now();\n\n// ============================================================================\n// 3. ENTERPRISE CIRCUIT BREAKER\n// ============================================================================\ntype BreakerState = \"CLOSED\" | \"OPEN\" | \"HALF_OPEN\";\ninterface Breaker {\n failCount: number;\n state: BreakerState;\n nextAttempt: number;\n}\nconst breakers = new Map<string, Breaker>();\nconst MAX_FAILS = 2;\nconst BREAK_DURATION_MS = 25_000;\nconst HARD_TIMEOUT_MS = 12_000;\n\nfunction getBreaker(name: string): Breaker {\n if (!breakers.has(name)) {\n breakers.set(name, { failCount: 0, state: \"CLOSED\", nextAttempt: 0 });\n }\n return breakers.get(name)!;\n}\n\nfunction checkBreaker(name: string): boolean {\n const b = getBreaker(name);\n if (b.state === \"OPEN\" && Date.now() >= b.nextAttempt) {\n b.state = \"HALF_OPEN\";\n }\n return b.state === \"HALF_OPEN\" ||\n b.state === \"CLOSED\";\n}\n\nfunction reportBreaker(name: string, success: boolean) {\n const b = getBreaker(name);\n if (success) {\n b.failCount = 0;\n b.state = \"CLOSED\";\n } else {\n b.failCount++;\n if (\n b.failCount >= MAX_FAILS ||\n b.state === \"HALF_OPEN\"\n ) {\n b.state = \"OPEN\";\n b.nextAttempt = Date.now() + BREAK_DURATION_MS;\n b.failCount = 0;\n }\n }\n}\n\n// ============================================================================\n// 4. CORE INFERENCE ENGINE\n// ============================================================================\nasync function invokeProvider(\n provider: Provider,\n targetModel: string,\n messages: any[],\n tools?: any[],\n toolChoice?: any,\n) {\n const key = Deno.env.get(provider.keyEnv)!;\n const url = `${provider.base}/chat/completions`;\n\n const payload: any = {\n model: targetModel,\n messages,\n temperature: 0.0,\n stream: false,\n };\n if (tools && tools.length > 0) {\n payload.tools = tools;\n if (toolChoice) payload.tool_choice = toolChoice;\n }\n\n const controller = new AbortController();\n const timeoutId = setTimeout(() => controller.abort(), HARD_TIMEOUT_MS);\n\n try {\n const t0 = now();\n const res = await fetch(url, {\n method: \"POST\",\n headers: {\n \"Authorization\": `Bearer ${key}`,\n \"Content-Type\": \"application/json\",\n \"HTTP-Referer\": \"https://val.town\",\n \"X-Title\": \"Ziggyware-OmniRouter\",\n },\n body: JSON.stringify(payload),\n signal: controller.signal,\n });\n clearTimeout(timeoutId);\n if (!res.ok) {\n throw new Error(\n `HTTP ${res.status}: ${await res.text().catch(() => \"Unknown\")}`,\n );\n }\n\n const json = await res.json();\n if (!json.choices?.[0]?.message) throw new Error(\"Malformed JSON response\");\n\n return {\n message: json.choices[0].message,\n latencyMs: Math.round(now() - t0),\n modelUsed: targetModel,\n };\n } catch (e: any) {\n clearTimeout(timeoutId);\n throw new Error(\n e.name === \"AbortError\" ? \"API_STALLED_TIMEOUT\" : e.message,\n );\n }\n}\n\nasync function routeInference(\n messages: any[],\n tools?: any[],\n toolChoice?: any,\n) {\n while (inflight >= MAX_CONCURRENCY) await sleep(10);\n inflight++;\n\n try {\n const savedModels = await getSavedModels();\n const enabledProviders = getProviderArray();\n\n if (enabledProviders.length === 0) {\n throw new Error(\"NO_PROVIDERS_CONFIGURED\");\n }\n\n for (const provider of enabledProviders) {\n if (!checkBreaker(provider.name)) continue;\n const targetModel = savedModels[provider.name] ||\n provider.fallbackModel;\n\n try {\n const {\n message,\n latencyMs,\n modelUsed,\n } = await invokeProvider(\n provider,\n targetModel,\n messages,\n tools,\n toolChoice,\n );\n\n reportBreaker(provider.name, true);\n\n const resultPayload = {\n message,\n provider: provider,\n model: modelUsed,\n latencyMs,\n };\n\n return resultPayload;\n } catch (e: any) {\n reportBreaker(provider.name, false);\n console.warn(\n `[SEVERED] ${provider.name} failed (${e.message}). Routing next...`,\n );\n }\n }\n throw new Error(\"ALL_PROVIDERS_EXHAUSTED_OR_OPEN\");\n } finally {\n inflight--;\n }\n}\n\nexport default async function handler(req: Request): Promise<Response> {\n const url = new URL(req.url);\n\n // CORS & Preflight\n if (req.method === \"OPTIONS\") {\n return new Response(null, {\n headers: {\n \"Access-Control-Allow-Origin\": \"*\",\n \"Access-Control-Allow-Headers\": \"*\",\n \"Access-Control-Allow-Methods\": \"POST, GET, OPTIONS\",\n },\n });\n }\n\n // Ensure DB\n await initDB();\n if (req.method === \"GET\") {\n // --- API: Get Status ---\n if (url.pathname === \"/api/providers\") {\n const savedModels = await getSavedModels();\n const status = getProviderArray().map((p) => ({\n ...p,\n hasKey: !!Deno.env.get(p.keyEnv),\n currentModel: savedModels[p.name] || p.fallbackModel,\n })).sort((a, b) => a.priority - b.priority);\n return Response.json(status);\n }\n\n // --- API: Proxy Fetch Models ---\n if (url.pathname === \"/api/models\") {\n const pName = url.searchParams.get(\"provider\");\n const p = getProviderArray().find((x) => x.name === pName);\n if (!p) {\n return Response.json({ error: \"Provider not found\" }, { status: 404 });\n }\n\n const key = Deno.env.get(p.keyEnv);\n if (!key) {\n return Response.json({ error: \"Missing API Key\" }, { status: 401 });\n }\n\n try {\n const res = await fetch(`${p.base}/models`, {\n headers: { \"Authorization\": `Bearer ${key}` },\n });\n if (!res.ok) throw new Error(`HTTP ${res.status}`);\n const data = await res.json();\n return Response.json(data.data || data);\n } catch (e) {\n return Response.json({ error: e.message, fallback: p.fallbackModel }, {\n status: 500,\n });\n }\n }\n\n return new Response(\n `\n <!DOCTYPE html>\n <html lang=\"en\">\n <head>\n <meta charset=\"UTF-8\">\n <meta name=\"viewport\" content=\"width=device-width, initial-scale=1.0\">\n <title>Ziggyware AI Chat\n \n \n \n \n \n \n \n

\n \n \n \n `,\n { headers: { \"Content-Type\": \"text/html\" } },\n );\n } else if (req.method === \"POST\") {\n // --- API: Update SQLite DB ---\n if (url.pathname === \"/api/update\") {\n const { provider, model } = await req.json();\n await sqlite.execute({\n sql:\n `INSERT INTO omni_router_config (provider, selected_model) VALUES (?, ?) ON CONFLICT(provider) DO UPDATE SET selected_model = excluded.selected_model`,\n args: [provider, model],\n });\n return Response.json({ success: true });\n }\n\n try {\n const body = await req.json();\n const messages = body.messages ||\n (body.prompt ? [{ role: \"user\", content: body.prompt }] : []);\n if (!messages.length) {\n return new Response(\n JSON.stringify({ error: \"Empty message payload\" }),\n { status: 400 },\n );\n }\n const { message, provider, model, latencyMs } = await routeInference(\n messages,\n body.tools,\n body.tool_choice,\n );\n\n const activeGuards = getProviderArray()\n .filter((p) => Deno.env.get(p.keyEnv))\n .map(\n (p) => ({ name: p.name, state: getBreaker(p.name).state }),\n );\n\n const sseId = `chatcmpl-omni-${crypto.randomUUID()}`;\n\n if (body.stream) {\n const encoder = new TextEncoder();\n\n const stream = new ReadableStream({\n start(controller) {\n const sendChunk = (\n delta: any,\n finish_reason: string | null = null,\n ) => {\n controller.enqueue(encoder.encode(`data: ${\n JSON.stringify({\n id: sseId,\n object: \"chat.completion.chunk\",\n created: Math.floor(Date.now() / 1000),\n model: model,\n choices: [{ index: 0, delta, finish_reason }],\n _omni_meta: {\n provider: JSON.stringify(provider),\n latency_ms: latencyMs,\n inflight_requests: inflight,\n circuit_states: activeGuards,\n },\n })\n }\\n\\n`));\n };\n\n // 1. Role chunk\n sendChunk({ role: \"assistant\", content: \"\" });\n\n // 2. Content chunk\n if (message.content) sendChunk({ content: message.content });\n\n // 3. Tools chunk\n if (message.tool_calls) {\n sendChunk({ tool_calls: message.tool_calls });\n }\n\n // 4. Finish chunk\n sendChunk({}, message.tool_calls ? \"tool_calls\" : \"stop\");\n\n // 5. End Stream\n controller.enqueue(encoder.encode(\"data: [DONE]\\n\\n\"));\n controller.close();\n },\n });\n\n return new Response(stream, {\n headers: {\n \"Content-Type\": \"text/event-stream\",\n \"Cache-Control\": \"no-cache\",\n \"Connection\": \"keep-alive\",\n \"Access-Control-Allow-Origin\": \"*\",\n },\n });\n }\n\n return new Response(\n JSON.stringify({\n id: sseId,\n object: \"chat.completion\",\n created: Math.floor(Date.now() / 1000),\n model: model,\n choices: [{\n index: 0,\n message: message,\n finish_reason: message.tool_calls ? \"tool_calls\" : \"stop\",\n }],\n _omni_meta: {\n provider: provider,\n latency_ms: latencyMs,\n inflight_requests: inflight,\n circuit_states: activeGuards,\n },\n }),\n {\n status: 200,\n headers: {\n \"Content-Type\": \"application/json\",\n \"Access-Control-Allow-Origin\": \"*\",\n },\n },\n );\n } catch (e: any) {\n return new Response(\n JSON.stringify({\n error: {\n message: \"(error).\",\n details: e.message,\n },\n }),\n {\n status: 503,\n headers: {\n \"Content-Type\": \"application/json\",\n \"Access-Control-Allow-Origin\": \"*\",\n },\n },\n );\n }\n }\n\n return new Response(\"Not Found\", { status: 404 });\n}",[],[596],"886d4d01-fdfb-4473-a256-a9ccc595575c","_inference-router-dev.tsx","01a0a661-f049-7779-b407-733a997b9c05",["D",1789497831502],"import { getProviderArray } from \"./inference-provider.ts\";\nimport { all, initDB, type Row, run, sql } from \"./db.ts\";\n\ntype Provider = Row<\"provider\">;\n\nconst MAX_CONCURRENCY = 3;\nconst MAX_FAILS = 2;\nconst BREAK_DURATION_MS = 25_000;\nconst HARD_TIMEOUT_MS = 120_000;\nconst INFERENCE_TEMPERATURE = 0.0;\n\n// ============================================================================\n// 2. STATE MANAGEMENT (SQLite) & CACHING\n// ============================================================================\n// omni_router_config is already part of db.ts's shared Schema/DDL — initDB()\n// from db.ts creates it (and every other table) idempotently. The local\n// CREATE TABLE here was a second, drifting source of truth for the same\n// table and has been removed rather than kept \"just in case\" the two diverge.\n\nasync function getSavedModels(): Promise> {\n const r = await all(\n \"omni_router_config\",\n sql`SELECT * FROM omni_router_config`,\n );\n if (!r.ok) {\n console.error(\"getSavedModels failed:\", r.error);\n return {};\n }\n const map: Record = {};\n for (const row of r.value) map[row.provider] = row.selected_model;\n return map;\n}\n\nlet inflight = 0;\n\nconst sleep = (ms: number) => new Promise((r) => setTimeout(r, ms));\nconst now = () => performance.now();\n\n// ============================================================================\n// 3. ENTERPRISE CIRCUIT BREAKER\n// ============================================================================\ntype BreakerState = \"CLOSED\" | \"OPEN\" | \"HALF_OPEN\";\ninterface Breaker {\n failCount: number;\n state: BreakerState;\n nextAttempt: number;\n}\nconst breakers = new Map();\n\nfunction getBreaker(name: string): Breaker {\n if (!breakers.has(name)) {\n breakers.set(name, { failCount: 0, state: \"CLOSED\", nextAttempt: 0 });\n }\n return breakers.get(name)!;\n}\n\nfunction checkBreaker(name: string): boolean {\n const b = getBreaker(name);\n if (b.state === \"OPEN\" && Date.now() >= b.nextAttempt) {\n b.state = \"HALF_OPEN\";\n }\n return b.state === \"HALF_OPEN\" || b.state === \"CLOSED\";\n}\n\nfunction reportBreaker(name: string, success: boolean) {\n const b = getBreaker(name);\n if (success) {\n b.failCount = 0;\n b.state = \"CLOSED\";\n } else {\n b.failCount++;\n if (b.failCount >= MAX_FAILS || b.state === \"HALF_OPEN\") {\n b.state = \"OPEN\";\n b.nextAttempt = Date.now() + BREAK_DURATION_MS;\n b.failCount = 0;\n }\n }\n}\n\n// ============================================================================\n// 4. CORE INFERENCE ENGINE\n// ============================================================================\nasync function invokeProvider(\n provider: Provider,\n targetModel: string,\n messages: any[],\n tools?: any[],\n toolChoice?: any,\n temperature: number = 0.0,\n maxTokens: number = 1024000,\n) {\n const key = Deno.env.get(provider.keyEnv!)!;\n const url = `${provider.base}/chat/completions`;\n\n const payload: any = {\n model: targetModel,\n messages,\n temperature: temperature,\n stream: false,\n max_tokens: maxTokens,\n };\n if (tools && tools.length > 0) {\n payload.tools = tools;\n if (toolChoice) payload.tool_choice = toolChoice;\n }\n\n const controller = new AbortController();\n const timeoutId = setTimeout(() => controller.abort(), HARD_TIMEOUT_MS);\n\n try {\n const t0 = now();\n const res = await fetch(url, {\n method: \"POST\",\n headers: {\n \"Authorization\": `Bearer ${key}`,\n \"Content-Type\": \"application/json\",\n \"HTTP-Referer\": \"https://val.town\",\n \"X-Title\": \"Ziggyware-OmniRouter\",\n },\n body: JSON.stringify(payload),\n signal: controller.signal,\n });\n clearTimeout(timeoutId);\n if (!res.ok) {\n throw new Error(\n `HTTP ${res.status}: ${await res.text().catch(() => \"Unknown\")}`,\n );\n }\n\n const json = await res.json();\n if (!json.choices?.[0]?.message) throw new Error(\"Malformed JSON response\");\n\n return {\n message: json.choices[0].message,\n latencyMs: Math.round(now() - t0),\n modelUsed: targetModel,\n provider: provider,\n };\n } catch (e: any) {\n clearTimeout(timeoutId);\n throw new Error(\n e.name === \"AbortError\" ? \"API_STALLED_TIMEOUT\" : e.message,\n );\n }\n}\n\nasync function routeInference(\n messages: any[],\n tools?: any[],\n toolChoice?: any,\n temperature: number = 0.0,\n maxTokens: number = 1024000,\n) {\n while (inflight >= MAX_CONCURRENCY) await sleep(10);\n inflight++;\n\n try {\n const savedModels = await getSavedModels();\n const enabledProviders = getProviderArray()\n .filter((p) => Deno.env.get(p.keyEnv!))\n .sort((a, b) => a.priority - b.priority);\n\n if (enabledProviders.length === 0) {\n throw new Error(\"NO_PROVIDERS_CONFIGURED\");\n }\n\n for (const prov of enabledProviders) {\n if (!checkBreaker(prov.name)) continue;\n const targetModel = savedModels[prov.name] || prov.fallbackModel!;\n\n try {\n const { message, latencyMs, modelUsed, provider } =\n await invokeProvider(\n prov,\n targetModel,\n messages,\n tools,\n toolChoice,\n temperature,\n maxTokens,\n );\n\n reportBreaker(provider.name, true);\n\n return { message, provider, model: modelUsed, latencyMs };\n } catch (e: any) {\n reportBreaker(prov.name, false);\n console.warn(\n `[SEVERED] ${prov.name} failed (${e.message}). Routing next...`,\n );\n }\n }\n throw new Error(\"ALL_PROVIDERS_EXHAUSTED_OR_OPEN\");\n } finally {\n inflight--;\n }\n}\n\nexport default async function handler(req: Request): Promise {\n const url = new URL(req.url);\n\n // CORS & Preflight\n if (req.method === \"OPTIONS\") {\n return new Response(null, {\n headers: {\n \"Access-Control-Allow-Origin\": \"*\",\n \"Access-Control-Allow-Headers\": \"*\",\n \"Access-Control-Allow-Methods\": \"POST, GET, OPTIONS\",\n },\n });\n }\n\n const init = await initDB();\n if (!init.ok) {\n return Response.json({ error: `DB init failed: ${init.error}` }, {\n status: 500,\n });\n }\n\n if (req.method === \"GET\") {\n if (url.pathname === \"/api/providers\") {\n const savedModels = await getSavedModels();\n const status = getProviderArray().map((p) => ({\n ...p,\n hasKey: !!Deno.env.get(p.keyEnv!),\n currentModel: savedModels[p.name] || p.fallbackModel,\n })).sort((a, b) => a.priority - b.priority);\n return Response.json(status);\n }\n\n if (url.pathname === \"/api/models\") {\n const pName = url.searchParams.get(\"provider\");\n const p = getProviderArray().find((x) => x.name === pName);\n if (!p) {\n return Response.json({ error: \"Provider not found\" }, { status: 404 });\n }\n const key = Deno.env.get(p.keyEnv!);\n if (!key) {\n return Response.json({ error: \"Missing API Key\" }, { status: 401 });\n }\n\n try {\n const res = await fetch(`${p.base}/models`, {\n headers: { \"Authorization\": `Bearer ${key}` },\n });\n if (!res.ok) throw new Error(`HTTP ${res.status}`);\n const data = await res.json();\n return Response.json(data.data || data);\n } catch (e: any) {\n return Response.json({ error: e.message, fallback: p.fallbackModel }, {\n status: 500,\n });\n }\n }\n\n return new Response(\n `\n \n \n \n \n \n Ziggyware Router Matrix\n \n \n \n \n \n \n \n

Invariant Lab

Multi‑provider claim graphs · invariants · disagreement spectra · ?mode=registry for JSON

\n \n

Technique

${TECHNIQUES[0].name}

${TECHNIQUES[0].goal}

Strata

Syntactic: ${\n TECHNIQUES[0].strata.syntactic\n }
Semantic: ${TECHNIQUES[0].strata.semantic}
Structural: ${\n TECHNIQUES[0].strata.structural\n }

Bayesian Δ & Invariants

${\n TECHNIQUES[0].bayesianDelta\n }

${x}

\n\n\n`;\n\n return new Response(html, {\n status: 200,\n headers: { \"Content-Type\": \"text/html; charset=utf-8\" },\n });\n}",[],[638],"ea7395c3-7068-415a-9d70-79b400a99904","README.md","01a0a661-f049-7779-b407-5b764a0def22",["D",1789497831502],"# free-ai — omni router\n\nOpenAI-compatible endpoint that fans out across every configured provider with per-instance circuit breaking, model-level fallback, and live model discovery. Endpoint: `https://router.val.run`.\n\n## Call it\n\n```ts\nimport { omni } from \"https://esm.town/v/ziggyware/free-ai/omni-client.ts\";\nconst r = await omni.chat([{ role: \"user\", content: \"hi\" }], { model: \"coder\" });\nfor await (const tok of omni.stream(msgs)) process.stdout.write(tok);\n```\n\nOr any OpenAI SDK with `baseURL: \"https://router.val.run/v1\"`. Provenance comes back in `_omni_meta` (JSON) and `X-Omni-Provider` / `X-Omni-Model` headers (both modes).\n\n## `model` field\n\n| value | routes to |\n|---|---|\n| `auto` / `best` / empty | every free instance, **best model first** (catalog `QUALITY` prior; ties → vendor priority) |\n| `fast` | every instance in vendor priority (latency) order, catalog model order |\n| `coder` `reasoning` `long` `vision` | tagged vendors first, quality-ranked; `coder` boosts ids matching /coder/ |\n| `groq` | that vendor's instances (all keys) |\n| `groq#1` | one key slot |\n| `groq:openai/gpt-oss-120b` or `groq/openai/gpt-oss-120b` | that vendor, forced upstream id |\n\nUnknown values → 404, never silently rerouted.\n\nAll standard OpenAI params pass through (`tools`, `tool_choice`, `response_format`, `temperature`, `max_tokens`, `reasoning_effort`, `stop`, `top_p`, `seed`, …). Reasoning models (gpt-oss, qwen3.x) spend `max_tokens` on thinking first — send ≥256 or `reasoning_effort: \"low\"`, or you get `finish_reason: \"length\"` with empty content.\n\n## Keys (env / secrets)\n\nFor each vendor's `keyEnv` (see `catalog-free.ts`, `catalog-paid.ts`): `GROQ_API_KEY`, `GROQ_API_KEY1` … `GROQ_API_KEY9`, and any of those may hold `k1,k2,k3`. Every key becomes an instance (`groq`, `groq#1`, …) with its own cooldown. `cloudflare` also needs `CF_ACCOUNT_ID`.\n\n| env | effect |\n|---|---|\n| `OMNI_PROVIDERS` | JSON array of vendor rows, merged by `vendor` — add a vendor or override `models`/`priority`/`base` |\n| `OMNI_DISABLE` | `vendor,vendor` to skip |\n| `OMNI_QUALITY` | `{\"regex\":score}` — override/add quality scores (0–100) used by `auto` |\n| `OMNI_CLIENT_KEYS` | `k1,k2` — when set, callers must send `Authorization: Bearer k` (or `?key=`); `/` and `/health` stay open |\n\n## Failure policy\n\n| upstream | action |\n|---|---|\n| 404 / model-not-found / tier-not-allowed | mark that model dead 1 h, try next model; after the list, pick from live `/models` by the vendor's `prefer` regex and remember it |\n| 413 / \"Request too large\" / \"Limit N, Requested M\" | shrink `max_tokens` to fit and retry the same model; an output ceiling (\"OTPM\", \"max_tokens\") is remembered 15 min per model (`caps` in `/health`) |\n| 429 | cool the instance for `Retry-After` (≤15 min) or 60 s |\n| 401 / 403 | cool the instance 1 h |\n| 5xx / timeout (120 s) / network | breaker: 2 strikes → 25 s |\n\nAfter the first pass over every vendor, the router retries up to 2 more rounds — immediately for transient failures, or after waiting for the soonest cooldown when that fits the deadline. With `omni_deadline_ms` set, a vendor that cannot finish in the time left is skipped while a faster one is queued, and `max_tokens` is capped to what the vendor can stream before the deadline (a `finish_reason: \"length\"` reply beats a timeout — the client continues it).\n\nState lives in `omni_state` (SQLite) so it survives isolate recycling. `GET /health` shows what is cooling, dead, resolved, and capped.\n\n## Routes\n\n`GET /health` · `GET /v1/models` · `GET /api/providers` · `GET /api/models?provider=` · `POST /v1/chat/completions` (also `/`, `/chat/completions`) · `POST /api/update {provider, model}` binds a vendor's default · `GET /` matrix UI.\n",[],[645],["D",1777220541701],["D",1789508444194],{"_11":12,"_124":73},{"_11":12,"_120":-5,"_72":73,"_30":71,"_121":122,"_123":24},{"_38":39,"_13":16,"_32":35,"_111":75,"_112":62,"_113":114,"_115":79,"_116":75,"_117":41,"_118":14,"_119":12},"ctx",{"_51":-5,"_145":689,"_690":691,"_692":693,"_694":24,"_695":24,"_696":24,"_697":143,"_698":-5,"_699":-7,"_700":-7,"_701":26,"_702":24,"_703":24,"_704":-5,"_705":24,"_706":24,"_707":24,"_708":-5},"ENV",{"_660":661,"_662":663,"_664":63,"_665":666,"_667":668,"_669":670,"_671":672,"_673":674,"_675":676,"_677":678,"_679":680,"_681":682,"_683":684,"_685":686,"_687":688},"REACT_APP_SENTRY_DSN","https://586020d86f05489db63a554690acf1e9@o1398359.ingest.sentry.io/6724499","RENDER_GIT_COMMIT","fdb67791413f4386b4dbf18f354029a617f6160f","RENDER_GIT_BRANCH","POSTHOG_API_TOKEN","phc_nmoyuAkStEZuUp69rYlg1DZgrIt19Gx0Y4Hwf8tGgJP","POSTHOG_API_HOST","https://wart.val.town","WEB_API_TEMPLATE","https://{val}.web.val.run","APP_URL","https://www.val.town","MODULE_URL","https://esm.town","VANITY_WEB_API_TEMPLATE","https://{label}.val.run/","DENOLS_URL","https://lsp.val.town","NODE_ENV","production","CLOUDFLARE_TURNSTILE_SITE_KEY","0x4AAAAAACtaIZJH7s9rBMZ6","GOOGLE_CLIENT_ID","875989696899-jtsqkhlbdv5jrlsk8631p7ipjg3c6o62.apps.googleusercontent.com","STRIPE_PUBLISHABLE_KEY","pk_live_51N2zvWEpZTTLipIkVEwXrUG9AT0yeOOEYxszarZT1jpsZV58Ly6wlg3SltZpHN26CUalOBmUSY7adQRQiSkCZz2500qSXNkf8Y","API_URL","https://api.val.town",[],"preferencesKeybindings","default","preferencesTheme","light","preferencesFixers","preferencesCopilot","preferencesSelectionAi","preferencesCloseBrackets","preferencesChatPanelState","preferencesDprint","preferencesDenofmt","notificationCount","flagDenols","isBusinessOrgMember","businessOrgHandle","isBusinessOrgAdmin","isHigherTownieTier","navCollapsed","currentOrgId"]

History · prompts.ts