[{"_1":2,"_3":-5,"_4":-5},"loaderData",{"_5":6,"_7":8,"_9":10},"actionData","errors","root",{"_880":881,"_882":883},"routes/_app.x.$handle.$name",{"_98":97,"_32":45,"_106":88,"_27":107,"_13":14,"_108":-5,"_11":109,"_110":111,"_112":113,"_114":115,"_116":117,"_20":-5,"_118":-5,"_119":-5,"_120":25,"_121":122,"_123":124,"_125":-5,"_126":-5,"_127":128},"routes/_app.x.$handle.$name.branch.$branch.version.$version.diff",{"_11":12,"_13":14,"_15":16,"_17":-5,"_18":-5},"project",{"_43":44,"_32":45,"_38":46,"_47":-5,"_48":25,"_49":50,"_51":50,"_52":53,"_54":-5,"_18":55,"_56":-5,"_57":25,"_58":-5,"_59":-5,"_60":61,"_62":63,"_64":65,"_66":67,"_68":69,"_70":-5,"_71":72,"_73":72,"_74":75,"_76":75,"_77":75,"_78":72,"_79":72,"_80":81,"_82":83,"_84":85,"_86":25},"version",1526,"changes",[19],"aiChatInfo","userId",{"_20":21,"_22":23,"_24":25,"_26":25,"_27":28,"_29":28},"file",{"_30":31,"_32":28,"_33":34,"_35":37,"_38":39,"_40":14,"_41":42,"_24":25},"previous",{"_30":31,"_32":28,"_33":34,"_35":36},"deleted",false,"deletedHere","path","router-core.ts","latestPath","fileId","6a4ef8c6-ab4d-11f1-9c11-1607ee4eb77e","name","fileType","script","content","import {\n getProviderArray,\n type ProviderInstance,\n type Resolved,\n resolveTargets,\n} from \"./inference-provider.ts\";\nimport { quality } from \"./catalog.ts\";\nimport { discover } from \"./router-discover.ts\";\nimport {\n attempt,\n estimateInputTokens,\n type Fail,\n inputBound,\n outputLimit,\n shrinkMaxTokens,\n} from \"./router-attempt.ts\";\nimport {\n boundFor,\n capFor,\n capInFor,\n COOL,\n coolFor,\n coolInfo,\n isCool,\n isDead,\n isDeadOn,\n loadState,\n markDead,\n markDeadOn,\n noteFailure,\n noteStat,\n noteSuccess,\n resolvedFor,\n setCap,\n setCapIn,\n setResolved,\n soonestCooldown,\n statFor,\n tpsFor,\n unbind,\n} from \"./router-state.ts\";\n// OMNI_KEY_POLICY: \"depth\" (default) exhausts every key slot of a vendor before the next vendor;\n// \"breadth\" tries slot 0 of every vendor first, then slot 1 of every vendor, …\nconst res429 = (e: Error) => /^HTTP 429/.test(e.message);\n// …or \"rr\": spread load — the least-used slot of a vendor goes first. Per-request override: omni_key_policy.\n/** Upper bound on vendor calls a single request may make while walking the roster: 40-line failure lists were 40 upstream calls\n * in ~50 s, which is what trips edge blocks and per-second limits. Override with OMNI_MAX_ATTEMPTS. */\nconst MAX_ATTEMPTS = Math.max(\n 6,\n Number(Deno.env.get(\"OMNI_MAX_ATTEMPTS\") ?? 12) || 12,\n);\n/** Real attempts one request may spend on a single vendor (all its key slots together) before moving to the next vendor:\n * five OpenRouter keys all hitting the same upstream 429 must not consume the whole request budget while together/nvidia sit untried. */\n/** Soft cap on one attempt while other candidates remain (ms). Env OMNI_ATTEMPT_CAP_MS. */\nconst ATTEMPT_CAP_MS = Math.max(\n 8_000,\n Number(Deno.env.get(\"OMNI_ATTEMPT_CAP_MS\") ?? 25_000) || 25_000,\n);\nconst MAX_PER_VENDOR = Math.max(\n 2,\n Number(Deno.env.get(\"OMNI_MAX_PER_VENDOR\") ?? 3) || 3,\n);\n/** Minimum spacing between calls to the same instance, per vendor (ms). Free tiers that limit requests per second. */\nconst MIN_GAP_MS: Record = {\n mistral: 1100,\n nvidia_nim: 700,\n siliconflow: 700,\n together: 350,\n huggingface: 500,\n};\nconst KEY_POLICY =\n ([\"breadth\", \"rr\"].includes(Deno.env.get(\"OMNI_KEY_POLICY\") ?? \"\")\n ? Deno.env.get(\"OMNI_KEY_POLICY\")\n : \"depth\") as \"depth\" | \"breadth\" | \"rr\";\nconst VENDOR_ORDER = (Deno.env.get(\"OMNI_VENDOR_ORDER\") ?? \"\").split(\",\").map((\n s,\n) => s.trim()).filter(Boolean);\n\nexport const MAX_CONCURRENCY = 3;\nexport const HARD_TIMEOUT_MS = 180_000;\nexport let inflight = 0;\nconst sleep = (ms: number) => new Promise((r) => setTimeout(r, ms));\n\nexport type { RouteOpts } from \"./router-attempt.ts\";\nimport type { RouteOpts } from \"./router-attempt.ts\";\nconst etaMs = (p: ProviderInstance, o: RouteOpts) =>\n (p.ttft ?? 1500) +\n ((o.expectTokens ?? (Number(o.extra?.max_tokens) || 1000)) / (p.tps ?? 60)) *\n 1000;\nexport type RouteHit = {\n provider: ProviderInstance;\n model: string;\n latencyMs: number;\n res: Response;\n attempts: number;\n sentMax?: number;\n};\nconst TOOL_SAFE =\n /gpt-oss|deepseek|kimi|glm|gemini|claude|gpt-5|groq|grok|command-a|mistral/i; // reliable structured tool calls\nconst score = (id: string, boost?: RegExp, tools = false) =>\n quality(id) + (boost?.test(id) ? 6 : 0) +\n (tools && TOOL_SAFE.test(id) ? 5 : 0);\n// bound (UI) model leads; then quality or catalog order; models whose learned output\n// ceiling is below the expected length go last; discovered id last of all.\nfunction candidates(\n p: ProviderInstance,\n r: Resolved,\n o: RouteOpts = {},\n): string[] {\n if (r.forcedModel) return [r.forcedModel];\n const tools = !!(o.extra?.tools as unknown[] | undefined)?.length;\n const list = r.rank === \"quality\"\n ? [...p.models].sort((a, b) =>\n score(b, r.boost, tools) - score(a, r.boost, tools)\n )\n : [...p.models];\n const need = o.expectTokens ?? 0;\n const fitsCap = (m: string) => capFor(p.vendor, m) >= need;\n const ordered = [...list.filter(fitsCap), ...list.filter((m) => !fitsCap(m))];\n const seq = [boundFor(p.vendor), ...ordered, resolvedFor(p.vendor)].filter((\n m,\n ): m is string => !!m);\n return [...new Set(seq)].filter((m) =>\n !isDead(p.vendor, m) && !isDeadOn(p.name, m)\n ); // vendor-wide (404) or this key's daily quota for that model\n}\n\nexport async function routeInference(\n messages: unknown[],\n o: RouteOpts = {},\n): Promise {\n while (inflight >= MAX_CONCURRENCY) await sleep(10);\n inflight++;\n const errors: string[] = [];\n let attempts = 0;\n let spent = 0;\n const spentBy: Record =\n {}; /* real attempts per vendor this request */ /* real inference attempts (rate/transient/size), for MAX_ATTEMPTS */\n const t0 = performance.now();\n const left = () =>\n (o.deadlineMs ?? HARD_TIMEOUT_MS) - (performance.now() - t0);\n try {\n await loadState();\n const live = getProviderArray();\n if (!live.length) {\n throw Object.assign(new Error(\"NO_PROVIDERS_CONFIGURED\"), {\n status: 503,\n });\n }\n const resolved = resolveTargets(o.model, live);\n const { forcedModel } = resolved;\n const inTok = estimateInputTokens(messages, o.extra);\n const tooBig = new Set(); // vendors whose tier cannot take this prompt (learned this request)\n if (!resolved.targets.length) {\n throw Object.assign(new Error(`UNKNOWN_MODEL: ${o.model}`), {\n status: 404,\n });\n }\n for (let round = 0; round < 3; round++) {\n if (round) {\n const wait = soonestCooldown(resolved.targets.map((p) => p.name));\n if (wait > left() - 3000) break;\n if (wait > 0) await sleep(wait + 50);\n await loadState(true);\n errors.push(`--- retry round ${round + 1}`);\n for (const k of Object.keys(spentBy)) {\n spentBy[k] =\n 0; /* per-vendor budget is per round; the global cap still holds */\n }\n }\n for (const p of resolved.targets) {\n if (isCool(p.name)) {\n const c = coolInfo(p.name);\n errors.push(\n `${p.name}: cooling ${\n Math.ceil(((c?.until ?? 0) - Date.now()) / 1000)\n }s — ${c?.reason ?? \"\"}`.slice(0, 160),\n );\n }\n }\n let targets = resolved.targets.filter((p) => !isCool(p.name));\n // per-request exclusion (the app excludes an instance that just returned an empty/garbage reply); ignored if it would leave nothing\n if (o.exclude?.length) {\n const kept = targets.filter((p) =>\n !o.exclude!.includes(p.name)\n ); /* instance names only: slot 0 is named like its vendor */\n if (kept.length) {\n for (const p of targets) {\n if (!kept.includes(p)) {\n errors.push(`${p.name}: excluded by caller`);\n }\n }\n targets = kept;\n }\n }\n // free before paid; those that can make the deadline before those that can't; then quality (auto) or latency priority (fast)\n const fits = (p: ProviderInstance) =>\n !o.deadlineMs || etaMs(p, o) <= o.deadlineMs;\n const best = new Map(\n targets.map((p) => [\n p.name,\n resolved.rank === \"quality\"\n ? Math.max(\n -1,\n ...candidates(p, resolved, o).map((m) =>\n score(\n m,\n resolved.boost,\n !!(o.extra?.tools as unknown[] | undefined)?.length,\n )\n ),\n )\n : 0,\n ]),\n );\n const policy = o.keyPolicy ?? KEY_POLICY;\n const used = (p: ProviderInstance) => statFor(p.name)?.req ?? 0;\n // Explicit vendor order (OMNI_VENDOR_ORDER env or per-request omni_vendor_order): listed vendors go first, in that\n // order, every key slot of one before the next vendor, and the deadline-fit heuristic does not demote them — the\n // user chose them; a slow vendor gets a capped max_tokens and the app continues the reply. Unlisted vendors keep the\n // quality/priority ranking behind them.\n const order = o.vendorOrder ?? VENDOR_ORDER;\n const rankOf = (p: ProviderInstance) => {\n const i = order.indexOf(p.vendor);\n return i === -1 ? order.length : i;\n };\n targets = [...targets].sort((a, b) =>\n (+!!a.paid - +!!b.paid) || rankOf(a) - rankOf(b) ||\n (rankOf(a) < order.length ? 0 : (+!fits(a) - +!fits(b))) ||\n (policy === \"breadth\" ? a.slot - b.slot : 0) ||\n best.get(b.name)! - best.get(a.name)! || a.priority - b.priority ||\n (policy === \"rr\" ? used(a) - used(b) : 0) || a.slot - b.slot\n );\n for (const [idx, p] of targets.entries()) {\n if (left() < 1500) break;\n if (tooBig.has(p.vendor)) continue; // every key slot of a vendor shares the tier's prompt ceiling\n if (p.maxIn && inTok > p.maxIn) {\n errors.push(\n `${p.name}: skipped (prompt ≈${inTok} tok > maxIn ${p.maxIn})`,\n );\n continue;\n }\n // a vendor that cannot finish in the time left is skipped while a faster one is still queued behind it\n if (\n o.deadlineMs && rankOf(p) === order.length && etaMs(p, o) > left() &&\n targets.slice(idx + 1).some((q) => etaMs(q, o) <= left())\n ) {\n errors.push(\n `${p.name}: skipped (eta ${Math.round(etaMs(p, o) / 1000)}s > ${\n Math.round(left() / 1000)\n }s left)`,\n );\n continue;\n }\n const shrinkCount = new Map();\n const models = candidates(p, resolved, o).filter((m) =>\n capInFor(p.vendor, m) >= inTok\n );\n if (!models.length && !forcedModel) {\n errors.push(\n `${p.name}: skipped (prompt ≈${inTok} tok over every model's learned input limit)`,\n );\n continue;\n }\n let stop = false;\n if (spent >= MAX_ATTEMPTS) {\n errors.push(\n `attempt cap ${MAX_ATTEMPTS} reached — remaining instances not tried`,\n );\n break;\n }\n if ((spentBy[p.vendor] ?? 0) >= MAX_PER_VENDOR) {\n errors.push(\n `${p.name}: skipped (${MAX_PER_VENDOR} attempts already spent on ${p.vendor} this request)`,\n );\n continue;\n }\n // Pacing: free tiers with a per-second limit (mistral: 1 req/s) fail on bursts, and a swarm is a burst. Space calls to the same\n // instance by its vendor's minimum gap — a short wait beats a 429 that cools the key for a minute.\n const gap = MIN_GAP_MS[p.vendor] ?? 0;\n const sinceLast = Date.now() - (statFor(p.name)?.last ?? 0);\n if (gap && sinceLast < gap) {\n if (left() < gap - sinceLast + 3000) {\n errors.push(`${p.name}: paced (last call ${sinceLast} ms ago)`);\n continue;\n }\n await new Promise((r) => setTimeout(r, gap - sinceLast));\n }\n for (let i = 0; i < models.length + 1 && !stop; i++) {\n const model = i < models.length\n ? models[i]\n : (forcedModel ? null : await discover(p));\n if (!model) break;\n if (i === models.length) await setResolved(p.vendor, model);\n if (left() < 1500) {\n errors.push(\"DEADLINE\");\n stop = true;\n break;\n }\n if (\n spent >= MAX_ATTEMPTS || (spentBy[p.vendor] ?? 0) >= MAX_PER_VENDOR\n ) {\n stop = true;\n break;\n }\n attempts++;\n const t0 = performance.now();\n try {\n const tps = tpsFor(p.vendor, model) ?? p.tps ?? 60;\n const capMax = Math.min(\n capFor(p.vendor, model),\n o.deadlineMs && !o.stream\n ? Math.max(\n 1024,\n Math.floor(((left() - (p.ttft ?? 1500)) / 1000) * tps * 0.85),\n )\n : Infinity,\n );\n // One attempt may not eat the whole deadline while other instances are still queued: cap it so a slow model\n // leaves room for at least one more try elsewhere (floor 8 s; the last candidate gets everything that is left).\n const others = idx < targets.length - 1 || i < models.length - 1;\n const perAttempt = others\n ? Math.min(\n left(),\n Math.max(8_000, Math.min(left() * 0.6, ATTEMPT_CAP_MS)),\n )\n : left();\n const res = await attempt(\n p,\n model,\n messages,\n o,\n Math.min(HARD_TIMEOUT_MS, perAttempt),\n capMax,\n );\n await Promise.all([\n noteSuccess(p.name),\n noteStat(p.name, true, model),\n ]);\n return {\n provider: p,\n model,\n latencyMs: Math.round(performance.now() - t0),\n res,\n attempts,\n sentMax: Number(res.headers.get(\"x-omni-sent-max\")) ||\n Number(o.extra?.max_tokens) || undefined,\n };\n } catch (e: any) {\n errors.push(`${p.name}/${model}: ${e.message}`);\n await noteStat(p.name, false, model, e.message);\n // quota/auth/model failures are instant rejections and do not count as spent capacity; everything else does\n if (!/^(quota|auth|model)$/.test(String((e as Fail).kind))) {\n spent++;\n spentBy[p.vendor] = (spentBy[p.vendor] ?? 0) + 1;\n }\n switch ((e as Fail).kind) {\n case \"size\": {\n const inLim = inputBound(e.message);\n if (inLim !== null) { // prompt-bound: no max_tokens shrink can help; remember the ceiling, skip the whole vendor this request\n if (inLim > 0) await setCapIn(p.vendor, model, inLim - 64);\n tooBig.add(p.vendor);\n errors.push(\n `${p.name}: prompt too large for this tier (≈${inTok} tok${\n inLim ? ` > ${inLim}` : \"\"\n }) — skipping vendor`,\n );\n stop = true;\n break;\n }\n const lim = outputLimit(e.message);\n if (lim) await setCap(p.vendor, model, Math.max(256, lim - 64));\n const shrinks = shrinkCount.get(model) ?? 0;\n shrinkCount.set(model, shrinks + 1);\n if (shrinks < 2 && shrinkMaxTokens(o, e.message, e.sentMax)) {\n i--;\n errors.push(`shrunk max_tokens→${o.extra!.max_tokens}`);\n }\n break;\n }\n case \"model\":\n if (forcedModel) {\n stop = true;\n break;\n }\n await markDead(p.vendor, model);\n if (model === boundFor(p.vendor)) await unbind(p.vendor);\n break;\n case \"rate\": {\n // OpenRouter relays an UPSTREAM per-model limit (\"Provider returned error\" / \"temporarily rate-limited upstream\"): every key\n // sees the same wall, so park the model vendor-wide for 10 min and try this key's next model instead of burning the other keys.\n if (\n /provider returned error|rate-limited upstream|temporarily rate.?limited/i\n .test(e.message)\n ) {\n await markDead(p.vendor, model, Date.now() + 6_000);\n errors.push(\n `${p.name}: ${model} rate-limited upstream (10 min), trying next model`,\n );\n break;\n }\n await coolFor(\n p.name,\n e.retryAfterMs ?? COOL.rate,\n \"rate \" + e.message.slice(0, 80),\n );\n stop = true;\n break;\n }\n case \"auth\":\n await coolFor(\n p.name,\n COOL.auth,\n \"auth \" + e.message.slice(0, 80),\n );\n stop = true;\n break;\n // quota: this KEY is spent until its reset; sibling slots (p.vendor#n) are next in `targets` under depth policy\n case \"quota\": {\n const until = Math.max(\n Date.now() + 6_000,\n (e as Fail).resetAt ?? Date.now() + 6 * 1_000,\n );\n // Groq-style per-MODEL daily quota (\"Rate limit reached for model `x` … per day\"): only that model on this key is spent; try the key's other models.\n const perModel = e.message.match(\n /for model [`'\"]?([\\w./:-]+)/i,\n );\n if (perModel && res429(e)) {\n await markDeadOn(p.name, model, until);\n errors.push(\n `${p.name}: ${model} spent until reset, trying next model`,\n );\n break;\n }\n await coolFor(\n p.name,\n until - Date.now(),\n \"quota \" + e.message.slice(0, 160),\n );\n stop = true;\n break;\n }\n default:\n // A TIMEOUT is the MODEL being too slow for the budget, not the key being broken: park that model vendor-wide for\n // 30 min (every key sees the same throughput) and try this key's next model. Other transient errors move to the next key.\n if (e.message === \"TIMEOUT\" && !forcedModel) {\n await markDead(p.vendor, model, Date.now() + 1_800);\n errors.push(\n `${p.name}: ${model} timed out — parked 30 min, trying next model`,\n );\n break;\n }\n await noteFailure(p.name, e.message.slice(0, 60));\n stop = true;\n }\n }\n }\n }\n }\n throw Object.assign(new Error(\"ALL_PROVIDERS_EXHAUSTED_OR_COOLING\"), {\n status: 503,\n errors,\n attempts,\n });\n } finally {\n inflight--;\n }\n}","import { getProviderArray, type ProviderInstance, type Resolved, resolveTargets } from \"./inference-provider.ts\";\nimport { quality } from \"./catalog.ts\";\nimport { discover } from \"./router-discover.ts\";\nimport { attempt, estimateInputTokens, type Fail, inputBound, outputLimit, shrinkMaxTokens } from \"./router-attempt.ts\";\nimport { boundFor, capFor, capInFor, tpsFor, COOL, coolFor, coolInfo, isCool, isDead, isDeadOn, loadState, markDead, markDeadOn, noteFailure, noteStat, noteSuccess, resolvedFor, setCap, setCapIn, setResolved, soonestCooldown, statFor, unbind } from \"./router-state.ts\";\n// OMNI_KEY_POLICY: \"depth\" (default) exhausts every key slot of a vendor before the next vendor;\n// \"breadth\" tries slot 0 of every vendor first, then slot 1 of every vendor, …\nconst res429 = (e: Error) => /^HTTP 429/.test(e.message);\n// …or \"rr\": spread load — the least-used slot of a vendor goes first. Per-request override: omni_key_policy.\n/** Upper bound on vendor calls a single request may make while walking the roster: 40-line failure lists were 40 upstream calls\n * in ~50 s, which is what trips edge blocks and per-second limits. Override with OMNI_MAX_ATTEMPTS. */\nconst MAX_ATTEMPTS = Math.max(6, Number(Deno.env.get(\"OMNI_MAX_ATTEMPTS\") ?? 12) || 12);\n/** Real attempts one request may spend on a single vendor (all its key slots together) before moving to the next vendor:\n * five OpenRouter keys all hitting the same upstream 429 must not consume the whole request budget while together/nvidia sit untried. */\n/** Soft cap on one attempt while other candidates remain (ms). Env OMNI_ATTEMPT_CAP_MS. */\nconst ATTEMPT_CAP_MS = Math.max(8_000, Number(Deno.env.get(\"OMNI_ATTEMPT_CAP_MS\") ?? 25_000) || 25_000);\nconst MAX_PER_VENDOR = Math.max(2, Number(Deno.env.get(\"OMNI_MAX_PER_VENDOR\") ?? 3) || 3);\n/** Minimum spacing between calls to the same instance, per vendor (ms). Free tiers that limit requests per second. */\nconst MIN_GAP_MS: Record = { mistral: 1100, nvidia_nim: 700, siliconflow: 700, together: 350, huggingface: 500 };\nconst KEY_POLICY = ([\"breadth\", \"rr\"].includes(Deno.env.get(\"OMNI_KEY_POLICY\") ?? \"\") ? Deno.env.get(\"OMNI_KEY_POLICY\") : \"depth\") as \"depth\" | \"breadth\" | \"rr\";\nconst VENDOR_ORDER = (Deno.env.get(\"OMNI_VENDOR_ORDER\") ?? \"\").split(\",\").map((s) => s.trim()).filter(Boolean);\n\nexport const MAX_CONCURRENCY = 3;\nexport const HARD_TIMEOUT_MS = 60_000;\nexport let inflight = 0;\nconst sleep = (ms: number) => new Promise((r) => setTimeout(r, ms));\n\nexport type { RouteOpts } from \"./router-attempt.ts\";\nimport type { RouteOpts } from \"./router-attempt.ts\";\nconst etaMs = (p: ProviderInstance, o: RouteOpts) => (p.ttft ?? 1500) + ((o.expectTokens ?? (Number(o.extra?.max_tokens) || 1000)) / (p.tps ?? 60)) * 1000;\nexport type RouteHit = { provider: ProviderInstance; model: string; latencyMs: number; res: Response; attempts: number; sentMax?: number };\nconst TOOL_SAFE = /gpt-oss|deepseek|kimi|glm|gemini|claude|gpt-5|grok|command-a|mistral/i; // reliable structured tool calls\nconst score = (id: string, boost?: RegExp, tools = false) => quality(id) + (boost?.test(id) ? 6 : 0) + (tools && TOOL_SAFE.test(id) ? 5 : 0);\n// bound (UI) model leads; then quality or catalog order; models whose learned output\n// ceiling is below the expected length go last; discovered id last of all.\nfunction candidates(p: ProviderInstance, r: Resolved, o: RouteOpts = {}): string[] {\n if (r.forcedModel) return [r.forcedModel];\n const tools = !!(o.extra?.tools as unknown[] | undefined)?.length;\n const list = r.rank === \"quality\" ? [...p.models].sort((a, b) => score(b, r.boost, tools) - score(a, r.boost, tools)) : [...p.models];\n const need = o.expectTokens ?? 0;\n const fitsCap = (m: string) => capFor(p.vendor, m) >= need;\n const ordered = [...list.filter(fitsCap), ...list.filter((m) => !fitsCap(m))];\n const seq = [boundFor(p.vendor), ...ordered, resolvedFor(p.vendor)].filter((m): m is string => !!m);\n return [...new Set(seq)].filter((m) => !isDead(p.vendor, m) && !isDeadOn(p.name, m)); // vendor-wide (404) or this key's daily quota for that model\n}\n\nexport async function routeInference(messages: unknown[], o: RouteOpts = {}): Promise {\n while (inflight >= MAX_CONCURRENCY) await sleep(10);\n inflight++;\n const errors: string[] = [];\n let attempts = 0;\n let spent = 0;\n const spentBy: Record = {}; /* real attempts per vendor this request */ /* real inference attempts (rate/transient/size), for MAX_ATTEMPTS */\n const t0 = performance.now();\n const left = () => (o.deadlineMs ?? HARD_TIMEOUT_MS) - (performance.now() - t0);\n try {\n await loadState();\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 { forcedModel } = resolved;\n const inTok = estimateInputTokens(messages, o.extra);\n const tooBig = new Set(); // vendors whose tier cannot take this prompt (learned this request)\n if (!resolved.targets.length) throw Object.assign(new Error(`UNKNOWN_MODEL: ${o.model}`), { status: 404 });\n for (let round = 0; round < 3; round++) {\n if (round) {\n const wait = soonestCooldown(resolved.targets.map((p) => p.name));\n if (wait > left() - 3000) break;\n if (wait > 0) await sleep(wait + 50);\n await loadState(true);\n errors.push(`--- retry round ${round + 1}`); for (const k of Object.keys(spentBy)) spentBy[k] = 0; /* per-vendor budget is per round; the global cap still holds */\n }\n for (const p of resolved.targets) if (isCool(p.name)) { const c = coolInfo(p.name); errors.push(`${p.name}: cooling ${Math.ceil(((c?.until ?? 0) - Date.now()) / 1000)}s — ${c?.reason ?? \"\"}`.slice(0, 160)); }\n let targets = resolved.targets.filter((p) => !isCool(p.name));\n // per-request exclusion (the app excludes an instance that just returned an empty/garbage reply); ignored if it would leave nothing\n if (o.exclude?.length) { const kept = targets.filter((p) => !o.exclude!.includes(p.name)); /* instance names only: slot 0 is named like its vendor */ if (kept.length) { for (const p of targets) if (!kept.includes(p)) errors.push(`${p.name}: excluded by caller`); targets = kept; } }\n // free before paid; those that can make the deadline before those that can't; then quality (auto) or latency priority (fast)\n const fits = (p: ProviderInstance) => !o.deadlineMs || etaMs(p, o) <= o.deadlineMs;\n const best = new Map(targets.map((p) => [p.name, resolved.rank === \"quality\" ? Math.max(-1, ...candidates(p, resolved, o).map((m) => score(m, resolved.boost, !!(o.extra?.tools as unknown[] | undefined)?.length))) : 0]));\n const policy = o.keyPolicy ?? KEY_POLICY;\n const used = (p: ProviderInstance) => statFor(p.name)?.req ?? 0;\n // Explicit vendor order (OMNI_VENDOR_ORDER env or per-request omni_vendor_order): listed vendors go first, in that\n // order, every key slot of one before the next vendor, and the deadline-fit heuristic does not demote them — the\n // user chose them; a slow vendor gets a capped max_tokens and the app continues the reply. Unlisted vendors keep the\n // quality/priority ranking behind them.\n const order = o.vendorOrder ?? VENDOR_ORDER;\n const rankOf = (p: ProviderInstance) => { const i = order.indexOf(p.vendor); return i === -1 ? order.length : i; };\n targets = [...targets].sort((a, b) => (+!!a.paid - +!!b.paid) || rankOf(a) - rankOf(b) || (rankOf(a) < order.length ? 0 : (+!fits(a) - +!fits(b))) || (policy === \"breadth\" ? a.slot - b.slot : 0) || best.get(b.name)! - best.get(a.name)! || a.priority - b.priority || (policy === \"rr\" ? used(a) - used(b) : 0) || a.slot - b.slot);\n for (const [idx, p] of targets.entries()) {\n if (left() < 1500) break;\n if (tooBig.has(p.vendor)) continue; // every key slot of a vendor shares the tier's prompt ceiling\n if (p.maxIn && inTok > p.maxIn) { errors.push(`${p.name}: skipped (prompt ≈${inTok} tok > maxIn ${p.maxIn})`); continue; }\n // a vendor that cannot finish in the time left is skipped while a faster one is still queued behind it\n if (o.deadlineMs && rankOf(p) === order.length && etaMs(p, o) > left() && targets.slice(idx + 1).some((q) => etaMs(q, o) <= left())) { errors.push(`${p.name}: skipped (eta ${Math.round(etaMs(p, o) / 1000)}s > ${Math.round(left() / 1000)}s left)`); continue; }\n const shrinkCount = new Map();\n const models = candidates(p, resolved, o).filter((m) => capInFor(p.vendor, m) >= inTok);\n if (!models.length && !forcedModel) { errors.push(`${p.name}: skipped (prompt ≈${inTok} tok over every model's learned input limit)`); continue; }\n let stop = false;\n if (spent >= MAX_ATTEMPTS) { errors.push(`attempt cap ${MAX_ATTEMPTS} reached — remaining instances not tried`); break; }\n if ((spentBy[p.vendor] ?? 0) >= MAX_PER_VENDOR) { errors.push(`${p.name}: skipped (${MAX_PER_VENDOR} attempts already spent on ${p.vendor} this request)`); continue; }\n // Pacing: free tiers with a per-second limit (mistral: 1 req/s) fail on bursts, and a swarm is a burst. Space calls to the same\n // instance by its vendor's minimum gap — a short wait beats a 429 that cools the key for a minute.\n const gap = MIN_GAP_MS[p.vendor] ?? 0; const sinceLast = Date.now() - (statFor(p.name)?.last ?? 0);\n if (gap && sinceLast < gap) { if (left() < gap - sinceLast + 3000) { errors.push(`${p.name}: paced (last call ${sinceLast} ms ago)`); continue; } await new Promise((r) => setTimeout(r, gap - sinceLast)); }\n for (let i = 0; i < models.length + 1 && !stop; i++) {\n const model = i < models.length ? models[i] : (forcedModel ? null : await discover(p));\n if (!model) break;\n if (i === models.length) await setResolved(p.vendor, model);\n if (left() < 1500) { errors.push(\"DEADLINE\"); stop = true; break; }\n if (spent >= MAX_ATTEMPTS || (spentBy[p.vendor] ?? 0) >= MAX_PER_VENDOR) { stop = true; break; }\n attempts++;\n const t0 = performance.now();\n try {\n const tps = tpsFor(p.vendor, model) ?? p.tps ?? 60;\n const capMax = Math.min(capFor(p.vendor, model), o.deadlineMs && !o.stream ? Math.max(1024, Math.floor(((left() - (p.ttft ?? 1500)) / 1000) * tps * 0.85)) : Infinity);\n // One attempt may not eat the whole deadline while other instances are still queued: cap it so a slow model\n // leaves room for at least one more try elsewhere (floor 8 s; the last candidate gets everything that is left).\n const others = idx < targets.length - 1 || i < models.length - 1;\n const perAttempt = others ? Math.min(left(), Math.max(8_000, Math.min(left() * 0.6, ATTEMPT_CAP_MS))) : left();\n const res = await attempt(p, model, messages, o, Math.min(HARD_TIMEOUT_MS, perAttempt), capMax);\n await Promise.all([noteSuccess(p.name), noteStat(p.name, true, model)]);\n return { provider: p, model, latencyMs: Math.round(performance.now() - t0), res, attempts, sentMax: Number(res.headers.get(\"x-omni-sent-max\")) || Number(o.extra?.max_tokens) || undefined };\n } catch (e: any) {\n errors.push(`${p.name}/${model}: ${e.message}`);\n await noteStat(p.name, false, model, e.message);\n // quota/auth/model failures are instant rejections and do not count as spent capacity; everything else does\n if (!/^(quota|auth|model)$/.test(String((e as Fail).kind))) { spent++; spentBy[p.vendor] = (spentBy[p.vendor] ?? 0) + 1; }\n switch ((e as Fail).kind) {\n case \"size\": {\n const inLim = inputBound(e.message);\n if (inLim !== null) { // prompt-bound: no max_tokens shrink can help; remember the ceiling, skip the whole vendor this request\n if (inLim > 0) await setCapIn(p.vendor, model, inLim - 64);\n tooBig.add(p.vendor); errors.push(`${p.name}: prompt too large for this tier (≈${inTok} tok${inLim ? ` > ${inLim}` : \"\"}) — skipping vendor`); stop = true; break;\n }\n const lim = outputLimit(e.message); if (lim) await setCap(p.vendor, model, Math.max(256, lim - 64));\n const shrinks = (shrinkCount.get(model) ?? 0); shrinkCount.set(model, shrinks + 1);\n if (shrinks < 2 && shrinkMaxTokens(o, e.message, e.sentMax)) { i--; errors.push(`shrunk max_tokens→${o.extra!.max_tokens}`); } break;\n }\n case \"model\":\n if (forcedModel) { stop = true; break; }\n await markDead(p.vendor, model);\n if (model === boundFor(p.vendor)) await unbind(p.vendor);\n break;\n case \"rate\": {\n // OpenRouter relays an UPSTREAM per-model limit (\"Provider returned error\" / \"temporarily rate-limited upstream\"): every key\n // sees the same wall, so park the model vendor-wide for 10 min and try this key's next model instead of burning the other keys.\n if (/provider returned error|rate-limited upstream|temporarily rate.?limited/i.test(e.message)) { await markDead(p.vendor, model, Date.now() + 600_000); errors.push(`${p.name}: ${model} rate-limited upstream (10 min), trying next model`); break; }\n await coolFor(p.name, e.retryAfterMs ?? COOL.rate, \"rate \" + e.message.slice(0, 80)); stop = true; break;\n }\n case \"auth\": await coolFor(p.name, COOL.auth, \"auth \" + e.message.slice(0, 80)); stop = true; break;\n // quota: this KEY is spent until its reset; sibling slots (p.vendor#n) are next in `targets` under depth policy\n case \"quota\": {\n const until = Math.max(Date.now() + 60_000, (e as Fail).resetAt ?? Date.now() + 6 * 3_600_000);\n // Groq-style per-MODEL daily quota (\"Rate limit reached for model `x` … per day\"): only that model on this key is spent; try the key's other models.\n const perModel = e.message.match(/for model [`'\"]?([\\w./:-]+)/i);\n if (perModel && res429(e)) { await markDeadOn(p.name, model, until); errors.push(`${p.name}: ${model} spent until reset, trying next model`); break; }\n await coolFor(p.name, until - Date.now(), \"quota \" + e.message.slice(0, 160)); stop = true; break;\n }\n default:\n // A TIMEOUT is the MODEL being too slow for the budget, not the key being broken: park that model vendor-wide for\n // 30 min (every key sees the same throughput) and try this key's next model. Other transient errors move to the next key.\n if (e.message === \"TIMEOUT\" && !forcedModel) { await markDead(p.vendor, model, Date.now() + 1_800_000); errors.push(`${p.name}: ${model} timed out — parked 30 min, trying next model`); break; }\n await noteFailure(p.name, e.message.slice(0, 60)); stop = true;\n }\n }\n }\n }\n }\n throw Object.assign(new Error(\"ALL_PROVIDERS_EXHAUSTED_OR_COOLING\"), { status: 503, errors, attempts });\n } finally { inflight--; }\n}\n","createdAt",["D",1789251211713],"branchVersion","branchId","1a44bbd4-418c-11f1-af64-42b51c65c3df","id","1a430fbe-418c-11f1-af64-42b51c65c3df","free-ai",["D",1777220541701],"deletedAt","legacyEnvVars","privacy","public","httpPrivacy","description","Free ai chat with inference router.","imageUrl","2c6d46b0-54dd-11f0-a86c-6fbc2b29331a","pinnedFileId","pinPreview","pinMigrated","screenshotVersion","likeCount",0,"orgId","850db142-aa5e-4914-bceb-1e2e085ae4c5","tursoName","1a430fbe418c11f1af6442b51c65c3df","tursoHostname","1a430fbe418c11f1af6442b51c65c3df-val-town.aws-us-east-2.turso.io","tursoMeasuredAt",["D",1789508444194],"pinned","userPerm","r","valPerm","blobPerm","rw","emailPerm","sqlitePerm","projectPerm","telemetryPerm","user",{"_98":97,"_105":101},"org",{"_98":97,"_99":-5,"_100":101,"_18":55,"_102":103,"_104":25},"branch",{"_13":87,"_32":88,"_43":42,"_89":44,"_90":-5,"_91":-5,"_92":-5,"_93":-5,"_94":-5,"_95":45,"_96":97},"isOwned",1682,"main","projectId","forkedBranchId","forkedBranchVersion","forkedProjectPrivacy","forkedProjectId","forkedProjectOrgId","projectName","userHandle","ziggyware","handle","displayName","avatar","https://img.clerk.com/eyJ0eXBlIjoicHJveHkiLCJzcmMiOiJodHRwczovL2ltYWdlcy5jbGVyay5kZXYvb2F1dGhfZ2l0aHViL2ltZ18yekJHRDE5SGVoUHUzUENQaXY2MTRkT3h0cncifQ","tier","free","canRestrictVals","profileImageUrl","branchName","","adoptOrgId",{"_43":44,"_32":45,"_38":875,"_47":-5,"_48":25,"_49":50,"_51":50,"_52":53,"_54":-5,"_18":55,"_56":-5,"_57":25,"_58":-5,"_59":-5,"_60":61,"_62":63,"_64":65,"_66":67,"_68":876,"_70":-5,"_71":72,"_73":72,"_74":75,"_76":75,"_77":75,"_78":72,"_79":72,"_80":877,"_82":878,"_84":879,"_86":25},"tree",[638,639,640,641,642,643,644,645,646,647,648,649,650,651,652,653,654,655,656,657,658,659,660,661,662,663,664,665,666,667,668,669,670,671,672,673,674,675,676,677,678,679,680,681,682,683,684,685,686,687,688,689,690,691,692,693,694,695,696],"files",[638,639,640,641,642,643,644,645,646,647,648,649,650,651,652,653,654,655,656,657,658,659,660,661,662,663,664,665,666,667,668,669,670,671,672,673,674,675,676,677,678,679,680,681,682,683,684,685,686,687,688,689,690,691,692,693,694,695,696],"counts",{"_627":61,"_628":195,"_629":61,"_630":631,"_632":633},"contents",[129,130,131,132,133,134,135,136,137,138,139,140,141,142,143,144,145,146,147,148,149,150,151,152,153,154,155,156,157,158,159,160,161,162,163,164,165,166,167,168,169,170,171,172,173,174,175,176,177,178,179,180,181,182,183,184,185,186,187],"upstreamProject","upstreamBranch","emptyEnvVars","flags",[],"hasHttpFiles",true,"grants","orgs","customDomains",{},{"_43":620,"_32":621,"_30":622,"_191":-5,"_41":42,"_40":623,"_193":124,"_38":624,"_13":204,"_33":20,"_35":625,"_24":25,"_197":-5,"_27":626},{"_43":613,"_32":614,"_30":615,"_191":-5,"_41":42,"_40":616,"_193":124,"_38":617,"_13":339,"_33":34,"_35":618,"_24":25,"_197":-5,"_27":619},{"_43":606,"_32":607,"_30":608,"_191":-5,"_41":42,"_40":609,"_193":124,"_38":610,"_13":498,"_33":34,"_35":611,"_24":25,"_197":-5,"_27":612},{"_43":599,"_32":600,"_30":601,"_191":-5,"_41":42,"_40":602,"_193":124,"_38":603,"_13":324,"_33":34,"_35":604,"_24":25,"_197":-5,"_27":605},{"_43":592,"_32":593,"_30":594,"_191":-5,"_41":42,"_40":595,"_193":124,"_38":596,"_13":195,"_33":34,"_35":597,"_24":25,"_197":-5,"_27":598},{"_43":584,"_32":585,"_30":586,"_191":-5,"_41":42,"_40":587,"_193":124,"_38":588,"_13":589,"_33":34,"_35":590,"_24":25,"_197":-5,"_27":591},{"_43":575,"_32":576,"_30":577,"_191":-5,"_41":42,"_40":578,"_193":124,"_38":579,"_13":580,"_33":34,"_35":581,"_24":25,"_197":582,"_27":583},{"_43":567,"_32":568,"_30":569,"_191":-5,"_41":42,"_40":570,"_193":124,"_38":571,"_13":264,"_33":34,"_35":572,"_24":25,"_197":573,"_27":574},{"_43":560,"_32":561,"_30":562,"_191":-5,"_41":42,"_40":563,"_193":124,"_38":564,"_13":296,"_33":34,"_35":565,"_24":25,"_197":-5,"_27":566},{"_43":553,"_32":554,"_30":555,"_191":-5,"_41":42,"_40":556,"_193":124,"_38":557,"_13":204,"_33":34,"_35":558,"_24":25,"_197":-5,"_27":559},{"_43":545,"_32":546,"_30":547,"_191":-5,"_41":42,"_40":548,"_193":124,"_38":549,"_13":550,"_33":34,"_35":551,"_24":25,"_197":-5,"_27":552},{"_43":538,"_32":539,"_30":540,"_191":-5,"_41":42,"_40":541,"_193":25,"_38":542,"_13":304,"_33":34,"_35":543,"_24":25,"_197":-5,"_27":544},{"_43":531,"_32":532,"_30":533,"_191":-5,"_41":42,"_40":534,"_193":25,"_38":535,"_13":195,"_33":34,"_35":536,"_24":25,"_197":-5,"_27":537},{"_43":524,"_32":525,"_30":526,"_191":-5,"_41":42,"_40":527,"_193":25,"_38":528,"_13":61,"_33":34,"_35":529,"_24":25,"_197":-5,"_27":530},{"_43":517,"_32":518,"_30":519,"_191":-5,"_41":42,"_40":520,"_193":124,"_38":521,"_13":195,"_33":34,"_35":522,"_24":25,"_197":-5,"_27":523},{"_43":510,"_32":511,"_30":512,"_191":-5,"_41":42,"_40":513,"_193":25,"_38":514,"_13":324,"_33":34,"_35":515,"_24":25,"_197":-5,"_27":516},{"_43":501,"_32":502,"_30":503,"_191":-5,"_41":42,"_40":504,"_193":25,"_38":505,"_13":506,"_33":249,"_35":507,"_24":25,"_197":508,"_27":509},{"_43":493,"_32":494,"_30":495,"_191":-5,"_41":42,"_40":496,"_193":25,"_38":497,"_13":498,"_33":34,"_35":499,"_24":25,"_197":-5,"_27":500},{"_43":486,"_32":487,"_30":488,"_191":-5,"_41":42,"_40":489,"_193":124,"_38":490,"_13":316,"_33":34,"_35":491,"_24":25,"_197":-5,"_27":492},{"_43":479,"_32":480,"_30":481,"_191":-5,"_41":42,"_40":482,"_193":124,"_38":483,"_13":195,"_33":34,"_35":484,"_24":25,"_197":-5,"_27":485},{"_43":471,"_32":472,"_30":473,"_191":-5,"_41":42,"_40":474,"_193":124,"_38":475,"_13":476,"_33":34,"_35":477,"_24":25,"_197":-5,"_27":478},{"_43":463,"_32":464,"_30":465,"_191":-5,"_41":42,"_40":466,"_193":25,"_38":467,"_13":468,"_33":34,"_35":469,"_24":25,"_197":-5,"_27":470},{"_43":456,"_32":457,"_30":458,"_191":-5,"_41":42,"_40":459,"_193":124,"_38":460,"_13":304,"_33":34,"_35":461,"_24":25,"_197":-5,"_27":462},{"_43":449,"_32":450,"_30":451,"_191":-5,"_41":42,"_40":452,"_193":124,"_38":453,"_13":61,"_33":34,"_35":454,"_24":25,"_197":-5,"_27":455},{"_43":442,"_32":443,"_30":444,"_191":-5,"_41":42,"_40":445,"_193":124,"_38":446,"_13":61,"_33":34,"_35":447,"_24":25,"_197":-5,"_27":448},{"_43":435,"_32":436,"_30":437,"_191":-5,"_41":42,"_40":438,"_193":124,"_38":439,"_13":61,"_33":34,"_35":440,"_24":25,"_197":-5,"_27":441},{"_43":428,"_32":429,"_30":430,"_191":-5,"_41":42,"_40":431,"_193":124,"_38":432,"_13":61,"_33":34,"_35":433,"_24":25,"_197":-5,"_27":434},{"_43":421,"_32":422,"_30":423,"_191":-5,"_41":42,"_40":424,"_193":124,"_38":425,"_13":61,"_33":34,"_35":426,"_24":25,"_197":-5,"_27":427},{"_43":414,"_32":415,"_30":416,"_191":-5,"_41":42,"_40":417,"_193":124,"_38":418,"_13":304,"_33":34,"_35":419,"_24":25,"_197":-5,"_27":420},{"_43":406,"_32":407,"_30":408,"_191":-5,"_41":42,"_40":409,"_193":25,"_38":410,"_13":411,"_33":34,"_35":412,"_24":25,"_197":-5,"_27":413},{"_43":399,"_32":400,"_30":401,"_191":-5,"_41":42,"_40":402,"_193":124,"_38":403,"_13":61,"_33":34,"_35":404,"_24":25,"_197":-5,"_27":405},{"_43":391,"_32":392,"_30":393,"_191":-5,"_41":42,"_40":394,"_193":25,"_38":395,"_13":396,"_33":34,"_35":397,"_24":25,"_197":-5,"_27":398},{"_43":383,"_32":384,"_30":385,"_191":-5,"_41":42,"_40":386,"_193":25,"_38":387,"_13":388,"_33":34,"_35":389,"_24":25,"_197":-5,"_27":390},{"_43":376,"_32":377,"_30":378,"_191":-5,"_41":42,"_40":379,"_193":25,"_38":380,"_13":316,"_33":34,"_35":381,"_24":25,"_197":-5,"_27":382},{"_43":369,"_32":370,"_30":371,"_191":-5,"_41":42,"_40":372,"_193":25,"_38":373,"_13":195,"_33":34,"_35":374,"_24":25,"_197":-5,"_27":375},{"_43":362,"_32":363,"_30":364,"_191":-5,"_41":42,"_40":365,"_193":124,"_38":366,"_13":61,"_33":34,"_35":367,"_24":25,"_197":-5,"_27":368},{"_43":355,"_32":356,"_30":357,"_191":-5,"_41":42,"_40":358,"_193":124,"_38":359,"_13":195,"_33":34,"_35":360,"_24":25,"_197":-5,"_27":361},{"_43":349,"_32":350,"_30":351,"_191":-5,"_41":42,"_40":352,"_193":124,"_38":353,"_13":61,"_33":34,"_35":107,"_24":25,"_197":-5,"_27":354},{"_43":342,"_32":343,"_30":344,"_191":-5,"_41":42,"_40":345,"_193":124,"_38":346,"_13":304,"_33":34,"_35":347,"_24":25,"_197":-5,"_27":348},{"_43":334,"_32":335,"_30":336,"_191":-5,"_41":42,"_40":337,"_193":124,"_38":338,"_13":339,"_33":34,"_35":340,"_24":25,"_197":-5,"_27":341},{"_43":327,"_32":328,"_30":329,"_191":-5,"_41":42,"_40":330,"_193":124,"_38":331,"_13":61,"_33":34,"_35":332,"_24":25,"_197":-5,"_27":333},{"_43":319,"_32":320,"_30":321,"_191":-5,"_41":42,"_40":322,"_193":124,"_38":323,"_13":324,"_33":34,"_35":325,"_24":25,"_197":-5,"_27":326},{"_43":311,"_32":312,"_30":313,"_191":-5,"_41":42,"_40":314,"_193":124,"_38":315,"_13":316,"_33":34,"_35":317,"_24":25,"_197":-5,"_27":318},{"_43":307,"_32":28,"_30":31,"_191":-5,"_41":42,"_40":14,"_193":25,"_38":308,"_13":309,"_33":34,"_35":37,"_24":25,"_197":-5,"_27":310},{"_43":299,"_32":300,"_30":301,"_191":-5,"_41":42,"_40":302,"_193":124,"_38":303,"_13":304,"_33":34,"_35":305,"_24":25,"_197":-5,"_27":306},{"_43":291,"_32":292,"_30":293,"_191":-5,"_41":42,"_40":294,"_193":25,"_38":295,"_13":296,"_33":34,"_35":297,"_24":25,"_197":-5,"_27":298},{"_43":283,"_32":284,"_30":285,"_191":-5,"_41":42,"_40":286,"_193":124,"_38":287,"_13":288,"_33":34,"_35":289,"_24":25,"_197":-5,"_27":290},{"_43":274,"_32":275,"_30":276,"_191":-5,"_41":42,"_40":277,"_193":25,"_38":278,"_13":279,"_33":249,"_35":280,"_24":25,"_197":281,"_27":282},{"_43":267,"_32":268,"_30":269,"_191":-5,"_41":42,"_40":270,"_193":124,"_38":271,"_13":61,"_33":34,"_35":272,"_24":25,"_197":-5,"_27":273},{"_43":259,"_32":260,"_30":261,"_191":-5,"_41":42,"_40":262,"_193":25,"_38":263,"_13":264,"_33":249,"_35":265,"_24":25,"_197":-5,"_27":266},{"_43":253,"_32":254,"_30":255,"_191":-5,"_41":42,"_40":256,"_193":124,"_38":257,"_13":61,"_33":34,"_35":107,"_24":25,"_197":-5,"_27":258},{"_43":244,"_32":245,"_30":246,"_191":-5,"_41":42,"_40":247,"_193":124,"_38":248,"_13":219,"_33":249,"_35":250,"_24":25,"_197":251,"_27":252},{"_43":237,"_32":238,"_30":239,"_191":-5,"_41":42,"_40":240,"_193":25,"_38":241,"_13":234,"_33":34,"_35":242,"_24":25,"_197":-5,"_27":243},{"_43":229,"_32":230,"_30":231,"_191":-5,"_41":42,"_40":232,"_193":124,"_38":233,"_13":234,"_33":34,"_35":235,"_24":25,"_197":-5,"_27":236},{"_43":222,"_32":223,"_30":224,"_191":-5,"_41":42,"_40":225,"_193":25,"_38":226,"_13":61,"_33":34,"_35":227,"_24":25,"_197":-5,"_27":228},{"_43":214,"_32":215,"_30":216,"_191":-5,"_41":42,"_40":217,"_193":25,"_38":218,"_13":219,"_33":34,"_35":220,"_24":25,"_197":-5,"_27":221},{"_43":207,"_32":208,"_30":209,"_191":-5,"_41":42,"_40":210,"_193":124,"_38":211,"_13":61,"_33":34,"_35":212,"_24":25,"_197":-5,"_27":213},{"_43":199,"_32":200,"_30":201,"_191":-5,"_41":42,"_40":202,"_193":25,"_38":203,"_13":204,"_33":34,"_35":205,"_24":25,"_197":-5,"_27":206},{"_43":188,"_32":189,"_30":190,"_191":-5,"_41":42,"_40":192,"_193":25,"_38":194,"_13":195,"_33":34,"_35":196,"_24":25,"_197":-5,"_27":198},"1ec24029-80f6-4e64-9b87-a38489963e60","ui-timeline.ts","6a4fbfcc-acb5-11f1-93b9-1607ee4eb77e","parentId",1488,"latest",["D",1789154863585],2,"// Run timeline: ONE renderer for everything the AI did in a turn — live (fed by\n// ?progress events) and from history (fed by message_meta.toolEvents). Collapsed to a\n// summary line; each row expands to the request args and the response, pretty-printed.\nexport const timelineStyle: string = `\n.tl{margin:4px 0 8px;border:1px solid var(--line);background:var(--s0);font-size:11px;}\n.tl-h{display:flex;gap:8px;align-items:center;padding:4px 8px;cursor:pointer;color:var(--dim);letter-spacing:.06em;user-select:none;}\n.tl-h::before{content:\"▸\";color:var(--acc);font-size:10px;} .tl.open .tl-h::before{content:\"▾\";}\n.tl-h .tl-sum{flex:1;overflow:hidden;text-overflow:ellipsis;white-space:nowrap;} .tl-h .tl-n{color:var(--dim2);font-size:10px;}\n.tl-body{display:none;border-top:1px solid var(--line);} .tl.open .tl-body{display:block;}\n.tl-row{display:flex;gap:8px;align-items:center;padding:3px 8px;border-bottom:1px solid var(--line);cursor:pointer;} .tl-row:last-child{border-bottom:0;} .tl-row:hover{background:var(--s1);}\n.tl-k{font-size:9px;letter-spacing:.12em;text-transform:uppercase;padding:1px 6px;border:1px solid var(--line2);color:var(--dim);white-space:nowrap;min-width:52px;text-align:center;}\n.tl-k.ok{border-color:rgba(198,255,127,.35);color:var(--acc2);} .tl-k.err{border-color:rgba(255,107,107,.4);color:var(--err);} .tl-k.sys{border-color:transparent;color:var(--dim2);} .tl-k.live{border-color:var(--acc);color:var(--acc);}\n.tl-d{flex:1;overflow:hidden;text-overflow:ellipsis;white-space:nowrap;color:var(--text);} .tl-t{color:var(--dim2);font-size:10px;white-space:nowrap;}\n.tl-x{display:none;padding:6px 8px 8px;background:var(--bg);border-bottom:1px solid var(--line);} .tl-row.open+.tl-x{display:grid;grid-template-columns:1fr 1fr;gap:8px;} @media (max-width:800px){.tl-row.open+.tl-x{grid-template-columns:1fr;}}\n.tl-x h5{font-size:9px;letter-spacing:.14em;text-transform:uppercase;color:var(--dim2);margin:0 0 4px;font-weight:500;display:flex;justify-content:space-between;} .tl-x h5 button{background:none;border:0;color:var(--dim2);font:inherit;cursor:pointer;} .tl-x h5 button:hover{color:var(--acc);}\n.tl-x pre{margin:0;white-space:pre-wrap;word-break:break-word;font-size:11px;line-height:1.45;color:var(--dim);max-height:320px;overflow:auto;} .tl-x pre.full{max-height:none;}\n.tl-x .tl-more{color:var(--acc);cursor:pointer;font-size:10px;margin-top:4px;display:inline-block;}\n.tl.live .tl-h{color:var(--acc);}\n.tl-st{font-size:9px;letter-spacing:.12em;text-transform:uppercase;min-width:56px;text-align:right;color:var(--dim2);}\n.tl-row.st-done .tl-st{color:var(--acc2);} .tl-row.st-failed .tl-st{color:var(--err);} .tl-row.st-repeated .tl-st{color:var(--warn);} .tl-row.st-running .tl-st{color:var(--acc);} .tl-row.st-running .tl-st::after{content:\"\";display:inline-block;width:6px;height:6px;border-radius:50%;background:var(--acc);margin-left:5px;animation:blink 1s infinite;}\n.tl-row.st-failed .tl-k{border-color:rgba(255,107,107,.4);color:var(--err);} .tl-row.st-done .tl-k{border-color:rgba(198,255,127,.3);color:var(--acc2);} .tl-row[data-kind=step] .tl-k{border-color:transparent;color:var(--dim2);}\n.tl-plan{display:flex;gap:4px;flex-wrap:wrap;padding:5px 8px;border-bottom:1px solid var(--line);} .tl-pl{font-size:9px;letter-spacing:.12em;text-transform:uppercase;padding:1px 8px;border:1px solid var(--line2);color:var(--dim2);} .tl-pl::before{content:\"○ \";} .tl-pl.st-running{border-color:var(--acc);color:var(--acc);} .tl-pl.st-running::before{content:\"● \";} .tl-pl.st-done{border-color:rgba(198,255,127,.3);color:var(--acc2);} .tl-pl.st-done::before{content:\"✓ \";} .tl-pl.st-failed{border-color:var(--err);color:var(--err);} .tl-pl.st-failed::before{content:\"✗ \";} .tl-pl.st-skipped{opacity:.5;} .tl-pl.st-skipped::before{content:\"– \";}\n`;\n\nexport const timelineScript: string = String.raw`\nconst TL = (() => {\n const LABEL = { web_search: 'search', eval_js: 'eval', create_artifact: 'build', update_artifact: 'edit', read_artifact: 'read', search_artifact: 'grep', memory_recall: 'recall', memory_write: 'remember', name_session: 'name', session_clear: 'clear', session_delete: 'delete' };\n const esc = (s) => String(s).replace(/[&<>\"]/g, (c) => ({ '&': '&', '<': '<', '>': '>', '\"': '"' }[c]));\n const short = (v) => { if (v == null) return ''; if (typeof v === 'string') return v; try { return JSON.stringify(v); } catch { return String(v); } };\n const pretty = (v) => { if (v == null) return ''; if (typeof v === 'string') return v; try { return JSON.stringify(v, null, 2); } catch { return String(v); } };\n function desc(ev) {\n const a = ev.args || {}, r = ev.result;\n switch (ev.tool) {\n case 'web_search': return 'search: ' + (a.query || '');\n case 'eval_js': return 'eval: ' + String(a.code || a.expression || '').slice(0, 80).replace(/\\s+/g, ' ');\n case 'create_artifact': return 'built ' + (a.title || (r && r.title) || '') + (a.files ? ' (' + (1 + (a.files.length || 0)) + ' files)' : '');\n case 'update_artifact': { const n = (a.edits && a.edits.length) || 0, f = (a.files && a.files.length) || 0; return 'edit #' + (a.id ?? (r && r.id) ?? '') + (n ? ' · ' + n + ' edit' + (n > 1 ? 's' : '') : '') + (f ? ' · ' + f + ' file' + (f > 1 ? 's' : '') : '') + (a.delete ? ' · delete' : ''); }\n case 'search_artifact': return 'grep: ' + (a.query || '') + (r && r.matches != null ? ' → ' + r.matches + ' hits' : '');\n case 'read_artifact': return 'read #' + (a.id ?? '') + (a.all ? ' (all files)' : a.path ? ' ' + a.path : '');\n case 'memory_recall': return 'recall: ' + (a.query || '');\n case 'memory_write': return 'remember: ' + String(a.fact || a.value || '').slice(0, 80);\n case 'name_session': return 'title: ' + (a.name || '');\n default: return ev.tool + ' ' + short(a).slice(0, 80);\n }\n }\n function stageText(ev) {\n if (ev.type === 'pass') return 'model call ' + ev.n + '/' + ev.of;\n if (ev.type === 'continue') return 'reply cut at the vendor cap after ' + (ev.chars ? ev.chars.toLocaleString() + ' chars' : 'a partial reply') + ' — continuing (' + ev.n + ')' + (ev.tail ? '\\n…' + ev.tail : '');\n if (ev.type === 'retry') return 'retry ' + ev.n + (ev.error ? ' — ' + ev.error : '');\n if (ev.type === 'thinking') return 'thinking…';\n return ev.label || ev.text || ev.type;\n }\n function summary(events) {\n const tools = events.filter(e => e.tool);\n const errs = tools.filter(e => e.error && e.error !== 'repeated').length;\n const parts = [tools.length + ' tool' + (tools.length === 1 ? '' : 's')];\n const files = tools.filter(e => /artifact/.test(e.tool) && !e.error).length; if (files) parts.push(files + ' artifact op' + (files > 1 ? 's' : ''));\n if (errs) parts.push(errs + ' failed');\n const conts = events.filter(e => e.type === 'continue').length; if (conts) parts.push(conts + ' continued');\n const retries = events.filter(e => e.type === 'retry').length; if (retries) parts.push(retries + ' retr' + (retries > 1 ? 'ies' : 'y'));\n return parts.join(' · ');\n }\n function pane(title, value) {\n const s = pretty(value), cut = s.length > 4000;\n const wrap = document.createElement('div');\n const h = document.createElement('h5'); h.innerHTML = esc(title) + (s ? ' ' : '');\n const pre = document.createElement('pre'); pre.textContent = cut ? s.slice(0, 4000) + '\\n…' : s;\n wrap.appendChild(h); wrap.appendChild(pre);\n if (h.querySelector('button')) h.querySelector('button').onclick = (e) => { e.stopPropagation(); navigator.clipboard.writeText(s); h.querySelector('button').textContent = 'copied'; };\n if (cut) { const m = document.createElement('span'); m.className = 'tl-more'; m.textContent = 'show all ' + s.length.toLocaleString() + ' chars'; m.onclick = (e) => { e.stopPropagation(); pre.textContent = s; pre.classList.add('full'); m.remove(); }; wrap.appendChild(m); }\n return wrap;\n }\n // status vocabulary: waiting · running · done · failed · repeated · skipped\n function statusOf(ev, isLast, live) {\n if (ev.tool) return ev.error === 'repeated' || (ev.result && ev.result.repeated) ? 'repeated' : ev.error ? 'failed' : 'done';\n if (ev.type === 'retry') return 'failed';\n return live && isLast ? 'running' : 'done';\n }\n function row(ev, status) {\n const r = document.createElement('div'); r.className = 'tl-row st-' + status; r.dataset.kind = ev.tool ? 'tool' : 'step';\n const k = document.createElement('span'); k.className = 'tl-k'; k.textContent = ev.tool ? (LABEL[ev.tool] || ev.tool.replace(/^rojs_/, 'rojs ').slice(0, 10)) : (ev.type === 'pass' ? 'model' : ev.type === 'stage' ? 'step' : ev.type);\n const d = document.createElement('span'); d.className = 'tl-d'; d.textContent = ev.tool ? desc(ev) + (ev.error && ev.error !== 'repeated' ? ' — ' + ev.error : '') : stageText(ev);\n const st = document.createElement('span'); st.className = 'tl-st'; st.textContent = status;\n const t = document.createElement('span'); t.className = 'tl-t'; t.textContent = ev.ms ? (ev.ms / 1000).toFixed(1) + 's' : ev.hop != null && ev.tool ? 'hop ' + (ev.hop + 1) : '';\n r.appendChild(k); r.appendChild(d); r.appendChild(st); r.appendChild(t);\n if (!ev.tool) return [r];\n const x = document.createElement('div'); x.className = 'tl-x';\n x.appendChild(pane('request · ' + ev.tool, ev.args)); x.appendChild(pane(ev.error && ev.error !== 'repeated' ? 'error' : 'response', ev.error && ev.error !== 'repeated' ? ev.error : ev.result));\n r.onclick = () => r.classList.toggle('open');\n return [r, x];\n }\n /** Build a timeline. Rendering is INCREMENTAL: rows already on screen are kept (so an expanded row stays expanded\n * while live events stream in); only new events are appended and the previous last step is re-labelled done.\n * opts.live: expanded, elapsed clock, checklist of planned stages (opts.plan: ['plan','answer','verify']). */\n function make(events, opts = {}) {\n const root = document.createElement('div'); root.className = 'tl' + (opts.live ? ' live open' : (opts.open ? ' open' : ''));\n const h = document.createElement('div'); h.className = 'tl-h';\n const sum = document.createElement('span'); sum.className = 'tl-sum'; const clock = document.createElement('span'); clock.className = 'tl-n';\n h.appendChild(sum); h.appendChild(clock); h.onclick = () => root.classList.toggle('open');\n const body = document.createElement('div'); body.className = 'tl-body';\n let plan = null;\n if (opts.live && opts.plan && opts.plan.length) { plan = document.createElement('div'); plan.className = 'tl-plan'; plan.innerHTML = opts.plan.map(p => '' + esc(p) + '').join(''); root.appendChild(plan); }\n root.appendChild(h); root.appendChild(body);\n const t0 = Date.now(); let rendered = 0, lastStepRow = null, timer = null;\n const api = {\n root, events: [],\n set(evs) {\n api.events = evs;\n if (evs.length < rendered) { body.innerHTML = ''; rendered = 0; lastStepRow = null; } // not append-only: rebuild\n for (let i = rendered; i < evs.length; i++) {\n const ev = evs[i]; if (ev.type === 'thinking') continue;\n if (lastStepRow) { lastStepRow.classList.remove('st-running'); lastStepRow.classList.add('st-done'); lastStepRow.querySelector('.tl-st').textContent = 'done'; lastStepRow = null; }\n const status = statusOf(ev, i === evs.length - 1, !!opts.live);\n const els = row(ev, status); for (const el of els) body.appendChild(el);\n if (status === 'running') lastStepRow = els[0];\n if (plan && ev.type === 'stage') { const key = String(ev.label || '').toLowerCase(); plan.querySelectorAll('.tl-pl').forEach(pl => { const p = pl.dataset.p; if (key.startsWith(p) || (p === 'answer' && /answer|summar/.test(key)) || (p === 'plan' && key.startsWith('plan')) || (p === 'verify' && key.startsWith('verif'))) { plan.querySelectorAll('.tl-pl.st-running').forEach(o => { o.className = 'tl-pl st-done'; }); pl.className = 'tl-pl st-running'; } }); }\n }\n rendered = evs.length;\n const now = evs.filter(e => e.type !== 'thinking').slice(-1)[0];\n sum.textContent = opts.live ? ((opts.label || 'working') + (now ? ' · now: ' + (now.tool ? desc(now) : stageText(now)) : '') + ' · ' + summary(evs)) : summary(evs);\n if (!opts.live) clock.textContent = evs.length ? evs.length + ' step' + (evs.length > 1 ? 's' : '') : '';\n },\n label(l) { opts.label = l; api.set(api.events); },\n finish(ok) { clearInterval(timer); if (lastStepRow) { lastStepRow.classList.remove('st-running'); lastStepRow.classList.add(ok ? 'st-done' : 'st-failed'); lastStepRow.querySelector('.tl-st').textContent = ok ? 'done' : 'failed'; } if (plan) plan.querySelectorAll('.tl-pl').forEach(pl => { if (pl.classList.contains('st-running')) pl.className = 'tl-pl ' + (ok ? 'st-done' : 'st-failed'); else if (pl.classList.contains('st-waiting')) pl.className = 'tl-pl st-skipped'; }); },\n };\n if (opts.live) timer = setInterval(() => { clock.textContent = Math.round((Date.now() - t0) / 1000) + 's'; }, 1000);\n api.set(events || []);\n return api;\n }\n return { make, summary, desc };\n})();\nwindow.TL = TL;\n`;\n","label",[189],"77f22f22-05db-4711-b208-8baba0bda84d","ui-style-settings.ts","c77c9562-abcf-11f1-a463-1607ee4eb77e",1395,["D",1789003062000],4,"// Layout column, themes, density, settings drawer, thinking/tool/error blocks.\nexport const settingsStyle: string = `\n:root{--content-w:760px;--fs:13px;}\nbody[data-theme=\"black\"]{--bg:#000;--s0:#050505;--s1:#0a0a0a;--s2:#111;--s3:#181818;--line:#1f1f1f;--line2:#2a2a2a;}\nbody[data-theme=\"paper\"]{--bg:#f4f4f1;--s0:#ececea;--s1:#e4e4e1;--s2:#dcdcd8;--s3:#d2d2ce;--line:#c9c9c4;--line2:#b8b8b3;--text:#15171c;--dim:#4d5260;--dim2:#7a7f8c;--acc2:#3b7d00;}\nbody[data-theme=\"paper\"] .user .msg-body{color:#15171c;}\nbody[data-theme=\"paper\"] .hljs-title,body[data-theme=\"paper\"] .hljs-title.function_,body[data-theme=\"paper\"] .hljs-title.class_,body[data-theme=\"paper\"] .hljs-section,body[data-theme=\"paper\"] .hljs-name{color:#15171c;}\nbody[data-theme=\"paper\"] .hljs-variable,body[data-theme=\"paper\"] .hljs-template-variable,body[data-theme=\"paper\"] .hljs-params,body[data-theme=\"paper\"] .hljs-property{color:#3a3f4c;}\nbody[data-theme=\"paper\"] .hljs-attribute,body[data-theme=\"paper\"] .hljs-selector-id,body[data-theme=\"paper\"] .hljs-selector-class{color:#1f4f9a;}\nbody[data-theme=\"paper\"] .hljs-number,body[data-theme=\"paper\"] .hljs-symbol,body[data-theme=\"paper\"] .hljs-meta{color:#9a5a00;}\nbody[data-font=\"sans\"] .msg-body,body[data-font=\"sans\"] #inp,body[data-font=\"sans\"] .md p,body[data-font=\"sans\"] .md li{font-family:var(--sans);}\n.msg-body,#inp{font-size:var(--fs);}\nbody[data-density=\"compact\"] .msg-wrap{padding:5px 0;} body[data-density=\"compact\"] .md p{margin-bottom:.4em;}\nbody.no-ts .msg-ts{display:none;} body.no-tel .telemetry{display:none;} body.no-tools .tool-chain{display:none;}\nbody.code-wrap .cb pre code{white-space:pre-wrap;word-break:break-word;}\n\n/* reading column: messages, composer and typing row share one centered width */\n#chat>.msg-wrap,#typing,#empty,.bar-inner,.bar-meta{max-width:var(--content-w);margin-left:auto;margin-right:auto;width:100%;}\n\n/* thinking block */\n.think{margin:4px 0 8px;border:1px solid var(--line);background:var(--s0);}\n.think summary{cursor:pointer;list-style:none;padding:4px 8px;font-size:10px;letter-spacing:.1em;text-transform:uppercase;color:var(--dim);display:flex;gap:8px;align-items:center;}\n.think summary::before{content:\"▸\";color:var(--acc);} .think[open] summary::before{content:\"▾\";}\n.think .think-body{padding:6px 10px 8px;font-size:calc(var(--fs) - 1px);line-height:1.55;color:var(--dim);white-space:pre-wrap;word-break:break-word;border-top:1px solid var(--line);max-height:320px;overflow:auto;}\nbody[data-thinking=\"hidden\"] .think{display:none;}\n#live-think{font-size:11px;color:var(--dim2);white-space:pre-wrap;max-height:64px;overflow:hidden;margin-top:4px;line-height:1.4;}\n\n.trunc{font-size:10px;color:var(--warn);letter-spacing:.08em;text-transform:uppercase;display:flex;align-items:center;gap:6px;margin-top:4px;}\n/* error card */\n.err-card{border:1px solid var(--err);background:rgba(255,107,107,.06);padding:8px 10px;}\n.err-card .err-h{color:var(--err);font-size:11px;letter-spacing:.08em;text-transform:uppercase;margin-bottom:4px;}\n.err-card .err-m{font-size:12px;white-space:pre-wrap;word-break:break-word;}\n.err-card details{margin-top:6px;} .err-card summary{cursor:pointer;font-size:10px;color:var(--dim);letter-spacing:.08em;text-transform:uppercase;}\n.err-card pre{margin:6px 0 0;font-size:11px;white-space:pre-wrap;word-break:break-word;color:var(--dim);max-height:260px;overflow:auto;}\n.err-card .err-actions{margin-top:6px;display:flex;gap:6px;}\n\n/* settings drawer */\n#settings{position:fixed;top:0;right:0;height:100dvh;width:min(380px,100vw);background:var(--s0);border-left:1px solid var(--line);z-index:40;display:flex;flex-direction:column;transition:transform .18s;}\n#settings.closed{transform:translateX(100%);}\n#st-scrim{position:fixed;inset:0;background:rgba(0,0,0,.35);z-index:39;display:none;} #st-scrim.on{display:block;}\n#st-hdr{display:flex;align-items:center;justify-content:space-between;padding:10px 12px;border-bottom:1px solid var(--line);font-size:11px;letter-spacing:.18em;text-transform:uppercase;color:var(--acc);}\n#st-body{overflow:auto;padding:6px 12px 24px;}\n#st-body section{padding:10px 0;border-bottom:1px solid var(--line);}\n#st-body h4{font-size:10px;letter-spacing:.14em;text-transform:uppercase;color:var(--dim2);margin:0 0 8px;font-weight:500;}\n#st-body label{display:flex;align-items:center;justify-content:space-between;gap:10px;font-size:12px;color:var(--text);padding:4px 0;}\n#st-body label>input[type=checkbox]{order:-1;margin:0 8px 0 0;accent-color:var(--acc);} #st-body label:has(input[type=checkbox]){justify-content:flex-start;}\n#st-body select,#st-body input[type=text],#st-body textarea{background:var(--s1);color:var(--text);border:1px solid var(--line2);font:inherit;font-size:12px;padding:4px 6px;min-width:150px;}\n#st-body textarea{width:100%;resize:vertical;line-height:1.5;}\n#st-body input[type=range]{flex:1;accent-color:var(--acc);} #st-body output{min-width:36px;text-align:right;color:var(--dim);font-size:11px;}\n#st-body input[type=color]{width:36px;height:22px;border:1px solid var(--line2);background:none;padding:0;}\n#st-tools label{padding:2px 0;}\n.pr-hint{font-size:10px;color:var(--dim2);letter-spacing:.04em;margin:2px 0 4px;flex:1;}\n#pr-text{font-family:var(--mono);font-size:11px;line-height:1.45;min-height:160px;} #pr-text.custom{border-color:var(--acc);}\n.sk{border:1px solid var(--line);margin:4px 0;padding:0 8px 4px;} .sk summary{cursor:pointer;list-style:none;display:flex;gap:8px;padding:5px 0;font-size:11px;} .sk .sk-name{color:var(--text);flex:1;} .sk .sk-mode{color:var(--dim2);font-size:10px;white-space:nowrap;overflow:hidden;text-overflow:ellipsis;max-width:55%;} .sk textarea{font-family:var(--mono);font-size:11px;margin:4px 0;} .sk label{padding:2px 0;} .sk label>input[type=text]{flex:1;min-width:0;}\n.st-row{display:flex;gap:6px;flex-wrap:wrap;align-items:center;}\ntable.rt{width:100%;border-collapse:collapse;font-size:10px;margin-top:6px;} table.rt th{text-align:left;color:var(--dim2);font-weight:500;letter-spacing:.1em;text-transform:uppercase;font-size:9px;padding:2px 4px;border-bottom:1px solid var(--line);} table.rt td{padding:2px 4px;border-bottom:1px solid var(--line);color:var(--dim);} table.rt td.cool{color:var(--warn);} .st-btn{background:none;border:1px solid var(--line2);color:var(--dim);font:inherit;font-size:10px;letter-spacing:.1em;text-transform:uppercase;padding:4px 8px;cursor:pointer;} .st-btn:hover{border-color:var(--acc);color:var(--acc);}\n`;\n",[200],"1e8e8db7-ea81-46a6-b774-05c9cbc5e81d","ui-style-md.ts","b8f94408-abb2-11f1-81cd-1607ee4eb77e",1282,["D",1788891952093],"// Styles for Markdown bodies, code blocks (hljs token map on the house palette),\n// tables, math, the progress line, model chips and the stop control.\nexport const mdStyle: string = `\n.msg-body.md{white-space:normal;}\n.md p{margin:0 0 .7em;} .md p:last-child{margin-bottom:0;}\n.md h1,.md h2,.md h3,.md h4{font-family:var(--mono);letter-spacing:.04em;margin:.9em 0 .4em;color:var(--text);line-height:1.3;}\n.md h1{font-size:16px;} .md h2{font-size:14px;} .md h3{font-size:13px;color:var(--acc);} .md h4{font-size:12px;color:var(--dim);}\n.md ul,.md ol{margin:0 0 .7em 1.3em;} .md li{margin:.15em 0;} .md li>ul,.md li>ol{margin-bottom:0;}\n.md blockquote{border-left:2px solid var(--line2);padding:.1em 0 .1em .8em;margin:0 0 .7em;color:var(--dim);}\n.md hr{border:0;border-top:1px solid var(--line);margin:.9em 0;}\n.md a{color:var(--acc);text-decoration:none;border-bottom:1px solid rgba(79,140,255,.35);} .md a:hover{border-bottom-color:var(--acc);}\n.md code{font-family:var(--mono);font-size:12px;background:var(--s2);border:1px solid var(--line);padding:0 4px;border-radius:var(--r);color:var(--acc2);}\n.md .ck{display:inline-block;width:10px;height:10px;border:1px solid var(--line2);margin:0 6px -1px 0;} .md .ck.on{background:var(--acc2);border-color:var(--acc2);}\n.md .tbl{overflow-x:auto;margin:0 0 .7em;border:1px solid var(--line);}\n.md table{border-collapse:collapse;font-size:12px;min-width:100%;} .md th,.md td{padding:5px 9px;border-bottom:1px solid var(--line);text-align:left;vertical-align:top;}\n.md th{background:var(--s1);color:var(--dim);font-weight:500;letter-spacing:.06em;text-transform:uppercase;font-size:10px;} .md tr:last-child td{border-bottom:0;}\n.md .mx-d{display:block;text-align:center;margin:.5em 0;} .md .mx:not(.done){color:var(--dim);}\n.md .katex{font-size:1.05em;}\n\n.cb{margin:.6em 0;border:1px solid var(--line);background:var(--s0);}\n.cb-h{display:flex;align-items:center;gap:10px;padding:4px 8px;border-bottom:1px solid var(--line);background:var(--s1);font-size:10px;letter-spacing:.08em;text-transform:uppercase;color:var(--dim);}\n.cb-lang{color:var(--acc);} .cb-n{color:var(--dim2);margin-right:auto;}\n.cb-copy{background:none;border:1px solid var(--line2);color:var(--dim);font:inherit;letter-spacing:.08em;text-transform:uppercase;padding:1px 6px;cursor:pointer;} .cb-copy:hover{border-color:var(--acc);color:var(--acc);}\n.cb pre{margin:0;padding:9px 10px;overflow-x:auto;font-size:12px;line-height:1.55;tab-size:2;}\n.cb pre code{background:none;border:0;padding:0;color:var(--text);font-size:inherit;white-space:pre;}\n.hljs{background:transparent;color:var(--text);}\n.hljs-comment,.hljs-quote{color:var(--dim2);font-style:italic;}\n.hljs-keyword,.hljs-selector-tag,.hljs-literal,.hljs-built_in,.hljs-type{color:var(--acc);}\n.hljs-string,.hljs-regexp,.hljs-addition,.hljs-attr,.hljs-template-tag{color:var(--acc2);}\n.hljs-number,.hljs-symbol,.hljs-bullet,.hljs-meta,.hljs-link{color:var(--warn);}\n.hljs-title,.hljs-title.function_,.hljs-title.class_,.hljs-section,.hljs-name{color:#e4e8f4;font-weight:500;}\n.hljs-variable,.hljs-template-variable,.hljs-params,.hljs-property{color:#c4cbe0;}\n.hljs-attribute,.hljs-selector-id,.hljs-selector-class,.hljs-selector-attr,.hljs-selector-pseudo{color:#9fd0ff;}\n.hljs-deletion{color:var(--err);} .hljs-emphasis{font-style:italic;} .hljs-strong{font-weight:600;}\n\n#progress{font-size:10px;color:var(--dim);letter-spacing:.08em;text-transform:uppercase;margin-left:8px;white-space:nowrap;overflow:hidden;text-overflow:ellipsis;max-width:60vw;}\n#progress .tk{color:var(--acc);}\n.bar-modes{display:flex;gap:4px;margin-left:auto;}\n.mode{background:none;border:1px solid var(--line);color:var(--dim2);font:inherit;font-size:9px;letter-spacing:.1em;text-transform:uppercase;padding:2px 7px;cursor:pointer;}\n.mode:hover{border-color:var(--acc);color:var(--acc);} .mode.on{border-color:var(--acc);color:var(--acc);background:var(--s1);}\n#send.stop{background:var(--err);border-color:var(--err);}\n.msg-wrap.streaming .msg-body::after{content:\"▌\";color:var(--acc);animation:blink 1s steps(2) infinite;} @keyframes blink{to{opacity:0;}}\n.telemetry{font-size:9px;color:var(--dim2);letter-spacing:.06em;margin-left:auto;}\n`;\n",[208],"1f66d685-05ee-48a5-92b1-10c951c23b5d","ui-settings.ts","c73730bc-abcf-11f1-a6fb-1607ee4eb77e",1474,["D",1789154322787],12,"// Settings drawer: markup + logic. State lives in localStorage(\"omni.settings\");\n// `SETTINGS.current()` is read by main-script on every send, and appearance is\n// applied as CSS variables / body classes so the page re-skins without reload.\nexport const settingsHtml: string = `\n

settings

appearance

\n width \n font \n size \n theme \n accent \n density \n timestamps\n telemetry tags\n wrap long code lines\n

thinking & tools

\n thinking \n show tool chain\n live thinking while generating\n

router

\n router url \n router key \n vendor order \n key rotation \n

model

\n routing \n pin \n reasoning effort \n temperature \n max_tokens policy \n fixed max_tokens \n history sent to the model \n expected length \n in-turn continues \n auto-continue across turns \n auto-continue unfinished work (rounds) \n keep retrying when providers are exhausted (min) \n swarm builders in parallel \n work until I say /finalize — unlimited retries and continues; after each completed round, ship the next upgrades\n auto-compact history when it outgrows the context budget\n two top-tier examples in every prompt (few-shot; off saves input tokens on tight free tiers)\n app-scale asks go through /build automatically (architect → parallel builders → integrate)\n inline preview on artifact cards\n turn budget (s) \n

governor

\n plan before answering (complex asks)\n verification pass\n ensemble on high-stakes\n long-term memory\n

tools

custom instructions

\n \n

prompts

\n edit \n

\n \n

import

skills

Instruction packs the model receives as binding rules. always = every turn; auto = when the message matches the keywords (comma-separated) or a /regex/i.

input

\n send with \n auto-scroll to new replies\n

data

\n`;\n\nexport const settingsScript: string = String.raw`\nconst SETTINGS = (() => {\n const KEY = 'omni.settings';\n const DEF = { width: '760', font: 'mono', fontSize: 13, theme: 'midnight', accent: '#4f8cff', density: 'cozy', timestamps: true, telemetry: true, codeWrap: false,\n thinking: 'collapsed', showTools: true, liveThinking: true, mode: 'auto', pin: '', reasoning: '', temp10: -1, maxTokens: 0, continueMax: 3, budgetS: 52,\n plan: true, verify: true, ensemble: true, memory: true, tools: null, system: '', sendKey: 'enter', autoScroll: true, prompts: {}, skills: [],\n routerUrl: '', routerKey: '', maxMode: 'auto', expect: 'auto', autoContinue: 6, swarm: 3, cardPreview: true, autoSwarm: true, examples: true, untilStopped: false, autoCompact: true, keyPolicy: '', contextChars: 16000, retryMinutes: 5, vendorOrder: 'openrouter', autoComplete: 4 };\n const PRESETS = {\n 'single-file html': { name: 'single-file html', mode: 'auto', when: 'html, webgl, canvas, shader, game, app, visuali', body: 'Deliver browser artifacts as ONE self-contained index.html: inline all CSS in '; });\n const base = '';\n html = /]*>/i.test(html) ? html.replace(/]*>/i, (m) => m + base + HOOK) : base + HOOK + html;\n return html;\n }\n function reload(fromServer) {\n if (!cur) return;\n if (!fromServer && liveCb.checked) { const doc = buildLive(); if (doc !== null) { frame.srcdoc = doc; return; } }\n frame.removeAttribute('srcdoc'); frame.src = '/artifact/' + cur.id + '/?v=' + Date.now();\n }\n function scheduleLive() { if (!liveCb.checked) return; clearTimeout(liveTimer); liveTimer = setTimeout(() => { clearConsole(); reload(false); }, E.liveDelay); }\n function scheduleSave() { if (!E.autoSave) return; clearTimeout(saveTimer); saveTimer = setTimeout(() => save(), E.autoSaveDelay); }\n\n // ─── file tree ──────────────────────────────────────────────────────────────\n function renderTree() {\n treeB.innerHTML = '';\n const root = {};\n for (const f of [...cur.files].sort((a, b) => a.path.localeCompare(b.path))) { const parts = f.path.split('/'); let n = root; for (const p of parts.slice(0, -1)) n = n[p] = n[p] || {}; n[parts[parts.length - 1]] = f; }\n const walk = (node, host, prefix, depth) => {\n const names = Object.keys(node).sort((a, b) => (typeof node[a].path === 'string') - (typeof node[b].path === 'string') || a.localeCompare(b));\n for (const name of names) {\n const v = node[name];\n if (typeof v.path === 'string') { const d = document.createElement('div'); d.className = 'tr-f' + (v.path === active ? ' on' : '') + (cur.dirty.has(v.path) ? ' dirty' : ''); d.style.paddingLeft = (8 + depth * 12) + 'px'; d.textContent = name; d.title = v.path; d.onclick = () => showFile(v.path); host.appendChild(d); }\n else { const dir = prefix + name + '/'; const d = document.createElement('div'); d.className = 'tr-d' + (closedDirs.has(dir) ? ' closed' : ''); d.style.paddingLeft = (6 + depth * 12) + 'px'; d.textContent = name; d.onclick = () => { closedDirs.has(dir) ? closedDirs.delete(dir) : closedDirs.add(dir); renderTree(); }; host.appendChild(d); const kids = document.createElement('div'); kids.className = 'tr-kids'; host.appendChild(kids); walk(v, kids, dir, depth + 1); }\n }\n };\n walk(root, treeB, '', 0);\n }\n function showFile(p) {\n if (!cur) return; const f = cur.files.find(x => x.path === p); if (!f) return;\n active = p;\n if (isBin(p, f.content)) { bin.hidden = false; ta.hidden = true; gutter.hidden = true; if (cm) cm.getWrapperElement().style.display = 'none'; bin.innerHTML = /^data:image\\//.test(f.content) ? '' : 'binary · ' + Math.round(f.content.length * 0.75 / 1024) + ' KB'; if (bin.querySelector('img')) bin.querySelector('img').src = f.content; }\n else { bin.hidden = true; if (cm) { cm.getWrapperElement().style.display = ''; } else { ta.hidden = false; gutter.hidden = !E.lineNumbers || !!E.wrap; } ED.mode(p); ED.set(f.content); }\n renderTree(); updState(); updPos(); ED.refresh();\n }\n function onEdit() { if (!cur || !active) return; const f = cur.files.find(x => x.path === active); if (!f || isBin(active, f.content)) return; f.content = ED.get(); cur.dirty.add(active); renderTree(); updState(); renderGutter(); lintLocal(); scheduleLive(); scheduleSave(); }\n function renderGutter() { if (cm || !E.lineNumbers || E.wrap) return; const n = ta.value.split('\\n').length; if (gutter.dataset.n == n) return; gutter.dataset.n = n; let s = ''; for (let i = 1; i <= n; i++) s += i + '\\n'; gutter.textContent = s; }\n ta.addEventListener('scroll', () => { gutter.scrollTop = ta.scrollTop; });\n function updState() { const f = cur.files.find(x => x.path === active); state.textContent = (f ? (isBin(f.path, f.content) ? 'binary' : f.content.length + ' chars · ' + f.content.split('\\n').length + ' lines') : '') + (cur.dirty.size ? ' · ' + cur.dirty.size + ' unsaved' : ''); }\n function updPos() { const [l, c, s] = ED.cursor(); pos.textContent = l + ':' + c + (s ? ' (' + s + ' sel)' : ''); }\n let lintIssues = []; const autoFixed = {}; // artifact id → rounds used\n function showLint(issues) {\n lintIssues = issues || [];\n if (lintIssues.length) { lint.hidden = false; $('ap-lint-n').textContent = lintIssues.length + ' lint issue' + (lintIssues.length > 1 ? 's' : ''); $('ap-lint-b').textContent = lintIssues.join('\\n'); maybeAutoFix(); }\n else lint.hidden = true;\n }\n function lintVars() { return { ...vars(), lint: lintIssues.join('\\n') }; }\n function fixLint() { if (!cur || !lintIssues.length || typeof sendWith !== 'function') return; const text = tpl(E.pLint, lintVars()); askSt.textContent = 'fixing lint'; sendWith(text, focusPayload()); }\n function copyLint() { navigator.clipboard.writeText(lintIssues.join('\\n')); }\n function maybeAutoFix() {\n if (!E.autoFixLint || !cur || !lintIssues.length || (window.inflightTurn && window.inflightTurn())) return;\n const key = cur.id + ':' + lintIssues.join('|');\n autoFixed[cur.id] = autoFixed[cur.id] || { rounds: 0, last: '' };\n const st = autoFixed[cur.id]; if (st.rounds >= 2 || st.last === key) return;\n st.rounds++; st.last = key; setTimeout(fixLint, 400);\n }\n function lintLocal() { if (!E.lintOnType || !active || !/\\.m?js$/.test(active)) return; try { new Function(ED.get().replace(/^\\s*import\\s[\\s\\S]*?from\\s*[\"'][^\"']+[\"']\\s*;?/gm, '').replace(/^\\s*export\\s*\\{[\\s\\S]*?\\}\\s*;?/gm, '').replace(/^\\s*export\\s+(default\\s+)?/gm, '')); showLint(null); } catch (e) { showLint([active + ': ' + e.message]); } }\n // textarea fallback keys\n ta.addEventListener('input', onEdit); ta.addEventListener('keyup', updPos); ta.addEventListener('click', updPos);\n ta.addEventListener('keydown', (e) => {\n const s = ta.selectionStart, en = ta.selectionEnd, ind = E.softTabs ? ' '.repeat(E.tabSize) : '\\t';\n if ((e.ctrlKey || e.metaKey) && e.key.toLowerCase() === 's') { e.preventDefault(); save(); }\n else if ((e.ctrlKey || e.metaKey) && e.key === 'Enter') { e.preventDefault(); reload(true); }\n else if ((e.ctrlKey || e.metaKey) && e.key.toLowerCase() === 'f') { e.preventDefault(); findToggle(true); }\n else if (e.key === 'Tab') { e.preventDefault(); ta.setRangeText(ind, s, en, 'end'); ta.dispatchEvent(new Event('input')); }\n else if (e.key === 'Enter' && E.autoIndent) { const line = ta.value.slice(0, s).split('\\n').pop(); const base = line.match(/^\\s*/)[0] + (/[{(\\[:]\\s*$/.test(line) ? ind : ''); e.preventDefault(); ta.setRangeText('\\n' + base, s, en, 'end'); ta.dispatchEvent(new Event('input')); }\n });\n\n // ─── find / replace ─────────────────────────────────────────────────────────\n const fbox = $('ap-find'), fin = $('ap-find-in'), frep = $('ap-find-rep'), fn = $('ap-find-n'), fall = $('ap-find-all');\n function findToggle(open) { const o = open ?? fbox.hidden; fbox.hidden = !o; if (o) { fin.value = ED.sel() || fin.value; fin.focus(); fin.select(); } else ED.focus(); }\n function rx() { try { return new RegExp(fin.value, 'gi'); } catch { return new RegExp(fin.value.replace(/[.*+?^$\\{\\}()|[\\]\\\\]/g, '\\\\$&'), 'gi'); } }\n function findNext(dir) {\n if (!fin.value) return;\n if (fall.checked) { const hits = cur.files.filter(f => !isBin(f.path, f.content) && rx().test(f.content)); fn.textContent = hits.length + ' files'; const i = hits.findIndex(f => f.path === active); const nxt = hits[(i + (dir > 0 ? 1 : hits.length - 1)) % Math.max(1, hits.length)]; if (nxt && nxt.path !== active) { showFile(nxt.path); } }\n const v = ED.get(), re = rx(); const all = [...v.matchAll(re)]; fn.textContent = (fall.checked ? fn.textContent + ' · ' : '') + all.length + ' here'; if (!all.length) return;\n const from = cm ? cm.indexFromPos(dir > 0 ? cm.getCursor('to') : cm.getCursor('from')) : (dir > 0 ? ta.selectionEnd : ta.selectionStart);\n const m = dir > 0 ? (all.find(x => x.index >= from) || all[0]) : ([...all].reverse().find(x => x.index < from) || all[all.length - 1]);\n if (cm) { cm.setSelection(cm.posFromIndex(m.index), cm.posFromIndex(m.index + m[0].length)); cm.scrollIntoView(cm.posFromIndex(m.index), 60); } else { ta.focus(); ta.setSelectionRange(m.index, m.index + m[0].length); }\n }\n function replaceOne() { const sel = ED.sel(); if (!sel || !rx().test(sel)) { findNext(1); return; } if (cm) cm.replaceSelection(sel.replace(rx(), frep.value)); else { ta.setRangeText(sel.replace(rx(), frep.value), ta.selectionStart, ta.selectionEnd, 'end'); ta.dispatchEvent(new Event('input')); } if (cm) onEdit(); findNext(1); }\n function replaceAll() {\n if (!fin.value) return; let n = 0;\n const targets = fall.checked ? cur.files.filter(f => !isBin(f.path, f.content)) : cur.files.filter(f => f.path === active);\n for (const f of targets) { const out = f.content.replace(rx(), (m) => { n++; return frep.value; }); if (out !== f.content) { f.content = out; cur.dirty.add(f.path); } }\n fn.textContent = n + ' replaced'; showFile(active); scheduleLive();\n }\n fin.addEventListener('keydown', (e) => { if (e.key === 'Enter') findNext(e.shiftKey ? -1 : 1); if (e.key === 'Escape') findToggle(false); });\n\n // ─── save / files ───────────────────────────────────────────────────────────\n async function save() {\n if (!cur || !cur.dirty.size) return true;\n const files = [...cur.dirty].map(p => cur.files.find(x => x.path === p)).filter(Boolean).map(f => ({ path: f.path, content: f.content }));\n const r = await post('?artifact_save_many', { id: cur.id, session: cur.session, files });\n if (!r.ok) { state.textContent = 'save failed: ' + r.status; return false; }\n const j = await r.json(); for (const f of files) { cur.dirty.delete(f.path); cur.orig[f.path] = f.content; }\n showLint(j.issues); renderTree(); updState(); state.textContent += ' · saved'; clearConsole(); reload(true); return true;\n }\n function ask(label, def) { return new Promise((res) => { const box = $('ap-prompt'), inp = $('ap-prompt-in'); inp.placeholder = label; inp.value = def || ''; box.hidden = false; inp.focus(); inp.select();\n const done = (v) => { box.hidden = true; $('ap-prompt-ok').onclick = $('ap-prompt-no').onclick = null; inp.onkeydown = null; res(v); };\n $('ap-prompt-ok').onclick = () => done(inp.value.trim()); $('ap-prompt-no').onclick = () => done(null); inp.onkeydown = (e) => { if (e.key === 'Enter') done(inp.value.trim()); if (e.key === 'Escape') done(null); }; }); }\n async function addFile() { const p = await ask('path, e.g. src/game/enemy.js'); if (!p || !cur) return; const clean = p.replace(/^\\/+/, ''); if (!cur.files.some(f => f.path === clean)) cur.files.push({ path: clean, content: '' }); cur.dirty.add(clean); showFile(clean); ED.focus(); }\n async function renameFile() {\n if (!cur || !active || /^index\\.html?$/.test(active)) return; const p = await ask('new path', active); if (!p || p === active) return;\n const f = cur.files.find(x => x.path === active); const old = active;\n await post('?artifact_save_many', { id: cur.id, session: cur.session, files: [{ path: p, content: f.content }] });\n await post('?artifact_delete_file', { id: cur.id, path: old });\n f.path = p; cur.dirty.delete(old); cur.orig[p] = f.content; showFile(p); reload(true);\n }\n async function deleteFile() {\n if (!cur || !active || /^index\\.html?$/.test(active)) return; const ok = await ask('type DELETE to remove ' + active); if (ok !== 'DELETE') return;\n await post('?artifact_delete_file', { id: cur.id, path: active });\n cur.files = cur.files.filter(f => f.path !== active); cur.dirty.delete(active); showFile(cur.files[0].path); reload(true);\n }\n function revert() { if (!cur || !active) return; const f = cur.files.find(x => x.path === active); f.content = cur.orig[active] ?? f.content; cur.dirty.delete(active); showFile(active); scheduleLive(); }\n function copy() { if (cur && active) navigator.clipboard.writeText(cur.files.find(x => x.path === active)?.content ?? ''); }\n async function loadZip() { if (!window.JSZip) await new Promise((res, rej) => { const s = document.createElement('script'); s.src = 'https://cdnjs.cloudflare.com/ajax/libs/jszip/3.10.1/jszip.min.js'; s.onload = res; s.onerror = rej; document.head.appendChild(s); }); return window.JSZip; }\n async function zip() {\n if (!cur) return; const J = await loadZip(); const z = new J();\n for (const f of cur.files) { if (isBin(f.path, f.content)) { const m = f.content.match(/^data:[^;,]*;base64,(.*)$/s); if (m) z.file(f.path, m[1], { base64: true }); } else z.file(f.path, f.content); }\n const blob = await z.generateAsync({ type: 'blob' }); const a = document.createElement('a'); a.href = URL.createObjectURL(blob); a.download = (cur.title || 'artifact').replace(/[^\\w.-]+/g, '_') + '.zip'; a.click();\n }\n /** Upload loose files or a .zip into the open artifact (or create a new artifact from a zip when none is open). Text as-is, binaries as data: URLs. */\n async function upload(fileList) {\n const files = [];\n const readText = (f) => f.text(); const readData = (f) => new Promise((res) => { const r = new FileReader(); r.onload = () => res(r.result); r.readAsDataURL(f); });\n for (const f of fileList) {\n if (/\\.zip$/i.test(f.name)) { const J = await loadZip(); const z = await J.loadAsync(f); const names = Object.keys(z.files).filter(n => !z.files[n].dir && !/(^|\\/)(\\.|__MACOSX)/.test(n)); const strip = names.every(n => n.includes('/')) && new Set(names.map(n => n.split('/')[0])).size === 1 ? names[0].split('/')[0].length + 1 : 0;\n for (const n of names) { const path = n.slice(strip); if (!path) continue; if (isBin(path)) { const b64 = await z.files[n].async('base64'); files.push({ path, content: 'data:' + (mimeOf(path)) + ';base64,' + b64 }); } else files.push({ path, content: await z.files[n].async('string') }); } }\n else files.push({ path: f.name, content: isBin(f.name) ? await readData(f) : await readText(f) });\n }\n if (!files.length) return;\n if (!cur) { const r = await post('?artifact_save_many', { session: window.currentSession, title: files.find(f => /\\.zip$/i.test(f.path))?.path || 'Imported', files }); const j = await r.json(); if (j.id) { open(j.id, 'split'); listForSession(window.currentSession); } return; }\n for (const f of files) { const ex = cur.files.find(x => x.path === f.path); if (ex) ex.content = f.content; else cur.files.push(f); cur.dirty.add(f.path); }\n renderTree(); updState(); state.textContent = files.length + ' file' + (files.length > 1 ? 's' : '') + ' added (unsaved)'; if (!E.autoSave) scheduleLive(); else scheduleSave();\n }\n const mimeOf = (p) => ({ png: 'image/png', jpg: 'image/jpeg', jpeg: 'image/jpeg', gif: 'image/gif', webp: 'image/webp', ico: 'image/x-icon', woff: 'font/woff', woff2: 'font/woff2', ttf: 'font/ttf', otf: 'font/otf', mp3: 'audio/mpeg', ogg: 'audio/ogg', wav: 'audio/wav', mp4: 'video/mp4', webm: 'video/webm', wasm: 'application/wasm' })[(p.split('.').pop() || '').toLowerCase()] || 'application/octet-stream';\n const tree = $('ap-tree');\n ['dragenter', 'dragover'].forEach(ev => tree.addEventListener(ev, (e) => { e.preventDefault(); tree.classList.add('drag'); }));\n ['dragleave', 'drop'].forEach(ev => tree.addEventListener(ev, (e) => { e.preventDefault(); tree.classList.remove('drag'); if (ev === 'drop' && e.dataTransfer?.files?.length) upload(e.dataTransfer.files); }));\n\n // ─── jobs strip (swarm builds) ──────────────────────────────────────────────\n const jobs = {};\n function setJobs(list) { if (list === null) { jobsEl.hidden = true; jobsEl.innerHTML = ''; for (const k in jobs) delete jobs[k]; return; } for (const j of list) jobs[j.path] = j.st; jobsEl.hidden = false; jobsEl.innerHTML = Object.entries(jobs).map(([p, s]) => '' + p.split('/').pop() + '').join(''); }\n\n // ─── composer ───────────────────────────────────────────────────────────────\n function prefill(text) { askIn.value = text; askIn.dispatchEvent(new Event('input')); askIn.focus(); askIn.setSelectionRange(askIn.value.length, askIn.value.length); }\n function askEdit() { if (cur) prefill(tpl(E.pAsk, vars())); }\n function askFix() { if (cur) prefill(tpl(E.pFix, vars())); }\n function askSelection() { if (!cur) return; const v = vars(); if (!v.selection) { state.textContent = 'select some text first'; return; } prefill(tpl(E.pSel, v)); }\n function focusPayload() { if (!cur || el.classList.contains('closed')) return null; const v = vars(); return { focus: { id: cur.id, title: cur.title, file: active || '', files: cur.files.map(f => f.path), selection: (v.selection || '').slice(0, 4000), errors: lastErrors.slice(-E.errorsToAI) } }; }\n async function send() {\n const t = askIn.value.trim(); if (!t || !cur || typeof sendWith !== 'function') return;\n const fp = focusPayload(), text = (E.pPrefix ? tpl(E.pPrefix, vars()) + '\\n' : '') + t;\n if (E.saveBeforeAsk && cur.dirty.size) { askSt.textContent = 'saving'; if (!(await save())) return; }\n askIn.value = ''; askIn.dispatchEvent(new Event('input')); askGo.disabled = true; askSt.textContent = 'sending';\n sendWith(text, fp);\n }\n askIn.addEventListener('input', () => { askIn.style.height = 'auto'; askIn.style.height = Math.min(askIn.scrollHeight, 120) + 'px'; });\n askIn.addEventListener('keydown', (e) => { if (e.key === 'Enter' && (e.ctrlKey || e.metaKey)) { e.preventDefault(); send(); } });\n document.addEventListener('omni:typing', (e) => { if (el.classList.contains('closed')) return; askSt.textContent = e.detail.on ? (e.detail.label + (e.detail.progress ? ' · ' + e.detail.progress.replace(/<[^>]+>/g, '') : '')) : ''; });\n document.addEventListener('omni:turn', (e) => {\n askGo.disabled = false; askSt.textContent = e.detail.error ? 'error' : '';\n if (el.classList.contains('closed') || !E.focusReply) return;\n const f = e.detail.final; reply.hidden = false; $('ap-reply-t').textContent = e.detail.error ? 'error' : 'reply';\n replyB.innerHTML = e.detail.error ? '' + MD.esc(e.detail.error) + '' : MD.render(f && f.reply || '');\n if (!e.detail.error && window.MD && MD.enhance) MD.enhance(replyB);\n });\n function hideReply() { reply.hidden = true; }\n\n // ─── resize + keys ──────────────────────────────────────────────────────────\n const drag = (handle, onMove) => handle.addEventListener('pointerdown', (e) => { e.preventDefault(); handle.setPointerCapture(e.pointerId); frame.style.pointerEvents = 'none'; const mv = (ev) => onMove(ev); const up = () => { handle.removeEventListener('pointermove', mv); handle.removeEventListener('pointerup', up); frame.style.pointerEvents = ''; saveE(); }; handle.addEventListener('pointermove', mv); handle.addEventListener('pointerup', up); });\n drag($('ap-resize'), (ev) => { E.width = Math.round(Math.max(30, Math.min(100, (window.innerWidth - ev.clientX) / window.innerWidth * 100))); document.documentElement.style.setProperty('--ap-w', E.width + 'vw'); });\n drag($('ap-divider'), (ev) => { const body = $('ap-body'), r = body.getBoundingClientRect(); const horiz = getComputedStyle(body).flexDirection === 'row'; let pct = horiz ? (ev.clientX - r.left) / r.width * 100 : (ev.clientY - r.top) / r.height * 100; if (E.side !== 'left') pct = 100 - pct; E.split = Math.round(Math.max(15, Math.min(85, pct))); document.documentElement.style.setProperty('--ap-split', E.split + '%'); ED.refresh(); });\n liveCb.addEventListener('change', () => { E.live = liveCb.checked; saveE(); reload(!liveCb.checked); });\n const narrow = () => el.classList.toggle('narrow', window.innerWidth < 1000); window.addEventListener('resize', () => { narrow(); ED.refresh(); }); narrow();\n document.addEventListener('keydown', (e) => {\n if (el.classList.contains('closed')) return;\n if (e.key === 'Escape') { if (!fbox.hidden) findToggle(false); else if (!setEl.classList.contains('closed')) toggleSettings(false); else if (document.activeElement !== ta && !(cm && cm.hasFocus()) && $('ap-prompt').hidden) close(); }\n else if ((e.ctrlKey || e.metaKey) && e.key === ',') { e.preventDefault(); toggleSettings(); }\n });\n\n // ─── sidebar list ───────────────────────────────────────────────────────────\n async function listForSession(session) {\n const host = $('sb-artifacts'); if (!host) return;\n try { listCache = await (await fetch('?artifacts&session=' + encodeURIComponent(session))).json();\n host.innerHTML = listCache.length ? '

artifacts

' : '';\n for (const a of listCache) { const b = document.createElement('button'); b.className = 'sb-art'; b.textContent = '#' + a.id + ' ' + a.title; b.onclick = () => open(a.id); host.appendChild(b); }\n } catch {}\n }\n applyE();\n return { open, close, restore, view, reload, save, addFile, renameFile, deleteFile, revert, copy, zip, upload, askEdit, askFix, askSelection, send, hideReply, clearConsole, hideConsole, listForSession, toggleSettings, resetPrompts, resetSettings,\n findToggle, findNext, replaceOne, replaceAll, fixLint, copyLint, jobs: setJobs, tpl: (k, v) => tpl(E['p' + k[0].toUpperCase() + k.slice(1)] || PDEF['p' + k[0].toUpperCase() + k.slice(1)], v), list: () => listCache, lastErrors: () => lastErrors.slice(-E.errorsToAI),\n focusPayload, settings: () => E, setSettings: (o) => { if (o && typeof o === 'object') { E = { ...DEF, ...o }; saveE(); } }, current: () => cur };\n})();\nwindow.ART = ART;\n`;\n",[238],"f7688a5e-462d-11f1-b542-42b51c65c3df","text-chunker.ts","1093f85e-462b-11f1-8725-42b51c65c3df",888,["D",1777729866332],"http","// ─────────────────────────────────────────────────────────────────────────────\n// CHUNK ENGINE · mobile-first · Val Town Hono/Deno\n// ─────────────────────────────────────────────────────────────────────────────\n\n// ── Constants ────────────────────────────────────────────────────────────────\n\nconst DEFAULT_CHUNK = 10_000;\nconst MAX_BOUND = 1_000_000;\n\nconst CFG = Object.freeze({\n NORMALIZE: true,\n SAFE_SLICE: true,\n HTML_ESCAPE: true,\n JSON_ESCAPE: true,\n});\n\nconst ESC: Readonly> = Object.freeze({\n \"&\": \"&\",\n \"<\": \"<\",\n \">\": \">\",\n '\"': \""\",\n \"'\": \"'\",\n});\n\n// ── Pure utilities ────────────────────────────────────────────────────────────\n\nfunction escH(s: string): string {\n return CFG.HTML_ESCAPE ? s.replace(/[&<>\"']/g, (c) => ESC[c]) : s;\n}\n\nfunction escJSON(v: unknown): string {\n return JSON.stringify(v)\n .replace(/<\\/script>/gi, \"<\\\\/script>\")\n .replace(/\n

\n Input ▾\n

\n \n

\n \n \n \n

\n\n\n

\n Chunking ▾\n

\n Size\n \n

\n Overlap\n \n

\n Min Length\n \n

\n Boundary Snap\n \n

\n Token Model\n \n

\n \n

\n \n Collapse whitespace\n \n \n \n \n \n \n Reverse order\n \n \n \n \n \n \n NFKC normalize\n \n \n \n \n \n \n Strip BOM\n \n \n \n \n \n

\n\n\n

\n Wrap & Describe ▾\n

\n Global Prefix\n \n

\n Global Postfix\n \n

\n Per-chunk Prefix Template\n \n

\n Per-chunk Postfix Template\n \n

\n Templates: {i} = 1-based index · {n} = total · {s} = start offset · {e} = end offset · {chars} = length · {tokens} = est. tokens · {hash} = 8-char fingerprint\n

\n \n Show chunk descriptions\n \n \n \n \n \n \n Base64 output\n \n \n \n \n \n

\n Description template\n \n

\n\n\n\n\n

\n\n\n\n\n\n`;\n}\n\n// ── Handler ───────────────────────────────────────────────────────────────────\n\nconst handler = async (req: Request): Promise => {\n let text = \"\";\n let reqSize = 0;\n\n if (req.method === \"POST\") {\n const parsed = await parse(req);\n text = parsed.text;\n reqSize = parsed.size;\n }\n\n const size = clamp(reqSize > 0 ? reqSize : DEFAULT_CHUNK, 1, MAX_BOUND);\n const chunks = chunk(text, size);\n const meta = buildMeta(chunks, text.length, size);\n const payload: Payload = Object.freeze({ chunks, meta });\n\n return new Response(html(text, payload), {\n headers: { \"content-type\": \"text/html; charset=utf-8\" },\n });\n};\n\nexport default handler;","text-chunker",[245],"0b992977-7e91-405a-8e51-7642b04aaad8","test.txt","e40c4538-ad82-11f1-82c6-1607ee4eb77e",1432,["D",1789091310842],[254],"11743adc-45e9-11f1-b6ab-42b51c65c3df","test-inference.ts","b9598324-41ca-11f1-898b-42b51c65c3df",752,["D",1777700274460],48,"const INFERENCE_URL = \"https://inference-router-dev.val.run\";\n\nexport default async function (): Promise {\n const res = await fetch(INFERENCE_URL, {\n method: \"POST\",\n headers: { \"Content-Type\": \"application/json\" },\n body: JSON.stringify({\n messages: [\n { role: \"user\", content: \"Hello — testing inference.\" },\n ],\n }),\n });\n\n /*example response:\n\n {\"id\":\"chatcmpl-omni-1600e4d2-90ed-4b01-ade3-3ae3a1e3b239\",\"object\":\"chat.completion\",\"created\":1777700178,\"model\":\"qwen/qwen3-32b\",\"choices\":[{\"index\":0,\"message\":{\"role\":\"assistant\",\"content\":\"\\nOkay, the user sent \\\"Hello — testing inference.\\\" Let me break this down. First, they're greeting me, which is standard. The phrase \\\"testing inference\\\" is key here. They probably want to see how I handle a query that's not straightforward. Maybe they're checking if I can infer the intent behind a vague message.\\n\\nI should consider possible scenarios. Are they testing if I can ask clarifying questions? Or maybe they want to see if I can generate a response without more context. Since the message is short, I need to respond in a way that invites them to provide more details. \\n\\nI should acknowledge their greeting and then ask what specific aspect of inference they're testing. That way, I can tailor my response to their actual needs. Also, making sure the tone is friendly and helpful is important. Let me check if there's any other nuance I might be missing. Maybe they're a developer testing the model's capabilities? Or a student trying to understand how inference works in practice. \\n\\nEither way, the best approach is to be open and ask for clarification. That way, I can provide a more accurate and useful response. I'll go with that.\\n\\n\\nHello! It seems like you're testing how I handle inference or interpretation. Could you clarify what specific aspect you'd like me to demonstrate or explain? For example: \\n\\n- Are you testing my ability to infer intent from ambiguous input? \\n- Do you want an example of logical inference? \\n- Are you curious about how I process context or generate responses? \\n\\nLet me know how I can assist! 😊\"},\"finish_reason\":\"stop\"}],\"_omni_meta\":{\"provider\":{\"id\":3,\"name\":\"groq\",\"base\":\"https://api.groq.com/openai/v1\",\"keyEnv\":\"GROQ_API_KEY\",\"fallbackModel\":\"llama-3.3-70b-versatile\",\"priority\":3},\"latency_ms\":936,\"inflight_requests\":0,\"circuit_states\":[{\"name\":\"openrouter_free\",\"state\":\"CLOSED\"},{\"name\":\"sambanova\",\"state\":\"CLOSED\"},{\"name\":\"groq\",\"state\":\"CLOSED\"},{\"name\":\"github_models\",\"state\":\"CLOSED\"},{\"name\":\"mistral\",\"state\":\"CLOSED\"},{\"name\":\"siliconflow\",\"state\":\"CLOSED\"},{\"name\":\"deepinfra\",\"state\":\"CLOSED\"},{\"name\":\"cerebras\",\"state\":\"CLOSED\"}]}}\n */\n\n return res //?.choices?.[0]?.message?.content,\n ; //provider: res?._omni_meta?.providerId ?? null,});\n}",[260],"7e5924a8-ced9-424b-a38b-31dc646bae61","salvage.ts","d274047a-ae16-11f1-8e73-1607ee4eb77e",1480,["D",1789154846829],"import { salvageArtifactArgs } from \"./app-helpers.ts\";\nconst cut = '{\"id\": 59, \"files\": [{\"path\": \"style.css\", \"content\": \":root{--bg:#000}\\\\nbody{margin:0}\"}, {\"path\": \"store.js\", \"content\": \"export function createStore(r, i) {\\\\n let s = i;\\\\n return { get: () => s };\\\\n}\"}, {\"path\": \"render.js\", \"content\": \"import { get } from \\'./store.js\\';\\\\nconst c = document.getElementById(\\'c\\');\\\\nfunction draw() { c.getContext';\nconst r = salvageArtifactArgs(cut);\nconsole.log(JSON.stringify({ saved: r?.saved, cut: r?.cut, id: r?.args.id, store: (r?.args.files as any)?.[1]?.content.slice(0, 30) }));\nconst cut2 = '{\"title\": \"Novel AI Chat\", \"kind\": \"html\", \"content\": \"

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",[268],"a2ceaabd-1aab-4d96-bc36-68acd85ebc8f","router.tsx","334f4f70-7c3a-11f1-857e-1607ee4eb77e",1513,["D",1789221063505],38,"import {\n getCatalogStatus,\n getInstanceRoster,\n getProviderArray,\n type ProviderInstance,\n} from \"./inference-provider.ts\";\nimport { initDB } from \"./db.ts\";\nimport { inflight } from \"./router-core.ts\";\nimport {\n bindModel,\n loadState,\n resetCooldowns,\n stateView,\n} from \"./router-state.ts\";\nimport {\n clientAuthorized,\n CORS,\n handleChatCompletions,\n json,\n} from \"./router-openai.ts\";\n\nconst OPENAI_PATHS = new Set([\n \"/\",\n \"/chat/completions\",\n \"/v1/chat/completions\",\n]);\nconst OPEN_GET = new Set([\"/\", \"/health\", \"/source\"]);\n\nasync function listUpstreamModels(p: ProviderInstance): Promise {\n const headers: Record = { ...(p.headers ?? {}) };\n if (p.key) headers.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 if (req.method === \"OPTIONS\") {\n return new Response(null, {\n headers: {\n ...CORS,\n \"Access-Control-Allow-Headers\": \"*\",\n \"Access-Control-Allow-Methods\": \"POST, GET, OPTIONS\",\n },\n });\n }\n const init = await initDB();\n if (!init.ok) {\n return json({ error: `DB init failed: ${(init.error ?? \"\")}` }, 500);\n }\n if (\n !(req.method === \"GET\" && OPEN_GET.has(url.pathname)) &&\n !clientAuthorized(req)\n ) {\n return json({\n error: {\n message:\n \"UNAUTHORIZED: send Authorization: Bearer \",\n },\n }, 401);\n }\n\n if (req.method === \"GET\") {\n if (url.pathname === \"/source\") {\n return new Response(null, {\n status: 302,\n headers: {\n Location: \"https://www.val.town/x/ziggyware/free-ai/code/router.tsx\",\n },\n });\n }\n if (url.pathname === \"/health\") {\n await loadState();\n const live = getProviderArray();\n const s = stateView();\n return json({\n status: live.length ? \"ok\" : \"degraded\",\n providers_configured: live.length,\n vendors: [...new Set(live.map((p) => p.vendor))],\n cooling: s.cooling,\n dead_models: s.dead_models,\n resolved: s.resolved,\n caps: s.caps,\n caps_in: s.caps_in,\n tps: s.tps,\n inflight_requests: inflight,\n auth: !!Deno.env.get(\"OMNI_CLIENT_KEYS\"),\n key_policy: Deno.env.get(\"OMNI_KEY_POLICY\") ?? \"depth\",\n vendor_order: (Deno.env.get(\"OMNI_VENDOR_ORDER\") ?? \"\").split(\",\").map((\n s,\n ) => s.trim()).filter(Boolean),\n instances: getInstanceRoster().map((i) => {\n const c = s.cooling.find((x) => x.name === i.name);\n const st = s.stat[i.name];\n return {\n ...i,\n cooling: c ? { reason: c.reason, ms_left: c.ms_left } : null,\n requests: st?.req ?? 0,\n ok: st?.ok ?? 0,\n failed: st?.fail ?? 0,\n last_used: st?.last ?? null,\n last_model: st?.model ?? null,\n last_error: st?.err || null,\n };\n }),\n });\n }\n if (url.pathname === \"/api/providers\") {\n await loadState();\n const s = stateView();\n return json(\n getCatalogStatus().map((p) => ({\n ...p,\n currentModel: s.bound[p.vendor] || p.models.find((m) =>\n !s.dead_models.includes(`${p.vendor}:${m}`)\n ) || s.resolved[p.vendor] || p.fallbackModel,\n cooling: s.cooling.filter((c) => c.name.split(\"#\")[0] === p.vendor),\n })),\n );\n }\n if (url.pathname === \"/api/models\") {\n const pName = url.searchParams.get(\"provider\");\n const p = getProviderArray().find((x) =>\n x.name === pName || x.vendor === pName\n );\n if (!p) return json({ error: \"Provider not found or no key\" }, 404);\n try {\n return json(await listUpstreamModels(p));\n } catch (e: any) {\n return json({ error: e.message, fallback: p.fallbackModel }, 502);\n }\n }\n // OpenAI-style list: aliases, vendors, then every upstream id as \":\"\n if (url.pathname === \"/v1/models\") {\n await loadState();\n const live = getProviderArray();\n const s = stateView();\n const seen = new Set();\n const rows: Record[] = [];\n const add = (\n id: string,\n owned_by: string,\n extra: Record = {},\n ) => {\n if (!seen.has(id)) {\n seen.add(id);\n rows.push({ id, object: \"model\", owned_by, ...extra });\n }\n };\n for (\n const a of [\"auto\", \"fast\", \"coder\", \"reasoning\", \"long\", \"vision\"]\n ) add(a, \"omni\");\n for (const p of live) {\n add(p.vendor, p.vendor, {\n bound: s.bound[p.vendor] || s.resolved[p.vendor] || p.fallbackModel,\n slots: live.filter((x) => x.vendor === p.vendor).length,\n });\n }\n const lists = await Promise.all(\n live.filter((p) => p.slot === 0).map(async (p) => {\n try {\n const l = await listUpstreamModels(p);\n return Array.isArray(l)\n ? l.map((m: any) =>\n `${p.vendor}:${typeof m === \"string\" ? m : m.id}`\n )\n : [];\n } catch {\n return [];\n }\n }),\n );\n for (const id of lists.flat()) add(id, id.split(\":\")[0]);\n return json({ object: \"list\", data: rows });\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/reset\") { // clear cooldowns: {name?: \"groq\" | \"groq#2\" | \"\"}\n const b = await req.json().catch(() => ({}));\n await loadState(true);\n return json({\n ok: true,\n ...(await resetCooldowns(typeof b?.name === \"string\" ? b.name : \"\")),\n });\n }\n if (url.pathname === \"/api/update\") {\n const { provider, model } = await req.json();\n if (typeof provider !== \"string\" || typeof model !== \"string\") {\n return json({ error: \"provider and model required\" }, 400);\n }\n const r = await bindModel(provider.split(\"#\")[0], model);\n return r.ok ? json({ success: true }) : json({ error: r.error }, 500);\n }\n if (!OPENAI_PATHS.has(url.pathname)) {\n return new Response(\"Not Found\", { status: 404 });\n }\n let body: unknown;\n try {\n body = await req.json();\n } catch {\n return json({ error: { message: \"Invalid JSON body\" } }, 400);\n }\n return await handleChatCompletions(body);\n }\n return new Response(\"Not Found\", { status: 404 });\n}","router",[275],"41835688-4fca-4c8c-aae8-4a3428e7bbd1","router-state.ts","b109aa6c-ab4e-11f1-8d3d-1607ee4eb77e",1520,["D",1789222232524],11,"// 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}",[284],"8637bc74-437e-4a9f-9a13-8d93c241f0d9","router-openai.ts","6a8c10a8-ab4d-11f1-aa95-1607ee4eb77e",1491,["D",1789154871933],7,"// OpenAI chat.completions surface. Streams are piped byte-for-byte from\n// upstream; failover happens only before the first upstream byte.\n// Client auth is opt-in: set OMNI_CLIENT_KEYS=\"k1,k2\" and callers must send\n// Authorization: Bearer . Unset ⇒ open endpoint (current behaviour).\nimport { inflight, type RouteHit, routeInference, type RouteOpts } from \"./router-core.ts\";\nimport { noteTps, stateView, tpsFor } from \"./router-state.ts\";\n\n// OpenAI params forwarded verbatim (tools included). Reasoning models (gpt-oss, qwen3.x)\n// spend max_tokens on reasoning first — send ≥256 or reasoning_effort:\"low\".\n// Router-only params (not forwarded): omni_deadline_ms — must finish within this;\n// omni_expect_tokens — expected output length, used with vendor tps to pick a\n// model that can make the deadline (quality-first among those that fit).\nconst PASS = [\"temperature\", \"max_tokens\", \"max_completion_tokens\", \"tools\", \"tool_choice\", \"response_format\", \"reasoning_effort\", \"stop\", \"top_p\", \"seed\", \"presence_penalty\", \"frequency_penalty\", \"parallel_tool_calls\", \"user\"];\nexport const CORS = { \"Access-Control-Allow-Origin\": \"*\" };\nexport const json = (body: unknown, status = 200, extra: Record = {}) =>\n new Response(JSON.stringify(body), { status, headers: { \"Content-Type\": \"application/json\", ...CORS, ...extra } });\n\nexport function clientAuthorized(req: Request): boolean {\n const keys = (Deno.env.get(\"OMNI_CLIENT_KEYS\") ?? \"\").split(\",\").map((s) => s.trim()).filter(Boolean);\n if (!keys.length) return true;\n const h = req.headers.get(\"authorization\") ?? \"\";\n const tok = h.replace(/^Bearer\\s+/i, \"\").trim() || new URL(req.url).searchParams.get(\"key\") || \"\";\n return keys.includes(tok);\n}\n\nconst sanitize = (p: RouteHit[\"provider\"]) => ({ id: p.name, name: p.name, vendor: p.vendor, base: p.base, keyEnv: p.keyEnv, slot: p.slot, priority: p.priority, fallbackModel: p.fallbackModel });\n\nconst meta = (hit: RouteHit) => ({\n provider: sanitize(hit.provider), model: hit.model, latency_ms: hit.latencyMs, attempts: hit.attempts, max_tokens_sent: hit.sentMax ?? null, instance: hit.provider.name, slot: hit.provider.slot,\n inflight_requests: inflight, cooling: stateView().cooling,\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 extra: Record = {};\n for (const k of PASS) if (body[k] !== undefined && body[k] !== null) extra[k] = body[k];\n const opts: RouteOpts = { model: body.model, stream: !!body.stream, extra, deadlineMs: +body.omni_deadline_ms || undefined, expectTokens: +body.omni_expect_tokens || undefined, keyPolicy: [\"depth\", \"breadth\", \"rr\"].includes(body.omni_key_policy) ? body.omni_key_policy : undefined,\n exclude: Array.isArray(body.omni_exclude) ? body.omni_exclude.filter((v: unknown) => typeof v === \"string\").slice(0, 24) : undefined,\n vendorOrder: typeof body.omni_vendor_order === \"string\" ? body.omni_vendor_order.split(\",\").map((s: string) => s.trim().toLowerCase()).filter(Boolean).slice(0, 12) : Array.isArray(body.omni_vendor_order) ? body.omni_vendor_order.filter((v: unknown) => typeof v === \"string\").slice(0, 12) : undefined };\n let hit: RouteHit;\n try {\n hit = await routeInference(messages, opts);\n } catch (e: any) {\n return json({ error: { message: e.message, details: e.errors ?? null, attempts: e.attempts ?? 0 } }, e.status ?? 503);\n }\n const m = meta(hit);\n const prov = { \"X-Omni-Provider\": hit.provider.name, \"X-Omni-Model\": hit.model };\n\n if (!opts.stream) {\n const up = await hit.res.json().catch(() => ({}));\n if (!up?.choices?.[0]?.message) return json({ error: { message: \"MALFORMED_UPSTREAM\", details: [`${hit.provider.name}/${hit.model}`] } }, 502);\n const outTok = Number(up?.usage?.completion_tokens); if (outTok) await noteTps(hit.provider.vendor, hit.model, outTok, hit.latencyMs);\n return json({ ...up, model: up.model ?? hit.model, _omni_meta: { ...m, tokens_out: outTok || null, tps_learned: Math.round(tpsFor(hit.provider.vendor, hit.model) ?? 0) || null } }, 200, prov);\n }\n\n const enc = new TextEncoder();\n const upstream = hit.res.body ?? new ReadableStream({ start(c) { c.close(); } });\n const out = new ReadableStream({\n async start(c) {\n c.enqueue(enc.encode(`: omni ${JSON.stringify(m)}\\n\\n`)); // SSE comment: SDKs ignore, humans can read\n const r = upstream.getReader();\n try {\n for (;;) { const { done, value } = await r.read(); if (done) break; c.enqueue(value); }\n } catch (e: any) {\n c.enqueue(enc.encode(`data: ${JSON.stringify({ error: { message: `UPSTREAM_STREAM_ABORTED: ${e.message}` } })}\\n\\ndata: [DONE]\\n\\n`));\n }\n c.close();\n },\n cancel() { hit.res.body?.cancel().catch(() => {}); },\n });\n return new Response(out, { headers: { \"Content-Type\": \"text/event-stream\", \"Cache-Control\": \"no-cache\", ...prov, ...CORS } });\n}\n",[292],"3ba4d79f-be8e-41dc-a05d-9f0d06316c50","router-discover.ts","b14c41f6-ab4e-11f1-8fec-1607ee4eb77e",1406,["D",1789017793310],1,"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",[300],"dad245c5-9d62-4061-85b4-c2618218bf47",["D",1789251211713],27,[28],"f4612f85-5e74-4631-bd53-239e820c0dee","router-attempt.ts","c60156b4-abcf-11f1-b2a4-1607ee4eb77e",1445,["D",1789091784441],8,"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",[312],"907e461a-ab53-4642-a09e-fdbb24410980","rate-limit-tracker.ts","2ed0f30a-aae2-11f1-ad3f-1607ee4eb77e",1219,["D",1788802411124],3,"// 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}",[320],"14c474e8-2e93-4d58-a954-adbffc7c5b8b","quality-report.ts","fc3a82a2-ad8a-11f1-9829-1607ee4eb77e",1457,["D",1789094787388],"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",[328],"fada88c3-bc24-40d1-af47-b564ac76d0c4","prompts.ts","2f259634-abe2-11f1-966c-1607ee4eb77e",1487,["D",1789154860496],10,"// Every prompt the app sends, in one place. Each is a template with {{vars}};\n// users override any of them per request (settings.prompts[key]) and the UI\n// fetches this table from GET ?prompts to show defaults + placeholders.\nimport { PROMPT_EXAMPLES } from \"./prompt-examples.ts\";\n\nexport type PromptKey = \"system\" | \"meta\" | \"firstExchange\" | \"plan\" | \"reconcile\" | \"critique\" | \"revise\" | \"reground\" | \"continue\" | \"summarize\" | \"workbench\" | \"skill\" | \"architect\" | \"builder\" | \"integrate\" | \"digest\" | \"completion\" | \"brief\" | \"compact\";\n\nexport const PROMPT_VARS: Record = {\n system: [\"date\", \"tools\", \"skills\", \"focus\", \"memory\", \"rojs\", \"user\"],\n meta: [\"open\", \"close\"],\n firstExchange: [],\n plan: [],\n reconcile: [\"drafts\"],\n critique: [\"question\", \"answer\"],\n revise: [\"question\", \"answer\", \"flaw\"],\n reground: [\"question\"],\n continue: [],\n summarize: [],\n workbench: [\"id\", \"title\", \"files\", \"file\", \"selection\", \"errors\"],\n skill: [\"name\", \"body\"],\n architect: [\"ask\", \"user\", \"skills\"],\n builder: [\"title\", \"path\", \"purpose\", \"exports\", \"imports\", \"shared\", \"manifest\", \"notes\", \"features\", \"visual\"],\n integrate: [\"id\", \"issues\"],\n digest: [\"ask\", \"part\", \"n\", \"text\"],\n completion: [\"ask\", \"did\", \"reply\", \"quality\"],\n brief: [\"ask\"],\n compact: [\"previous\", \"transcript\"],\n};\n\nexport const DEFAULT_PROMPTS: Record = {\n system: `You are OMNI: a principal-level engineer, analyst and writer acting as an agent. Date {{date}}. Tools available this turn: {{tools}}.\n\nDoctrine\n1. Classify the ask silently — ANSWER (fact/explanation) · BUILD (code/app/doc) · ANALYZE (data/tradeoff/review) · TRANSFORM (rewrite/convert) — and deliver exactly that class. Never print the class label. No restating the question, no preamble, no closing offers, no \"certainly\".\n2. Density: every sentence load-bearing. Numbers carry units and provenance. Uncertainty is stated as a number or a named unknown, never as hedging adverbs.\n3. Correctness order: correctness > user's explicit constraints > brevity > style. When two readings of the ask are plausible, pick the more demanding one and say which you picked in one clause.\n4. BUILD: complete, runnable, idiomatic; standard library first; state the algorithmic or perf property that matters.\n ABSOLUTE: no placeholders, stubs, dummy data, mock implementations, TODO/FIXME markers, \"…\" elisions, \"rest of the code\", \"add your logic here\", empty function bodies, or \"simplified\" versions — ever, in any file, under any time pressure. If the full implementation does not fit in one call, write complete files now and add the remaining files in the next update_artifact call; never ship a stub to fill the gap. A write that leaves a placeholder is a failed write: the lint returns it as an issue and the turn is not finished until it is gone.\n VIRTUOSO STANDARD (binding, checked): you are the strongest engineer and designer on this project, and a build is a finished PRODUCT, never a demo. Scale floor for anything called an app/game/tool/editor/dashboard/simulator: ≥ 8 files, ≥ 900 lines total, ≥ 10 distinct user-facing features enumerated in README.md and all implemented. Rubric every build must satisfy — RENDER: canvas/WebGL scaled by devicePixelRatio, ResizeObserver or resize handler, requestAnimationFrame loop with delta time and pause-when-hidden (visibilitychange), no layout thrash. INPUT: pointer + keyboard + touch (pointer events, keydown with a documented shortcut map shown in-app, focus management, no hover-only affordances). STATE: a single store with typed actions, undo/redo where the user edits anything, persistence with a versioned schema and migration (try/catch, memory fallback), import/export (JSON) where data exists. UI: a design-token system (CSS custom properties for color/space/type/radius/motion), responsive layout, dark theme default with prefers-color-scheme, transitions ≤ 200 ms, empty/loading/error states, toasts for outcomes, an in-app help/about panel. ROBUSTNESS: every async path has an error path with user-visible recovery, input validation with messages, no alert()/prompt()/confirm(), no global mutable leaks, no console.log left in shipped code (a real logger with levels if logging exists). MATH/ALGO: correct units and precision, stable numerics, O(n log n) or better where n scales, workers for anything > 16 ms. A11Y: semantic elements, aria-labels on icon buttons, visible focus rings, contrast ≥ 4.5:1. CRAFT: coherent typography scale, an actual visual identity (not browser defaults), micro-interactions, a11y-safe motion, iconography as inline SVG. Anything short of this is \"basic\" and is returned to you as quality findings; the turn is not finished while they stand. Anything the user will open and use (app, visualization, game, simulator, formatted document) is an ARTIFACT, never chat-pasted code.\n Artifact model — a project of files: \\`content\\` = entry index.html; \\`files\\` = siblings referenced by RELATIVE path (\n \n \n \n \n\n`;\n",[392],"f4de80e9-04a6-4602-817c-5664568fd8f5","lint-portability.ts","f80151d4-ad8a-11f1-bcb5-1607ee4eb77e",1456,["D",1789094780295],"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",[400],"fefa6f1c-e2c2-451a-8699-878655f4fb5f","inference-provider.ts","f5b23648-43b6-11f1-91e9-42b51c65c3df",1407,["D",1789017793755],51,"// Instances = vendor × key. Keys per vendor come from every one of:\n// , 1 … 9 (numbered slots)\n// comma-separated values inside any of those (\"k1,k2,k3\")\n// so N accounts at one vendor = N instances, each with its own breaker/cooldown.\n// OMNI_PROVIDERS: JSON array of Vendor rows merged by `vendor` (add or override).\n// OMNI_DISABLE: comma list of vendors to skip.\nimport { CATALOG, type Vendor } from \"./catalog.ts\";\n\nexport type ProviderInstance = Vendor & {\n name: string; // \"\" or \"#\"\n slot: number;\n key: string | null;\n keyEnv: string | null; // resolved env name for this slot\n fallbackModel: string;\n};\n\nconst KEY_SLOTS = 10;\nconst MAX_KEYS = 10; // per vendor, after slot expansion + comma split\nconst validKey = (k: string) => /^[\\x21-\\x7e]{8,}$/.test(k); // header-safe ASCII, no whitespace\nconst interp = (s: string) => s.replace(/\\$\\{(\\w+)\\}/g, (m, v) => Deno.env.get(v) ?? m); // unresolved stays, row dropped below\n\nlet cache: { vendors: Vendor[]; at: number } | null = null;\nexport const invalidateVendors = () => { cache = null; };\nexport function getVendors(): Vendor[] {\n if (cache && Date.now() - cache.at < 60_000) return cache.vendors;\n const byVendor = new Map(CATALOG.map((v) => [v.vendor, { ...v }]));\n try {\n const extra = JSON.parse(Deno.env.get(\"OMNI_PROVIDERS\") ?? \"[]\") as Partial[];\n for (const e of extra) {\n if (!e.vendor) continue;\n byVendor.set(e.vendor, { ...(byVendor.get(e.vendor) ?? { base: \"\", keyEnv: null, models: [], priority: 50 }), ...e } as Vendor);\n }\n } catch (e) { console.warn(\"OMNI_PROVIDERS unparsable:\", e); }\n const disabled = new Set((Deno.env.get(\"OMNI_DISABLE\") ?? \"\").split(\",\").map((s) => s.trim()).filter(Boolean));\n const vendors = [...byVendor.values()].filter((v) => !disabled.has(v.vendor) && v.base && v.models.length)\n .map((v) => ({ ...v, base: interp(v.base).replace(/\\/$/, \"\") }))\n .filter((v) => !/\\$\\{/.test(v.base))\n .sort((a, b) => a.priority - b.priority);\n cache = { vendors, at: Date.now() };\n return vendors;\n}\n\nfunction keysFor(v: Vendor): { keyEnv: string; key: string }[] {\n if (!v.keyEnv) return [];\n const out: { keyEnv: string; key: string }[] = [];\n for (let i = 0; i < KEY_SLOTS; i++) {\n const env = v.keyEnv + (i ? String(i) : \"\");\n for (const key of (Deno.env.get(env) ?? \"\").split(\",\").map((s) => s.trim()).filter(Boolean)) {\n if (validKey(key)) out.push({ keyEnv: env, key });\n else console.warn(`[omni] ${env}: skipping malformed key (${key.length} chars)`);\n }\n }\n return out.slice(0, MAX_KEYS);\n}\n\nexport function getProviderArray(): ProviderInstance[] {\n const out: ProviderInstance[] = [];\n for (const v of getVendors()) {\n const mk = (slot: number, keyEnv: string | null, key: string | null): ProviderInstance =>\n ({ ...v, name: slot ? `${v.vendor}#${slot}` : v.vendor, slot, keyEnv, key, fallbackModel: v.models[0] });\n if (!v.keyEnv) { out.push(mk(0, null, null)); continue; }\n keysFor(v).forEach((k, i) => out.push(mk(i, k.keyEnv, k.key)));\n }\n return out;\n}\n\n/** One row per vendor×key instance for /health; key shown as env name + last 4 chars only. */\nexport function getInstanceRoster() {\n return getProviderArray().map((p) => ({ name: p.name, vendor: p.vendor, slot: p.slot, keyEnv: p.keyEnv, key: p.key ? \"…\" + p.key.slice(-4) : null, paid: !!p.paid }));\n}\n/** Status rows for UI/health; never carries key material. */\nexport function getCatalogStatus() {\n const live = getProviderArray();\n return getVendors().map((v) => ({\n ...v, name: v.vendor, fallbackModel: v.models[0],\n hasKey: !v.keyEnv || live.some((i) => i.vendor === v.vendor),\n keySlots: live.filter((i) => i.vendor === v.vendor).length,\n }));\n}\n\n/** Request `model` →\n * \"\"|\"auto\"|\"best\" every instance, best model quality first (see catalog QUALITY)\n * \"fast\" every instance, vendor priority (latency) order\n * \"coder\"|\"reasoning\"|\"fast\"|\"long\"|\"vision\" instances whose vendor carries the tag, then the rest\n * \"\" that vendor's instances; \"#n\" one instance\n * \":\" that vendor, forced upstream id\n * \"/\" same as above (OpenRouter-style spelling)\n * unknown → [] */\nexport type Rank = \"quality\" | \"priority\";\nexport type Resolved = { targets: ProviderInstance[]; forcedModel: string | null; rank: Rank; boost?: RegExp };\nexport function resolveTargets(requested: string | undefined, live: ProviderInstance[]): Resolved {\n const none: Resolved = { targets: [], forcedModel: null, rank: \"priority\" };\n if (!requested || requested === \"auto\" || requested === \"best\") return { targets: live, forcedModel: null, rank: \"quality\" };\n if (requested === \"fast\") return { targets: live, forcedModel: null, rank: \"priority\" };\n if (/^(coder|reasoning|long|vision)$/.test(requested)) {\n const tagged = live.filter((p) => p.tags?.includes(requested));\n return { targets: [...tagged, ...live.filter((p) => !tagged.includes(p))], forcedModel: null, rank: \"quality\", boost: requested === \"coder\" ? /coder|code/i : undefined };\n }\n const exact = live.filter((p) => p.name === requested || p.vendor === requested);\n if (exact.length) return { targets: exact, forcedModel: null, rank: \"quality\" };\n const m = requested.match(/^([a-z0-9_-]+)[:/](.+)$/i);\n if (!m) return none;\n const byVendor = live.filter((p) => p.vendor === m[1]);\n return byVendor.length ? { targets: byVendor, forcedModel: m[2], rank: \"priority\" } : none;\n}\n",[407],"40685802-fdec-4000-9169-037c8f481110","e2e9.mjs","05b43346-ad8b-11f1-a3e8-1607ee4eb77e",1479,["D",1789154842693],"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",[415],"c2b7c429-6d10-4a1d-aeec-5d1d722a80c9","e2e15.mjs","a8c67bd6-ae1a-11f1-af4d-1607ee4eb77e",1498,["D",1789156494898],"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",[422],"7f748b60-3e4c-4785-b13e-42adf7bead90","e2e14.mjs","cfb60a80-ae16-11f1-a6aa-1607ee4eb77e",1478,["D",1789154842238],"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",[429],"5c9d1215-2136-4673-8adf-575bc51b1b9a","e2e13.mjs","cf59ed68-ae16-11f1-a59a-1607ee4eb77e",1477,["D",1789154841633],"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",[436],"7f593e81-d60f-407b-abe1-3280cc43debf","e2e12.mjs","ceec123e-ae16-11f1-b8bc-1607ee4eb77e",1476,["D",1789154840909],"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",[443],"c98a97b9-4cfd-4562-806c-462d1abcde70","e2e11.mjs","03081450-ad8b-11f1-bc95-1607ee4eb77e",1459,["D",1789094798792],"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",[450],"eb553536-e6b3-498a-8ea3-1da08f738838","e2e10.mjs","008bd9f0-ad8b-11f1-ac72-1607ee4eb77e",1475,["D",1789154839726],"// 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",[457],"20feeac4-d28f-472a-b0b3-91f1b78e6a48","db.ts","0b4ab11c-43c9-11f1-8043-42b51c65c3df",1381,["D",1789001806726],66,"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 artifact: {\n id: number;\n session: string;\n title: string;\n kind: \"html\" | \"svg\" | \"markdown\";\n content: string;\n ts: number;\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 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 UNIQUE(session, title))`,\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];\n\nconst initDB = () =>\n attempt(() =>\n sqlite.batch([\n `PRAGMA foreign_keys = ON`,\n ...Object.values(DDL),\n ...INDEXES,\n ])\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}, ${ts})\n ON CONFLICT(id) DO UPDATE SET name = excluded.name, ts = excluded.ts`);\n\nexport {\n all,\n type Bound,\n type DbError,\n Err,\n initDB,\n insertMessage,\n insertMessageSafe,\n messagesFor,\n Ok,\n one,\n providerById,\n providersByPriority,\n pruneSession,\n raw,\n type Result,\n type Row,\n run,\n type Schema,\n sql,\n type Table,\n touchSession,\n unwrap,\n upsertMemory,\n};",[464],"f8bca865-5d85-4915-8172-61ec0a2caa65","catalog.ts","b04a98e8-ab4e-11f1-a8c4-1607ee4eb77e",1410,["D",1789018830497],5,"// Vendor catalog. One row per OpenAI-compatible endpoint. `models` is the\n// ordered fallback list; `prefer` picks a replacement from the vendor's live\n// /models when every listed id 404s. `${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",[472],"6b6be713-93fa-47ed-bafe-4eeb65f5abbe","catalog-paid.ts","b0d2b336-ab4e-11f1-a8c4-1607ee4eb77e",1463,["D",1789096031038],"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",[480],"682fd4d5-fa55-46d5-93e0-f51806c1f9c8","catalog-free.ts","b08f5190-ab4e-11f1-b3f3-1607ee4eb77e",1462,["D",1789096030770],"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",[487],"21af926b-06b2-4c30-b28a-c3a89ed5301d","artifacts.ts","4b6b75ba-ac6f-11f1-a399-1607ee4eb77e",1494,["D",1789156468768],9,"// 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",[494],"c25cf08e-64cf-4d3a-b68a-1fd61882e28b","app.tsx","a03409c6-79e4-11f1-bc29-1607ee4eb77e",1525,["D",1789251207587],69,"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 { 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// CONSTANTS\n// ════════════════════════════════════════════════════════════════════════════\n\n// Hop/pass budget governor caps. A single cheap query spends 1 inference pass;\n// escalation spends from a bounded pool so \"brilliant\" never becomes \"slow\".\nconst BUDGET = {\n loopHops: 8, // max tool-hops inside agentLoop (unchanged from original)\n planPass: 1, // deliberation pass (mechanism 5)\n ensembleSamples: 3, // self-consistency fan-out on high-stakes (mechanism 6)\n verifyPass: 1, // adversarial self-refutation (mechanism 1)\n regroundPass: 1, // abstention→retrieval escalation (mechanism 2)\n hardCeilingPasses: 14, // absolute upper bound on total inference calls / turn\n};\n\n// Optional distinct model for the verifier so self-critique doesn't share the\n// generator's exact blind spots. If the router ignores `model`, this is inert.\nconst VERIFIER_MODEL: string | null = null; // e.g. \"different-model-id\"\n\n// ════════════════════════════════════════════════════════════════════════════\n// RATE LIMIT + SANITIZE\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(); // read once, as 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// LEGACY BLOB MEMORY (kept for backward compat; no longer the ratchet path)\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// Deliberately does NOT reuse touchSession (which also bumps ts) — renaming\n// a session shouldn't move it to the top of a recency-sorted list.\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 // Messages first: message.session REFERENCES session(id) under\n // PRAGMA foreign_keys = ON, so the child rows must go before the parent.\n await run(sql`DELETE FROM message WHERE session = ${id}`);\n await run(sql`DELETE FROM session WHERE id = ${id}`);\n}\n\n// pruneSession(id, 0): the \"keep newest 0\" subquery returns no rows, so\n// `id NOT IN (∅)` is true for every row — this clears the session without\n// a second hand-written DELETE statement.\nasync function clearSession(id: string): Promise<void> {\n await pruneSession(id, 0);\n}\n\n// ════════════════════════════════════════════════════════════════════════════\n// MESSAGE CRUD\n// ════════════════════════════════════════════════════════════════════════════\nasync function addMsg(\n role: Row<\"message\">[\"role\"],\n content: string,\n providerId: string | null = null,\n session = \"default\",\n meta: Record<string, unknown> | null = null,\n): Promise<void> {\n const r = await insertMessageSafe({\n role,\n content,\n ts: Date.now(),\n session,\n providerId,\n });\n if (!r.ok) { console.error(`addMsg failed (session=${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); // newest first; folded turns live in the summary\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; // always keep the last exchange\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 // Backfill: assistant rows written before message_meta existed get synthetic create_artifact events for artifacts created in their window.\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// ARTIFACTS — small standalone apps (html/svg/markdown) the agent builds\n// in-chat. Uniqueness is (session, title): re-using a title overwrites the\n// same artifact rather than accumulating duplicates as the agent iterates.\n// ════════════════════════════════════════════════════════════════════════════\n// ════════════════════════════════════════════════════════════════════════════\n// TOOLS (registry unchanged in spirit; memory tools now hit the fact store)\n// ════════════════════════════════════════════════════════════════════════════\ninterface ToolCtx {\n session: string;\n turn?: { artifactId?: number; artifactTitle?: string }; // dedupe: one artifact per turn, later calls update it\n // Per-request tool registry: built-ins plus whatever MCP servers exposed at\n // discovery time. Threaded through the loop instead of read from module\n // globals so MCP availability can vary turn-by-turn without racing.\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: ${k}. Must be html, svg, or markdown.` };\n // one artifact per turn: a second call re-targets the first (title may be refined)\n const t = ctx.turn?.artifactTitle ?? String(title);\n const r = await saveArtifact(ctx.session, t, k, String(content), Array.isArray(files) ? (files as ArtFile[]) : []);\n if (!r.ok) return { error: r.error };\n if (ctx.turn) { ctx.turn.artifactId = r.value.id; ctx.turn.artifactTitle = r.value.title; }\n const lf = await listFiles(r.value.id);\n const issues = lf ? lintArtifact(lf.files) : [];\n const quality = lf ? qualityReport(lf.files) : [];\n return { ok: true, id: r.value.id, url: `/artifact/${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 = ${String(title)} ORDER BY ts DESC LIMIT 1`);\n if (r.ok && r.value) aid = r.value.id;\n }\n if (!aid) return { error: \"artifact not found; pass id or title (call read_artifact first)\" };\n const ops: { path: string; find?: string; replace?: string; content?: string }[] = [];\n if (Array.isArray(edits)) for (const e of edits as any[]) if (e?.path && typeof e.find === \"string\") ops.push({ path: String(e.path), find: e.find, replace: String(e.replace ?? \"\") });\n if (Array.isArray(files)) for (const f of files as any[]) if (f?.path && typeof f.content === \"string\") ops.push({ path: String(f.path), content: f.content });\n if (typeof content === \"string\") ops.push({ path: String(path || \"index.html\"), content });\n else if (typeof find === \"string\") ops.push({ path: String(path || \"index.html\"), find, replace: String(replace ?? \"\") });\n const applied: string[] = [], failed: string[] = [];\n // apply edits to in-memory copies so several edits to one file compose, then write once per file\n const buf = new Map<string, string>();\n for (const op of ops) {\n if (op.content !== undefined) { buf.set(op.path, op.content); applied.push(`${op.path}: replaced (${op.content.length} chars)`); continue; }\n let cur = buf.get(op.path);\n if (cur === undefined) { const r = await readFile(aid, op.path); if (!r) { failed.push(`${op.path}: file not found`); 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// Compose the per-request registry: built-ins win name collisions, because a\n// remote MCP server must never be able to shadow memory/session primitives.\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 };\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>; 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// Per-request user settings (from the UI's settings panel). Every field optional; defaults here.\nconst TURN_BUDGET_MS = 52_000; // free plan: val killed at 60 s (\"Val timed out after 1m\"); keep 8 s for DB writes + response\nclass PassMeter {\n private used = 0;\n private t0 = Date.now();\n public reasoning: string[] = [];\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 // Never overshoot the val's wall clock: a call that cannot get 8 s does not start (the turn returns what it has;\n // the client continues it next turn). The old Math.max(8000, …) forced 8 s even with 2 s left → \"Val timed out after 1m\"\n // and everything in flight was lost.\n if (this.timeLeft < 8_000) throw Object.assign(new Error(\"TURN_BUDGET: no time left for another model call\"), { budget: true });\n let r = await callInference(messages, tools, model ?? this.defaultModel, this.timeLeft - 3_000, gen);\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 // Auto-continue: a reply cut at max_tokens is resumed in place (no tool calls, budget permitting).\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 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.emit({ type: \"retry\", n, error: `continuation failed, keeping ${r.message.content.length} chars: ${String((e as Error).message).slice(0, 120)}` });\n break; // keep the partial reply rather than failing the whole turn\n }\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 // progress guard: a continuation that adds < 200 chars is the cap being too small for this vendor, not a long reply — stop looping\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// ════════════════════════════════════════════════════════════════════════════\n// META CONTROL CHANNEL\n// The final answer emits a trailing fenced block the USER never sees; the\n// governor parses it for {confidence, stakes, abstain}. Kept out of prose so\n// the visible reply stays clean.\n// ════════════════════════════════════════════════════════════════════════════\n\n// ════════════════════════════════════════════════════════════════════════════\n// AGENT LOOP (tool-agnostic; unchanged interface, threads the PassMeter)\n// ════════════════════════════════════════════════════════════════════════════\ninterface AgentResult {\n reply: string;\n sessionName?: string;\n toolEvents: ToolEvent[];\n meta: InferResult[\"meta\"];\n truncated?: boolean;\n}\n// Some open models (Qwen on Groq) emit tool calls as text instead of the\n// structured field. Recover both shapes:\n// <tool_call>{\"name\":\"x\",\"arguments\":{…}}</tool_call>\n// <tool_call><function=x><parameter=k>v</parameter>…</function></tool_call>\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>(); // tool+args → result, for the repeat guard\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 // Once an artifact exists the model only writes a short wrap-up; a small expectation lets slower vendors qualify.\n const built = toolEvents.some((e) => /artifact/.test(e.tool) && !e.error);\n let resp: InferResult | null;\n try {\n // After the first write, each further call is one file (≤ ~200 lines ≈ 3k tokens) or the wrap-up: a bounded expectation lets slower vendors qualify.\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 // Work already done by tools must not be lost to a failed wrap-up pass or an exhausted turn budget: report it, and mark the\n // turn truncated so the client continues in a fresh invocation instead of the val being killed mid-call.\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 // Empty/garbage guard: no tool call and nothing visible (gpt-oss-20b \"thinks\" 30 tokens and stops). Never accept it —\n // retry on a different instance (router excludes the one that produced it), twice, before giving up.\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 // \"empty\" must be judged on the VISIBLE part: a reply that is only the hidden META block is empty (qwen does this after a tool call)\n const visibleLen = (r: string) => parseMeta(r).visible.trim().length;\n const onlyMeta = reply && visibleLen(reply) === 0;\n if (onlyMeta) reply = \"\";\n // Placeholders left by this turn's writes: the turn is not finished. One nudge, then the lint issues ride along in the reply.\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 // Virtuoso gate: for app-scale asks the artifact must clear the quality report too (one nudge; the completion audit carries the rest across turns).\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 // A turn does not end while the artifact it just wrote is broken (syntax, missing export, dangling ref) or stubbed. Two nudges max.\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 // Arguments cut at max_tokens (finish_reason \"length\" with a tool call): keep every complete file, drop the cut one,\n // and tell the model exactly what to resend — one file per call. Progress is never thrown away.\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 ? `, ${sv.cut} was cut` : \"\"}` }); }\n else { console.log(\"function argument error: \" + String(err)); }\n }\n // aliases models reach for on their own\n if (!ctx.tools[toolName]) { const alias = TOOL_ALIASES[toolName]; if (alias && ctx.tools[alias]) toolName = alias; }\n\n const tool = ctx.tools[toolName];\n let output: unknown;\n let eventError: string | undefined;\n // Repeat guard: the same read twice in a turn returns the cached result plus a nudge — models that keep\n // re-reading instead of editing (gpt-oss, qwen) burn the whole budget otherwise.\n const sig = toolName + \":\" + JSON.stringify(args);\n const seenBefore = seen.get(sig);\n if (tool && seenBefore && /^(read_artifact|search_artifact|memory_recall|web_search)$/.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 } });\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 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: ${toolName}`;\n }\n if (!eventError) seen.set(sig, output);\n if (salvaged && output && typeof output === \"object\") output = { ...(output as object), TRUNCATED_CALL: `Your tool call was cut at max_tokens. Saved complete files: [${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 });\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// INTELLIGENCE STAGES (each gated by the governor; each spends from PassMeter)\n// ════════════════════════════════════════════════════════════════════════════\n\n// Mechanism 5 — deliberation: produce an explicit plan before acting.\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// Mechanism 6 — self-consistency: sample the answer path N times, reconcile.\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], 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// Mechanism 1 — adversarial verifier: self-refute, then revise if a flaw is found.\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 }; // never replace a full answer with a cut one\n}\n\n// Mechanism 2 — abstention→reground: if the model abstains or is under-confident\n// on a high-stakes ask, force a retrieval pass rather than let it confabulate.\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// ORCHESTRATOR — the cognition-budget governor wiring the stages together\n// ════════════════════════════════════════════════════════════════════════════\ninterface Cognition {\n reply: string;\n sessionName?: string;\n toolEvents: ToolEvent[];\n meta: InferResult[\"meta\"];\n reasoning: string[];\n truncated: boolean; // reply ended at max_tokens with no budget left to continue\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 // Design brief first: features (10–16), visual system, architecture, edge cases, acceptance — the single biggest lever\n // against \"basic\" output, because a model that has committed to 12 named features builds 12 features.\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 // Primary answer path — ensemble only when pre-estimate smells high-stakes AND\n // budget allows; otherwise a single loop pass.\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 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 }\n\n // Parse the hidden meta channel.\n let { visible, meta: ansMeta } = parseMeta(reply);\n reply = visible;\n\n // The model wrote a whole HTML app into the reply instead of calling the tool: save it as an artifact anyway.\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/${r.value.id}/`, title: r.value.title } }];\n reply = reply.replace(/```(?:html|htm)?\\s*\\n[\\s\\S]*?\\n```/i, `_Saved as artifact **${r.value.title}** — open it above._`);\n }\n }\n }\n\n // Mechanism 2 gate: abstain OR (high stakes AND low confidence) → reground.\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 // Mechanism 1 gate: verify non-trivial answers when budget remains.\n let verifierVerdict = \"SKIPPED\";\n const nonTrivial = complex || hiStakes || toolEvents.length > 0;\n // Verification is skipped once an artifact was built: a 400-token critique cannot judge a 5k-token program, and the pass costs 10–20 s of a 52 s turn.\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 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// CORE RUNTIME\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)) // a client retry after a 503: the ask is already stored\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 = ${session}`); // recency: a turn moves the session up the list\n\n const blob = settings.memory ? await getMemoryBlob() : null; // legacy blob, if present\n const recalled = settings.memory ? await retrieveFacts(text) : []; // surprise-salience recall\n const context = await getContext(session, 24, settings.contextChars);\n const registry0 = await buildToolRegistry();\n const allow = settings.tools ? new Set(settings.tools) : null;\n const registry = allow\n ? { ...registry0, tools: Object.fromEntries(Object.entries(registry0.tools).filter(([k]) => allow.has(k) || k === \"name_session\")),\n schemas: Object.values(registry0.tools).filter((t: any) => allow.has(t.schema?.function?.name) || t.schema?.function?.name === \"name_session\").map((t: any) => t.schema) }\n : registry0;\n\n const memoryHeader = [\n blob?.value ? `blob:${blob.value}` : null,\n recalled.length ? `recalled:\\n- ${recalled.join(\"\\n- \")}` : null,\n ].filter(Boolean).join(\"\\n\\n\") || \"∅\";\n\n // Advertise rojs only when discovery actually found it — a prompt promising\n // tools that aren't in the schema list just teaches the model to hallucinate\n // tool calls.\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 // COUNT(*) doesn't correspond to the message row shape, so it goes through\n // raw() rather than all(\"message\", ...) — same honesty contract, just an\n // explicit generic instead of a table to pin it to.\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 // Oversized ask (pasted logs, whole files): split across requests — parallel digests of each part, then the real\n // turn runs on the digests. This is what makes a 40k-char paste work on 7k-token free tiers.\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${d}`).join(\"\\n\\n\")}`;\n emit({ type: \"stage\", label: \"digests ready\" });\n }\n\n const result = await cognize(\n turnText,\n base,\n { session, tools: registry.tools, schemas: registry.schemas, turn: {} },\n isFirstExchange,\n emit,\n model,\n settings,\n );\n // Session title fallback: weaker models skip name_session, so derive one from the first user message.\n let sessionName = result.sessionName;\n if (!sessionName && !settings.continuation && (sess.name === \"New Chat\" || !sess.name)) {\n sessionName = deriveTitle(text);\n await renameSession(session, sessionName);\n }\n await addMsg(\"assistant\", result.reply, null, session, {\n provider: result.meta?.provider ?? null, model: result.meta?.model ?? null, instance: result.meta?.instance ?? null, reasoning: result.reasoning ?? [], truncated: !!result.truncated,\n continuation: settings.continuation, followup: settings.followup, maxTokensSent: result.meta?.maxTokensSent ?? null, maxTokensAsked: result.meta?.maxTokensAsked ?? null, finishReason: result.meta?.finishReason ?? null, tokensOut: result.meta?.tokensOut ?? null,\n telemetry: result.telemetry ?? null, toolEvents: slimEvents(result.toolEvents ?? []), skills: skillsOn.map((k) => k.name),\n });\n // Completion audit: did this turn finish the ASK, or only its first pass? Cheap fast-tier call; the client\n // auto-continues with the remaining items until done (bounded by its autoComplete setting).\n let completion: { done: boolean; remaining: string[] } | null = null;\n const worked = (result.toolEvents ?? []).some((e) => /artifact|eval_js|rojs_/.test(e.tool));\n if (!settings.continuation && (worked || BUILD_RX.test(text)) && !result.truncated) {\n try {\n const did = (result.toolEvents ?? []).map((e) => `${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 // Compaction hint: when the live (unfolded) history exceeds twice the context budget, the client runs POST ?compact\n // as its own invocation after this reply lands — never inside the turn's 60 s.\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// HTTP HANDLER (branch style preserved; +telemetry surfaced on the q path)\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: ${init.error}`, { status: 500 });\n }\n\n // /artifact/<id>/<path> — multi-file artifacts served path-style so relative references resolve\n const am = url.pathname.match(/^\\/artifact\\/(\\d+)(?:\\/(.*))?$/);\n if (am) {\n const id = Number(am[1]);\n if (am[2] === undefined) return new Response(null, { status: 302, headers: { Location: `/artifact/${id}/` } });\n if (method === \"GET\") {\n const f = await readFile(id, am[2] || \"\");\n if (!f) return new Response(\"not found\", { status: 404 });\n const path = am[2] || (f.kind === \"html\" ? \"index.html\" : f.kind === \"svg\" ? \"index.svg\" : \"index.md\");\n return serveFile(url.origin, path, f.content);\n }\n }\n const compacted = await handleCompact(req, url, (messages, settings) => callInference(messages, undefined, \"fast\", 45_000, { expect: 1400, maxTokens: 2000, temperature: 0.2, router: settings.router, keyPolicy: settings.keyPolicy, vendorOrder: settings.vendorOrder }).then((r) => String(r.message?.content ?? \"\")));\n if (compacted) return compacted;\n if (method === \"GET\" && url.searchParams.has(\"artifact_files\")) {\n const r = await listFiles(Number(url.searchParams.get(\"artifact_files\")));\n if (!r) return Response.json({ error: \"not found\" }, { status: 404 });\n return Response.json({ id: r.artifact.id, title: r.artifact.title, kind: r.artifact.kind, session: r.artifact.session, files: r.files, issues: lintArtifact(r.files) });\n }\n const swarm = await handleSwarm(req, url, { PassMeter, deriveTitle, clean, ensureSession });\n if (swarm) return swarm;\n if (method === \"POST\" && url.searchParams.has(\"artifact_save_many\")) {\n const b = await req.json().catch(() => ({}));\n const id = Number.isInteger(b.id) ? b.id : null;\n if (id) { const a = await listFiles(id); if (!a) return Response.json({ error: \"not found\" }, { status: 404 }); }\n const r = await saveMany(String(b.session ?? \"default\"), id, String(b.title ?? \"\"), Array.isArray(b.files) ? b.files : []);\n return r ? Response.json({ ok: true, ...r, url: `/artifact/${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 // POST {q, session, model} → JSON result. Progress is NOT streamed: Val Town\n // terminates a streamed body ~10 s after headers go out (observed 11.0 s ×3,\n // keepalives don't help), so the turn runs to completion here while the UI\n // polls ?progress=<session>&since=<n> against the in-memory event log.\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 // App-scale builds default to the coder tier (openrouter/hf/mistral/zai/dashscope coder-tagged models); groq's small models produce \"basic\" apps.\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 // Progress is persisted (not just in-memory): Val Town may serve the poll from a\n // different isolate than the one running the turn. Writes are serialized.\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(() => {}); }; // never let a progress write reject unhandled\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}","freeai",[502],"95066c14-3dd8-4a90-b6bc-fa735d33c151","app-swarm.ts","40d1c4cc-acda-11f1-b796-1607ee4eb77e",1439,["D",1789091642883],"// 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 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 return Response.json({ ok: true, path, chars: content.length, lines: content.split(\"\\n\").length, truncated: r?.finishReason === \"length\", issues: own, written: !!w, meta: r?.meta ?? null });\n } catch (e: any) {\n return Response.json({ error: { message: `build ${path} failed: ${String(e?.message ?? e).slice(0, 300)}` }, 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} → upsert (state.status \"done\" deletes the row).\n if (url.searchParams.has(\"job\")) {\n if (method === \"GET\") {\n const session = String(url.searchParams.get(\"session\") ?? \"\");\n const rows = unwrap(await all<\"omni_state\">(\"omni_state\", sql`SELECT * FROM omni_state WHERE key LIKE ${\"job:\" + session + \":%\"} ORDER BY ts DESC LIMIT 20`), []);\n const jobs = rows.map((r) => { try { return { ...JSON.parse(r.value), ts: r.ts }; } catch { return null; } }).filter((j) => j && j.status !== \"done\");\n return Response.json({ jobs });\n }\n if (method === \"POST\") {\n const b = await req.json().catch(() => ({}));\n const session = String(b.session ?? \"\"), id = Number(b.artifactId), state = b.state ?? {};\n if (!session || !Number.isInteger(id)) return Response.json({ error: { message: \"bad session/artifactId\" } }, { status: 400 });\n const key = `job:${session}:${id}`;\n if (state.status === \"done\" || state.status === \"cancelled\") { await run(sql`DELETE FROM omni_state WHERE key = ${key}`); return Response.json({ ok: true, deleted: true }); }\n const value = JSON.stringify({ session, artifactId: id, manifest: state.manifest ?? null, done: [].concat(state.done ?? []).map(String).slice(0, 200), failed: [].concat(state.failed ?? []).map(String).slice(0, 200), status: String(state.status ?? \"building\"), title: String(state.title ?? \"\"), ask: String(state.ask ?? \"\").slice(0, 2000) }).slice(0, 60_000);\n await run(sql`INSERT INTO omni_state (key, value, ts) VALUES (${key}, ${value}, ${Date.now()}) ON CONFLICT(key) DO UPDATE SET value = excluded.value, ts = excluded.ts`);\n return Response.json({ ok: true });\n }\n }\n return null;\n}\n",[511],"65e5dd38-8866-49c9-8459-21d968ef0e6c","app-settings.ts","405c3590-acda-11f1-ac59-1607ee4eb77e",1466,["D",1789150837612],"// 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",[518],"cab61056-500a-421a-a122-ad30bf87dff2","app-infer.ts","d82c8fe0-ae16-11f1-98dc-1607ee4eb77e",1482,["D",1789154856421],"// 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; attempts?: number; retries?: number; instance?: string; maxTokensSent?: number | null; maxTokensAsked?: number; finishReason?: string; tokensOut?: number | null; tokensIn?: number | null };\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 for (let retry = 0; retry < 4; retry++) {\n if (retry) {\n // Prompt-size exhaustion (\"prompt ≈N tok > maxIn\", \"too large\", \"over every model's … limit\"): the vendors that are\n // left cannot take this prompt, so make the prompt fit — drop oldest history, trim tool payloads — and try again at once.\n const sizeBound = /prompt ≈\\d+ tok|too large|input limit|maxIn/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 lastErr = Object.assign(new Error(`router ${res.status}: ${detail}`), { status: res.status, html });\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 const omni = json._omni_meta;\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, attempts: omni?.attempts, 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 };\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",[525],"41533c86-5a57-477e-ae86-425031a9fa64","app-helpers.ts","e15c4b62-ad82-11f1-b116-1607ee4eb77e",1481,["D",1789154854814],"// 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\";\nexport interface ToolEvent { hop: number; tool: string; args: Record<string, unknown>; result?: unknown; error?: string }\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 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, 20).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 evs.map((e) => ({ hop: e.hop, tool: e.tool, args: trim(e.args) as Record<string, unknown>, result: trim(e.result), ...(e.error ? { error: e.error } : {}) }));\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 const at = raw.indexOf('\"files\"');\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 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 === \"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",[532],"cb9a7c12-6079-4513-b12b-1150cd789b28","app-compact.ts","7c7e2310-ae0d-11f1-b403-1607ee4eb77e",1472,["D",1789154314199],"// 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 > ${uptoPrev} ORDER BY id ASC`), []);\n if (rows.length <= KEEP_TAIL + 1) return { ok: false, before: 0, after: prev?.chars ?? 0, covered: 0, uptoId: uptoPrev, reason: \"nothing old enough to fold\" };\n const fold = rows.slice(0, rows.length - KEEP_TAIL);\n const before = fold.reduce((n, r) => n + r.content.length, 0);\n const transcript = fold.map((m) => `[#${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 next: Compact = { summary: summary.slice(0, 12_000), uptoId: fold[fold.length - 1].id, ts: Date.now(), covered: (prev?.covered ?? 0) + fold.length, chars: summary.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",[539],"a41cb04f-4a34-49bc-927c-6182c66feae3","_valtown-workspace.ts","9f0f3fb4-4610-11f1-a0e9-42b51c65c3df",1198,["D",1783676130488],14,"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};",[546],"6afd3ca7-07a8-46e7-b281-28b288595709","_valtown-file-editor.ts","9646c13a-467f-11f1-8bfd-42b51c65c3df",1197,["D",1783676124604],"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}",[554],"08041a77-a517-40c5-93ab-9dbd1e5fd18d","_valtown-api.ts","351d250c-467f-11f1-9311-42b51c65c3df",1196,["D",1783676119609],"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};",[561],"05156804-4768-43fc-b7f5-c39c53f573f3","_inference-router.ts","58cb4aa2-425f-11f1-bc49-42b51c65c3df",1193,["D",1783675978922],"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}","inference-router",[568],"273340ae-9854-4838-aaed-d7c30d3e5c9d","_inference-router-dev.tsx","676ac11c-4595-11f1-99e3-42b51c65c3df",1191,["D",1783675963454],20,"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}

Diff · v1526