OpenAI Spend — Current Month

\n \n \n \n \n \n \n \n \n \n \n \n \n {stats.map((row) => (\n \n \n \n \n \n \n \n \n ))}\n \n

Handle	Tier	Requests	Spent	Budget	Remaining
{row.handle}	{row.tier}	{row.request_count}	${row.spent.toFixed(4)}	${row.budget.toFixed(2)}	${(row.budget - row.spent).toFixed(4)}

\n \n ,\n );\n});\n\nexport default oauthMiddleware(app.fetch);\n",[],[229],"ec92ad01-966e-4a1e-8e73-306ab025f4a6","README.md","019fb4ea-df76-7108-899d-52a56e2a413c",24,["D",1785449814411],"# std/openai - [Docs ↗](https://docs.val.town/std/openai)\n\nUse OpenAI's API with [`std/openai`](https://www.val.town/v/std/openai) — no API\nkey needed. The SDK wraps the OpenAI client and routes requests through Val\nTown's proxy, which tracks spend in real dollars.\n\n## Budgets\n\n| Tier | Monthly Budget |\n|------|---------------|\n| Free | $1 |\n| Pro | $10 |\n| Business | $40 |\n\nBudgets follow the val's owning org and reset on the 1st of each month (UTC).\nWhen you exceed your budget, the proxy returns `402` with your spend details.\n\nAny model in the pricing table is available to all users — the dollar budget is\nthe only limit. Streaming is not supported.\n\n## Basic Usage\n\n```ts title=\"Example\" val\nimport { OpenAI } from \"https://esm.town/v/std/openai/main.ts\";\n\nconst openai = new OpenAI();\n\nconst completion = await openai.chat.completions.create({\n messages: [\n { role: \"user\", content: \"Say hello in a creative way\" },\n ],\n model: \"gpt-5.6-luna\",\n max_completion_tokens: 100,\n});\n\nconsole.log(completion.choices[0].message.content);\n```\n\n## Bring Your Own API Key\n\nThe free proxy is great for getting started, but you may outgrow the budget.\nThe easiest upgrade path is to bring your own OpenAI API key — just set\n`OPENAI_API_KEY` and use the `npm:openai` client directly when it's available,\nfalling back to `ValTownOpenAI` otherwise. This way your val works with or\nwithout a key.\n\n```ts title=\"Bring Your Own Key\" val\nimport { ValTownOpenAI } from \"https://esm.town/v/std/openai/main.ts\";\nimport OpenAI from \"npm:openai\";\n\n// Use your own key if available, otherwise the Val Town proxy.\nconst apiKey = Deno.env.get(\"OPENAI_API_KEY\");\nconst openai = apiKey ? new OpenAI({ apiKey }) : new ValTownOpenAI({});\n\nconst completion = await openai.chat.completions.create({\n model: \"gpt-5.6-luna\",\n messages: [{ role: \"user\", content: \"Say hello in one word.\" }],\n});\n\nconsole.log(completion.choices[0].message.content);\n```\n\nSee the runnable example:\n[`examples/bring-your-own-key.ts`](https://www.val.town/x/std/openai/code/examples/bring-your-own-key.ts)\n\n## Responses API & Web Search\n\nThe Responses API (`responses.create`) is fully supported through the Val Town\nproxy. It enables the **web_search** tool, which lets the model browse the web\nin real time — incredibly effective for enrichment, research, and anything where\nfresh data beats stale training knowledge.\n\n```ts title=\"Web Search\" val\nimport { ValTownOpenAI } from \"https://esm.town/v/std/openai/main.ts\";\n\nconst openai = new ValTownOpenAI({});\n\nconst response = await openai.responses.create({\n model: \"gpt-5.6-luna\",\n tools: [{ type: \"web_search\" }],\n instructions:\n \"You research companies. Given a company name, reply in 2-3 short plain-text lines: what they do, their stage, and why a developer might care. No markdown.\",\n input: \"Val Town\",\n});\n\nconsole.log(response.output_text);\n```\n\nSee the runnable example:\n[`examples/web-search.ts`](https://www.val.town/x/std/openai/code/examples/web-search.ts)\n\n## Structured Outputs\n\nInstead of asking the model to \"return JSON\" and manually parsing the response,\npass a JSON schema in `text.format`. The model is constrained to output JSON\nthat matches your schema exactly — no fence stripping, no regex, no try/catch on\nmalformed JSON. This is the idiomatic approach for classification, scoring,\nextraction, routing, or any task where you need typed data back.\n\n```ts title=\"Structured Outputs\" val\nimport { ValTownOpenAI } from \"https://esm.town/v/std/openai/main.ts\";\n\nconst openai = new ValTownOpenAI({});\n\nconst response = await openai.responses.create({\n model: \"gpt-5.6-luna\",\n instructions:\n \"You classify user feedback. Given a piece of feedback, classify it and write a one-sentence summary.\",\n input:\n \"The app keeps logging me out every 5 minutes, it's really frustrating.\",\n text: {\n format: {\n type: \"json_schema\",\n name: \"classification\",\n schema: {\n type: \"object\",\n properties: {\n category: {\n type: \"string\",\n enum: [\"bug\", \"feature_request\", \"praise\", \"question\"],\n },\n priority: {\n type: \"string\",\n enum: [\"low\", \"medium\", \"high\"],\n },\n summary: { type: \"string\" },\n },\n required: [\"category\", \"priority\", \"summary\"],\n additionalProperties: false,\n },\n },\n },\n});\n\n// output_text is guaranteed valid JSON matching the schema.\nconst result = JSON.parse(response.output_text);\nconsole.log(result);\n```\n\nSee the runnable example:\n[`examples/structured-outputs.ts`](https://www.val.town/x/std/openai/code/examples/structured-outputs.ts)\n\n## Images\n\nTo send an image to ChatGPT, the easiest way is by converting it to a data URL,\nwhich is easiest to do with\n[@stevekrouse/fileToDataURL](https://www.val.town/v/stevekrouse/fileToDataURL).\n\n## OpenAI Agents SDK\n\nThe [OpenAI Agents SDK](https://openai.github.io/openai-agents-js/) works with\n`std/openai` too, including hosted tools like web search:\n\n```ts title=\"Agents Example\" val\nimport { ValTownOpenAI } from \"https://esm.town/v/std/openai/main.ts\";\nimport {\n Agent,\n run,\n setDefaultOpenAIClient,\n setTracingDisabled,\n webSearchTool,\n} from \"npm:@openai/agents\";\n\nsetTracingDisabled(true); // tracing requires your own OpenAI key\nsetDefaultOpenAIClient(new ValTownOpenAI({}));\n\nconst agent = new Agent({\n name: \"Researcher\",\n tools: [webSearchTool()],\n});\n\nconst result = await run(agent, \"What's new in Val Town?\");\nconsole.log(result.finalOutput);\n```\n\n## Available Models\n\nPricing per 1M tokens (Standard tier). See the full table in\n[`pricing.ts`](https://www.val.town/x/std/openai/code/pricing.ts).\n\n| Model | Input | Cached Input | Output |\n|-------|-------|-------------|--------|\n| gpt-5.6-luna | $0.20 | $0.02 | $1.20 |\n| gpt-5.4-nano | $0.20 | $0.02 | $1.25 |\n| gpt-5.4-mini | $0.75 | $0.075 | $4.50 |\n| gpt-5.6-terra | $2.00 | $0.20 | $12.00 |\n| gpt-5.4 | $2.50 | $0.25 | $15.00 |\n| gpt-5.6-sol | $5.00 | $0.50 | $30.00 |\n| gpt-5.5 | $5.00 | $0.50 | $30.00 |\n| gpt-5.4-pro | $30.00 | — | $180.00 |\n| gpt-5.5-pro | $30.00 | — | $180.00 |\n\nIf these limits are too low, you can [bring your own API key](#bring-your-own-api-key).\n\n[📝 Edit docs](https://github.com/val-town/val-town-docs/edit/main/src/content/docs/std/openai.mdx)\n",[],[238],"8f07f093-d16a-43e0-b337-331c5d9c71df","AGENTS.md","019fb4ea-df76-7108-899d-4c0be1bf7963",["D",1785449822893],"# std/openai\n\nClient SDK for using OpenAI's API through Val Town's proxy. No API key needed.\n\n## File Map\n\n| File | Type | Description |\n|------|------|-------------|\n| `main.ts` | script | `ValTownOpenAI` class — extends OpenAI SDK to use Val Town's proxy automatically |\n| `main.tsx` | script | Backward compatibility re-export from `main.ts` |\n| `pricing.ts` | script | Model pricing table ($/1M tokens) + budget tiers + cost calculation |\n| `ledger.ts` | script | Dollar-based spend tracking in project-scoped SQLite (openai_spend + openai_monthly_budget tables) |\n| `proxy/main.ts` | http | **New proxy** — authenticates users, enforces dollar budgets, records actual spend |\n| `dashboard.tsx` | http | Spend dashboard — per-subject monthly spend vs budget |\n| `tests/integration.ts` | script | Integration tests (run on demand with `run_file`) |\n| `examples/web-search.ts` | script | Example: web search via Responses API |\n| `examples/structured-outputs.ts` | script | Example: structured outputs with JSON schema |\n| `examples/bring-your-own-key.ts` | script | Example: using your own OpenAI API key with fallback to proxy |\n| `README.md` | file | User-facing documentation |\n| `AGENTS.md` | file | This file |\n\n## Architecture (dollar-rate-limiting branch)\n\nThe proxy has been moved from `std/openaiproxy` into this val (`proxy/main.ts`) to use project-scoped SQLite.\n\n### How rate limiting works now\n\n1. User's val calls OpenAI via the SDK → proxy authenticates via `/v1/me` → identifies the val's owning org + tier\n2. Pre-flight check: has this subject (user or org) exceeded their monthly dollar budget?\n3. Budgets (per calendar month, UTC):\n - free: $1/month\n - pro: $10/month\n - business: $40/month\n4. If over budget → 402 with remaining-budget info\n5. Proxy to OpenAI, parse response usage, calculate cost in microdollars\n6. Record spend (fire-and-forget, doesn't add to request latency)\n\n### What was removed (vs old openaiproxy)\n\n- ❌ Request count limits (was 100/day) — dollars are the only guard now\n- ❌ Model gating (was: free users forced to cheap models, pro users get 10 expensive requests) — any priced model is allowed, budget is the guard\n- ❌ Streaming support — rejected with 400 (simplifies usage extraction)\n- ❌ Unknown models allowed — now rejected with 400 (never free-by-accident)\n\n### Cost calculation\n\n- Prices stored as USD per 1M tokens in `pricing.ts` (readable, easy to update)\n- Cost calculated in **microdollars** (integer) to avoid floating-point issues\n- Formula: `cost_microdollars = round(tokens × price_per_1M)` (works because $/1M × tokens = microdollars directly)\n\n### SQLite tables\n\n- `openai_spend` — append-only ledger: subject, model, token counts, cost, timestamp\n- `openai_monthly_budget` — aggregate: subject_id × month → spent_microdollars (for fast pre-flight checks)\n\nBoth use project-scoped SQLite (`std/sqlite/main.ts`), not the deprecated global DB.\n\n## Environment Variables\n\n- `OPENAI_API_KEY` — OpenAI API key (needs to be added to this val before cutover)\n- `OPENAI_API_ORG` — OpenAI organization ID (needs to be added to this val before cutover)\n\n## Cutover plan\n\n1. ✅ Build proxy on `dollar-rate-limiting` branch\n2. ⬜ Add `OPENAI_API_KEY` + `OPENAI_API_ORG` env vars to this val\n3. ⬜ Merge branch to main\n4. ⬜ Set custom subdomain `std-openaiproxy` on `proxy/main.ts` (with `force: true`)\n5. ⬜ Archive `std/openaiproxy`\n\n## Related Vals\n\n- [std/openaiproxy](https://www.val.town/x/std/openaiproxy) — Old proxy (to be archived after cutover)\n",[],[246],"10a5dd82-9daf-46d5-bc45-6c5059055f00","integration.ts","56cdfbe8-8c5e-11f1-a52a-1607ee4eb77e","622d9d61-914e-4390-9281-9d8b62dbb1a3",18,["D",1785449381929],3,"import { httpEndpoint } from \"https://esm.town/v/std/utils/index.ts\";\nimport { sqlite } from \"https://esm.town/v/std/sqlite/main.ts\";\n\nconst endpoint = await httpEndpoint(\"proxy/main.ts\");\nconst token = Deno.env.get(\"valtown\");\n\ntype TestResult = { name: string; passed: boolean; error?: string };\nconst results: TestResult[] = [];\n\nasync function test(name: string, fn: () => Promise) {\n try {\n await fn();\n results.push({ name, passed: true });\n console.log(`✅ ${name}`);\n } catch (e) {\n const msg = e instanceof Error ? e.message : String(e);\n results.push({ name, passed: false, error: msg });\n console.log(`❌ ${name}: ${msg}`);\n }\n}\n\nfunction assert(condition: boolean, message: string) {\n if (!condition) throw new Error(message);\n}\n\n// --- Backward compatibility tests (same API surface as old proxy) ---\n\n// 1. Unsupported path returns 404\nawait test(\"unsupported path returns 404\", async () => {\n const res = await fetch(`${endpoint}/v1/foo`);\n assert(res.status === 404, `Expected 404, got ${res.status}`);\n});\n\n// 2. No auth returns 401\nawait test(\"no auth returns 401\", async () => {\n const res = await fetch(`${endpoint}/v1/chat/completions`, {\n method: \"POST\",\n headers: { \"Content-Type\": \"application/json\" },\n body: JSON.stringify({ model: \"gpt-5.6-luna\", messages: [] }),\n });\n assert(res.status === 401, `Expected 401, got ${res.status}`);\n});\n\n// 3. Invalid auth returns 401\nawait test(\"invalid auth returns 401\", async () => {\n const res = await fetch(`${endpoint}/v1/chat/completions`, {\n method: \"POST\",\n headers: {\n \"Content-Type\": \"application/json\",\n \"Authorization\": \"Bearer invalid-token\",\n },\n body: JSON.stringify({ model: \"gpt-5.6-luna\", messages: [] }),\n });\n assert(res.status === 401, `Expected 401, got ${res.status}`);\n});\n\n// 4. Chat completions succeeds (backward compat: same API surface)\nawait test(\"chat completions succeeds\", async () => {\n const res = await fetch(`${endpoint}/v1/chat/completions`, {\n method: \"POST\",\n headers: {\n \"Content-Type\": \"application/json\",\n \"Authorization\": `Bearer ${token}`,\n },\n body: JSON.stringify({\n model: \"gpt-5.6-luna\",\n messages: [{ role: \"user\", content: \"Say hello in one word\" }],\n max_completion_tokens: 10,\n }),\n });\n const body = await res.text();\n assert(res.status === 200, `Expected 200, got ${res.status}: ${body}`);\n const data = JSON.parse(body);\n assert(data.choices?.[0]?.message?.content, \"No message content in response\");\n // Verify spend header is present\n const remaining = res.headers.get(\"x-vt-openai-spend-remaining\");\n assert(remaining !== null, \"Missing x-vt-openai-spend-remaining header\");\n});\n\n// 5. Responses API succeeds (backward compat: same API surface)\nawait test(\"responses API succeeds\", async () => {\n const res = await fetch(`${endpoint}/v1/responses`, {\n method: \"POST\",\n headers: {\n \"Content-Type\": \"application/json\",\n \"Authorization\": `Bearer ${token}`,\n },\n body: JSON.stringify({\n model: \"gpt-5.6-luna\",\n input: \"Say hello in one word\",\n }),\n });\n const body = await res.text();\n assert(res.status === 200, `Expected 200, got ${res.status}: ${body}`);\n const data = JSON.parse(body);\n const outputMsg = data.output?.find((o: { type: string }) => o.type === \"message\");\n const text = outputMsg?.content?.find((c: { type: string }) => c.type === \"output_text\")?.text;\n assert(text, `No output text in response. Body: ${body.slice(0, 500)}`);\n});\n\n// 6. SDK backward compatibility — ValTownOpenAI works end-to-end\nawait test(\"ValTownOpenAI SDK works end-to-end\", async () => {\n const { ValTownOpenAI } = await import(\"../main.ts\");\n const client = new ValTownOpenAI({});\n const response = await client.chat.completions.create({\n model: \"gpt-5.6-luna\",\n messages: [{ role: \"user\", content: \"Say hello in one word\" }],\n max_completion_tokens: 10,\n });\n assert(\n response.choices[0].message.content,\n \"No message content from SDK response\",\n );\n});\n\n// --- New behavior tests ---\n\n// 7. Streaming rejected with 400\nawait test(\"streaming rejected with 400\", async () => {\n const res = await fetch(`${endpoint}/v1/chat/completions`, {\n method: \"POST\",\n headers: {\n \"Content-Type\": \"application/json\",\n \"Authorization\": `Bearer ${token}`,\n },\n body: JSON.stringify({\n model: \"gpt-5.6-luna\",\n messages: [{ role: \"user\", content: \"Hi\" }],\n stream: true,\n }),\n });\n assert(res.status === 400, `Expected 400, got ${res.status}`);\n const data = await res.json();\n assert(\n data.error?.toLowerCase().includes(\"stream\"),\n `Expected streaming error, got: ${data.error}`,\n );\n});\n\n// 8. Unknown model rejected with 400\nawait test(\"unknown model rejected with 400\", async () => {\n const res = await fetch(`${endpoint}/v1/chat/completions`, {\n method: \"POST\",\n headers: {\n \"Content-Type\": \"application/json\",\n \"Authorization\": `Bearer ${token}`,\n },\n body: JSON.stringify({\n model: \"gpt-999-turbo\",\n messages: [{ role: \"user\", content: \"Hi\" }],\n }),\n });\n assert(res.status === 400, `Expected 400, got ${res.status}`);\n const data = await res.json();\n assert(\n data.error?.includes(\"pricing table\"),\n `Expected pricing error, got: ${data.error}`,\n );\n assert(\n Array.isArray(data.available) && data.available.length > 0,\n \"Expected available models list\",\n );\n});\n\n// 9. Spend recorded in SQLite after successful request\nawait test(\"spend recorded in SQLite\", async () => {\n // Give the fire-and-forget write time to complete\n await new Promise((r) => setTimeout(r, 2000));\n\n const result = await sqlite.execute({\n sql: `SELECT COUNT(*) as count, SUM(cost) as total\n FROM openai_spend\n WHERE timestamp > datetime('now', '-5 minutes')`,\n args: [],\n });\n const count = Number(result.rows[0][0]);\n const total = Number(result.rows[0][1] ?? 0);\n assert(count > 0, \"No spend records found in the last 5 minutes\");\n assert(total > 0, \"Spend total is zero — cost was not recorded\");\n console.log(` Found ${count} spend records, total cost: ${total.toFixed(8)}`);\n});\n\n// Summary\nconst passed = results.filter((r) => r.passed).length;\nconst failed = results.length - passed;\nconsole.log(\n `\\n${passed}/${results.length} tests passed${failed ? `, ${failed} failed` : \"\"}`,\n);\n",[],[262,253],"tests","df6180ff-2070-4390-aa62-b2d7df767d75",7,["D",1785447223616],"directory",[193],[262],"d5598d81-9b81-47d0-8120-d3f895fead29","7fb4e69e-8c5d-11f1-88fd-1607ee4eb77e","95bed124-9bf3-479e-9c36-fdd2f6dcc960",22,["D",1785449814395],"import { parseBearerString } from \"https://esm.town/v/andreterron/parseBearerString\";\nimport { API_URL } from \"https://esm.town/v/std/API_URL?v=5\";\nimport { hasPricing, PRICING, calculateCost } from \"../pricing.ts\";\nimport {\n initSchema,\n checkBudget,\n recordSpend,\n type Subject,\n} from \"../ledger.ts\";\n\n/**\n * OpenAI API proxy with dollar-based rate limiting.\n *\n * Replaces std/openaiproxy. Moved into std/openai to use project-scoped SQLite.\n *\n * Simplifications vs old proxy:\n * - No streaming (rejected with 400)\n * - No model gating — any priced model is allowed, budget is the guard\n * - No request count limits — dollars are the only limit\n * - Unknown models rejected (400) — never free-by-accident\n *\n * Budgets (per calendar month, UTC):\n * - free: $1\n * - pro: $10\n * - business: $40\n *\n * Budget follows the val's owning org (from /v1/me).\n *\n * Env vars (add to std/openai before cutover):\n * - OPENAI_API_KEY\n * - OPENAI_API_ORG\n */\n\nconst allowedPathnames = [\n \"/\",\n \"/v1/chat/completions\",\n \"/v1/responses\",\n \"/v1/embeddings\",\n];\n\nlet schemaInitialized = false;\n\nexport default async function(req: Request): Promise {\n if (!schemaInitialized) {\n await initSchema();\n schemaInitialized = true;\n }\n\n const { pathname, search } = new URL(req.url);\n\n if (pathname === \"/\") {\n return new Response(\"OK\");\n }\n\n if (!allowedPathnames.includes(pathname)) {\n return new Response(\"Path not supported\", { status: 404 });\n }\n\n // --- Authenticate ---\n const authHeader = req.headers.get(\"Proxy-Authorization\") ||\n req.headers.get(\"Authorization\");\n const token = authHeader ? parseBearerString(authHeader) : undefined;\n const meRes = await fetch(`${API_URL}/v1/me`, {\n headers: { Authorization: `Bearer ${token}` },\n });\n if (!meRes.ok) {\n return new Response(\"Unauthorized\", { status: 401 });\n }\n const me = await meRes.json();\n\n // --- Determine budget subject (the val's owning org) ---\n // /v1/me returns the org directly when the token belongs to an org-owned val:\n // { id, type: \"org\", username: \"std\", tier: \"pro\", ... }\n // For personal vals, type is \"user\".\n const subject: Subject = me.type === \"org\"\n ? {\n id: me.id,\n type: \"org\",\n handle: me.username,\n tier: me.tier ?? \"free\",\n }\n : {\n id: me.id,\n type: \"user\",\n handle: me.username,\n tier: me.tier ?? \"free\",\n };\n\n // --- Pre-flight budget check ---\n const budget = await checkBudget(subject);\n if (budget.exceeded) {\n return jsonResponse(\n {\n error: \"Monthly budget exceeded\",\n spent: formatUSD(budget.spent),\n budget: formatUSD(budget.budget),\n resets: \"next calendar month (UTC)\",\n },\n 402,\n { \"x-vt-openai-spend-remaining\": \"0\" },\n );\n }\n\n // --- Parse request body ---\n let body: Record;\n try {\n body = await req.json();\n } catch {\n return jsonResponse({ error: \"Invalid JSON body\" }, 400);\n }\n\n // --- Reject streaming ---\n if (body.stream === true) {\n return jsonResponse(\n { error: \"Streaming is not supported. Set stream: false or omit it.\" },\n 400,\n );\n }\n\n // --- Reject unknown models ---\n const model = body.model as string | undefined;\n if (!model || !hasPricing(model)) {\n return jsonResponse(\n {\n error: `Model '${model}' is not in the pricing table.`,\n available: Object.keys(PRICING),\n },\n 400,\n );\n }\n\n // --- Proxy to OpenAI ---\n const url = new URL(\".\" + pathname, \"https://api.openai.com\");\n url.search = search;\n\n const headers = new Headers(req.headers);\n headers.set(\"Host\", url.hostname);\n headers.set(\"Authorization\", `Bearer ${Deno.env.get(\"OPENAI_API_KEY\")}`);\n headers.set(\"OpenAI-Organization\", Deno.env.get(\"OPENAI_API_ORG\") ?? \"\");\n headers.delete(\"Proxy-Authorization\");\n\n const openAIRes = await fetch(url, {\n method: req.method,\n headers,\n body: JSON.stringify(body),\n redirect: \"manual\",\n });\n\n // Copy response headers (fetch response headers are immutable in Deno)\n const responseHeaders = new Headers(openAIRes.headers);\n responseHeaders.delete(\"openai-organization\");\n\n // --- Pass through non-200 responses without recording spend ---\n if (openAIRes.status !== 200) {\n return new Response(openAIRes.body, {\n status: openAIRes.status,\n headers: responseHeaders,\n });\n }\n\n // --- Parse response to extract usage ---\n const responseText = await openAIRes.text();\n let responseData: Record;\n try {\n responseData = JSON.parse(responseText);\n } catch {\n // Can't parse — return as-is\n return new Response(responseText, {\n status: openAIRes.status,\n headers: responseHeaders,\n });\n }\n\n // --- Extract usage and record spend ---\n const usage = extractUsage(pathname, responseData);\n let cost = 0;\n if (usage) {\n cost = calculateCost(\n model,\n usage.input,\n usage.cached,\n usage.output,\n );\n // Fire-and-forget — don't add to request latency\n recordSpend(subject, model, usage.input, usage.cached, usage.output)\n .catch((e) => console.error(\"Failed to record spend:\", e));\n }\n\n // --- Return response with spend header ---\n const remainingAfter = Math.max(\n 0,\n budget.remaining - cost,\n );\n const res = new Response(responseText, {\n status: openAIRes.status,\n headers: responseHeaders,\n });\n res.headers.set(\"x-vt-openai-spend-remaining\", String(remainingAfter));\n return res;\n}\n\n/**\n * Extract token usage from different OpenAI API response formats.\n *\n * Chat completions: usage.prompt_tokens, usage.completion_tokens,\n * usage.prompt_tokens_details.cached_tokens\n * Responses API: usage.input_tokens, usage.output_tokens,\n * usage.input_tokens_details.cached_tokens\n * Embeddings: usage.prompt_tokens (output is always 0)\n */\nfunction extractUsage(\n pathname: string,\n data: Record,\n): { input: number; cached: number; output: number } | null {\n const usage = data.usage as Record | undefined;\n if (!usage) return null;\n\n if (pathname === \"/v1/responses\") {\n const inputDetails = usage.input_tokens_details as\n | Record\n | undefined;\n return {\n input: Number(usage.input_tokens) || 0,\n cached: Number(inputDetails?.cached_tokens) || 0,\n output: Number(usage.output_tokens) || 0,\n };\n }\n\n // Chat completions + embeddings share the same shape\n const promptDetails = usage.prompt_tokens_details as\n | Record\n | undefined;\n return {\n input: Number(usage.prompt_tokens) || 0,\n cached: Number(promptDetails?.cached_tokens) || 0,\n output: Number(usage.completion_tokens) || 0,\n };\n}\n\nfunction jsonResponse(\n body: unknown,\n status: number,\n extraHeaders?: Record,\n): Response {\n const headers: Record = {\n \"Content-Type\": \"application/json\",\n ...extraHeaders,\n };\n return new Response(JSON.stringify(body), { status, headers });\n}\n\nfunction formatUSD(amount: number): string {\n return `${amount.toFixed(2)}`;\n}\n",[],[277,18],"proxy","1743ee63-eb5b-464e-86fa-568997e2f840",4,["D",1785446862743],[191],[277],"24f4c05a-6867-497e-95f6-1413499358d7","web-search.ts","019fb4ea-df76-7108-899d-476b76b86df2","019fb4ea-df76-7108-899d-48c5ac28bf1c",["D",1785446719355],"// Web search via the Responses API — incredibly effective for enrichment.\n//\n// The Responses API (responses.create) supports the web_search tool, which lets\n// the model browse the web in real time. This is dramatically more effective than\n// passing static text — the model can look up people, companies, products, and\n// current events on its own.\n//\n// Great use cases: lead enrichment, company research, summarizing a topic,\n// fact-checking, or anything where fresh data beats stale training knowledge.\n\nimport { ValTownOpenAI } from \"../main.ts\";\n\nconst openai = new ValTownOpenAI({});\n\nconst response = await openai.responses.create({\n model: \"gpt-5.6-luna\",\n tools: [{ type: \"web_search\" }],\n instructions:\n \"You research companies. Given a company name, reply in 2-3 short plain-text lines: what they do, their stage, and why a developer might care. No markdown.\",\n input: \"Val Town\",\n});\n\nconsole.log(response.output_text);\n",[],[291,284],"examples","c9a57e15-5ed3-4b2e-8e25-1e0d4fff5aef","structured-outputs.ts","019fb4ea-df76-7108-899d-43a30fa43e51",["D",1785446719355],"// Structured outputs via the Responses API — reliable, typed JSON.\n//\n// Instead of asking the model to \"return JSON\" and manually parsing/free-text\n// cleaning the response, pass a JSON schema in `text.format`. The model is\n// constrained to output JSON that matches your schema exactly — no fence\n// stripping, no regex, no try/catch on malformed JSON.\n//\n// This is the idiomatic approach for any task where you need the model's answer\n// as structured data: classification, scoring, extraction, routing decisions, etc.\n\nimport { ValTownOpenAI } from \"../main.ts\";\n\nconst openai = new ValTownOpenAI({});\n\nconst response = await openai.responses.create({\n model: \"gpt-5.6-luna\",\n instructions:\n \"You classify user feedback. Given a piece of feedback, classify it and write a one-sentence summary.\",\n input: \"The app keeps logging me out every 5 minutes, it's really frustrating.\",\n text: {\n format: {\n type: \"json_schema\",\n name: \"classification\",\n schema: {\n type: \"object\",\n properties: {\n category: {\n type: \"string\",\n enum: [\"bug\", \"feature_request\", \"praise\", \"question\"],\n },\n priority: {\n type: \"string\",\n enum: [\"low\", \"medium\", \"high\"],\n },\n summary: { type: \"string\" },\n },\n required: [\"category\", \"priority\", \"summary\"],\n additionalProperties: false,\n },\n },\n },\n});\n\n// output_text is guaranteed valid JSON matching the schema.\nconst result = JSON.parse(response.output_text);\nconsole.log(result);\n",[],[291,293],"cc3de97b-e809-46ab-b4f5-ada6b8808bb2","bring-your-own-key.ts","019fb4ea-df75-7236-add3-8098418c5d6a",["D",1785446719355],"// Bring your own OpenAI API key — or use the built-in Val Town proxy.\n//\n// If you set the OPENAI_API_KEY environment variable, all requests go directly\n// to OpenAI with no rate limits or proxy restrictions. If you don't, requests\n// fall back to ValTownOpenAI, which routes through our proxy (no key needed,\n// subject to free-tier quotas).\n//\n// This is the recommended pattern for production vals: start on the free proxy,\n// then add an API key when you outgrow the limits.\n\nimport { ValTownOpenAI } from \"../main.ts\";\nimport OpenAI from \"npm:openai\";\n\n// Use your own key if available, otherwise the Val Town proxy.\nconst apiKey = Deno.env.get(\"OPENAI_API_KEY\");\nconst openai = apiKey ? new OpenAI({ apiKey }) : new ValTownOpenAI({});\n\nconst completion = await openai.chat.completions.create({\n model: \"gpt-5.6-luna\",\n messages: [{ role: \"user\", content: \"Say hello in one word.\" }],\n});\n\nconsole.log(completion.choices[0].message.content);\n",[],[291,300],"f1080d92-5fb2-424f-b9c2-d33b4de7414e",["D",1785446719355],[187,188,189],[291],["D",1746043648993],["D",1789561254403],{"_11":12,"_139":135},{"_11":12,"_134":-5,"_74":135,"_30":-5,"_136":137,"_138":24},{"_39":125,"_13":16,"_33":36,"_126":86,"_127":63,"_128":65,"_129":91,"_130":86,"_131":102,"_132":14,"_133":12},"ctx",{"_52":-5,"_159":348,"_349":350,"_351":352,"_353":24,"_354":24,"_355":24,"_356":89,"_357":-5,"_358":-7,"_359":-7,"_360":26,"_361":24,"_362":24,"_363":-5,"_364":24,"_365":24,"_366":24,"_367":-5},"ENV",{"_319":320,"_321":322,"_323":64,"_324":325,"_326":327,"_328":329,"_330":331,"_332":333,"_334":335,"_336":337,"_338":339,"_340":341,"_342":343,"_344":345,"_346":347},"REACT_APP_SENTRY_DSN","https://586020d86f05489db63a554690acf1e9@o1398359.ingest.sentry.io/6724499","RENDER_GIT_COMMIT","fdb67791413f4386b4dbf18f354029a617f6160f","RENDER_GIT_BRANCH","POSTHOG_API_TOKEN","phc_nmoyuAkStEZuUp69rYlg1DZgrIt19Gx0Y4Hwf8tGgJP","POSTHOG_API_HOST","https://wart.val.town","WEB_API_TEMPLATE","https://{val}.web.val.run","APP_URL","https://www.val.town","MODULE_URL","https://esm.town","VANITY_WEB_API_TEMPLATE","https://{label}.val.run/","DENOLS_URL","https://lsp.val.town","NODE_ENV","production","CLOUDFLARE_TURNSTILE_SITE_KEY","0x4AAAAAACtaIZJH7s9rBMZ6","GOOGLE_CLIENT_ID","875989696899-jtsqkhlbdv5jrlsk8631p7ipjg3c6o62.apps.googleusercontent.com","STRIPE_PUBLISHABLE_KEY","pk_live_51N2zvWEpZTTLipIkVEwXrUG9AT0yeOOEYxszarZT1jpsZV58Ly6wlg3SltZpHN26CUalOBmUSY7adQRQiSkCZz2500qSXNkf8Y","API_URL","https://api.val.town",[],"preferencesKeybindings","default","preferencesTheme","light","preferencesFixers","preferencesCopilot","preferencesSelectionAi","preferencesCloseBrackets","preferencesChatPanelState","preferencesDprint","preferencesDenofmt","notificationCount","flagDenols","isBusinessOrgMember","businessOrgHandle","isBusinessOrgAdmin","isHigherTownieTier","navCollapsed","currentOrgId"]

History · main.ts