AGENTS.md@packages/provider-openai git:20260928.6369adaA

Immutable. This exact content never changes and is served at /api/v1/blob/f9603614e6aeacc9.

# provider-openai

## Purpose

OpenAI adapter — implements `ProviderAdapter` for OpenAI's Responses API using the `web_search` tool (the current GA tool, chosen over the legacy `web_search_preview`; its new knobs — e.g. `filters.allowed_domains` — stay off because allow-listing would defeat measuring who actually gets cited). Extracts cited domains from URL annotations.

## Key Files

| File | Role |
|------|------|
| `src/adapter.ts` | Exports `openaiAdapter` — the `ProviderAdapter` object |
| `src/normalize.ts` | Core logic: `validateConfig`, `healthcheck`, `buildTrackedQueryRequest`, `executeTrackedQuery`, `parseTrackedQueryResponse`, `normalizeResult`, `generateText` |
| `src/types.ts` | OpenAI-specific config and response types |
| `src/index.ts` | Re-exports public API |

## Patterns

All provider packages follow the same 4-file structure and implement the same `ProviderAdapter` interface from `@ainyc/canonry-contracts`:

- **`validateConfig(config)`** — verify API key and model are valid
- **`healthcheck(config)`** — test connectivity to the provider
- **`executeTrackedQuery(input)`** — send a tracked query and capture raw response with web search results
  - It is `buildTrackedQueryRequest` (the exact wire body) → the SDK call → `parseTrackedQueryResponse` (usage and stop reason included). Batch dispatch reuses both halves, so change the request or its reading there, never inline in `executeTrackedQuery`; `test/tracked-query-request.test.ts` pins the built body against the wire.
- **`normalizeResult(raw)`** — convert provider-specific response to standard `NormalizedQueryResult`
- **`generateText(config, prompt)`** — general-purpose text generation

Retrieval contract: every tracked query runs `search-required-v1` (`OPENAI_RETRIEVAL_CONTRACT` in `src/normalize.ts`): verbatim query, no `instructions`, `tool_choice: "required"` with `web_search` as the only tool. Adding any other tool to that request makes `required` stop forcing a search, so it is a contract change. `retrievalStatus` is read from `web_search_call` output items; a missing/empty output or an unfinished response without a search call is `unknown`, never `not-used`. Details and the history of the mislabelled `native-auto-v1` rows: `docs/providers/openai.md` → "Retrieval contract".

Note: The OpenAI `web_search` API returns fewer/different results than the ChatGPT UI search. That gap is structural (the API does not carry logged-in user context, conversation history, or personalization) and is addressed separately by a planned browser provider (drive the ChatGPT UI itself), not by switching API tool flavors.

## Common Mistakes

- **Not normalizing grounding sources to standard `CitedSource` format** — each provider returns different shapes.
- **Not handling rate limits** — implement retry with exponential backoff for 429 responses.
- **Forgetting to export from `adapter.ts`** — the provider registry imports the adapter object.

## See Also

- `docs/providers/openai.md` — OpenAI-specific API quirks
- `docs/providers/README.md` — provider system overview
- `packages/contracts/src/provider.ts` — `ProviderAdapter` interface definition