fix: OpenAI-compatible provider calls wrong endpoint (/responses), breaking self-hosted chat backends

[fabicarvano]

Issue 1 — Chat/streaming broken against any self-hosted OpenAI-compatible backend

OpenAICompatibleProvider calls the AI SDK client as a bare function (this.client(model)), which the SDK routes to the Responses API (/v1/responses). Any self-hosted backend that only implements /v1/chat/completions (Ollama, vLLM, LiteLLM, etc.) 404s on every chat/streaming request.

The end user only sees a generic "An error occurred." — the real cause (unsupported path /v1/responses) only shows up in server logs.

Fix: use this.client.chat(model), which targets /v1/chat/completions as intended. Embeddings are unaffected (already used textEmbeddingModel() correctly).

Issue 2 — Claw tools and scheduler sweep failing silently in the workers Docker image

In apps/workers/Dockerfile, the generated package.json for @chatbot/ai has no exports map, unlike @chatbot/shared and @chatbot/agent-studio in the same file. This breaks subpath imports like @chatbot/ai/tools/built-in-registry at runtime inside the container.

Observed effect: Claw's web-search, web-fetch and URL-guard tools become unavailable, and the claw-scheduler sweep fails every 30 seconds (Cannot find module '@chatbot/ai/tools/built-in-registry'), with no visible signal to the user.

Fix: add the same exports map already used for @chatbot/agent-studio.

How this was found

Both issues were confirmed by running the full stack in Docker end-to-end (build, agent create/publish, real inference call against an OpenAI-compatible backend) — not just by reading the code.

[kartikmanimuthu]

Thanks for this one, @fabicarvano — the diagnosis is right and it's a real bug: openai('model-id') routes to the Responses API, and openai.chat() is the documented way to target /v1/chat/completions for backends that only implement that endpoint. I confirmed streamChat was the only affected path — embed/embedBatch already used textEmbeddingModel() — and the mock update is the right way to pin the routing.

The Dockerfile exports fix is the other half of this and it genuinely matters: libs/claw-studio/src/agent/web-tools.ts and browser-session.ts import @chatbot/ai/tools/*, and claw-studio ships inside the workers image, so the missing exports map really did take those tools offline at runtime. Good catch on both, and thanks for confirming it end-to-end in Docker rather than by inspection.

Two non-blocking follow-ups I'll file separately rather than hold this up:

• The same provider defaults embeddingDimensions to 3072, which can't satisfy the vector(1024) columns — same class of bug as known issue #15.
• "./*": "./src/*.js" also resolves @chatbot/ai/src/...; harmless, but explicit subpaths (./tools/*) would keep the public surface intentional.

Approving — thanks for the careful root-cause work.