[mkotelnikov]

What

• packages/fake-openai (new, private): a scripted OpenAI-compatible endpoint.
  • It serves chat completions, streamed or not, as text or tool calls. Text streams a word at a time; tool calls stream as OpenAI streams them.
  • It also serves /v1/embeddings and /v1/models, with CORS open, and records every request.
  • A CLI runs it for celld dev (node packages/fake-openai/src/cli.ts 8787).
  • The Flue example's tests now use it.
• packages/agent-examples (new, private), both tested on cellb (Node) against the fake model:
  • pi: pi's agent loop (@earendil-works/pi-agent-core + pi-ai) in a Durable Object with one tool. A prompt calls the tool and gets the follow-up reply; the transcript persists in storage and a second prompt continues it.
    • CellD's examples/pi needs PiHarness from a pkg.pr.new preview of agents, which no published release contains and which some networks block, so this version uses pi's published packages instead.
  • opencode: CellD's examples/opencode, with an endpoint provider (@ai-sdk/openai-compatible, base URL from OPENCODE_BASE_URL) instead of opencode's free registry model. All 8 SDK steps succeed and the reply is the fake model's.
• cellb fixes these examples exposed:
  • A Response used as another Response's init (new Response(body, response)) kept only status: the spread dropped the prototype getters headers and statusText. opencode's SDK does this and failed with ClientError: UnsupportedContentType. Added a regression test in p07w.
  • Node bundles now resolve packages as celld deploy does: conditions workerd, worker, browser, then module and main. The neutral platform has no main fields, so packages with only main (opencode's @smithy/*) did not resolve.
  • A containers config is reported in unsupported instead of being silently ignored (test in p13).
• Docs: a "Features at a glance" table in docs/celld-compatibility.md against CellD, covering WebSockets, Workers AI, Python, containers and agent frameworks; docs/gaps.md is updated.
• CI runs the fake model's and the agent examples' tests.

Tests

Suite	Result
fake-openai	3 passed
agent-examples	2 passed (pi about 4 s; opencode about 37 s, mostly bundling its SDK)
flue-example on Node	6 passed
flue-example in the browser	8 passed
cellb unit tests	353 passed, 6 expected fail
cellb interop	123 passed, 6 expected fail
Chromium	209 passed, 15 skipped

Typecheck and Biome are clean.

Note

The Node bundler change overlaps the pending switch to @statewalker/webrun-bundler (branch ccr-ab3c0d30-6du8do-webrun-bundler), which resolves these fields itself. Whichever merges second needs a small merge in node-project.ts.

---

Generated by Claude Code