@statewalker/agent-examples

Coding-agent frameworks running in Durable Objects on CellB, each against a fake OpenAI-compatible model (@statewalker/fake-openai), so the whole thing runs offline once the packages are installed:

Example	Framework	What it shows
`pi/`	pi’s agent loop (`@earendil-works/pi-agent-core`, `@earendil-works/pi-ai`)	a prompt runs a tool call and its follow-up; the transcript persists in the object’s storage
`opencode/`	opencode’s workerd SDK (`@opencode-ai/sdk/workerd`)	celld’s opencode example: the SDK boots on the object’s SQLite, creates a session, and the model answers a prompt
`../flue`	Flue (`@flue/runtime`)	a whole Flue app, built for Cloudflare, on Node and in browsers

The package is private: it is a test of CellB, not a library.

Why it exists

celld ships pi and opencode examples, but they need the network: pi depends on an unreleased preview build of the agents package and OpenRouter, opencode on a free model from opencode’s own registry. These versions take the model from an endpoint named in the project’s vars, which the tests point at the fake model, and they run on CellB as celld runs them.

How to use it

pnpm install            # in examples/ (their own workspace; cellb from npm)
pnpm test               # in this directory: each example on cellb (Node), against the fake model

To run one against a real endpoint, set its vars (PI_BASE_URL, PI_MODEL, PI_API_KEY; OPENCODE_BASE_URL, OPENCODE_MODEL, OPENCODE_API_KEY) in wrangler.jsonc or .dev.vars, and start it with celld dev pi or load it with loadNodeProject. Against the fake model:

node ../fake-openai/src/cli.ts 8787     # the vars' defaults point here

Examples

What the tests prove:

pi: GET /?q=… runs the agent: the model is offered the where_am_i tool and calls it, the tool answers, and the second model call carries the result, which becomes the reply (The tool said: a Durable Object on cellb). A second prompt continues the same transcript, restored from storage.
opencode: GET / runs celld’s probe (model.default, provider.list, health.get, session.create, session.get, session.prompt, session.wait, message.list); every step succeeds and the reply is the fake model’s answer to hello world, from provider endpoint, model fake-model.

Internals

Differences from celld’s examples

pi uses pi’s published packages instead of the PiHarness of agents pull request 2197 (only on pkg.pr.new, which may expire, and unreachable from some networks). The object holds a pi Agent with one tool; the transcript is stored under the messages key.
opencode adds config to OpenCodeWorkerd.create(): a provider endpoint (@ai-sdk/openai-compatible with baseURL from OPENCODE_BASE_URL) and model: "endpoint/<OPENCODE_MODEL>". Everything else is celld’s code.

What surprised us

The opencode SDK builds a Response from another (new Response(body, response)). CellB’s Response lost the headers of such an init, so the SDK’s client saw no content-type and failed with ClientError: UnsupportedContentType; it now reads the init as Fetch does.
Packages with only a main field did not resolve in CellB’s Node bundles (esbuild’s neutral platform has no main fields): opencode’s dependencies (@smithy/*) failed to build. Node bundles now resolve as celld deploy does: conditions workerd, worker, browser, then module and main.
The opencode bundle is large (about 430,000 lines): its first load takes about half a minute.

Dependencies

@earendil-works/pi-agent-core and @earendil-works/pi-ai (pi), @opencode-ai/sdk (pinned to the build celld’s example uses); @statewalker/cellb, @statewalker/fake-openai and vitest for the tests.

License

MIT; opencode/index.js is adapted from celld’s example (Apache-2.0).