opencode-ccgo-provider/AGENTS.md

# AGENTS.md

Instructions for LLM coding agents working in this repository.

## Project overview

`opencode-commandcode-provider` is a native AI SDK provider that connects opencode to the
CommandCode `/alpha/generate` API. It implements `LanguageModelV3` from `@ai-sdk/provider@3.0.8`
and is loaded by opencode from `dist/index.js` via a `file://` npm spec.

`server.py` in this directory is the **reference implementation** of the same translation
(an OpenAI-compatible HTTP proxy). When behavior is ambiguous, `server.py` is the source of
truth for the upstream wire shape. This package is the TypeScript port; it does not run Python.

- Language/tooling: TypeScript, ESM, `tsc`.
- Runtime target: opencode's provider loader, `@ai-sdk/provider@3.0.8` (V3 specification).
- No runtime dependencies. `@ai-sdk/provider` is imported with `import type` only; the built
  output has no imports of its own.

## Commands

Run from the repository root.

```powershell
npm install        # dev deps: typescript, @types/node, @ai-sdk/provider
npm run typecheck  # tsc --noEmit
npm run build      # tsc -> dist/
npm run smoke      # live request against CommandCode (needs a key)
npm run sync-models # regenerate provider.commandcode.models from the live catalog
```

There is no unit-test suite. `scripts/smoke.mjs` is the end-to-end check. `scripts/sync-models.mjs`
is the model-catalog generator (needs `dist/` first; writes the opencode config unless `--out`/`--dry-run`).

After any change under `src/`, run `npm run typecheck`, then `npm run build`, then `npm run smoke`.
A change is not done until it typechecks and the smoke test passes.

## Layout

```
src/index.ts      Public exports: createCommandCode (named) and default.
src/model.ts      CommandCodeLanguageModel: HTTP, retries, doStream/doGenerate, event->part mapping.
src/transform.ts  LanguageModelV3CallOptions -> /alpha/generate envelope (returns a JSON string).
src/events.ts     Async iterator over the NDJSON/SSE response body.
src/usage.ts      finish event -> LanguageModelV3Usage; finish-reason unification.
src/redact.ts     Credential scrubbing for error surfaces.
src/constants.ts  Defaults, paths, headers, passthrough params, static config block.
scripts/smoke.mjs Live end-to-end check.
scripts/sync-models.mjs Catalog -> provider.commandcode.models generator (with vision probing).
server.py         Reference Python proxy (do not modify unless explicitly asked).
```

## Architecture / data flow

```
opencode
  └─ provider.languageModel(id).doStream(options)
       ├─ transform(options, id) ──► JSON envelope
       ├─ fetch POST {baseURL}/alpha/generate   (retry 429/5xx)
       ├─ iterateEvents(response.body)          (NDJSON lines)
       └─ map each event ──► LanguageModelV3StreamPart*
            ... always ending with a `finish` part
```

`doGenerate` calls `doStream` and drains the resulting stream into a
`LanguageModelV3GenerateResult`.

## Conventions

- ESM only. Imports of local files **must** end in `.js` (NodeNext resolution), e.g.
  `import { transform } from "./transform.js";`.
- `strict` TypeScript is on, including `noUncheckedIndexedAccess` and `verbatimModuleSyntax`.
  Type-only imports must use `import type`.
- Do **not** add comments unless they explain a non-obvious decision. Existing comments document
  upstream quirks and are intentional; match that spirit, not volume.
- Keep `transform.ts` output in parity with `server.py` `transform()`. If you change the envelope,
  explain which `server.py` behavior motivated it.
- Preserve the single public factory `createCommandCode`. `index.ts` re-exports it as both a named
  and the default export.
- The package must stay runtime-dependency-free. Do not add a runtime `import` of a package that
  would need `node_modules` at load time; opencode imports `dist/index.js` directly.

## Invariants and gotchas

These are load-bearing. Breaking one causes silent failures in opencode.

1. **Exactly one `create*` export reachable from the entrypoint.** opencode loads a custom npm
   provider by importing it, choosing the first export whose name starts with `create`, and calling
   it as `fn({ name: providerID, ...options })`. The result must expose `languageModel(id)`.
2. **`params.stream` is always `true`.** Sending `stream: false` makes upstream answer
   "Proxy use detected. This endpoint only serves CLI." Non-streaming clients are served by
   buffering the V3 stream in `doGenerate`.
3. **`tool-call` stream parts carry `input` as a stringified JSON string**, not an object
   (`LanguageModelV3ToolCall.input: string`). Serialize with `JSON.stringify`. Passing an object
   breaks opencode's tool execution with errors like `K.input.trim is not a function`.
4. **Always emit a terminal `finish` part.** If upstream ends without a `finish` event, synthesize
   one from the accumulated state. A missing finish surfaces as a truncated-response error.
5. **Drop unpaired tool ids.** Any `tool-call` without a matching `tool-result` (or vice versa) must
   be omitted from the envelope; upstream rejects the whole request otherwise.
6. **Usage is nested in V3.** Fill `inputTokens.{total,noCache,cacheRead,cacheWrite}` and
   `outputTokens.{total,text,reasoning}` — not flat `promptTokens`/`completionTokens`.
7. **`supportedUrls` is `{}` and stays empty.** CommandsCode URLs are not fetched by the SDK.
8. **Redact before surfacing errors.** Any upstream error text passed to the client must go through
   `redact()`.
9. **Replay assistant reasoning as a `{type:"reasoning"}` content part.** DeepSeek thinking mode
   rejects any request carrying `tools` whose prior assistant turns omit their reasoning. In
   `transform.ts`, forward each assistant `reasoning` part as `{type:"reasoning", text, signature?}`;
   do not drop it and do not fabricate empty reasoning. `server.py` predates this requirement and is
   not the guide here.
10. **`file://` npm specs bypass install.** opencode imports `dist/index.js` directly, so the repo
   must be rebuilt for opencode to see source changes.
11. **`/models` is config-driven, not provider-driven.** opencode builds the model list from
    `provider.commandcode.models` in `opencode.json` and never asks a custom `file://` provider to
    discover models (only internal providers can register `discoverModels`). Refresh the map with
    `npm run sync-models`; do not expect a discovery hook in `src/` to populate it.

## Change workflow

1. Read `server.py` for the behavior being ported and note the relevant function.
2. Make the smallest change in the matching `src/` file.
3. `npm run typecheck` — fix all errors.
4. `npm run build`.
5. `npm run smoke` — verify text and, when relevant, tool calls.
6. If the change affects the opencode integration, verify with `opencode run` (see below).
7. Update `README.md` if options, features, or behavior changed. Keep `AGENTS.md` invariants
   current if you discover a new one.

Do not edit `dist/` by hand; it is generated. Do not modify `server.py` unless explicitly asked.

## Verification

Automated:

```powershell
npm run typecheck; npm run build; npm run smoke
```

Inside opencode (restart it first if config changed):

```powershell
opencode models commandcode
opencode run "Reply with exactly: pong" -m commandcode/deepseek/deepseek-v4-flash-vision-exp
opencode run "Use the glob tool to list *.mjs in this directory and report the filenames." -m commandcode/deepseek/deepseek-v4-flash-vision-exp
```

The third command confirms the full tool-call round trip (tool-call emitted, executed, result
replayed). Run it after touching `transform.ts` or the tool-call mapping in `model.ts`.

## Do / Don't

- **Do** keep the wire envelope identical to `server.py`.
- **Do** run typecheck, build, and smoke after every source change.
- **Do** prefer `{env:COMMANDCODE_API_KEY}` in docs and examples.
- **Don't** commit real API keys, `dist/`, `node_modules/`, `dump/`, or logs (all gitignored).
- **Don't** add runtime dependencies or make `dist` require `node_modules`.
- **Don't** force `stream:false`, send object `tool-call.input`, or omit the `finish` part.
- **Don't** commit unless the user explicitly asks.