feat: support Command Code Provider Chat Completions API

AGENTS.md

@@ -136,16 +136,18 @@ tsdown.config.ts      Build config (tsdown -> lib/, ESM, .d.ts + client.js).
   declaration never exists and `slots.inject` never fires — the registration
   silently does not happen; do not "harden" that into an error.
 - **Wire protocol** (reverse-engineered, command-code@1.28.4; re-verified unchanged against 1.44.0):
-  - `POST {apiBase}/alpha/generate` — body `{ config, memory, taste, skills, params: { model, messages, tools, system, max_tokens, temperature, stream, reasoning_effort? }, threadId }`.
-  - Image parts use the official CLI wire shape: `{ type: 'image', source: { type: 'base64', media_type, data } }` (NOT pi's `data:` data-URI form).
-  - Stream: SSE-ish JSONL events `text-delta | reasoning-start/delta/end | tool-call | tool-result | finish | error`.
+  - `POST {apiBase}/alpha/generate` — CLI transport body `{ config, memory, taste, skills, params: { model, messages, tools, system, max_tokens, temperature, stream, reasoning_effort? }, threadId }`. Used for Go-plan accounts (the only plan without Provider API access) and as the fallback when `/provider/v1/chat/completions` returns `upgrade_required`.
+  - `POST {apiBase}/provider/v1/chat/completions` — documented OpenAI-format transport with a flat body `{ model, messages, tools?, max_tokens, temperature, stream, reasoning_effort? }`. Used for accounts with Provider API access; historical reasoning is replayed as `reasoning_content`.
+  - Image parts use the official CLI wire shape on `/alpha/generate`: `{ type: 'image', source: { type: 'base64', media_type, data } }`; OpenAI `/provider/v1/chat/completions` uses `{ type: 'image_url', image_url: { url: 'data:...' } }`.
+  - CLI stream: SSE-ish JSONL events `text-delta | reasoning-start/delta/end | tool-call | tool-result | finish | error`.
+  - OpenAI stream: standard SSE chunks with `delta.reasoning` / `delta.reasoning_content` / `delta.content` / `delta.tool_calls`, a `finish_reason`, and optional `usage`.
   - Catalog: `GET {apiBase}/provider/v1/models` → `{ object: 'list', data: [{ id, name, context_length }] }`.
   - Web search: `POST {apiBase}/alpha/web-search` — body `{ query, numResults, allowedDomains?, blockedDomains? }` → `{ results: [{ title, url, snippet }] }`; same `Authorization: Bearer <key>` + `x-command-code-version` as generate.
   - Defaults: `apiBase = https://api.commandcode.ai`, `COMMAND_CODE_CLI_VERSION = '1.44.0'`.
 - **API key resolution order** (in `src/index.ts`): `config.apiKey` → credential ref `apiKeyEnv` (default `COMMANDCODE_API_KEY`, via the dsh credentials seam) → launch environment → official CLI auth file `~/.commandcode/auth.json`. **pi/OMP auth files are intentionally NOT scanned** — keep it that way.
 - **Multi-account rotation** (`src/accounts.ts` + the adapter's connect loop): the top-level key forms the `default` slot; `Config.accounts` (`[{ label, apiKeyEnv | apiKey }]`) adds more, in rotation order. Rotation is **passive**: a key is marked only on a real pre-stream rejection (429 → `unknown` cooldown, 401 → `disabled`), and the adapter's `rotateApiKey` hook re-sends the same request with the next account's key (safe: nothing streamed, the body is account-independent, `threadId` random per request — mid-stream failures NEVER rotate). When every account is marked, the pool probes `/alpha/billing/credits` per key (`probeFiveHourWindow`) to revive reset windows, else throws `RATE_LIMIT` naming the earliest `resetAt` (all-401 → `INVALID_CREDENTIAL`). State is keyed by API key, not slot — shared credentials share one mark. **Manual selection**: `Config.activeAccount` (a slot id) pins the serving account via the pool's `preferredId` seam + `selectActiveAccount()` (shared with the usage view's active badge); a pinned-but-exhausted or unknown id falls back to rotation order. **Model routing**: `Config.modelAccountRules` (`[{ models: string[], account }]`) lists catalog model ids per account slot; the request's model reaches key resolution (`resolveApiKey(connection, model)`), the pool's `modelAccountRules` seam re-reads rules per resolution, and `matchModelRule()`/`selectAccountForModel()` serve the routed account before preferred/rotation — an unusable routed account falls back, so the router is a hint, never a hard gate. The rules editor's model list comes from a Host-side `commandcode/models` Remote (adapter catalog, sorted), so the browser never calls the Command Code API; `SettingsPageApi.models` is optional, so legacy transports degrade to the empty-catalog state. Extra-account slot ids are the credential reference itself (`COMMANDCODE_API_KEY_2`, …) so a stored selection survives list reorders/removals; only literal-only composition entries keep positional `account-N` ids. The settings page edits `activeAccount` through the generic section-field machinery (a `<select>` bound to a text field) and `modelAccountRules` through its own rules card (staged rows like `accounts`, one `modelAccountRules` write). The picker's billing-access cache is per key. The usage Remote result is `CommandCodeAccountsReport` (`{ accounts: [...] }`); host and client ship in one bundle, so wire-shape changes need no migration — only synced edits in `src/usage-wire.ts`, `src/usage-remote.ts`, `src/client/usage.ts`, and `src/commands.ts`.
 - **Web search (`src/web-search.ts` + the optional `web` seam)**: the model-facing `web_search` tool (from `@deepseek-ai/dsh-tool-web`) is served by a `CommandCodeSearchProvider` registered as `commandcode` on `ctx.web` — same `Authorization: Bearer <key>` + `x-command-code-version` chain, same `apiBase`, so DSH's web search needs NO separate key/endpoint config (unlike `dsh-web-search-deepseek`, which needs its own Anthropic-compatible base). It POSTs `{ query, numResults, allowedDomains?, blockedDomains? }` to `/alpha/web-search` and maps `{ title, url, snippet }` → `WebSearchSource`. Registration rides `ctx.inject(['web'], ...)` exactly like `commands`/`typert`: the provider is registered only when the profile mounts the web service, and the fiber never activates otherwise (this stays an LLM-provider-only plugin without web). The pool's `resolveKey()` (rotation + auth-file revived) is reused, so search benefits from the same multi-account selection; the search endpoint is account-independent so no mid-flight rotation happens. **Selection**: whether the `commandcode` provider WINS over the shipped `deepseek-official` is `Config.webSearch` (default on). The web seam has NO public runtime selector, so the plugin writes its private `searchProviderId` field (read per call by `web.search()`) via `selectCommandCodeSearchProvider()` in `src/web-search.ts` — applied at boot AND on every settings change (the `installSection` `onChange` hook). That write depends on the runtime shape (a plain writable property, not `#private`); the durable alternative is the boot-time `searchProvider: commandcode` cordis patch. `dsh-web` is a `^0.1.2-rc.1` peer (kept external in tsdown); `tests/web-search.test.ts` pins the wire body, header, result mapping, the `WEB_ABORTED`/`WEB_PROVIDER_CREDENTIAL_MISSING`/`WEB_PROVIDER_ERROR` taxonomy, and the selection-field rewrite.
-- **StreamChunk contract** (dsh-llm): each block starts with `block-start`, deltas by `index`, ends with `block-end`; `usage` before `finish`; nothing after `finish`. Tool-call `arguments` are raw JSON strings. Reasoning blocks are intentionally NOT replayed into later turns (matches the CLI; private reasoning must not leak). Only tool calls with a paired tool result are replayed.
+- **StreamChunk contract** (dsh-llm): each block starts with `block-start`, deltas by `index`, ends with `block-end`; `usage` before `finish`; nothing after `finish`. Tool-call `arguments` are raw JSON strings. On the legacy `/alpha/generate` transport, reasoning blocks are intentionally NOT replayed into later turns (matches the CLI; private reasoning must not leak); on the `/provider/v1/chat/completions` transport they are passed back as `reasoning_content` for tool-loop continuity. Only tool calls with a paired tool result are replayed on both transports.
 - **Errors**: throw `LlmError` with stable codes. 401 → `INVALID_CREDENTIAL`; 429 → `RATE_LIMIT`; other HTTP → `PROVIDER_HTTP_ERROR` (403 body's `error.code`, e.g. `MODEL_NOT_IN_PLAN`, is parsed into the message). Unsupported options (`stop`) and image input throw `UNSUPPORTED_OPTION` / `UNSUPPORTED_CONTENT` rather than silently dropping.
 - **Adapter is cordis-free** by design: `src/adapter.ts` takes a per-request `options()` thunk + `resolveApiKey()` from the plugin entry, so settings changes reach the next request without re-registration. It also accepts an injectable `fetchImpl` for tests.
 - **Usage Remote (`commandcode/report`)**: the settings page's account card

README.md

@@ -151,7 +151,7 @@ When your deployment's dsh shell mounts the web capability (`@deepseek-ai/dsh-we
 - **Image input is model-gated** — only Vision models accept images; text-only models refuse them.
 - Switching to a text-only model in an image-bearing session is rejected by dsh — pick a model marked *`Image`* or remove the images first.
 - **No `stop` sequences** — requests carrying one fail.
-- Reasoning blocks are not replayed into later turns; only tool calls with a paired tool result are replayed.
+- On the legacy `/alpha/generate` transport, reasoning blocks are not replayed into later turns; on the `/provider/v1/chat/completions` transport, historical reasoning is passed back as `reasoning_content` so tool-calling loops can keep their chain of thought. Only tool calls with a paired tool result are replayed on both transports.
 - The model catalog is browsable without a key; chat requests need one.
 
 ## Permissions & privacy

README.zh-CN.md

@@ -149,7 +149,7 @@ llm-commandcode:
 - **图片输入按模型能力限制**：仅 Vision 模型接受图片，纯文本模型会直接拒绝。
 - 含图片的会话切换到纯文本模型会被 dsh 拒绝——请改选带 *`Image`* 标记的模型，或先移除图片。
 - **不支持 `stop` 序列**：携带它的请求会报错。
-- 推理块不会重放到后续轮次；只有带配对工具结果的工具调用会被重放。
+- 在旧版 `/alpha/generate` 传输中，推理块不会重放到后续轮次；在 `/provider/v1/chat/completions` 传输中，历史推理会以 `reasoning_content` 回传，以便工具调用循环保留思维链。两种传输都只重放带配对工具结果的工具调用。
 - 模型目录无需 key 即可浏览；对话请求需要 key。
 
 ## 权限与隐私