@mars-sea/dsh-commandcode-provider 0.1.6 → 0.1.9

package/CHANGELOG.md

@@ -4,6 +4,39 @@ All notable changes to this project will be documented in this file.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.1.9] - 2026-08-15
+
+### Added
+
+- **Image input for Vision-capable models.** Models the official Command Code registry lists with Vision (see `KNOWN_IMAGE_MODELS` in `src/adapter.ts`, synced from the [official model registry](https://commandcode.ai/docs/reference/cli/models)) now accept attached images: bytes resolve through the dsh attachment service (`ctx.attachments`) and are sent in the official CLI wire shape `{ type: 'image', source: { type: 'base64', media_type, data } }`. Text-only models (e.g. `deepseek/deepseek-v4-flash`) refuse images loudly (`UNSUPPORTED_CONTENT`) rather than silently dropping them; a request carrying images also requires the attachment service. The `CommandCodeAdapterDeps` seam gains an optional `resolveAttachments` resolver (used lazily, only when a request actually has images).
+- **The model picker now shows each Command Code model's image capability** (`listModels`/`resolveModel` return a `description`: *"Supports image input"* / *"Text only"*), so switching in an image-bearing session is informed instead of surprising.
+- **A client half for the bundle** (`dsh.client` + `exports["./client"]` → `lib/client.js`): it wraps the shared `session.selectModel` face and rewrites the harness's image-session `model-unavailable` rejection into a clear, actionable message — `当前会话已包含图片，而模型 <model> 不支持图片输入；请选择支持图片的模型，或先移除会话中的图片。` — while passing the error code and details through unchanged. The rejection itself is a deliberate `dsh-host-apiproxy` guard that cannot be relaxed from the plugin side; this makes it friendlier. Both READMEs document the behavior.
+
+## [0.1.8] - 2026-08-15
+
+### Fixed
+
+- **The 0.1.7 scoped-name fix crashed on boot with a YAML parse error.** 0.1.7 rewrote the patch row's `name` to `@mars-sea/dsh-commandcode-provider` but left it unquoted; a YAML scalar starting with `@` is an indicator and fails to parse (`YAMLException: bad indentation of a mapping entry`), so `dsh --dump-config` and every boot died. The value is now quoted: `name: "@mars-sea/dsh-commandcode-provider"` (verified against dsh's own js-yaml and a real profile boot).
+- **Transport failures now surface the real root cause.** The `TRANSPORT` error from a failed `fetch` (DNS, refused/reset connection, TLS, proxy, timeout) previously reported only the generic wrapper — `Command Code API request to .../alpha/generate failed` — while the actionable detail sat unused on `error.cause`. The web UI renders only the error message (not the cause chain), so users hit a wall of retry rows ("重试延迟"/"Retry delay") with no way to diagnose. The message now appends `errorChain(cause)`, so the failure reason names e.g. `connect ECONNREFUSED`, `ENOTFOUND`, `CERT_HAS_EXPIRED`, or the timeout abort.
+- **A stalled connection no longer hangs the turn.** `requestTimeoutMs` (default 60s) bounds the wait for the first response byte via `AbortSignal.timeout`, and `streamIdleTimeoutMs` (default 120s) treats a stream with no events as a dead connection — both fail as `TIMEOUT` with the duration instead of hanging until the OS socket timeout (which can be minutes). Configurable per profile in the `llm-commandcode` settings section.
+
+### Changed
+
+- `CommandCodeConnectionOptions` gains `requestTimeoutMs` and `streamIdleTimeoutMs`; both are optional in the `Config` schema and default to 60s/120s. New `DEFAULT_REQUEST_TIMEOUT_MS` / `DEFAULT_STREAM_IDLE_TIMEOUT_MS` exports.
+- Both READMEs document the new knobs and the transport-failure troubleshooting entry (notably: Node's fetch ignores `HTTP_PROXY`/`HTTPS_PROXY`, so proxy-dependent networks fail here while the browser works).
+- Both READMEs gain an **Updating** section: since the bundle patch layer is read from the installed package at boot, updating the package fixes the patch row automatically; the section covers npm/git/local update commands and the ≤0.1.6 hand-copied-patch caveat.
+- Both READMEs restructure the install docs: **npm is now the recommended install path** (one command, always the latest published release), GitHub moves below it, and the uninstall command is documented (use the scoped name `@mars-sea/dsh-commandcode-provider`, since pnpm records dependencies under the real package name).
+
+## [0.1.7] - 2026-08-15
+
+### Fixed
+
+- **Boot-crashing patch row for every install path** (`cordis.patch.yml`): the layer's `name` was the bare `dsh-commandcode-provider`, but the loader imports it as a module from the profile's `node_modules`, where pnpm only links the true scoped name `@mars-sea/dsh-commandcode-provider`. Any install (npm, GitHub, local path — all of which link the scoped name) failed at load with `ERR_MODULE_NOT_FOUND` and took the web app into a `Restart=on-failure` crash loop. The row now reads `name: @mars-sea/dsh-commandcode-provider`. Existing profiles that copied the old row (or an old README example) into their own `cordis.patch.yml` must update it the same way (see the Troubleshooting entry).
+
+### Changed
+
+- Both READMEs document the scoped `name` requirement, show the corrected `--dump-config` layer heading (`# == @mars-sea/dsh-commandcode-provider`), and use the scoped name in the `allowBuilds` example and the `remove` command.
+
 ## [0.1.6] - 2026-08-15
 
 ### Added

package/README.md

@@ -17,6 +17,7 @@ Unofficial [DeepSeek Harness](https://deepseek-harness.github.io/deepseek-harnes
 - A **Models-page card** ("Command Code") with an API-key field — credentials are stored through the dsh credentials service, same as the DeepSeek card.
 - **API key resolution** in this order: `config.apiKey` → credential reference `apiKeyEnv` (the web Models page writes it, default `COMMANDCODE_API_KEY`) → the launching environment → the official Command Code CLI auth file (`~/.commandcode/auth.json`, written by `command-code login`).
 - **Reasoning-effort support** for the models Command Code's catalog marks as such (e.g. `claude-opus-5`, `gpt-5.5`, `deepseek/deepseek-v4-pro`, …) via `KNOWN_EFFORTS`, matching the official command-code@1.26.0 bundled catalog.
+- **Image input for Vision-capable models**: models the official registry lists with Vision (e.g. `claude-sonnet-5`, `gpt-5.4`, `google/gemini-3.5-flash`, …) accept attached images, resolved through the dsh attachment service and sent in the official Command Code wire format. Text-only models (e.g. `deepseek/deepseek-v4-flash`, `zai-org/GLM-5.3`) refuse images loudly rather than silently dropping them.
 
 ## Getting an API key
 
@@ -31,11 +32,19 @@ cmd login        # macOS/Linux; native Windows: cmdc login
 
 ## Install
 
-### From GitHub (recommended)
+### From npm (recommended)
+
+The plugin is published to the npm registry as **`@mars-sea/dsh-commandcode-provider`** (the bare name `dsh-commandcode-provider` is taken by an unrelated package):
+
+```sh
+dsh plugin --profile web add @mars-sea/dsh-commandcode-provider
+```
+
+### From GitHub
 
 ```sh
 # Pin a release tag (recommended — readable and immutable)
-dsh plugin --profile web add github:Mars-Sea/dsh-commandcode-provider#v0.1.6
+dsh plugin --profile web add github:Mars-Sea/dsh-commandcode-provider#v0.1.8
 # Or pin any exact commit by its SHA
 dsh plugin --profile web add github:Mars-Sea/dsh-commandcode-provider#<full-commit-sha>
 ```
@@ -46,19 +55,11 @@ A git install fetches **sources**, so the package's `prepare` script builds `lib
 
 ```yaml
 allowBuilds:
-  'dsh-commandcode-provider@github:Mars-Sea/dsh-commandcode-provider#<full-commit-sha>': true
+  '@mars-sea/dsh-commandcode-provider@github:Mars-Sea/dsh-commandcode-provider#<full-commit-sha>': true
 ```
 
 and re-run the `add`. Only allow packages whose source you trust (and pin a commit).
 
-### From npm
-
-Published as **`@mars-sea/dsh-commandcode-provider`** (the bare name `dsh-commandcode-provider` is taken on the npm registry by an unrelated package):
-
-```sh
-dsh plugin --profile web add @mars-sea/dsh-commandcode-provider
-```
-
 ### From a local checkout
 
 ```sh
@@ -71,23 +72,52 @@ A local path install links the checkout as-is, so after changing `src/` re-run `
 
 ### What the install does
 
-`dsh plugin add` links the package into the profile, appends `dsh-commandcode-provider` to the profile's `dsh.profile.bundles`, and activates the `cordis.patch.yml` layer, which inserts:
+`dsh plugin add` links the package into the profile (pnpm records dependencies and links `node_modules` by the **true package name**, i.e. `@mars-sea/dsh-commandcode-provider`), appends that same name to the profile's `dsh.profile.bundles`, and activates the `cordis.patch.yml` layer, which inserts:
 
 ```yaml
 - insert:
     - id: llm-commandcode
-      name: dsh-commandcode-provider
+      name: "@mars-sea/dsh-commandcode-provider"
       config:
         apiKeyEnv: COMMANDCODE_API_KEY
 ```
 
+The `name` in the patch row must be the **full package specifier, quoted**: the loader imports it as a module and resolves it from the profile's `node_modules`, where pnpm only ever links the scoped name. A bare `dsh-commandcode-provider` fails with `ERR_MODULE_NOT_FOUND` and crashes the app on boot, and an unquoted `@mars-sea/...` fails YAML parsing (see [Troubleshooting](#troubleshooting)).
+
 Verify the composed layer, then (re)start the web app:
 
 ```sh
-dsh --profile web --dump-config          # shows a "# == dsh-commandcode-provider" layer
+dsh --profile web --dump-config          # shows a "# == @mars-sea/dsh-commandcode-provider" layer
 dsh web                                  # or restart your running instance
 ```
 
+## Updating
+
+The bundle's patch layer is read from the **installed package** at every boot, so updating the package brings in the fixed patch row automatically — you do not need to hand-edit `cordis.patch.yml` unless you copied its contents into your own profile layer.
+
+Update according to how you installed it:
+
+```sh
+# From npm (recommended): always the latest published release
+dsh plugin --profile web update @mars-sea/dsh-commandcode-provider
+
+# From GitHub pinned to a tag: point at the new tag
+# (no need to uninstall first — pnpm swaps the pinned revision in place,
+# and the bundle layer is re-read from the installed package on next boot)
+dsh plugin --profile web add github:Mars-Sea/dsh-commandcode-provider#v0.1.8
+
+# From a local checkout: pull the new code, rebuild, restart
+git -C /path/to/dsh-commandcode-provider pull
+npm run build --prefix /path/to/dsh-commandcode-provider
+dsh web
+```
+
+Then restart the web app (`dsh web`, or restart the service). Verify the running version with `dsh --profile web --dump-config` — the layer should show `name: '@mars-sea/dsh-commandcode-provider'`.
+
+> **Upgrading from ≤0.1.6** (or a broken hand-edited profile): the installed package's patch layer now carries the corrected, quoted `name`. If you previously *copied* the old patch row into your profile's own `cordis.patch.yml`, that copy still wins over the bundle layer — fix it manually to `name: "@mars-sea/dsh-commandcode-provider"` (see [Troubleshooting](#troubleshooting)) or remove it and let the bundle layer apply.
+
+> **To uninstall instead of upgrading** (e.g. you are on a broken pre-0.1.7 tag and want to start clean): `dsh plugin --profile web remove @mars-sea/dsh-commandcode-provider` (the scoped name — pnpm records the dependency under its real package name, so the bare `dsh-commandcode-provider` form does not match). This removes the dependency and its layer; your API key in the dsh credential store and `~/.commandcode/auth.json` are left untouched. Then install the current version with the npm or GitHub command above.
+
 ## Verify it works
 
 After restart, in the web UI: **Settings → Models** shows a **Command Code** card; the model picker lists the live catalog under **commandcode** (54 models at the time of writing). Send a message with a model your plan includes — the default `deepseek/deepseek-v4-flash` works on entry-level plans; open-weight models (DeepSeek/Qwen/Kimi/MiniMax) generally do, while frontier models (Claude/GPT/Gemini/Grok) may require Pro/Max plans or on-demand usage (see FAQ).
@@ -133,21 +163,31 @@ llm-commandcode:
   apiBase: https://api.commandcode.ai
   workingDir: /path/to/project     # reported to the API (project slug, config block)
   modelsCachePath: ~/.commandcode/models-cache.json
+  requestTimeoutMs: 60000          # max wait for the first response byte (default 60s)
+  streamIdleTimeoutMs: 120000      # stream stall before treated as a dead connection (default 120s)
 ```
 
 The composition-entry config (`cordis.patch.yml` / your profile `cordis.patch.yml`) accepts the same keys; a literal `apiKey` there takes precedence over the credential reference.
 
 ## Troubleshooting
 
+- **`Command Code API request to .../alpha/generate failed` and the turn keeps retrying (`重试延迟` / "Retry delay")** — this is a **transport-layer failure**: `fetch()` never received an HTTP response (not a 401/403/429, which would say "API error"). Since dsh's retry policy retries `TRANSPORT` twice with backoff, you'll see retry rows in the UI before the turn finally fails. Since 0.1.8 the failure reason shows the **real root cause** (e.g. `fetch failed: connect ECONNREFUSED`, `ENOTFOUND`, `CERT_HAS_EXPIRED`, `The operation was aborted due to timeout`). Common causes:
+  - **A proxy is required in your network.** Node's `fetch` (undici) does **not** read `HTTP_PROXY`/`HTTPS_PROXY` environment variables, so a browser/curl that goes through a system proxy works while dsh fails. Run dsh with the proxy configured for undici (e.g. `NODE_OPTIONS=--import undici` with a dispatcher, or a network-level route), or whitelist `api.commandcode.ai`.
+  - **The connection is being reset/throttled mid-request** (firewall, GFW-style interference, unstable Wi-Fi). The error message will name it (`socket hang up`, `ECONNRESET`, `UND_ERR_SOCKET`).
+  - **TLS interception** (corporate MITM) — `CERT_HAS_EXPIRED`/`DEPTH_ZERO_SELF_SIGNED_CERT` in the chain.
+  - A transient blip that a retry recovers from; if it persists every turn, it's environmental, not the API (the models endpoint and generate endpoint respond normally from healthy networks).
+- **A long generation stops mid-stream** — since 0.1.8 the adapter aborts a request that gets no response within `requestTimeoutMs` (60s default) and a stream that stalls past `streamIdleTimeoutMs` (120s default) instead of hanging forever. Both failures surface as `TIMEOUT` with the stall duration; tune the knobs in the `llm-commandcode` settings if your network is slow but stable.
+- **The web app crashes on boot with `ERR_MODULE_NOT_FOUND: Cannot find package 'dsh-commandcode-provider'`** — the patch row's `name` is the bare package name, but the loader imports it as a module from the profile's `node_modules`, where pnpm only links the scoped name `@mars-sea/dsh-commandcode-provider`. Pre-0.1.7 bundles shipped this wrong row, and the bug also bites when an old `cordis.patch.yml` example (or a cached profile layer) is copied by hand. Fix the row in your profile's `cordis.patch.yml` (or re-add the plugin) so it reads `name: "@mars-sea/dsh-commandcode-provider"` — note the **quotes**: an unquoted `@`-prefixed scalar fails YAML parsing (0.1.7 shipped that regression; 0.1.8 quotes it) — then restart.
 - **`MODEL_NOT_IN_PLAN` (403)** — the selected model is not in your Command Code plan. Pick an open-weight model (e.g. `deepseek/deepseek-v4-flash`) or upgrade. The error names the model and links the official docs.
 - **`MISSING_CREDENTIAL`** — no key anywhere. Store one via the Models page card, export `COMMANDCODE_API_KEY`, set `config.apiKey`, or run `command-code login`. The route stays registered and the catalog stays browsable without a key.
 - **The Models page card shows "not configured" but requests work** — the key came from `~/.commandcode/auth.json` (the `cmd login` fallback), not the dsh credential store. Paste it into the card once to make the card show as configured; both coexist fine.
 - **A reasoning model returns no visible text on short requests** — reasoning models (e.g. `deepseek/deepseek-v4-*`) consume output tokens on reasoning first; a small `maxTokens` can be exhausted before any visible text. This is normal.
-- **`allowBuilds` errors on `dsh plugin add` from git** — copy the exact package key pnpm printed (with the commit hash) into `pnpm-workspace.yaml` and re-run (see [Install](#from-github-recommended)).
+- **`allowBuilds` errors on `dsh plugin add` from git** — copy the exact package key pnpm printed (with the commit hash) into `pnpm-workspace.yaml` and re-run (see [Install](#from-github)).
 
 ## Notes & limitations
 
-- **Text-only for now**: image input throws `UNSUPPORTED_CONTENT` (wiring the attachment service to resolve image bytes is future work). The pi plugin's `MODEL_INPUT_MODALITIES` table is intentionally not claimed.
+- **Image input is model-gated**: only models the official Command Code registry lists with Vision accept images (see the `KNOWN_IMAGE_MODELS` snapshot in `src/adapter.ts`, synced from the [official model registry](https://commandcode.ai/docs/reference/cli/models)). The model picker annotates each Command Code model with *"Supports image input"* / *"Text only"*, so the capability is visible before you switch. Sending an image to a text-only model throws `UNSUPPORTED_CONTENT`. Command Code's own CLI falls back to a client-side *VISION* side-call for text-only models; this adapter does **not** reproduce that interactive feature — switch to a Vision-capable model instead. Image input also requires the dsh **attachment service** (`ctx.attachments`); without it, requests carrying images throw `UNSUPPORTED_CONTENT`.
+- **Switching to a text-only model in an image-bearing session is rejected by dsh itself** — a harness-level guard (`dsh-host-apiproxy`'s `selectModel` handler) refuses `model-unavailable` when the session history or the pending input already contains images and the target model does not declare `image` input. The rejection is intentional and cannot be relaxed from the plugin side (the picker rows this adapter provides are the input that makes the guard work — a text-only model correctly reports `inputModalities: ['text']`). What this bundle **does** do is make the message friendlier: its client half wraps `session.selectModel` and rewrites that rejection to `当前会话已包含图片，而模型 <model> 不支持图片输入；请选择支持图片的模型，或先移除会话中的图片。` (the error code and details pass through unchanged, so any caller switching on `error.code` keeps working). To keep using images, select a model the picker annotates *"Supports image input"*, or remove the images from the session first; alternatively an image-routing bundle (e.g. `@deepseek-ai/dsh-llm-image-routing`) can transparently route image turns to a vision fallback.
 - **No `stop` sequences**: the wire format has no stop field; requests carrying one throw `UNSUPPORTED_OPTION`.
 - Reasoning blocks are **not replayed** into later turns (matches the official CLI: prior private reasoning must not leak).
 - Only tool calls with a paired tool result are replayed into the conversation.
@@ -173,7 +213,7 @@ This plugin operates entirely within your dsh profile and your Command Code acco
 - **Uninstall** completely:
 
   ```sh
-  dsh plugin --profile web remove dsh-commandcode-provider
+  dsh plugin --profile web remove @mars-sea/dsh-commandcode-provider
   ```
 
   This removes the bundle dependency and its layer. Your API key in the dsh credential store and `~/.commandcode/auth.json` are left untouched (you can remove them manually if you want to revoke access).

package/README.zh-CN.md

@@ -17,6 +17,7 @@
 - **Models 页面卡片**（"Command Code"）带 API key 输入框——凭据通过 dsh 凭据服务存储，与 DeepSeek 卡片一致。
 - **API key 解析顺序**：`config.apiKey` → 凭据引用 `apiKeyEnv`（Web Models 页面写入，默认 `COMMANDCODE_API_KEY`）→ 启动环境变量 → 官方 Command Code CLI 认证文件（`~/.commandcode/auth.json`，由 `command-code login` 写入）。
 - **推理强度（reasoning-effort）支持**：针对 Command Code 目录中标为推理模型的模型（如 `claude-opus-5`、`gpt-5.5`、`deepseek/deepseek-v4-pro` 等），通过 `KNOWN_EFFORTS` 实现，与官方 command-code@1.26.0 内置目录一致。
+- **支持视觉模型的图片输入**：官方注册表中带 Vision 能力的模型（如 `claude-sonnet-5`、`gpt-5.4`、`google/gemini-3.5-flash` 等）可接收附加图片——通过 dsh 附件服务解析字节，并以官方 Command Code wire 格式发送。纯文本模型（如 `deepseek/deepseek-v4-flash`、`zai-org/GLM-5.3`）会明确拒绝图片而非静默丢弃。
 
 ## 获取 API key
 
@@ -31,11 +32,19 @@ cmd login        # macOS/Linux；Windows 原生版：cmdc login
 
 ## 安装
 
-### 从 GitHub 安装（推荐）
+### 从 npm 安装（推荐）
+
+插件发布在 npm 上，包名 **`@mars-sea/dsh-commandcode-provider`**（npm 上裸名 `dsh-commandcode-provider` 已被无关包占用）：
+
+```sh
+dsh plugin --profile web add @mars-sea/dsh-commandcode-provider
+```
+
+### 从 GitHub 安装
 
 ```sh
 # 推荐：锁定发布 tag（可读、不可变）
-dsh plugin --profile web add github:Mars-Sea/dsh-commandcode-provider#v0.1.6
+dsh plugin --profile web add github:Mars-Sea/dsh-commandcode-provider#v0.1.8
 # 或按完整 commit SHA 锁定任意提交
 dsh plugin --profile web add github:Mars-Sea/dsh-commandcode-provider#<完整-commit-sha>
 ```
@@ -46,19 +55,11 @@ git 安装会拉取**源码**，因此包的 `prepare` 脚本会在安装后构
 
 ```yaml
 allowBuilds:
-  'dsh-commandcode-provider@github:Mars-Sea/dsh-commandcode-provider#<完整-commit-sha>': true
+  '@mars-sea/dsh-commandcode-provider@github:Mars-Sea/dsh-commandcode-provider#<完整-commit-sha>': true
 ```
 
 然后重新运行 `add`。只允许信任其源码的包（并固定 commit）。
 
-### 从 npm 安装
-
-发布为 **`@mars-sea/dsh-commandcode-provider`**（npm 上裸名 `dsh-commandcode-provider` 已被无关包占用）：
-
-```sh
-dsh plugin --profile web add @mars-sea/dsh-commandcode-provider
-```
-
 ### 从本地检出安装
 
 ```sh
@@ -71,23 +72,51 @@ dsh plugin --profile web add /path/to/dsh-commandcode-provider
 
 ### 安装做了什么
 
-`dsh plugin add` 会将包链接到配置目录，把 `dsh-commandcode-provider` 追加到配置的 `dsh.profile.bundles`，并激活 `cordis.patch.yml` 层，其中插入：
+`dsh plugin add` 会将包链接到配置目录（pnpm 按**真实包名**记录依赖并链接 `node_modules`，即 `@mars-sea/dsh-commandcode-provider`），把同名包名追加到配置的 `dsh.profile.bundles`，并激活 `cordis.patch.yml` 层，其中插入：
 
 ```yaml
 - insert:
     - id: llm-commandcode
-      name: dsh-commandcode-provider
+      name: "@mars-sea/dsh-commandcode-provider"
       config:
         apiKeyEnv: COMMANDCODE_API_KEY
 ```
 
+patch 行里的 `name` 必须是**带引号的完整包名**：loader 会把它当作模块从 profile 的 `node_modules` 导入，而 pnpm 只会链接带 scope 的名字。写成裸名 `dsh-commandcode-provider` 会报 `ERR_MODULE_NOT_FOUND` 并在启动时崩溃；不引号的 `@mars-sea/...` 也会导致 YAML 解析失败（见[故障排查](#故障排查)）。
+
 验证合成后的层，然后（重新）启动 Web 应用：
 
 ```sh
-dsh --profile web --dump-config          # 会显示 "# == dsh-commandcode-provider" 层
+dsh --profile web --dump-config          # 会显示 "# == @mars-sea/dsh-commandcode-provider" 层
 dsh web                                  # 或重启你正在运行的实例
 ```
 
+## 更新
+
+bundle 的 patch 层在每次启动时都从**已安装的包**读取，所以更新包本身就会带入修复后的 patch 行——**不需要**手工改 `cordis.patch.yml`（除非你把它的内容复制到了自己 profile 的层里）。
+
+按安装方式选择更新命令：
+
+```sh
+# 从 npm 安装（推荐）：总是升到最新发布版本
+dsh plugin --profile web update @mars-sea/dsh-commandcode-provider
+
+# 从 GitHub 按 tag 安装：指向新 tag
+# （无需先卸载——pnpm 会就地替换固定的版本，下次启动时 bundle 层会从新安装的包重新读取）
+dsh plugin --profile web add github:Mars-Sea/dsh-commandcode-provider#v0.1.8
+
+# 从本地检出安装：拉取新代码、重新构建、重启
+git -C /path/to/dsh-commandcode-provider pull
+npm run build --prefix /path/to/dsh-commandcode-provider
+dsh web
+```
+
+然后重启 Web 应用（`dsh web`，或重启服务）。用 `dsh --profile web --dump-config` 验证运行的版本——层里应显示 `name: '@mars-sea/dsh-commandcode-provider'`。
+
+> **从 ≤0.1.6 升级**（或手改坏的 profile）：安装包自带的 patch 层现在已经带着修正后的带引号 `name`。如果你之前**手工复制**过旧的 patch 行到你 profile 自己的 `cordis.patch.yml`，那份拷贝会覆盖 bundle 层——请手动改成 `name: "@mars-sea/dsh-commandcode-provider"`（见[故障排查](#故障排查)），或删掉它让 bundle 层生效。
+
+> **想卸载而不是升级**（例如正卡在 0.1.7 之前坏掉的 tag，想干净重来）：`dsh plugin --profile web remove @mars-sea/dsh-commandcode-provider`（用 **scoped 名**——pnpm 按真实包名记录依赖，裸名 `dsh-commandcode-provider` 对不上）。这会移除依赖及其配置层；你在 dsh 凭据库和 `~/.commandcode/auth.json` 里的 API key 不受影响。然后用上面的 npm 或 GitHub 命令安装当前版本。
+
 ## 验证是否生效
 
 重启后，在 Web UI 中：**设置 → Models** 会显示 **Command Code** 卡片；模型选择器会在 **commandcode** 下列出实时目录（撰写本文时有 54 个模型）。发送一条消息，选择你套餐中包含的模型——默认的 `deepseek/deepseek-v4-flash` 适用于入门级套餐；开放权重模型（DeepSeek/Qwen/Kimi/MiniMax）通常都可用，而前沿模型（Claude/GPT/Gemini/Grok）可能需要 Pro/Max 套餐或按需计费（见 FAQ）。
@@ -133,21 +162,31 @@ llm-commandcode:
   apiBase: https://api.commandcode.ai
   workingDir: /path/to/project     # 上报给 API（项目 slug、配置块）
   modelsCachePath: ~/.commandcode/models-cache.json
+  requestTimeoutMs: 60000          # 等待首个响应字节的最长时间（默认 60s）
+  streamIdleTimeoutMs: 120000      # 流停顿超过该时长即视为死连接（默认 120s）
 ```
 
 组合入口配置（`cordis.patch.yml` / 你 profile 的 `cordis.patch.yml`）接受相同的键；那里的字面量 `apiKey` 优先于凭据引用。
 
 ## 故障排查
 
+- **`Command Code API request to .../alpha/generate failed`，且每次会话都在重试（看到"重试延迟"）** ——这是**传输层失败**：`fetch()` 根本没拿到 HTTP 响应（不是 401/403/429，那些会显示 "API error"）。dsh 的重试策略会重试 `TRANSPORT` 两次（指数退避），所以 UI 里会先出现重试行，最终才失败。0.1.8 起失败原因会显示**真实根因**（例如 `fetch failed: connect ECONNREFUSED`、`ENOTFOUND`、`CERT_HAS_EXPIRED`、`The operation was aborted due to timeout`）。常见原因：
+  - **你的网络需要代理**。Node 的 `fetch`（undici）**不读取** `HTTP_PROXY`/`HTTPS_PROXY` 环境变量，所以浏览器/curl 走系统代理能通，而 dsh 进程直连失败。需要给 dsh 配置 undici 代理（例如 `NODE_OPTIONS=--import undici` + dispatcher，或网络层路由），或把 `api.commandcode.ai` 加入白名单。
+  - **连接被中途重置/限速**（防火墙、GFW 类干扰、Wi-Fi 不稳）。错误消息会点名（`socket hang up`、`ECONNRESET`、`UND_ERR_SOCKET`）。
+  - **TLS 被中间人替换**（企业 MITM）——错误链里出现 `CERT_HAS_EXPIRED`/`DEPTH_ZERO_SELF_SIGNED_CERT`。
+  - 也可能是瞬时抖动，重试能恢复；如果每轮都失败，那是环境问题而非 API 问题（健康网络下 models 端点和 generate 端点都能正常响应）。
+- **长回答生成到一半中断** ——0.1.8 起，adapter 会在 `requestTimeoutMs`（默认 60s）内拿不到响应时中止请求，并在流停顿超过 `streamIdleTimeoutMs`（默认 120s）时判定为死连接，不再无限挂起。这两种失败都以 `TIMEOUT` 呈现并附带停顿时长；如果你的网络慢但稳定，可以在 `llm-commandcode` 设置里调大这两个值。
+- **Web 应用启动即崩溃，报 `ERR_MODULE_NOT_FOUND: Cannot find package 'dsh-commandcode-provider'`** ——patch 行的 `name` 写成了裸包名，但 loader 会把它当作模块从 profile 的 `node_modules` 导入，而 pnpm 只会链接带 scope 的名字 `@mars-sea/dsh-commandcode-provider`。0.1.7 之前的 bundle 自带的就是这行错误配置；手工复制旧的 `cordis.patch.yml` 示例（或缓存层）也会踩中。修复方法：把你 profile 的 `cordis.patch.yml` 中该行改成 `name: "@mars-sea/dsh-commandcode-provider"` ——注意**必须加引号**：不引号的 `@` 开头标量会导致 YAML 解析失败（0.1.7 就带了这处回归，0.1.8 已加引号）——然后重启。
 - **`MODEL_NOT_IN_PLAN` (403)** ——所选模型不在你的 Command Code 套餐内。选择一个开放权重模型（如 `deepseek/deepseek-v4-flash`）或升级套餐。错误信息会指明模型并附官方文档链接。
 - **`MISSING_CREDENTIAL`** ——任何地方都没有 key。通过 Models 页面卡片存储一个、`export COMMANDCODE_API_KEY`、设置 `config.apiKey`，或运行 `command-code login`。没有 key 时路由保持注册、目录保持可浏览。
 - **Models 页面卡片显示"未配置"但请求可用** ——key 来自 `~/.commandcode/auth.json`（`cmd login` 兜底），而不是 dsh 凭据存储。把它粘贴到卡片一次即可让卡片显示为已配置；两者可以共存。
 - **推理模型在短请求下不返回可见文本** ——推理模型（如 `deepseek/deepseek-v4-*`）会先消耗输出 token 进行推理；`maxTokens` 较小时可能在出现可见文本前就用完。这属于正常现象。
-- **git 安装时 `dsh plugin add` 报 `allowBuilds` 错误** ——把 pnpm 打印的确切包 key（含 commit hash）复制到 `pnpm-workspace.yaml` 并重新运行（见[从 GitHub 安装（推荐）](#从-github-安装推荐)）。
+- **git 安装时 `dsh plugin add` 报 `allowBuilds` 错误** ——把 pnpm 打印的确切包 key（含 commit hash）复制到 `pnpm-workspace.yaml` 并重新运行（见[从 GitHub 安装](#从-github-安装)）。
 
 ## 注意事项与限制
 
-- **目前仅支持文本**：图片输入会抛出 `UNSUPPORTED_CONTENT`（接入附件服务以解析图片字节是后续工作）。有意不声明 pi 插件的 `MODEL_INPUT_MODALITIES` 表。
+- **图片输入按模型能力限制**：只有官方 Command Code 注册表标记为 Vision 的模型接受图片（见 `src/adapter.ts` 中的 `KNOWN_IMAGE_MODELS` 快照，与[官方模型注册表](https://commandcode.ai/docs/reference/cli/models)同步）。模型选择器会为每个 Command Code 模型标注 *"Supports image input"* / *"Text only"*，切换前即可看出能力。向纯文本模型发送图片会抛出 `UNSUPPORTED_CONTENT`。官方 CLI 对纯文本模型会回退到客户端 *VISION* 副调用转文字；本适配器**不**复现该交互功能——请改用支持 Vision 的模型。图片输入还需要 dsh 的**附件服务**（`ctx.attachments`）；缺失时携带图片的请求会抛出 `UNSUPPORTED_CONTENT`。
+- **在含图片的会话里切换到纯文本模型会被 dsh 自身拒绝**——这是 harness 层的守卫（`dsh-host-apiproxy` 的 `selectModel` 处理器）：当会话历史或待处理输入已包含图片、而目标模型未声明 `image` 输入时，会返回 `model-unavailable`。该拒绝是刻意设计，无法从插件侧放宽（适配器提供的模型行正是让守卫生效的输入——纯文本模型如实上报 `inputModalities: ['text']`）。本 bundle **能**做的是让提示更友好：它的客户端插件会包装 `session.selectModel`，把这条拒绝改写为「当前会话已包含图片，而模型 `<model>` 不支持图片输入；请选择支持图片的模型，或先移除会话中的图片。」（错误码与 details 原样透传，按 `error.code` 分支的调用方不受影响）。要继续使用图片，请选择选择器中标注 *"Supports image input"* 的模型，或先清空会话中的图片；也可安装图片路由 bundle（如 `@deepseek-ai/dsh-llm-image-routing`）把图片轮透明路由到视觉回退模型。
 - **不支持 `stop` 序列**：线上格式没有 stop 字段；携带它的请求会抛出 `UNSUPPORTED_OPTION`。
 - 推理块**不会**重放到后续轮次（与官方 CLI 一致：先前的私有推理不得泄漏）。
 - 只有带配对工具结果的工具调用会被重放到对话中。
@@ -173,7 +212,7 @@ llm-commandcode:
 - **完全卸载**：
 
   ```sh
-  dsh plugin --profile web remove dsh-commandcode-provider
+  dsh plugin --profile web remove @mars-sea/dsh-commandcode-provider
   ```
 
   这会移除 bundle 依赖及其配置层。你在 dsh 凭据库和 `~/.commandcode/auth.json` 中的 API key 不会被改动（如需撤销访问权限，可手动删除）。

package/cordis.patch.yml

@@ -1,8 +1,15 @@
 # dsh-commandcode-provider bundle layer.
 # One plugin row: registers the `commandcode` provider route and its
 # Models-page card. All fields are optional — schema defaults apply.
+#
+# `name` must be the FULL package specifier as installed in the profile's
+# node_modules: the loader imports it as a module and pnpm links packages by
+# their true (scoped) name, so a bare `dsh-commandcode-provider` here fails to
+# resolve (ERR_MODULE_NOT_FOUND) and crashes the app on boot. Quote the value:
+# a YAML scalar starting with `@` is parsed as a directive/indicator and fails
+# to parse.
 - insert:
     - id: llm-commandcode
-      name: dsh-commandcode-provider
+      name: "@mars-sea/dsh-commandcode-provider"
       config:
         apiKeyEnv: COMMANDCODE_API_KEY

package/lib/client.js

@@ -0,0 +1,57 @@
+window.__ModuleLoader__.load({
+	id: "@mars-sea/dsh-commandcode-provider",
+	factory: (require) => {
+		var module = { exports: {} };
+		var exports = module.exports;
+		Object.defineProperty(exports, Symbol.toStringTag, { value: "Module" });
+		//#region src/client/index.ts
+		/** Whether a selectModel rejection is the harness's image-session gate. */
+		function isImageSessionRejection(result) {
+			return !result.result.ok && result.result.error.code === "model-unavailable" && result.result.error.message.includes("does not accept image input");
+		}
+		/** Wrap the shared sessions API so selectModel failures read friendlier. */
+		function withFriendlyImageError(sessions) {
+			const selectModel = sessions.selectModel.bind(sessions);
+			return {
+				...sessions,
+				selectModel: async (payload, signal) => {
+					const result = await selectModel(payload, signal);
+					if (!isImageSessionRejection(result)) return result;
+					const model = result.result.error.details?.model ?? payload.model;
+					return {
+						...result,
+						result: {
+							...result.result,
+							error: {
+								...result.result.error,
+								message: `当前会话已包含图片，而模型 ${model} 不支持图片输入；请选择支持图片的模型，或先移除会话中的图片。`
+							}
+						}
+					};
+				}
+			};
+		}
+		/**
+		* Client plugin body: install the selectModel wrapper on the connection's
+		* shared api. `inject: ['connection']` gates activation until the connection
+		* service is provided (the same pattern the harness's own client plugins
+		* use), and `connection.api.sessions` is a stable object the model-selection
+		* UI reads fresh on every call — so wrapping it once covers both the /model
+		* popup and the composer seat, across reconnects.
+		*/
+		function apply(ctx) {
+			const connection = ctx.get("connection");
+			if (connection === void 0) return;
+			connection.api.sessions = withFriendlyImageError(connection.api.sessions);
+		}
+		const inject = ["connection"];
+		//#endregion
+		exports.apply = apply;
+		exports.inject = inject;
+		exports.isImageSessionRejection = isImageSessionRejection;
+		exports.withFriendlyImageError = withFriendlyImageError;
+		return module.exports;
+	}
+});
+
+//# sourceMappingURL=client.js.map

package/lib/client.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"client.js","names":[],"sources":["../src/client/index.ts"],"sourcesContent":["/**\n * Browser half of the dsh-commandcode-provider bundle.\n *\n * The host rejects switching to a text-only model while the session already\n * contains images with a harness-level `model-unavailable` error\n * (`dsh-host-apiproxy`'s `session.selectModel` handler). That rejection is\n * intentional and cannot be relaxed from the plugin side — the adapter's\n * `inputModalities` is exactly what makes the guard work. What we CAN do is\n * make the error message friendlier: this client plugin wraps the shared\n * `connection.api.sessions.selectModel` face so a `model-unavailable`\n * rejection shows a clear, actionable hint (with the requested model name)\n * instead of the raw English harness message.\n *\n * The wrapper is deliberately narrow: only the `model-unavailable` code is\n * rewritten, only when the message matches the image-session gate, and only\n * the message text changes — the error code and details pass through\n * untouched so any caller that switches on `error.code` keeps working.\n *\n * The wire types are spelled structurally here (not imported from\n * `@deepseek-ai/dsh-host-apiproxy`) so this client bundle does not drag an\n * extra peer dependency into the package; the shapes are stable and the\n * client build inlines them anyway.\n */\n\nimport type { Context } from '@deepseek-ai/cordis'\n\n/** The `model-unavailable` error details: provider + model id. */\ninterface ModelUnavailableDetails {\n  provider: string\n  model: string\n}\n\n/** The narrow slice of the RPC error we need to inspect and rewrite. */\ninterface RpcErrorLike {\n  code: string\n  message: string\n  details?: ModelUnavailableDetails\n}\n\n/**\n * The narrow slice of a unary RPC result we need to inspect and rewrite.\n * The wire shape from `sessions.selectModel` (via `AbstractApiClient.callUnary`)\n * is the full envelope `{ rpcId, result: { ok, error? } }` — the error lives\n * under `result.result`, not at the top level. `RpcResultLike` models that.\n */\ninterface RpcResultLike {\n  rpcId: string\n  result:\n    | { ok: true; value?: unknown }\n    | { ok: false; error: RpcErrorLike }\n}\n\n/** One selectModel call: payload in, envelope out. */\ntype SelectModelCall = (\n  payload: { sessionId: string; provider: string; model: string; reasoningEffort?: string },\n  signal?: AbortSignal,\n) => Promise<RpcResultLike>\n\n/** The shared sessions wire face we wrap. */\ninterface SessionsLike {\n  selectModel: SelectModelCall\n}\n\n/** The connection handle shape we read `api.sessions` from. */\ninterface ConnectionLike {\n  api: { sessions: SessionsLike }\n}\n\n/** Whether a selectModel rejection is the harness's image-session gate. */\nexport function isImageSessionRejection(\n  result: RpcResultLike,\n): result is RpcResultLike & { result: { ok: false; error: RpcErrorLike } } {\n  return (\n    !result.result.ok &&\n    result.result.error.code === 'model-unavailable' &&\n    result.result.error.message.includes('does not accept image input')\n  )\n}\n\n/** Wrap the shared sessions API so selectModel failures read friendlier. */\nexport function withFriendlyImageError(sessions: SessionsLike): SessionsLike {\n  const selectModel = sessions.selectModel.bind(sessions)\n  return {\n    ...sessions,\n    selectModel: async (payload, signal) => {\n      const result = await selectModel(payload, signal)\n      if (!isImageSessionRejection(result)) return result\n      const model = result.result.error.details?.model ?? payload.model\n      return {\n        ...result,\n        result: {\n          ...result.result,\n          error: {\n            ...result.result.error,\n            message:\n              `当前会话已包含图片，而模型 ${model} 不支持图片输入；`\n              + '请选择支持图片的模型，或先移除会话中的图片。',\n          },\n        },\n      }\n    },\n  }\n}\n\n/**\n * Client plugin body: install the selectModel wrapper on the connection's\n * shared api. `inject: ['connection']` gates activation until the connection\n * service is provided (the same pattern the harness's own client plugins\n * use), and `connection.api.sessions` is a stable object the model-selection\n * UI reads fresh on every call — so wrapping it once covers both the /model\n * popup and the composer seat, across reconnects.\n */\nexport function apply(ctx: Context): void {\n  const connection = ctx.get('connection') as ConnectionLike | undefined\n  if (connection === undefined) return\n  connection.api.sessions = withFriendlyImageError(connection.api.sessions)\n}\n\nexport const inject: readonly string[] = ['connection']\n"],"mappings":";;;;;;;;EAqEA,SAAgB,wBACd,QAC0E;GAC1E,OACE,CAAC,OAAO,OAAO,MACf,OAAO,OAAO,MAAM,SAAS,uBAC7B,OAAO,OAAO,MAAM,QAAQ,SAAS,6BAA6B;EAEtE;;EAGA,SAAgB,uBAAuB,UAAsC;GAC3E,MAAM,cAAc,SAAS,YAAY,KAAK,QAAQ;GACtD,OAAO;IACL,GAAG;IACH,aAAa,OAAO,SAAS,WAAW;KACtC,MAAM,SAAS,MAAM,YAAY,SAAS,MAAM;KAChD,IAAI,CAAC,wBAAwB,MAAM,GAAG,OAAO;KAC7C,MAAM,QAAQ,OAAO,OAAO,MAAM,SAAS,SAAS,QAAQ;KAC5D,OAAO;MACL,GAAG;MACH,QAAQ;OACN,GAAG,OAAO;OACV,OAAO;QACL,GAAG,OAAO,OAAO;QACjB,SACE,iBAAiB,MAAM;OAE3B;MACF;KACF;IACF;GACF;EACF;;;;;;;;;EAUA,SAAgB,MAAM,KAAoB;GACxC,MAAM,aAAa,IAAI,IAAI,YAAY;GACvC,IAAI,eAAe,KAAA,GAAW;GAC9B,WAAW,IAAI,WAAW,uBAAuB,WAAW,IAAI,QAAQ;EAC1E;EAEA,MAAa,SAA4B,CAAC,YAAY"}

package/lib/index.d.ts

@@ -2,13 +2,34 @@ import z from "@deepseek-ai/schemastery";
 import { GenerateOptions, LlmAdapter, LlmModelInfo, LlmResolvedModelInfo, ResolvedRetryPolicy, StreamChunk } from "@deepseek-ai/dsh-llm";
 import { CredentialRef } from "@deepseek-ai/dsh-credentials";
 import { Context } from "@deepseek-ai/cordis";
+import { AttachmentStore } from "@deepseek-ai/dsh-attachment";
 import { CommandDefinition } from "@deepseek-ai/dsh-commands";
 //#region src/adapter.d.ts
 declare const KNOWN_EFFORTS: Readonly<Record<string, readonly string[]>>;
+/**
+ * Models whose Capabilities include Vision, per the official Command Code
+ * model registry (`https://commandcode.ai/docs/reference/cli/models`, generated
+ * from the same registry as `cmd --list-models` / the `/model` picker).
+ *
+ * The Provider API does not expose modality metadata, so this snapshot is the
+ * source of truth for image-input gating. Command Code's own CLI falls back to
+ * a client-side VISION side-call for text-only models; this adapter does not
+ * reproduce that interactive feature, so images sent to a model outside this
+ * list are refused loudly (`UNSUPPORTED_CONTENT`) instead of being dropped or
+ * sent to a model that cannot read them.
+ *
+ * Keep in sync with the official registry when new models ship (see the
+ * dsh-commandcode-upstream skill).
+ */
+declare const KNOWN_IMAGE_MODELS: ReadonlySet<string>;
 declare const COMMAND_CODE_CLI_VERSION = "1.26.0";
 declare const DEFAULT_API_BASE = "https://api.commandcode.ai";
 declare const DEFAULT_GENERATE_MAX_TOKENS = 64000;
 declare const DEFAULT_MAX_OUTPUT_TOKENS = 65536;
+/** Head-of-request timeout: how long to wait for the first response byte. */
+declare const DEFAULT_REQUEST_TIMEOUT_MS = 60000;
+/** Stream idle timeout: a generation that stalls this long is a dead connection. */
+declare const DEFAULT_STREAM_IDLE_TIMEOUT_MS = 120000;
 declare function projectSlugFromPath(pathName: string): string;
 /** Read a usable Command Code credential from the official CLI auth file. */
 declare function resolveAuthFileApiKey(): string | undefined;
@@ -20,7 +41,17 @@ interface CommandCodeConnectionOptions {
   workingDir: string;
   /** Model catalog cache path. */
   modelsCachePath: string;
+  /** Milliseconds to wait for the generate response's first byte (default 60s). */
+  requestTimeoutMs: number;
+  /** Milliseconds a stream may stall before it is treated as a dead connection (default 120s). */
+  streamIdleTimeoutMs: number;
 }
+/**
+ * Resolve the durable attachment service, or undefined when the host does not
+ * provide one. Called lazily only when a request actually carries images, so a
+ * text-only request never depends on the attachment seam.
+ */
+type ResolveAttachments = () => AttachmentStore | undefined;
 /** Everything the adapter needs beyond the request itself. */
 interface CommandCodeAdapterDeps<C extends CommandCodeConnectionOptions = CommandCodeConnectionOptions> {
   /** Resolve the current connection facts (fresh per request, settings-aware). */
@@ -29,6 +60,8 @@ interface CommandCodeAdapterDeps<C extends CommandCodeConnectionOptions = Comman
   resolveApiKey: (connection: C) => Promise<string>;
   /** HTTP transport override (tests); defaults to the global `fetch`. */
   fetchImpl?: typeof fetch;
+  /** Resolve the optional durable attachment service for image input (tests); defaults to none. */
+  resolveAttachments?: ResolveAttachments;
 }
 /** Account identity from `/alpha/whoami`. */
 interface CommandCodeAccount {
@@ -80,6 +113,7 @@ declare class CommandCodeAdapter<C extends CommandCodeConnectionOptions = Comman
   private readonly deps;
   private catalog;
   private readonly fetchImpl;
+  private readonly resolveAttachments;
   constructor(deps: CommandCodeAdapterDeps<C>);
   /**
    * Command Code is a metered subscription API: 429 (rate limit) and 5xx
@@ -141,6 +175,10 @@ interface Config {
   workingDir?: string;
   /** Model catalog cache path; defaults to `~/.commandcode/models-cache.json`. */
   modelsCachePath?: string;
+  /** Milliseconds to wait for the generate response's first byte; defaults to 60s. */
+  requestTimeoutMs?: number;
+  /** Milliseconds a stream may stall before being treated as a dead connection; defaults to 120s. */
+  streamIdleTimeoutMs?: number;
 }
 declare const Config: z<Config>;
 /** One resolution's complete request facts: connection plus credential reference. */
@@ -156,5 +194,5 @@ interface ResolvedCommandCodeOptions extends CommandCodeConnectionOptions {
 declare function resolveAdapterOptions(config: Config): ResolvedCommandCodeOptions;
 declare function apply(ctx: Context, config: Config): void;
 //#endregion
-export { COMMAND_CODE_CLI_VERSION, CommandCodeAdapter, type CommandCodeAdapterDeps, type CommandCodeCommandDeps, type CommandCodeConnectionOptions, type CommandCodeUsageReport, Config, DEFAULT_API_BASE, DEFAULT_GENERATE_MAX_TOKENS, DEFAULT_MAX_OUTPUT_TOKENS, DEFAULT_MODELS_CACHE_PATH, KNOWN_EFFORTS, PROVIDER, ResolvedCommandCodeOptions, apply, applyCommands, commandDefinition, inject, name, projectSlugFromPath, resolveAdapterOptions, resolveAuthFileApiKey };
+export { COMMAND_CODE_CLI_VERSION, CommandCodeAdapter, type CommandCodeAdapterDeps, type CommandCodeCommandDeps, type CommandCodeConnectionOptions, type CommandCodeUsageReport, Config, DEFAULT_API_BASE, DEFAULT_GENERATE_MAX_TOKENS, DEFAULT_MAX_OUTPUT_TOKENS, DEFAULT_MODELS_CACHE_PATH, DEFAULT_REQUEST_TIMEOUT_MS, DEFAULT_STREAM_IDLE_TIMEOUT_MS, KNOWN_EFFORTS, KNOWN_IMAGE_MODELS, PROVIDER, type ResolveAttachments, ResolvedCommandCodeOptions, apply, applyCommands, commandDefinition, inject, name, projectSlugFromPath, resolveAdapterOptions, resolveAuthFileApiKey };
 //# sourceMappingURL=index.d.ts.map