@cjhyy/code-shell 0.7.1 → 0.8.0

package/CHANGELOG.md

@@ -8,6 +8,8 @@ breaking.
 
 ## [Unreleased]
 
+## [0.8.0] - 2026-08-05
+
 ### Added
 
 #### Unified input attachments
@@ -59,13 +61,16 @@ breaking.
 
 #### Browser automation + built-in browser panel
 
-- CDP-driven automation of the built-in browser (no Playwright/JS-injection):
-  semantic `browser_observe` / `browser_act` / `browser_navigate` tools, image
-  & vision observation, multi-tab support, and link/image URL extraction. The
-  whole capability is a single toggle — turning it off removes both the tools
-  and their prompt text. Stale snapshots are folded to save tokens.
-- The agent can open the browser panel itself; selections echo back into tool
-  cards; screenshots echo to the stream as clickable thumbnails.
+- Semantic `browser_observe` / `browser_act` / `browser_navigate` tools now use
+  a task-owned background target in the built-in browser profile by default,
+  sharing that profile's login state without silently taking over a tab the
+  user opened. The exact same target can be revealed for login or human
+  takeover.
+- User-opened built-in tabs and regular Chrome tabs require an explicit claim;
+  dedicated Playwright remains an explicit isolated option for scheduled or
+  unattended work. Image/vision observation, multi-tab control, URL extraction,
+  trace visibility, and stale-snapshot folding are available across the runtime
+  control plane.
 
 #### Credentials: cookie login, multi-account, inject
 
@@ -123,6 +128,26 @@ breaking.
 
 ### Fixed
 
+- Link Actions now preserve the dedicated `link` credential-access purpose
+  through the desktop bridge, so provider-owned credentials remain
+  non-agent-exposable while still being usable by the approved Link tool.
+- Browser device authorization now encrypts the access token, refresh token,
+  token and refresh expiries, token endpoint, and public client ID together.
+  Only the access token crosses the Link execution boundary; expired or invalid
+  connections are hidden from Link Actions and surfaced in the UI for a clean
+  reconnect.
+- Link connection cards isolate state per provider and connection method,
+  cancel pending device flows on close/unmount, clear abandoned local secrets,
+  derive gateway totals from the live channel list, and leave browser-login
+  status retryable after a transient status-check failure.
+- Clean-checkout CI now builds workspace exports before package-boundary tests,
+  resolves the agent stdio source subpath during typechecking, recognizes the
+  packaged Chrome extension globals, and gives coverage tests the same timeout
+  budget as the full suite.
+- Desktop packaging now derives its workspace build order from one tested list,
+  including the Web, Server, and Chat packages introduced by the monorepo
+  split. npm publication also waits for all three native installer jobs, so a
+  failed desktop package cannot leave a half-published release.
 - Addressed the latest core/desktop architecture review findings around stream
   routing, permission/session scoping, and event coalescing, reducing
   wrong-session routing, duplicated events, and hard-to-cancel operations in
@@ -174,11 +199,66 @@ breaking.
 
 ### Changed
 
+- The core installer, marketplace, onboarding, and updater implementation
+  exports have moved to the `/internal` subpath. Embedders importing these
+  implementation details must update their import path for 0.8.0.
+- The short-lived personal todo UI has been replaced by Mimi follow-ups. Existing
+  `userData/pet/todos.json` files are intentionally left untouched and are not
+  auto-imported; users of that preview should back up or copy any remaining text
+  manually.
 - Refreshed beta-release documentation and repository hygiene: README badges and
   status copy now reflect the public beta track, obsolete handoff notes are
   archived under `docs/archive/`, and generated build artifacts are covered by
   `.gitignore`.
 
+## [0.7.1] - 2026-07-11
+
+Hardening release. No breaking changes from 0.7.0 — 71 of the 119 commits are
+fixes, concentrated in usage accounting, tool failure semantics, permission
+boundaries and session lifecycle.
+
+### Added
+
+- **Full-context session forks** — fork a session and keep the entire context
+  instead of restarting from a summary.
+- **MCP HTTP OAuth** — the login and token-refresh flow is now complete, so
+  OAuth-protected HTTP MCP servers stay authenticated across restarts.
+- Send images from the mobile remote.
+- Jump from a `DriveAgent` job straight to its CLI session.
+- Browser panel address-bar actions.
+
+### Fixed
+
+- **Unified usage accounting** — billing and visibility are now split, and
+  aux / Arena / sub-agent / `maxTurns` usage is recovered instead of being
+  dropped. Image token estimation was also corrected.
+- **Tool failure and abort semantics** are enforced consistently: a failed tool
+  reports as failed, and an abort propagates instead of being swallowed.
+- **Permission modes defer to run boundaries**, so a mode change mid-run can no
+  longer apply to a tool call that was already in flight.
+- **Session close is serialized** with close-only generation fencing, removing a
+  race between closing a session and its in-flight work.
+- Hook stdin is protected against `EPIPE`, and transcript flush failures are
+  surfaced rather than silently lost.
+- Approvals are bounded and plugin denials propagate correctly.
+- Background media jobs are hardened: timed-out video URLs are retained, stream
+  fetches are capped, and MCP image names are unique.
+- Duplicate mobile uploads are serialized; review requests are scoped by
+  workspace.
+
+## [0.7.0] - 2026-07-10
+
+### Added
+
+- **Quick chat dock panel** — a lightweight always-available chat surface in the
+  desktop app.
+- **Live CC room transcripts** stream into the desktop client.
+- **Goal judge reads run evidence** (tool results, progress, previous gaps) on
+  the primary model, so goal evaluation is grounded in what the run actually did.
+- `DriveAgent` job inspection and cancellation.
+- `DriveAgent` accepts an optional per-job model override for claude/codex.
+- OAuth MCP credentials.
+
 ## [0.6.0-rc.14]
 
 ### Fixed
@@ -221,7 +301,7 @@ plus one private Electron POC:
   render layer, the `code-shell` CLI bin, themes, key bindings, all
   terminal-only utilities. Imports core; nothing else imports it.
 - **`@cjhyy/code-shell`** (meta) — thin re-export shim. `import { Engine }
-  from "@cjhyy/code-shell"` still works for SDK users; the `code-shell`
+from "@cjhyy/code-shell"` still works for SDK users; the `code-shell`
   bin loads the TUI CLI. Ships three files total: `index.js`, `index.d.ts`,
   `cli.js`.
 - **`@cjhyy/code-shell-desktop`** (private) — Electron POC. Currently a
@@ -282,7 +362,7 @@ plugin `hooks.json` files.
   user's prompt text on `user_prompt_submit`). Registry aggregation is
   last-write-wins for input/prompt, newline-joined for additionalContext.
 - **`settings.json` shell hooks.** New `hooks: [{event, command,
-  matcher?, timeout_ms?, cwd?}]` schema. Engine spawns the command per
+matcher?, timeout_ms?, cwd?}]` schema. Engine spawns the command per
   emit with the CC wire protocol — stdin is the HookContext JSON,
   stdout is a HookResult JSON, exit 2 = deny with stderr surfaced to
   the model. `CODESHELL_HOOK_EVENT` / `CODESHELL_HOOK_CWD` env vars
@@ -301,7 +381,7 @@ plugin `hooks.json` files.
   `${CODESHELL_PLUGIN_ROOT}` across every text file, so CC-authored
   plugin scripts that branch on host detection pick the codeshell
   path. `InstallResult.varRewrite` reports counts and `/plugin
-  install` prints a `rewrote: N file(s)` line so the modification is
+install` prints a `rewrote: N file(s)` line so the modification is
   visible.
 - **`ToolContext.hooks`.** Tools that need to emit lifecycle events
   (currently `notification` on background sub-agent terminal states)
@@ -365,6 +445,7 @@ see git log between v0.1.6 and v0.2.0 for the detailed change set.)
 ## [0.1.6] - 2026-05-13
 
 ### Added
+
 - **Built-in skills bundled with the package.** The skill scanner now also
   reads `skills-builtin/` shipped inside the installed npm package. First
   built-in is `codeshell-help` — answers questions about code-shell itself
@@ -377,6 +458,7 @@ see git log between v0.1.6 and v0.2.0 for the detailed change set.)
   block, since unattended automations have no human to retry the task.
 
 ### Fixed
+
 - `code-shell run` now resolves API keys / baseUrl / model from
   `settings.providers[]` and `settings.models[]` in addition to the legacy
   `settings.model.*` mirror — matches what Engine reconciles at startup,
@@ -395,8 +477,9 @@ see git log between v0.1.6 and v0.2.0 for the detailed change set.)
 ## [0.1.5] - 2026-05-13
 
 ### Added
+
 - **Auto-installing updater.** Background update check probes `npm config get
-  prefix` for write permission; if writable, registers a detached `npm i -g`
+prefix` for write permission; if writable, registers a detached `npm i -g`
   on process exit so the next launch picks up the new version. File lock at
   `~/.code-shell/.update.lock` (5min stale-takeover) prevents concurrent
   installs from corrupting each other. Disable with `DISABLE_AUTOUPDATER=1`,
@@ -421,7 +504,7 @@ see git log between v0.1.6 and v0.2.0 for the detailed change set.)
   reminder; a third re-read is hard-blocked. Four+ consecutive read-only
   calls (or three+ silent turns) inject a strategy-change reminder.
 - **Per-session verbose recorder.** Dev-only JSONL trace under `log/<date>/
-  session-<sid>.jsonl` capturing every LLM request/response, tool call, and
+session-<sid>.jsonl` capturing every LLM request/response, tool call, and
   engine event. Gated on `CODE_SHELL_DEV=1` / `--debug` / running from src.
   7-day retention, argv secrets redacted, per-record output clipped to 256 KB.
 - **Pasted-noise sanitizer.** Engine rejects tasks that are >70% ANSI/box
@@ -434,6 +517,7 @@ see git log between v0.1.6 and v0.2.0 for the detailed change set.)
   progress instead of a stale snapshot from the last completed run.
 
 ### Changed (breaking)
+
 - `SessionStatus` is now `"active" | "paused" | TerminalReason` instead of
   `"active" | "paused" | "completed" | "errored"`. `state.json` records the
   raw terminal reason (e.g. `aborted_streaming`, `model_error`,
@@ -442,6 +526,7 @@ see git log between v0.1.6 and v0.2.0 for the detailed change set.)
   need to widen its match.
 
 ### Changed
+
 - API key resolution centralized into `resolveApiKey()` (one canonical
   fallback chain: option → settings → all provider env vars). Replaces five
   drifting copies in `repl.ts`, `run.ts`, `runs.ts`, `main.ts`,
@@ -454,6 +539,7 @@ see git log between v0.1.6 and v0.2.0 for the detailed change set.)
 ## [0.1.4] - 2026-05-09
 
 ### Fixed
+
 - **Multi-model `apiKey` ignored on startup.** When `settings.json` defined
   a model in `models[]` (with its own `apiKey`/`baseUrl`) and the top-level
   `model.apiKey` was empty, the runtime fell through to
@@ -472,6 +558,7 @@ see git log between v0.1.6 and v0.2.0 for the detailed change set.)
 ## [0.1.0-alpha.1] - 2026-04-30
 
 ### Changed (breaking)
+
 - **Tool timeout system rewritten.** Removed the hardcoded
   `LEGACY_LONG_TIMEOUT_TOOLS = {Agent, Arena}` whitelist in
   `tool-system/registry.ts`. Tools now declare their own timeout via
@@ -491,6 +578,7 @@ see git log between v0.1.6 and v0.2.0 for the detailed change set.)
   (e.g. extended thinking).
 
 ### Added
+
 - **`Agent(run_in_background: true)`** — fire-and-forget sub-agents. Returns
   an `agent_id` immediately instead of blocking the parent turn. The agent
   runs detached in the same process; restarting loses its state.
@@ -502,6 +590,7 @@ see git log between v0.1.6 and v0.2.0 for the detailed change set.)
   registry for async agent handles.
 
 ### Notes
+
 - Cross-process / restart-survivable long tasks still belong to `RunManager`
   in `@cjhyy/code-shell/run`, not to `Agent(run_in_background)`. The split
   mirrors Claude Code's REPL/Agent-tool/Routines architecture.
@@ -509,6 +598,7 @@ see git log between v0.1.6 and v0.2.0 for the detailed change set.)
 ## [0.1.0-alpha.0] - 2026-04-28
 
 ### Added
+
 - **`IterativeArena` — multi-model authoring loop.** Pipeline: tournament v1
   (every participant writes a draft, the author merges anonymized drafts into
   v1) → critique-revise rounds (parallel critics produce anchored critiques
@@ -527,6 +617,7 @@ see git log between v0.1.6 and v0.2.0 for the detailed change set.)
 - `node >= 20` engine constraint alongside the existing `bun >= 1.3` hint.
 
 ### Fixed
+
 - TypeScript compilation is now clean (was 50+ errors).
   - Removed zombie `cli/transports/`, `cli/handlers/`, `cli/remoteIO.ts`,
     `cli/structuredIO.ts`, `cli/ndjsonSafeStringify.ts` — these were copied
@@ -571,11 +662,13 @@ see git log between v0.1.6 and v0.2.0 for the detailed change set.)
   Azure backend that rejects tool calls with 401.
 
 ### Removed
+
 - `TodoWrite` tool. It was a 6-line wrapper around `TaskCreate` and its
   presence confused models into using two equivalent interfaces. Use the
   `Task*` family (`TaskCreate`, `TaskUpdate`, `TaskList`, `TaskGet`, etc.).
 
 ### Changed
+
 - Arena research-phase prompts no longer hard-cap output at "exactly 3 to 6
   findings". The new instruction asks for "as many findings as the topic
   warrants — typically 5-15", with each `summary` field expected to be 80+
@@ -593,6 +686,7 @@ see git log between v0.1.6 and v0.2.0 for the detailed change set.)
   marked `[需调研]` / `[TBD: ...]`.
 
 ### Fixed
+
 - IterativeArena: `extractTag` now tolerates LLM responses where the
   closing `</v1_content>` / `</v_next_content>` marker is missing (e.g. due
   to max_tokens truncation). Previously the parser fell back to "use the

package/README.md

@@ -30,7 +30,7 @@ CodeShell is one orchestration engine wearing three faces:
 - an **Electron desktop app** with chat, file/browser/terminal/diff panels, model & credential management, an extensions marketplace, automation, and a phone remote, and
 - a **programmatic SDK** (`import { Engine } from "@cjhyy/code-shell"`) for embedding the engine in your own product.
 
-The core is deliberately **domain-agnostic**. The turn loop, context management, permissions, MCP integration, hooks, tasks, cron, sub-agents, sessions, and memory all stay generic; coding behavior is just a *preset* layered on top — not baked into the engine. (See `packages/core/CONTRIBUTING.md`: "core only carries mechanism, not policy.")
+The core is deliberately **domain-agnostic**. The turn loop, context management, permissions, MCP integration, hooks, tasks, cron, sub-agents, sessions, and memory all stay generic. Coding tools, git/worktree behavior, LSP, review, quota, and coding prompts live in the physically separate `@cjhyy/code-shell-capability-coding` package and are composed by the CLI/Desktop hosts. (See `packages/core/CONTRIBUTING.md`: "core only carries mechanism, not policy.")
 
 > Status: **0.6.x, entering beta**. The desktop app is the headline product; the CLI and SDK share the same core engine.
 
@@ -99,10 +99,11 @@ It gives you chat with streaming output, a side-by-side file / browser / termina
 
 ### Presets
 
-| Preset | Purpose | Extra tools |
-|--------|---------|-------------|
-| `general` | General orchestration, research, automation, long-running work | Core orchestration tools only |
-| `terminal-coding` | Terminal-native coding assistant | `EnterWorktree`, `ExitWorktree`, `NotebookEdit`, `LSP`, `Brief`, `Arena` |
+| Preset            | Purpose                                                        | Extra tools                                                              |
+| ----------------- | -------------------------------------------------------------- | ------------------------------------------------------------------------ |
+| `harness-min`     | Domain-neutral core default for embedding                      | Minimal filesystem, shell, MCP, memory, task and agent mechanisms        |
+| `general`         | General orchestration, research, automation, long-running work | Core orchestration tools only                                            |
+| `terminal-coding` | Terminal-native coding assistant                               | `EnterWorktree`, `ExitWorktree`, `NotebookEdit`, `LSP`, `Brief`, `Arena` |
 
 Presets select the system prompt, the built-in tool set, and permission defaults. Configure via the SDK, the CLI `--preset` flag, or settings.
 
@@ -127,8 +128,9 @@ Presets select the system prompt, the built-in tool set, and permission defaults
 
 ### Built-in tools
 
-The default `general` preset whitelists 48 built-ins. The CLI defaults to
-`terminal-coding`, which adds coding extras. Runtime guards may hide tools that
+Core by itself defaults to the minimal `harness-min` preset. The CLI and Desktop
+compose the coding capability package, whose host default is `terminal-coding`.
+Runtime guards may hide tools that
 need unavailable providers, credentials, cookies, or an active goal.
 
 - **File / workspace**: `Read`, `Write`, `Edit`, `ApplyPatch`, `Glob`, `Grep`
@@ -169,6 +171,14 @@ const codingEngine = new Engine({
 
 Everything is exported from the package root — `import { ... } from "@cjhyy/code-shell"` (or directly from `@cjhyy/code-shell-core`). There are no `/run`, `/arena`, or `/product` subpath entry points.
 
+Runnable SDK examples live in [`examples/`](examples/) — each runs directly
+with `bun run <file>` and supports `--dry-run` (scripted mock LLM) when no API
+key is configured:
+
+- [`examples/01-minimal-agent.ts`](examples/01-minimal-agent.ts) — one Engine, one run, streamed output
+- [`examples/02-approval-flow.ts`](examples/02-approval-flow.ts) — a custom `ApprovalBackend` gating tool calls
+- [`examples/03-in-process-transport.ts`](examples/03-in-process-transport.ts) — the recommended `createServer` / `createClient` factory pair
+
 ---
 
 ## Configuration
@@ -216,6 +226,7 @@ The OpenAI-compatible provider aborts any LLM stream idle for `CODESHELL_STREAM_
 At a high level, CodeShell routes CLI, headless, SDK, and desktop clients through the same engine runtime:
 
 - **Preset resolution** selects the system prompt, built-in tools, and permission defaults.
+- **Capability composition** lets products install their own tools, presets, prompt sections, file-history behavior, and session-workspace adapters without putting product policy in core.
 - **TurnLoop** coordinates model streaming, context assembly, tool execution, and lifecycle events.
 - **Tool system** hosts built-ins, MCP tools, permissions, hooks, and cancellation.
 - **Session / run layers** persist transcripts, state, tasks, automation runs, and memories.
@@ -225,7 +236,7 @@ In the desktop app, the Electron main process acts as an IPC service layer: it d
 Design principles:
 
 - **Core first** — the orchestration engine stays domain-agnostic.
-- **Presets over hardcoding** — coding behavior lives in configuration.
+- **Capabilities over hardcoding** — coding behavior lives in a separate package composed at the host boundary.
 - **Secure by default** — permission-gated actions and explicit approval flow; owner-only credential files.
 - **Long-running ready** — tasks, cron, sleep, sub-agents, and persistent goals are first-class.
 
@@ -235,10 +246,19 @@ Design principles:
 
 ```text
 packages/
-├── core/      # Engine, context, tools, MCP, hooks, sessions, runs, presets, memory
+├── link/      # Framework-independent Link provider manifests and authorization guides
+├── core/      # Domain-agnostic engine, context, MCP, hooks, sessions, runs, memory
+├── coding/    # Coding capability pack: tools, git/worktrees, LSP, review, prompt/presets
+├── arena/     # Optional multi-model Arena capability
+├── pet/       # Mimi behavior, DelegateWork, projection protocol, digital-human teams
+├── server/    # Headless HTTP/WebSocket host, rooms, uploads, mobile remote
+├── web/       # Browser-safe remote client state and SPA
 ├── tui/       # Terminal CLI, Ink-based UI, renderer, commands, approvals
 ├── desktop/   # Electron desktop client + agent worker bridge + mobile remote app
-└── cdp/       # Environment-agnostic CDP browser-action layer (no Playwright)
+├── cdp/       # Environment-agnostic CDP browser-action layer (no Playwright)
+└── chat/      # Standalone multi-channel chat gateway + optional CodeShell adapter
+
+package.json   # @cjhyy/code-shell compatibility meta package
 
 assets/       # README / product images (mascot, promo hero, Playwright desktop screenshots)
 
@@ -256,8 +276,8 @@ scripts/      # Build, release, and repo maintenance scripts
 
 ```bash
 bun install
-bun run build          # build core + tui + meta package
-bun run typecheck      # root core + tui check; currently not a clean gate
+bun run build          # build link + core + capability/host packages + meta package
+bun run typecheck      # root core + coding capability + tui + chat check
 bun test               # core / tui test suites
 
 # Desktop has its OWN typecheck and build (the root checks do NOT cover it):
@@ -266,7 +286,7 @@ bun run typecheck
 bun run build
 ```
 
-> Current caveat: `bun run typecheck` at the repo root reports a pre-existing test typing error in `packages/core/src/tool-system/builtin/drive-claude-code.test.ts:158`. Use it for signal, but do not treat it as a clean gate until that is fixed.
+> Current caveat: `bun run typecheck` at the repository root still reports pre-existing test-source typing diagnostics. The package builds and Desktop typecheck above are the clean gates for this change.
 
 `bun run dev` launches the desktop app. For the TUI in dev: `bun run dev:tui`.
 
@@ -276,6 +296,14 @@ bun run build
 
 ## Further reading
 
+- [Panel Apps v2 — project-bound UI, tools, and Skills](docs/panel-apps.md)
+- [Panel App starter template](https://github.com/cjhyy/codeshell-panel-apps/tree/main/templates/starter)
+- [Design Studio reference Panel App](https://github.com/cjhyy/codeshell-panel-apps/tree/main/apps/design-studio)
+- [Quant Lab reference Panel App](https://github.com/cjhyy/codeshell-panel-apps/tree/main/apps/quant-lab)
+- [Video editor reference plugin](examples/plugins/video-editor/README.md)
+- [Package boundaries & Pet split rationale](docs/architecture/12-package-boundaries-and-release-units.md)
+- [Plugin parity matrix](docs/architecture/13-plugin-parity-and-video-editor.md)
+- [Digital humans & Pet architecture](docs/architecture/14-digital-human-and-pet.md)
 - [Architecture & feature inventory](docs/architecture/README.md)
 - [Roadmap & TODO](docs/todo/README.md)
 - [Prior architecture documentation set (archived, pending rewrite)](docs/archive/architecture/README.md)
@@ -284,7 +312,7 @@ bun run build
 
 ## Acknowledgments
 
-The `ApplyPatch` tool (`packages/core/src/tool-system/builtin/apply-patch/`) is adapted from
+The `ApplyPatch` tool (`packages/coding/src/tools/apply-patch/`) is adapted from
 [OpenAI Codex `codex-rs/apply-patch`](https://github.com/openai/codex/tree/main/codex-rs/apply-patch),
 licensed under the Apache License 2.0. See `NOTICE.md` and `LICENSE-codex` in that directory
 for details, including the intentional behavioral divergence where our applier rolls back

package/README.zh-CN.md

@@ -30,7 +30,7 @@ CodeShell 是同一个编排引擎的三种形态：
 - **Electron 桌面应用**：提供聊天、文件/浏览器/终端/diff 面板、模型与凭证管理、扩展市场、自动化和手机远程控制；
 - **程序化 SDK** (`import { Engine } from "@cjhyy/code-shell"`)：可把引擎嵌入你自己的产品。
 
-核心引擎保持**领域无关**。turn loop、上下文管理、权限、MCP 集成、hooks、tasks、cron、sub-agents、sessions 和 memory 都是通用机制；“写代码助手”只是叠在引擎上的一个 preset，而不是写死在核心里。
+核心引擎保持**领域无关**。turn loop、上下文管理、权限、MCP 集成、hooks、tasks、cron、sub-agents、sessions 和 memory 都是通用机制。编码工具、git/worktree、LSP、review、quota 与编码 prompt 位于物理独立的 `@cjhyy/code-shell-capability-coding` 包，由 CLI/Desktop host 组合，不写进 core。
 
 > 状态：**0.6.x，进入 beta 阶段**。桌面应用是当前主产品；CLI 和 SDK 共享同一个核心引擎。
 
@@ -99,10 +99,11 @@ bun run dev          # 以开发模式启动桌面应用
 
 ### Presets
 
-| Preset | 用途 | 额外工具 |
-|--------|------|----------|
-| `general` | 通用编排、研究、自动化、长期任务 | 仅核心编排工具 |
-| `terminal-coding` | 终端原生编码助手 | `EnterWorktree`, `ExitWorktree`, `NotebookEdit`, `LSP`, `Brief`, `Arena` |
+| Preset            | 用途                               | 额外工具                                                                 |
+| ----------------- | ---------------------------------- | ------------------------------------------------------------------------ |
+| `harness-min`     | 供其他产品嵌入的领域无关 core 默认 | 最小文件、shell、MCP、memory、task 与 agent 机制                         |
+| `general`         | 通用编排、研究、自动化、长期任务   | 仅核心编排工具                                                           |
+| `terminal-coding` | 终端原生编码助手                   | `EnterWorktree`, `ExitWorktree`, `NotebookEdit`, `LSP`, `Brief`, `Arena` |
 
 Preset 决定 system prompt、内置工具集和权限默认值。可通过 SDK、CLI `--preset` 参数或 settings 配置。
 
@@ -127,8 +128,8 @@ Preset 决定 system prompt、内置工具集和权限默认值。可通过 SDK
 
 ### 内置工具
 
-默认 `general` preset 白名单包含 48 个内置工具。CLI 默认使用
-`terminal-coding` preset，并额外加入编码相关工具。运行时 guard 可能隐藏当前不可用的
+单独使用 core 时默认是最小的 `harness-min` preset。CLI 与 Desktop 会组合 coding
+capability 包，该宿主的默认 preset 是 `terminal-coding`。运行时 guard 可能隐藏当前不可用的
 provider、凭证、cookie 或 goal 相关工具。
 
 - **File / workspace**：`Read`, `Write`, `Edit`, `ApplyPatch`, `Glob`, `Grep`
@@ -214,6 +215,7 @@ OpenAI-compatible provider 会中止在 `CODESHELL_STREAM_IDLE_TIMEOUT_MS` 毫
 高层来看，CodeShell 让 CLI、headless、SDK 和桌面客户端都走同一个 engine runtime：
 
 - **Preset resolution** 选择 system prompt、内置工具和权限默认值；
+- **Capability composition** 允许产品在 core 之外注入工具、presets、prompt sections、文件历史行为与 session workspace adapter；
 - **TurnLoop** 协调模型流式输出、上下文组装、工具执行和生命周期事件；
 - **Tool system** 承载内置工具、MCP tools、permissions、hooks 和 cancellation；
 - **Session / run layers** 持久化 transcripts、state、tasks、automation runs 和 memories。
@@ -223,7 +225,7 @@ OpenAI-compatible provider 会中止在 `CODESHELL_STREAM_IDLE_TIMEOUT_MS` 毫
 设计原则：
 
 - **Core first**：编排引擎保持领域无关；
-- **Presets over hardcoding**：编码行为存在配置里；
+- **Capabilities over hardcoding**：编码行为放在独立 package 中，在 host 边界组合；
 - **Secure by default**：高影响动作默认经过权限门禁；
 - **Long-running ready**：tasks、cron、sleep、sub-agents 和 persistent goals 都是一等能力。
 
@@ -233,10 +235,19 @@ OpenAI-compatible provider 会中止在 `CODESHELL_STREAM_IDLE_TIMEOUT_MS` 毫
 
 ```text
 packages/
-├── core/      # Engine, context, tools, MCP, hooks, sessions, runs, presets, memory
-├── tui/       # Terminal CLI, Ink-based UI, renderer, commands, approvals
-├── desktop/   # Electron desktop client + agent worker bridge + mobile remote app
-└── cdp/       # Environment-agnostic CDP browser-action layer (no Playwright)
+├── link/      # 与框架无关的 Link provider 清单和授权指南
+├── core/      # 领域无关的 Engine、context、MCP、hooks、sessions、runs、memory
+├── coding/    # 编码 capability：tools、git/worktrees、LSP、review、prompt/presets
+├── arena/     # 可选的多模型 Arena capability
+├── pet/       # Mimi 行为、DelegateWork、投影协议、数字人团队
+├── server/    # Headless HTTP/WebSocket host、rooms、uploads、mobile remote
+├── web/       # 浏览器安全的远程客户端状态与 SPA
+├── tui/       # Terminal CLI、Ink UI、renderer、commands、approvals
+├── desktop/   # Electron desktop client、agent worker bridge、mobile remote app
+├── cdp/       # Environment-agnostic CDP browser-action layer (no Playwright)
+└── chat/      # 独立多渠道聊天网关与可选 CodeShell adapter
+
+package.json   # @cjhyy/code-shell 兼容 meta package
 
 assets/       # README / product images (mascot, promo hero, Playwright desktop screenshots)
 
@@ -254,8 +265,8 @@ scripts/      # Build, release, and repo maintenance scripts
 
 ```bash
 bun install
-bun run build          # build core + tui + meta package
-bun run typecheck      # root core + tui check; currently not a clean gate
+bun run build          # build link + core + capability/host packages + meta package
+bun run typecheck      # root core + coding capability + tui + chat check
 bun test               # core / tui test suites
 
 # Desktop has its OWN typecheck and build (root checks do NOT cover it):
@@ -264,7 +275,7 @@ bun run typecheck
 bun run build
 ```
 
-> 当前注意：repo root 的 `bun run typecheck` 会因 `packages/core/src/tool-system/builtin/drive-claude-code.test.ts:158` 的既有测试类型错误失败。它仍可提供信号，但在该错误修复前不要把它当成干净 gate。
+> 当前注意：repo root 的 `bun run typecheck` 仍会报告既有的测试源码类型诊断。本次改动以各 package build 与上面的 Desktop typecheck 作为干净 gate。
 
 `bun run dev` 会启动桌面应用。TUI 开发模式：`bun run dev:tui`。
 
@@ -275,6 +286,11 @@ bun run build
 ## 延伸阅读
 
 - [Architecture & feature inventory](docs/architecture/README.md)
+- [Plugin panels v1](docs/plugin-panels.md)
+- [视频剪辑参考插件](examples/plugins/video-editor/README.md)
+- [包边界与 Pet 拆分理由](docs/architecture/12-package-boundaries-and-release-units.md)
+- [插件能力差距矩阵](docs/architecture/13-plugin-parity-and-video-editor.md)
+- [数字人与 Pet 架构](docs/architecture/14-digital-human-and-pet.md)
 - [Roadmap & TODO](docs/todo/README.md)
 - [Prior architecture documentation set (archived, pending rewrite)](docs/archive/architecture/README.md)
 
@@ -282,7 +298,7 @@ bun run build
 
 ## 致谢
 
-`ApplyPatch` tool（`packages/core/src/tool-system/builtin/apply-patch/`）改编自 [OpenAI Codex `codex-rs/apply-patch`](https://github.com/openai/codex/tree/main/codex-rs/apply-patch)，使用 Apache License 2.0。详见该目录下的 `NOTICE.md` 和 `LICENSE-codex`。
+`ApplyPatch` tool（`packages/coding/src/tools/apply-patch/`）改编自 [OpenAI Codex `codex-rs/apply-patch`](https://github.com/openai/codex/tree/main/codex-rs/apply-patch)，使用 Apache License 2.0。详见该目录下的 `NOTICE.md` 和 `LICENSE-codex`。
 
 ## License
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@cjhyy/code-shell",
-  "version": "0.7.1",
+  "version": "0.8.0",
   "description": "Code Shell — meta package. Installs @cjhyy/code-shell-core and @cjhyy/code-shell-tui.",
   "type": "module",
   "main": "./dist/index.js",
@@ -16,22 +16,30 @@
   },
   "scripts": {
     "preinstall": "node scripts/check-node.cjs",
-    "build": "bun run --filter '@cjhyy/code-shell-core' build && bun run --filter '@cjhyy/code-shell-tui' build && bun run scripts/build-meta.ts",
+    "build": "bun run --filter '@cjhyy/code-shell-link' build && bun run --filter '@cjhyy/code-shell-core' build && bun run --filter '@cjhyy/code-shell-pet' build && bun run --filter '@cjhyy/code-shell-arena' build && bun run --filter '@cjhyy/code-shell-capability-coding' build && bun run --filter '@cjhyy/code-shell-cdp' build && bun run --filter '@cjhyy/code-shell-server' build && bun run --filter '@cjhyy/code-shell-web' build && bun run --filter '@cjhyy/code-shell-tui' build && bun run --filter '@cjhyy/code-shell-chat' build && bun run scripts/build-meta.ts",
     "dev": "bun run dev:desktop",
     "dev:desktop": "bun run --filter '@cjhyy/code-shell-desktop' dev",
     "dev:tui": "CODE_SHELL_DEV=1 CODESHELL_UI_PERF=1 bun run packages/tui/src/cli/main.ts",
     "dev:bigtranscript": "bun run scripts/render-bigtranscript-dev.ts",
-    "test": "bun test",
-    "test:watch": "bun test --watch",
+    "test": "bun test --timeout 30000",
+    "test:coverage": "bun test --timeout 30000 --config ./bunfig.coverage.toml packages/core/src/tool-system/builtin packages/core/src/tool-system/testing --coverage --coverage-reporter=text",
+    "test:e2e": "bun run build && bun run --cwd packages/desktop build && bun run --cwd packages/desktop test:e2e",
+    "smoke": "bun run build && bun run --cwd packages/desktop build && bun run --cwd packages/desktop smoke",
+    "test:core-exports": "bun run --filter '@cjhyy/code-shell-core' build && node scripts/smoke-core-exports.mjs",
+    "test:core-harness": "bun run --filter '@cjhyy/code-shell-core' build && node scripts/smoke-core-harness.mjs",
+    "test:package-release": "bun run scripts/package-release-smoke.ts",
+    "test:watch": "bun test --timeout 30000 --watch",
     "bench:render": "bun run bench/render-tail.bench.ts && bun run bench/render-streaming.bench.ts && bun run bench/render-spinner.bench.ts && bun run bench/render-wheel.bench.ts",
-    "typecheck": "tsc --noEmit",
+    "typecheck": "bun run --filter '@cjhyy/code-shell-chat' typecheck && tsc --noEmit",
     "lint": "eslint packages/",
     "lint:engine-bypass": "bash scripts/check-no-engine-bypass.sh",
+    "lint:workflow-test-paths": "node scripts/check-workflow-test-paths.mjs",
+    "lint:baseline": "node scripts/check-lint-baseline.mjs",
     "format": "prettier --write 'packages/**/*.ts'"
   },
   "dependencies": {
-    "@cjhyy/code-shell-core": "0.7.1",
-    "@cjhyy/code-shell-tui": "0.7.1"
+    "@cjhyy/code-shell-core": "0.8.0",
+    "@cjhyy/code-shell-tui": "0.8.0"
   },
   "workspaces": [
     "packages/*"