feat(session): preserve Claude turns and improve live session controls

.agents/skills/cc-remote-deploy/SKILL.md

@@ -30,6 +30,13 @@ and use the repository's immutable activation transactions. Lost connectivity
 means an unknown result: inspect the original operation before retrying. Do not
 overwrite live directories or restart the controlling Wrapper from itself.
 
+For Claude deployments, also read
+[persistent session installation and acceptance](../../../docs/claude-session-service.md).
+An independent SDK service must survive ordinary Wrapper upgrades. Check first
+migration, remaining in-process `/btw` turns, and deferred queues separately;
+wait for them to drain instead of interrupting work. Do not restart or replace
+an active SDK service to satisfy a version/readiness check.
+
 ## Codex CLI sharing is an acceptance check
 
 For every enabled Codex **Code** account, follow

.github/workflows/ci.yml

@@ -1,6 +1,7 @@
 name: CI
 
 on:
+  workflow_dispatch:
   workflow_call:
   pull_request:
   push:
@@ -40,82 +41,29 @@ jobs:
         with:
           name: ci-web-dist
           path: web/dist
-      - name: Test and lint
-        run: |
-          .venv/bin/python -m pytest
-          uvx --from ruff==0.15.13 ruff check cc_remote tests deploy
+      - name: Run pytest
+        run: .venv/bin/python -m pytest
 
   web:
-    name: Web
+    name: Web build
     runs-on: ubuntu-24.04
     permissions:
       contents: read
     steps:
       - uses: actions/checkout@v6
-      - uses: actions/setup-python@v6
-        with:
-          python-version: "3.13"
-      - name: Install uv for the model-free Viewer fixture
-        uses: astral-sh/setup-uv@08807647e7069bb48b6ef5acd8ec9567f424441b # v8.1.0
-        with:
-          version: "0.11.16"
-      - name: Install Viewer fixture runtime
-        run: |
-          uv venv --python 3.13 .venv
-          uv pip sync \
-            --python .venv/bin/python \
-            --require-hashes --only-binary=:all: --no-binary=http-ece \
-            requirements.lock
       - uses: actions/setup-node@v6
         with:
           node-version-file: ".nvmrc"
           cache: npm
           cache-dependency-path: web/package-lock.json
       - name: Install dependencies
-        run: |
-          npm --prefix web ci
-          npm --prefix web exec -- playwright install --with-deps chromium webkit
-      - name: Build, test, and lint
-        run: |
-          npm --prefix web run build
-          npm --prefix web run test:reliability
-          npm --prefix web run test:history-browser
-          npm --prefix web run test:viewer
-          npm --prefix web run lint
-      - name: Preserve Playwright failure evidence
-        if: failure()
-        uses: actions/upload-artifact@v4
-        with:
-          name: playwright-failure-${{ github.run_attempt }}
-          path: web/test-results
-          if-no-files-found: warn
-          retention-days: 3
+        run: npm --prefix web ci
+      - name: Build Web client
+        run: npm --prefix web run build
       - name: Preserve built Web client
         uses: actions/upload-artifact@v4
         with:
           name: ci-web-dist
           path: web/dist
           if-no-files-found: error
           retention-days: 1
-
-  deploy:
-    name: Deploy scripts
-    runs-on: ubuntu-24.04
-    permissions:
-      contents: read
-    steps:
-      - uses: actions/checkout@v6
-      - name: Validate scripts
-        run: |
-          bash -n \
-            deploy/install.sh \
-            deploy/install-relay.sh \
-            deploy/install-wrapper.sh \
-            deploy/setup-vps.sh
-          shellcheck -x \
-            deploy/install.sh \
-            deploy/install-relay.sh \
-            deploy/install-wrapper.sh \
-            deploy/setup-vps.sh \
-            deploy/setup_transaction.sh
-          git diff --check

AGENTS.md

@@ -41,6 +41,13 @@ and keep a decline or pending answer separate from deployment success. App
 attachment and optional App-control MCP tools are separate user choices.
 
 ## Critical constraints / traps
+- **Claude service lifetime**: when `CC_REMOTE_CLAUDE_SERVICE_SOCKET` is set,
+  regular Claude Code/Work SDK processes belong to the separate local service.
+  Wrapper shutdown detaches; explicit stop/reconnect/eviction still has its
+  deliberate native lifecycle. Never resubmit an accepted query during recovery,
+  treat a background Result as the human terminal, or restart the service during
+  an ordinary Wrapper deploy. See `docs/claude-session-service.md` for first
+  migration, remaining in-process tasks, queue drain and version boundaries.
 - **Drain footgun**: after `ClaudeSDKClient.interrupt()`, the SDK does NOT kill
   the session — the current turn's stream still emits a terminal
   `ResultMessage(subtype="error_during_execution")`. You MUST keep consuming
@@ -50,6 +57,12 @@ attachment and optional App-control MCP tools are separate user choices.
   to the terminal ResultMessage; state only returns to `idle` (and the next
   query is only accepted) after that break. Reject-while-busy prevents a second
   query racing the drain.
+- **Claude steering**: send `priority="next"` through streaming input, keeping
+  the one session reader. Rebind the visible turn only on the exact native user
+  UUID echo. A Result before an accepted input is consumed is intermediate;
+  the persistent service journals this distinction and commits the original
+  root turn identity. Explicit Stop uses `interrupt(cancel_queued=true)` when
+  advertised and pending inputs exist, then drains the real Result as above.
 - **cwd must match resume**: a session's jsonl lives at
   `~/.claude/projects/<cwd-with-/-as->/<uuid>.jsonl`. `ClaudeAgentOptions.cwd`
   MUST equal the original session's cwd or `resume` can't find it.
@@ -58,7 +71,7 @@ attachment and optional App-control MCP tools are separate user choices.
   interrupt+drain verification after any upgrade (`SdkHandle.preflight()` guards
   the exact verified patch at startup).
 - **Claude Code is the user's daily CLI, not the SDK bundle**: Claude Code
-  `>=2.1.258` is required and checked before a Claude session starts. The wrapper
+  `>=2.1.263` is required and checked before a Claude session starts. The wrapper
   defaults `CLAUDE_BIN` to `~/.local/bin/claude` and passes that path explicitly
   to the SDK. An empty value keeps this default; only another absolute path may
   override it. Keep that CLI updated and signed in before starting the wrapper.
@@ -93,7 +106,7 @@ attachment and optional App-control MCP tools are separate user choices.
   transport, never the caller's Origin. Uvicorn trusts forwarded transport
   metadata only from loopback Caddy. Never put tokens in URLs or protocol
   message bodies; logging redacts token/password fields.
-- **Protocol version gate**: current wire protocol v67 is declared by
+- **Protocol version gate**: current wire protocol v71 is declared by
   `PROTOCOL_VERSION` in both `protocol.py` and `web/src/protocol.ts`.
   `deserialize` hard-rejects a version mismatch, and
   `_Base` is `extra="forbid"`, so ANY protocol change must be deployed to all
@@ -219,11 +232,15 @@ attachment and optional App-control MCP tools are separate user choices.
   committing, verify the stored message and scope with
   `git log -1 --format=raw --stat`, then recheck
   `git status --short --branch`.
-- Before opening or updating **every** PR, run the complete local gate below;
-  a docs-only or apparently narrow change does not skip it unless the user
-  explicitly accepts that exception. Every command must exit zero. Expected
-  platform-defined test skips are allowed, but failures or missing tools must
-  be reported rather than silently bypassed.
+- Before opening or updating a maintainer-authored PR (including work prepared
+  by an agent for the maintainer), run the complete local gate below. A docs-only
+  or apparently narrow change does not skip it unless the user explicitly
+  accepts that exception. Every command must exit zero. Expected platform-defined
+  test skips are allowed; report failures or missing tools rather than bypassing
+  them. During development, use checks appropriate to the change.
+- Other contributors may open or update a PR without running the complete local
+  gate. Include the checks performed and any known validation gaps in the PR
+  description. Automatic CI still builds Web and runs pytest for every PR.
 - Run the Web gate with Node 24 (see `.nvmrc`), matching CI. Newer Node
   browser-like globals must not mask missing browser-environment guards.
 
@@ -232,8 +249,6 @@ attachment and optional App-control MCP tools are separate user choices.
 uvx --from ruff==0.15.13 ruff check cc_remote tests deploy
 npm --prefix web run build
 npm --prefix web run test:reliability
-npm --prefix web run test:history-browser
-npm --prefix web run test:viewer
 npm --prefix web run lint
 bash -n \
   deploy/install.sh \
@@ -249,9 +264,15 @@ shellcheck -x \
 git diff --check
 ```
 
-- `.github/workflows/ci.yml` repeats this gate for pushes and PRs. A local pass
-  is required before PR publication and does not replace green remote CI before
-  merge. These checks are zero-token; do not substitute a live model probe.
+- `.github/workflows/ci.yml` automatically runs only the Web build and pytest
+  on PRs and pushes to `master`. Release tags reuse the same CI before packaging
+  and publishing. Pytest waits only for the Web build artifact. Lint, front-end
+  reliability tests and shell checks remain part of the local gate above.
+- Playwright is not part of CI or the required local PR gate. Existing browser
+  tests remain available for explicitly requested diagnostics. When the
+  maintainer asks for PR acceptance, check out the requested revision, run the
+  application and verify the changed behavior; report the actual checks and
+  any remaining gaps. The automated checks above do not call a live model.
 
 ## Run / test
 ```bash
@@ -260,7 +281,6 @@ python -m cc_remote.relay        # terminal 1 (set WEB_STATIC_DIR=web/dist to se
 python -m cc_remote.wrapper      # terminal 2 (on each machine running Claude/Codex)
 pytest                           # zero-token unit tests
 npm --prefix web run test:reliability
-npm --prefix web run test:viewer
 npm --prefix web run lint
 npm --prefix web run build
 ```

CHANGELOG.md

@@ -4,6 +4,41 @@
 
 ## Unreleased
 
+- Combine terminal queue controls and session recovery under protocol v71.
+  Upgrade Relay, Wrapper, Web and TUI together. TUI recovery replaces replayed
+  text without duplicating existing output.
+
+- Add an optional independent Claude session service: Wrapper reconnects recover
+  accepted turns and pending questions without resubmitting prompts. Include a
+  separate service installer and preserve account isolation.
+- Support native Claude steering during active work, retaining message, attachment
+  and token ownership. Reserve private BTW identities before launch, refresh live
+  context without overwriting newer readings, and retry busy session-list reads.
+- Replace replayed text prefixes and bound history summaries. Protocol v70
+  requires Wrapper, Relay and Web to be upgraded together.
+
+- Preview `.mmd` and `.mermaid` Work artifacts as diagrams, with a source view.
+
+- Keep Claude's internal recovery prompts hidden after compaction and rebuild
+  stale projections. Show autonomous replies after agent completion in separate
+  live process sections, with their own thinking, tools and working indicator.
+- Stream subagent text without waiting for whole messages. Recover agent-detail
+  read errors and timeouts, refresh active source-backed agents, and show each
+  agent's own running state.
+- Use a subdued text shimmer for active process and tool summaries. Present
+  readable command/search inputs and structured tool output, with raw data
+  available on demand.
+
+- Show native input/output token counts beside the working spark, with exact
+  counts and cache usage on click or tap. Use subdued counters that show
+  the latest readings directly. Keep per-turn ownership and
+  replace-only recovery snapshots across reconnects (protocol v68).
+
+
+- Request readable Claude thinking summaries in SDK sessions and show returned
+  text in the existing thinking timeline. Preserve native thinking mode,
+  token budget and effort across new sessions, resume and private forks.
+
 - Accept provider-native Claude model ids (e.g. `glm-5.2` behind a custom
   `ANTHROPIC_BASE_URL` gateway) as explicit model selections: they are handed
   to Claude Code, persisted in the private session store, and restored across
@@ -17,6 +52,23 @@
   TUI editing, cancellation and ordering controls. Reject stale queue snapshots
   and changes while a message is starting. Deploy Relay, Web, Wrapper and TUI
   together; protocol v66 clients cannot connect to protocol v67 services.
+- Replace Claude's persistent background-task panel with a compact composer
+  indicator. Open details above the trigger on desktop or in a centered mobile
+  dialog; native task completion removes finished items and closes the indicator
+  when no tasks remain, independently of the parent reply's completion.
+- Track Claude compaction through native status and boundary events, preserve
+  valid context readings during refresh failures, and show the applied automatic
+  compaction threshold separately from model capacity. Hide internal command
+  messages and keep completed-turn timestamps stable when rebuilding history.
+  Require Claude Code 2.1.263 or newer for the verified control APIs.
+- Keep account-specific Claude controls attached to the correct session and
+  omit transient Codex thread notifications that have no readable history from
+  the session list.
+- Show explicit timed-message tags, a rounded moving session outline and the
+  next scheduled send time. Tasks use account-scoped native queue receipts
+  independently of Goals (protocol v67). Session menus and task popovers
+  avoid the sidebar footer and the mobile visual viewport.
+
 - Backport shared improvements from the DSH branch without adding a third
   engine (protocol v66): rounded Claude/Codex Goal dialogs with native save
   confirmation and mobile keyboard recovery; directory links open `/open`, and

CHANGELOG_zh.md

@@ -4,6 +4,32 @@
 
 ## 未发布
 
+- 终端队列控制与会话恢复合并使用协议 v71，Relay、Wrapper、Web 和 TUI 需要
+  一起升级。TUI 恢复时替换已有文本，不再重复追加已显示的内容。
+
+- 新增可选的独立 Claude 会话服务：Wrapper 重连可恢复已接受的轮次和待回答问题，
+  不会重发提示词；提供独立服务安装器，并保持账号隔离。
+- 支持 Claude 运行中原生引导，保留消息、附件和 token 的轮次归属；启动前预留私有
+  BTW 会话标识，运行中刷新上下文时保留较新读数，会话列表忙碌时自动重试读取。
+- 重放时替换已有文本前缀，并限制历史摘要块数。协议升至 v70，Wrapper、Relay 和
+  Web 需要同步升级。
+
+- Work Artifacts 支持 `.mmd` / `.mermaid` 图表预览，并可切换查看源码。
+
+- 修复 Claude 压缩历史中内部续写提示泄漏为用户消息的问题，并重建旧缓存。
+  子代理返回后的主代理续写独立显示实时过程、思考、工具和运行提示。
+- 子代理流式文字及时显示，无需等待整段消息完成；详情读取支持错误恢复与超时
+  重试，运行中的原生记录自动刷新，并显示子代理自己的运行状态和小火花。
+- 处理过程和工具摘要使用低调的文字扫光；命令、搜索条件和结构化工具结果按
+  可读形式展示，原始数据可按需展开。
+
+- 在处理中的小火花旁以低调灰色直接显示原生输入/输出 token 计数，点击查看完整数字
+  和缓存用量；绑定当前轮次，重连恢复不重复累计（protocol v68）。
+
+
+- Claude SDK 会话明确请求思考摘要，返回内容显示在现有“思考”区域；新建、恢复
+  和私有分叉均保留原生思考模式、token 预算和强度。
+
 - 支持把提供商原生模型 ID（例如自定义 `ANTHROPIC_BASE_URL` 网关下的
   `glm-5.2`）作为显式模型选择：它会传给 Claude Code、写入私有会话记录，并在
   重连和冷恢复后保留。仅在转录或 `/context` 元数据中“观测到”的模型 ID 依旧
@@ -13,6 +39,18 @@
 - 新增服务端队列原子排序（protocol v67），TUI 编辑、取消及排序操作支持配置
   快捷键。拒绝基于过期队列的排序和消息启动期间的调整。Relay、Web、Wrapper
   与 TUI 必须一起升级；protocol v66 客户端不能连接 protocol v67 服务。
+- Claude 后台任务改为输入区工具栏中的紧凑胶囊；桌面向上展开详情，手机居中显示。
+  按原生任务状态移除已结束的任务，全部结束时自动关闭小窗并隐藏入口；助手回复
+  完成不会提前隐藏仍在运行的后台任务。
+- 按 Claude 原生状态与压缩边界更新压缩动画；上下文刷新失败时保留有效读数，
+  分别显示生效的自动压缩阈值和模型容量。历史中隐藏内部命令消息，避免后续压缩
+  改变已完成回答的时间戳。已验证的控制接口要求 Claude Code 2.1.263 或更高版本。
+- 修复多账号 Claude 会话控制的持久化归属；Codex 临时线程通知尚无可读历史时，
+  不再生成点开后为空的会话列表项。
+- 定时消息增加来源标签；待发送任务的会话卡片显示圆角流动光圈及下次发送时间。
+  按账号和原生队列回执关联，与 Goal 分开（protocol v67）；会话菜单和任务小窗
+  自动避让侧栏底栏及手机可视区域。
+
 - 整合 DSH 分支中的通用改进，保留 Claude/Codex 双引擎（protocol v66）：圆角
   Goal 小窗等待原生保存确认，恢复手机键盘收起后的布局；目录链接接入 `/open`，
   XLSX 可预览已保存的单元格、切换工作表并下载原文件。

CLAUDE.md

@@ -28,6 +28,15 @@ this optional step does not block core deployment. App-control MCP tools require
 a separate choice; sharing alone does not authorize them.
 
 ## Critical constraints / traps
+- **Claude service lifetime**: when `CC_REMOTE_CLAUDE_SERVICE_SOCKET` is set,
+  regular Claude Code/Work SDK processes belong to the separate local service.
+  Wrapper shutdown detaches; never resubmit an accepted query during recovery
+  or restart the service during an ordinary Wrapper deploy. See
+  `docs/claude-session-service.md` for first migration and drain boundaries.
+- **Claude steering**: send `priority="next"` through the sole streaming-input
+  reader and rebind only on the exact native user UUID echo. A Result before
+  an accepted input is consumed is intermediate. Explicit Stop cancels queued
+  inputs when supported, then drains the real terminal Result.
 - **Drain footgun**: after `ClaudeSDKClient.interrupt()`, the SDK does NOT kill
   the session — the current turn's stream still emits a terminal
   `ResultMessage(subtype="error_during_execution")`. You MUST keep consuming
@@ -88,7 +97,7 @@ a separate choice; sharing alone does not authorize them.
   `useLayoutEffect` is deliberately dependency-free — late virtualizer/image
   measurements settle without a React render, and constraining it to its read
   set reintroduces a full-viewport jump on touch release.
-- **Protocol version gate**: current wire protocol v67 is declared by
+- **Protocol version gate**: current wire protocol v71 is declared by
   `PROTOCOL_VERSION` in both `protocol.py` and `web/src/protocol.ts`.
   `deserialize` hard-rejects a version mismatch, and
   `_Base` is `extra="forbid"`, so ANY protocol change must be deployed to all