[webshell]retry session switch while closing - #8864
Merged
Merged
Conversation
Collaborator
Author
macOS Desktop regression report
|
Collaborator
|
✅ Qwen Triage finished — CI landed green on ✅ Qwen Triage 已完成 —— |
Collaborator
|
Thanks for the PR! Template looks good ✓ Problem: observed bug, not theory — the PR carries before/after screenshots and concrete repro steps (switching between two persisted sessions without waiting), and the daemon side confirms the mechanism: Direction: aligned — this is a reliability gap in the Web Shell session lifecycle, part of the #8092 effort. No direction concerns. Size: not applicable — Approach: scope feels right. Retrying the daemon's exact close-in-progress response on the existing reconnect backoff is the minimal client-side fix; the alternatives (daemon queuing loads while closing, or a new wire Risk: no elevated risk signals — neither changed file matches the revert-correlated path patterns. Moving on to code review. 🔍 中文说明感谢贡献！ 模板完整 ✓ 问题：已观测到的 bug，不是理论问题——PR 提供了 before/after 截图和具体复现步骤（在两个已保存会话间快速切换），daemon 侧也印证了机制：会话清理期间 方向：对齐——这是 Web Shell 会话生命周期的可靠性问题，属于 #8092 工作的一部分。无方向性顾虑。 规模：不适用—— 方案：范围合理。用现有的重连退避机制重试 daemon 明确的"关闭进行中"响应，是客户端侧的最小修复；备选方案（daemon 在关闭期间排队 load 请求、或为该 404 新增 wire 风险：无升级风险信号——两个改动文件均未命中与 revert 相关的高风险路径。 进入代码审查 🔍 — Qwen Code · qwen3.8-max Reviewed at |
Collaborator
Code reviewI traced the mechanism end to end before reading the diff, and the fix matches what I would have proposed. The daemon contract is real and the match is correct. The guard is narrow in the right way. The retry only fires for an explicitly requested session load ( The retry loop can't run away or leak. Backoff reuses The test pins the change. It drives the exact path: first No blockers found. One non-blocking observation: since this 404 carries no CI test evidenceUnattended run — I did not build or execute PR code; the evidence below is the PR's own CI, read via the API at the reviewed commit. At fetch time the gating unit suite ( Final CI results for
One row per check name (latest run); skipped checks omitted; failures sort first. / 每个检查名一行（取最新一次运行），省略 skipped，失败项排在最前。 The new unit test substantiates the retry mechanism itself, but what neither static review nor CI can settle is the race under real daemon teardown timing in the Desktop app — the before/after evidence is the author's, who verified on the macOS bundle only (Windows/Linux packages unvalidated per the PR body). Sandboxed verification would settle this: 中文说明代码审查：我先独立推演了机制再看 diff，修复方案与我的独立提议一致。daemon 侧契约真实存在且匹配正确—— 测试证据：无人值守运行——未构建或执行 PR 代码，以上为通过 API 读取的 PR 自身 CI。抓取时 ubuntu 单测套件与 Web-shell Visuals 仍在运行；macOS/Windows 矩阵任务按 — Qwen Code · qwen3.8-max Reviewed at |
Collaborator
|
Confidence: 4/5 — clean, minimal fix that matches my independent proposal; the only reservation is that real-timing race verification rests on the author's macOS run, and CI hasn't landed yet. Stepping back: this is exactly the kind of fix the Web Shell needs right now — a small, scary-looking bug (error toast + dead-end "session does not exist" screen during ordinary rapid switching) with a root cause that is genuinely transient, and a fix that treats it as transient instead of terminal. I proposed the same approach before reading the diff: catch the daemon's exact close-in-progress 404 on the requested-session path and ride it out on the existing reconnect backoff. The PR does precisely that, in 28 production lines, with no drive-by changes and a unit test that fails without the fix. The parts I pressured hardest all check out: the daemon actually emits this response (four throw sites in The 4 rather than 5: the end-to-end race under real teardown timing is substantiated by the author's before/after on macOS only; Windows/Linux packages are unvalidated, and the mocked unit test can't stand in for live timing. The Approval is deferred until CI lands green on 中文说明置信度：4/5 —— 干净、最小的修复，与我的独立提议一致；唯一的保留是真实时序下的竞态验证依赖作者的 macOS 运行结果，且 CI 尚未结束。 整体来看：这正是 Web Shell 当前需要的修复——一个看起来吓人的 bug（快速切换时出现错误提示和"会话不存在"死胡同页面），根因是真实的瞬态状态，修复方案也把它当作瞬态而非终止状态处理。我在看 diff 之前就提出了相同方案：在请求会话路径上捕获 daemon 明确的"关闭进行中" 404，用现有重连退避等待其完成。PR 正是这么做的，28 行生产代码，无夹带改动，且单测在没有该修复时会失败。我重点核查的部分均通过：daemon 确实会发出该响应（ — Qwen Code · qwen3.8-max Reviewed at |
Collaborator
🖼️ web-shell visual previewRendered against a mock daemon (no real backend): the PR base vs this PR head Screenshots · before / afterℹ️ No screenshot changed against the PR base — but this PR edits 1 render-shaping file:
Either the change has no visual effect (logic, plumbing, a state the scenarios never reach), or no scenario renders this UI — in which case the preview cannot see it, and an empty result is a coverage gap rather than a clean bill of health. To make it visible, add a scenario to Full-resolution recordings (.webm) are attached to the workflow run. — Qwen Code · web-shell visuals |
Contributor
Code Coverage Summary
CLI Package - Full Text ReportCore Package - Full Text ReportFor detailed HTML reports, please see the 'coverage-reports-22.x-ubuntu-latest' artifact from the main CI run. |
qwen-code-ci-bot
approved these changes
Aug 10, 2026
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
Collaborator
There was a problem hiding this comment.
LGTM, looks ready to ship — CI landed green after the review. ✅
qwen-code-ci-bot
previously requested changes
Aug 10, 2026
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
Collaborator
There was a problem hiding this comment.
Not explored to full depth (tool budget reached): This PR fixes a Web Shell race: switching back to a sessi...: verification of which sessionScope the shipping web-shell/desktop consumers pass to DaemonSessionProvider (reachability of finding 2).
中文说明
未探索到全部深度（达到工具调用预算）：This PR fixes a Web Shell race: switching back to a sessi...：verification of which sessionScope the shipping web-shell/desktop consumers pass to DaemonSessionProvider (reachability of finding 2)。
— qwen3.8-max via Qwen Code /review (v0.21.8)
Collaborator
Author
|
Final closeout for merged PR head
中文摘要PR 已以旧 head 合并。6 条越界或 mutation-only 建议已说明并关闭；经验证的一行恢复修复已转为 follow-up PR #8883，原 PR 的 review threads 已全部关闭。 |
yiliang114
changed the title
doudouOUC
approved these changes
Aug 10, 2026
doudouOUC
left a comment
doudouOUC
left a comment
Collaborator
There was a problem hiding this comment.
Reviewed exact head efd58dc7200eeadc741a90243d48b14ff5e177e2. No blocking findings: the prior watchdog/connecting Critical and the autoReconnect: false regression are fixed, superseding switches are covered, and the retry remains scoped to explicit action-driven loads as documented. Exact-head verification: 254/254 focused WebUI tests passed, package typecheck passed, changed-file Prettier and git diff --check passed, and the Web Shell E2E smoke passed. The Ubuntu aggregate job failed in unrelated CLI/Core tests during a broad burst of 15-second timeouts and shared temporary-root errors; the changed WebUI suite itself passed. Approving with two non-blocking cleanup/test-hardening suggestions.
yiliang114
enabled auto-merge
August 10, 2026 13:13
wenshao
reviewed
Aug 10, 2026
wenshao
left a comment
wenshao
left a comment
Collaborator
There was a problem hiding this comment.
Reviewed. Suggestions are inline. Not explored to full depth (tool budget reached): This PR retries explicit session switches in the Web Shel...: did not run the full DaemonSessionProvider.test.tsx / actions.test.ts suites (only the targeted tests), and did not inspect the sidebar currentSessionIdent…; This PR retries explicit session switches in the Web Shel...: did not run the three new tests or any vitest run — verdicts are static-analysis based.; This PR retries explicit session switches in the Web Shel...: the discriminator's contract with the real daemon (exact 404 body message and status for a closing session) is not verifiable from this repo alone; both the tes…; This PR retries explicit session switches in the Web Shel...: could not execute the new tests (worktree lacks node_modules ; vitest fails at config load on vite-plugin-dts ).; This PR retries explicit session switches in the Web Shel...: did not enumerate the web-shell (packages/web-shell) UI consumers of connection.status transitions in full — verified the status value 'disconnected' is pre-exi…, and 5 more.
中文说明
已审查。 建议见行内评论。 未探索到全部深度（达到工具调用预算）：This PR retries explicit session switches in the Web Shel...：did not run the full DaemonSessionProvider.test.tsx / actions.test.ts suites (only the targeted tests), and did not inspect the sidebar currentSessionIdent…；This PR retries explicit session switches in the Web Shel...：did not run the three new tests or any vitest run — verdicts are static-analysis based.；This PR retries explicit session switches in the Web Shel...：the discriminator's contract with the real daemon (exact 404 body message and status for a closing session) is not verifiable from this repo alone; both the tes…；This PR retries explicit session switches in the Web Shel...：could not execute the new tests (worktree lacks node_modules ; vitest fails at config load on vite-plugin-dts ).；This PR retries explicit session switches in the Web Shel...：did not enumerate the web-shell (packages/web-shell) UI consumers of connection.status transitions in full — verified the status value 'disconnected' is pre-exi…，另有 5 条。
— DeepSeek/deepseek-v4-flash via Qwen Code /review (v0.21.8)
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
Collaborator
There was a problem hiding this comment.
Reviewed. Suggestions are inline. Not explored to full depth (tool budget reached): Context: PR 8864 (QwenLM/qwen-code), round 2 review of a ...: did not verify whether reconnectAttempt is reset on the loop's successful-connect path after a closing-retry (i.e., whether attempts accumulated during closin…; Context: PR 8864 (QwenLM/qwen-code), round 2 review of a ...: did not trace reloadSession 's exact interaction with the new cleanup identity guard for the live-journal-repair case (repair-triggered pending load present wh…; Context: PR 8864 (QwenLM/qwen-code), round 2 review of a ...: didn't read the full context of the 4th closing site at bridge.ts:6457 (verified the message string is identical via grep, but not the surrounding admission log…; Context: PR 8864 (QwenLM/qwen-code), round 2 review of a ...: didn't verify the daemon-side HTTP error mapping code (the body shape {error, sessionId} is taken from the PR's own tests).; Context: PR 8864 (QwenLM/qwen-code), round 2 review of a ...: can't see the existing PR comments, so can't rule out overlap with issues already discussed., and 5 more.
中文说明
已审查。 建议见行内评论。 未探索到全部深度（达到工具调用预算）：Context: PR 8864 (QwenLM/qwen-code), round 2 review of a ...：did not verify whether reconnectAttempt is reset on the loop's successful-connect path after a closing-retry (i.e., whether attempts accumulated during closin…；Context: PR 8864 (QwenLM/qwen-code), round 2 review of a ...：did not trace reloadSession 's exact interaction with the new cleanup identity guard for the live-journal-repair case (repair-triggered pending load present wh…；Context: PR 8864 (QwenLM/qwen-code), round 2 review of a ...：didn't read the full context of the 4th closing site at bridge.ts:6457 (verified the message string is identical via grep, but not the surrounding admission log…；Context: PR 8864 (QwenLM/qwen-code), round 2 review of a ...：didn't verify the daemon-side HTTP error mapping code (the body shape {error, sessionId} is taken from the PR's own tests).；Context: PR 8864 (QwenLM/qwen-code), round 2 review of a ...：can't see the existing PR comments, so can't rule out overlap with issues already discussed.，另有 5 条。
— qwen3.8-max via Qwen Code /review (v0.21.8)
Collaborator
Maintainer verification: real local stack, A/B against baseI built and ran this PR against a real Revisions: PR head Harness
Disclosed fault injection (one env-gated
|
Base 0d56e50b64 — bug reproduced |
PR efd58dc720 — fixed |
|---|---|
 |
 |
| Error toast + “current session does not exist”, terminal — still there 30 s after the teardown finished | Session A opens with its full transcript; no toast, no missing-session screen |
Reproduced on base in 5 separate runs. (Two further runs did not reproduce: on a daemon that had
accumulated stale client-id registrations from earlier page loads, close-on-last-detach never fires, so
there is no closing window to race at all. Restarting the daemon between arms makes it deterministic.)
Request ledger, base — one attempt, then it gives up:
POST /session/ALPHA/detach -> 204 # switch away
POST /session/BETA/load -> 200
POST /session/BETA/detach -> 204 # switch back
POST /session/ALPHA/load -> 404 {"error":"... The session is closing; retry after close completes"}
# no further attempts, ever
Captured toast text (polled from the DOM so it can't be lost to auto-dismiss):
POST /session/:id/load: No session with id "e44bee82-…". The session is closing; retry after close completes
Request ledger, PR — retries on the existing backoff and wins:
POST /session/ALPHA/load -> 404 +0.00s
POST /session/ALPHA/load -> 404 +0.97s
POST /session/ALPHA/load -> 404 +2.97s
POST /session/ALPHA/load -> 404 +6.97s
POST /session/ALPHA/load -> 200 +15.03s <- teardown finished, session restored
Gaps ≈ 1 s / 2 s / 4 s / 8 s — the real jittered exponential ramp. A second run reproduced the same shape (2/2/5/8 s). Over the whole 20 s window the only user-visible message was the benign “session is loading, please wait before sending” hint.
Critical finding R1-1 (watchdog fires mid-retry) — verified in the real stack
Ran with --session-restore-timeout-ms 5000 (→ 20 s pending-load watchdog) against a 60 s teardown, so the watchdog fires while the retry loop is still going:
| Base — terminal on the first 404 | PR — bounded retries, then settles |
|---|---|
 |
 |
The PR made 5 attempts (+0/2/4/8/15 s) and then stopped at the watchdog — no request storm, and no permanent spinner: the UI lands in an idle, usable state. The actions.ts watchdog change is what makes that deterministic.
One residual, non-blocking: on that path the user gets no explicit message — after 20 s of waiting they just see an empty composer. Strictly better than base (a wrong, permanent error), but a one-line notice would be kinder.
Tests and mutation matrix
DaemonSessionProvider.test.tsx + actions.test.ts: 254/254 pass on head. I then applied 7 single-edit mutants to the PR's production code:
| Mutant | Result |
|---|---|
| M1 remove the whole closing-retry branch | KILLED |
M2 delete reconnectAttempt += 1; |
|
M3 delete the await delay(...) |
KILLED (2 tests) |
M4 delete if (pendingSessionLoadRef.current !== pendingLoad) return; |
|
M5 drop the autoReconnect && gate |
KILLED |
M6 revert the actions.ts watchdog disconnected-settle |
KILLED |
M7 revert the effectPendingSessionLoad cleanup identity guard |
KILLED |
The two survivors, with measurements rather than assertions:
- M2 — the backoff ramp is still untested. A stepping probe measured the exact delay before each retry: PR =
[10, 20, 40, 80] ms, mutant =[10, 10, 10, 10] ms. The live run above shows the ramp is real production behaviour (1→2→4→8 s), so this is exactly the regression R1-5 described: the new test'sadvanceTimersByTimeAsync(20)also covers a 10 ms delay, so it cannot separate the two. Asserting the elapsed gap (or using clearly separated delay settings) would kill it. - M4 — the superseded-switch guard has no test that reaches it, even with the new supersession case. In that scenario the old runner's
abort.signalalready rejects thedelay, so control never gets to the guard; the line is defence-in-depth rather than the thing under test. Either note that in a comment or add a case that actually reaches it.
Neither survivor is a defect in the shipped behaviour — both are coverage gaps around code that demonstrably works.
Verdict
The fix does what the description claims, in a real stack, driven by the real teardown and the real daemon 404. No regressions observed in either scenario, and the watchdog race flagged as Critical in round 1 settles deterministically. Good to merge from my side. Follow-ups worth doing (here or later): the M2 timing assertion, a comment or test for the M4 guard, and a user-visible notice on the watchdog path.
How to reproduce
# per arm: build the worktree's packages, then
QWEN_HOME=<isolated> QWEN_TEST_CLOSE_DELAY_MS=15000 \
node scripts/dev.js serve --port 4270 --workspace <fixture-repo>
QWEN_DAEMON_URL=http://127.0.0.1:4270 \
npm run dev --workspace=packages/web-shell -- --port 5302
# open two persisted sessions, then: click A -> click B -> click A again ~2s laterRestart the daemon between arms: a leftover client-id registration from a previous page keeps clientIds.size > 0, which suppresses close-on-last-detach and makes the race unreachable — that is why an unclean run can look green on base.
中文说明
维护者验证：本地真实环境，与 base 做 A/B 对比
我在本地用真实 qwen serve（真实 ACP 子进程、真实 Web Shell）构建并运行了这个 PR，端到端验证修复，并补上了上一轮 review 因工具预算耗尽而未能确认的可达性问题。修复有效，走的是真实触发路径和真实的 daemon 响应。 文末记录两个非阻塞的测试覆盖缺口。
版本： PR head efd58dc720 · base 0d56e50b64（与 main 的 merge-base）· macOS 26.6、Node v24.18.1、Chrome 151。
环境搭建
| 组成 | 说明 |
|---|---|
| Daemon | 用 tsx 直接从 PR 源码运行 qwen serve（scripts/dev.js serve），真实 ACP 子进程与真实会话 JSONL |
| Web Shell | 各自 worktree 里的 vite dev。serve 模式下 vite.config.ts 把 @qwen-code/webui 别名指向 ../webui/src，所以被测代码确实是各自那一侧的 DaemonSessionProvider |
| A/B | 同一个 daemon，两个 dev server：base :5301 vs PR :5302，两侧唯一差异就是这 4 个改动文件 |
| 账本 | 在 Web Shell 与 daemon 之间加了记录型反向代理，每次 POST /session/:id/load 及其响应体都独立于 devtools 落盘 |
| 隔离 | 独立 QWEN_HOME、一次性 git workspace、离线的 OpenAI 兼容 fixture 模型（会话需要真实 transcript；本修复与模型无关） |
已披露的故障注入（daemon 里一处受环境变量控制的 setTimeout，两侧完全一致，事后已还原）：closing 窗口真实存在但通常只有几毫秒，浏览器点击无法稳定命中；而它在生产里本就是无界的 —— 健康 channel 下 closeIfChildUnheld 传的是 agentCloseTimeoutMs: undefined，closeSessionImpl 无时间上限地 await notifyAgentSessionClose。因此我只是把这个已有窗口加宽：在 closeSessionImpl 中 entry.closing = true 之后加一处由 QWEN_TEST_CLOSE_DELAY_MS 控制的 await setTimeout(...)，设为 15 秒（watchdog 场景设 60 秒）。没有改动任何判定逻辑，两侧跑的是同一个打了补丁的 daemon，最终测试前已还原 packages/acp-bridge/src/bridge.ts（worktree 已确认干净，停在 efd58dc720）。其余全是产品自身行为：teardown 触发、404、状态码、文案。
可达性 —— 上一轮的未决问题
运行中的系统里，这条链路的三个环节都成立：
- 切换确实走被门控的 action 路径。 侧边栏点击 →
loadSidebarSession→sessionActions.loadSession(sessionId, { workspaceCwd })（packages/web-shell/client/App.tsx:6885-6901），因此loadingRequestedSession && pendingLoad?.sessionId === restoreSessionId成立。（Web Shell 不传sessionScope；新建会话在 provider 内部强制'thread'。） - teardown 是产品自身的 close-on-last-detach，不是我人为制造的。切走会发
POST /session/:id/detach；该会话的 SSE 订阅者断开后，maybeCloseIdleSession → closeIfChildUnheld → closeSessionImpl就会执行。daemon 日志：qwen serve: closing session "e44bee82…" (reason: last_client_detached)。 - 响应与客户端匹配的字符串逐字节一致 ——
SessionNotFoundError→error-response.ts:392-393，404 且不带code字段：{"error":"No session with id \"e44bee82-…\". The session is closing; retry after close completes","sessionId":"e44bee82-…"}
全程只有真实 UI 操作，没有伪造响应：点会话 A → 点会话 B → 2 秒后再点 A。
修复前 / 修复后
base 上共 5 次独立运行均复现。（另有 2 次未复现：daemon 若积累了此前页面残留的 client-id 注册，close-on-last-detach 根本不会触发，也就没有 closing 窗口可以竞争。两侧之间重启 daemon 即可确定复现。）
base 0d56e50b64：错误提示 + “当前会话不存在”，且是终止态 —— teardown 结束 30 秒后界面依然停在那里。请求账本只有一次尝试：
POST /session/ALPHA/detach -> 204 # 切走
POST /session/BETA/load -> 200
POST /session/BETA/detach -> 204 # 切回
POST /session/ALPHA/load -> 404 {"error":"... The session is closing; retry after close completes"}
# 之后再无任何尝试
抓到的 toast 文案（从 DOM 轮询获取，避免被自动消失漏掉）：POST /session/:id/load: No session with id "e44bee82-…". The session is closing; retry after close completes
PR efd58dc720：按既有退避重试并成功：
POST /session/ALPHA/load -> 404 +0.00s
POST /session/ALPHA/load -> 404 +0.97s
POST /session/ALPHA/load -> 404 +2.97s
POST /session/ALPHA/load -> 404 +6.97s
POST /session/ALPHA/load -> 200 +15.03s <- teardown 完成，会话恢复
间隔约 1 s / 2 s / 4 s / 8 s，即真实的带抖动指数退避；另一次运行得到同样形态（2/2/5/8 s）。整个 20 秒窗口内唯一的用户可见提示是无害的“会话正在加载，请稍后再发送。”。会话 A 最终带完整 transcript 打开，无错误提示、无会话不存在页面。
Critical 项 R1-1（watchdog 在重试中途触发）—— 已在真实环境验证
用 --session-restore-timeout-ms 5000（即 20 秒 pending-load watchdog）对 60 秒 teardown，让 watchdog 在重试循环进行中触发：
- PR：共 5 次尝试（+0/2/4/8/15 秒）后在 watchdog 处停止 —— 没有请求风暴，也没有永久 spinner，UI 进入空闲可用状态。让这一点变确定的正是
actions.ts的 watchdog 改动。 - base：单次 404 → 立刻进入终止的会话不存在页面。
一个非阻塞的遗留问题：这条路径上用户拿不到任何明确提示 —— 等了 20 秒之后只看到一个空的输入框。这仍严格优于 base（一个错误且永久的报错），但加一行提示会更友好。
测试与变异矩阵
DaemonSessionProvider.test.tsx + actions.test.ts 在 head 上 254/254 通过。随后我对 PR 的生产代码施加了 7 个单行/单块变异：
| 变异 | 结果 |
|---|---|
| M1 删除整个 closing 重试分支 | 被杀 |
M2 删除 reconnectAttempt += 1; |
|
M3 删除 await delay(...) |
被杀（2 个测试） |
M4 删除 if (pendingSessionLoadRef.current !== pendingLoad) return; |
|
M5 去掉 autoReconnect && 门控 |
被杀 |
M6 还原 actions.ts 的 watchdog disconnected 收敛 |
被杀 |
M7 还原 effectPendingSessionLoad 清理身份守卫 |
被杀 |
两个存活变异，附实测数据：
- M2 —— 退避递增仍未被测试覆盖。 用逐毫秒步进的探针测出每次重试前的精确延迟：PR =
[10, 20, 40, 80] ms，变异体 =[10, 10, 10, 10] ms。上面的真机运行说明递增是真实的生产行为（1→2→4→8 秒），所以这正是 R1-5 描述的回归：新测试里的advanceTimersByTimeAsync(20)同样覆盖 10ms 的延迟，无法区分两者。断言实际经过的间隔（或使用差异明显的 delay 配置）即可杀掉它。 - M4 —— 被取代切换的守卫没有测试能走到。 即使加了新的 supersession 用例也一样：那个场景下旧 runner 的
abort.signal已经让delay抛出，控制流根本到不了该守卫，所以这行属于纵深防御而非被测对象。建议要么加注释说明，要么补一个真正走到它的用例。
两个存活变异都不是已发布行为的缺陷，而是覆盖缺口 —— 相关代码本身经实测是正确的。
结论
修复达到了描述所声称的效果：在真实环境中，由真实 teardown 触发、面对真实的 daemon 404。两个场景都未观察到回归，round 1 标为 Critical 的 watchdog 竞态也能确定性收敛。从我这边看可以合并。 建议的后续（本 PR 或后续 PR 均可）：补 M2 的时序断言、为 M4 守卫加注释或测试、以及在 watchdog 路径上给用户一个可见提示。
复现方式： 每一侧先构建 worktree 的 packages，然后
QWEN_HOME=<isolated> QWEN_TEST_CLOSE_DELAY_MS=15000 \
node scripts/dev.js serve --port 4270 --workspace <fixture-repo>
QWEN_DAEMON_URL=http://127.0.0.1:4270 \
npm run dev --workspace=packages/web-shell -- --port 5302
# 打开两个已持久化的会话，然后：点 A → 点 B → 约 2 秒后再点 A两侧之间要重启 daemon：上一个页面残留的 client-id 注册会让 clientIds.size > 0，从而抑制 close-on-last-detach、使竞态不可达 —— 这也是不干净的运行在 base 上也可能看起来正常的原因。
Tested on
| OS | Status |
|---|---|
| 🍏 macOS 26.6 | ✅ verified locally (real daemon + real Web Shell + Chrome) |
| 🪟 Windows | |
| 🐧 Linux |
wenshao
approved these changes
Aug 10, 2026
This was referenced Aug 10, 2026
Collaborator
|
Released in v0.21.9. |
qwen-code-dev-bot
pushed a commit
to DragonnZhang/qwen-code
that referenced
this pull request
Aug 11, 2026
…osing retry (QwenLM#8884) * fix: add structured error code to SessionNotFoundError responses PR QwenLM#8864 retried session switches while the target session is closing, but relied on fragile string matching against the daemon's error message. This commit: 1. Adds a `code` property to `SessionNotFoundError` — automatically set to `'session_closing'` when the extra message mentions "closing", otherwise `'session_not_found'`. 2. Includes `code` in the HTTP JSON response body so clients can distinguish closing (transient) from genuinely missing sessions without depending on error message text. 3. Updates the WebUI retry check in `DaemonSessionProvider` to use `errorBody.code === 'session_closing'` instead of matching `endsWith('The session is closing; retry after close completes')`. 4. Fixes an inconsistent error message in `rewindSession` that used the short `'The session is closing'` without the retry suffix. Closes: QwenLM#8864 (follow-up) * fix(daemon): expose session closing code * docs(serve): document session closing codes * fix(acp): preserve closing code after restore waits * fix: restore class pin in bridge test and update error taxonomy - Add toBeInstanceOf(SessionNotFoundError) alongside toMatchObject to preserve the envelope type assertion - Document session_closing code in 18-error-taxonomy.md * chore: drop unrelated merge formatting
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What this PR does
Keeps an explicit session switch pending when the daemon reports that the target session is still closing, then retries it with the existing reconnect backoff. Watchdog expiry now settles a stranded switch as disconnected, a newer switch supersedes the old retry cleanly, and
autoReconnect={false}still fails fast. Other missing-session responses keep their current terminal behavior.Why it's needed
Switching away from a session starts an asynchronous teardown. Switching back before that teardown finishes returns a transient 404, but the Web Shell treated it as a permanently missing session and showed both an error toast and the “current session does not exist” screen.
Reviewer Test Plan
How to verify
Open two persisted sessions and switch between them repeatedly without waiting between switches. A transient closing response may occur internally, but the requested session should eventually open without an error toast or missing-session screen.
Evidence (Before & After)
Tested on
Environment (optional)
Locally built macOS Desktop app bundle and its packaged runtime with the production Web Shell.
Risk & Scope
Linked Issues
Related to #8092.
中文说明
这个 PR 做了什么
当 daemon 返回目标会话仍在关闭中时，保持显式的会话切换请求，并使用现有的重连退避机制重试。watchdog 到期时，滞留的切换会确定性进入 disconnected；新切换会干净地取代旧重试；
autoReconnect={false}仍会快速失败。其他会话不存在响应保持当前终止行为。为什么需要
切离一个会话会启动异步清理。如果在清理结束前切回，daemon 会返回临时 404，但 Web Shell 之前会把它当作会话永久不存在，同时显示错误提示和“当前会话不存在”页面。
Reviewer Test Plan
如何验证
打开两个已保存的会话，并在不额外等待的情况下反复切换。内部可能短暂出现 closing 响应，但目标会话最终应正常打开，且不显示错误提示或会话不存在页面。
证据（修复前后）
测试平台
环境（可选）
本地构建的 macOS Desktop APP bundle、其内置 runtime 和生产 Web Shell。
风险与范围
关联 Issue
关联 #8092。