fix(web-shell): keep a manual session name across /clear (#8977) - #9260
fix(web-shell): keep a manual session name across /clear (#8977)#9260yiliang114 wants to merge 43 commits into
Conversation
/clear starts a brand-new deferred session, so a name the user chose via /rename or the sidebar stays on the old session and the successor gets an auto-generated title after the first turn. Carry a manual name over to the successor session: - daemon rename publications now stamp titleSource:'manual' (the child-side title-update path already forwarded it; the HTTP rename path did not) - webui connection state tracks titleSource alongside displayName and clears it with the rest of the session-scoped fields on clear - the /clear handler stashes the manual name and re-applies it via renameSession once the new session attaches (ensureSessionForPrompt success path); auto-generated titles are never carried forward, and /new and /reset deliberately keep today's "start untitled" behavior Tests: bridge stamps titleSource on rename publications; mappers keep it from session_metadata_updated (manual/auto/unknown); connection clear drops it; App-level /clear carry-over positive + auto-title negative.
|
Re-run on request ( Template: complete ✓ — all required sections present, including the bilingual details block. Problem: observed bug, not theory. #8977 is a user report from v0.21.10: a manually named Web Shell session loses its name after Direction: aligned — a deliberately chosen session name is session identity, and losing it on Size: 2,085 additions / 96 deletions across 44 files — ~512 production logic lines, ~1,643 test lines, and a 26-line design doc. The change spans six packages (acp-bridge, cli, core, sdk-typescript, web-shell, webui) and touches Approach: unchanged from round 3 and still right — title and provenance are a tuple, and every path that touches the title (cold-restore routes, ACP dispatch, branch, keepalive, live coordinator, SDK, WebUI provider) must carry both fields or the tuple desyncs. Cutting 80% of the files would reintroduce the reported bug through a surviving path. Nothing to split out; the design doc under Risk: one Stage-1e match — Moving on to code review. 🔍 中文说明应 **模板：**完整 ✓ —— 所有必需小节齐全，包含中文 details。 **问题：**已观测到的 bug，不是理论问题。#8977 是 v0.21.10 上的用户报告：手动命名的 Web Shell 会话在 **方向：**对齐——用户主动选择的会话名就是会话身份，在 **规模：**44 个文件 +2,085/-96，其中约 512 行生产逻辑、约 1,643 行测试、26 行设计文档。跨 6 个包（acp-bridge、cli、core、sdk-typescript、web-shell、webui）并触及 **方案：**与第 3 轮判断一致且仍然正确——标题与来源是一个元组，任何触及标题的路径（冷恢复路由、ACP dispatch、branch、keepalive、live coordinator、SDK、WebUI provider）都必须同时携带两个字段，否则元组失同步。砍掉 80% 的文件只会让幸存路径重新引入该 bug。无可拆分内容； **风险：**Stage-1e 命中一项—— 进入代码审查。🔍 — Qwen Code · qwen3.8-max Reviewed at |
|
Round-6 review at Round-5 findings: all three verified fixed at this head.
Nothing new surfaced in the rest of the diff: the carry-over ordering in Testing — the PR's own CI at this head, fetched once via the API (no PR code executed by this gate). At fetch time the primary unit gate had not finished; the table is updated in place by the finalize job once CI settles. One red check, classified below. 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 Not verified: no local build/test execution — this gate never runs PR code; the author's closeout claims (mutation-verified regressions, Sandboxed verification would settle the remaining behavioural claim: 中文说明第 6 轮审查（head 第 5 轮三项发现均已在本 head 验证修复：
diff 其余部分未发现新问题： **测试——本 head 上 PR 自身的 CI，经 API 一次性获取（门禁不执行 PR 代码）。**获取时主测试门禁尚未完成；表格由 finalize 任务在 CI 落定后就地更新。一项红色检查，分类见下。
未验证：不做本地构建/测试执行——本门禁从不运行 PR 代码；作者收尾声明（mutation 验证的回归、server.test.ts 1048 通过、App.test.tsx 501/501、tsc/eslint 干净）属于作者自述，独立确认依赖 CI 与下方沙箱通道。 沙箱验证可以补齐剩下的行为性主张： — Qwen Code · qwen3.8-max Reviewed at |
|
Confidence: 3/5 — the code itself would score higher: all three round-5 findings are verifiably fixed at this head with regression tests that actually pin them, and six rounds have worn the design down to the minimal fix. What keeps this at 3 is not code — it is one human-gated contract decision that a bot must not make, and a primary CI gate that had not finished when this pass ran. Stepping back: this is the cleanest this PR has ever been. My independent read of the problem — persist title provenance as a tuple, restore it before attach on every cold path, carry a Why defer instead of approve, plainly:
⏸️ Deferring to @wenshao — two human calls are needed: (a) the R4-1 daemon-contract choice above, and (b) confirming the Live Host lockfile failure (owned by #9491's import) should not block this PR once the unit suite lands green. Once R4-1 is decided and CI is green on the reviewed head, re-run 中文说明退一步看：这是这个 PR 历轮以来状态最好的一次。我对问题的独立方案——把标题与来源作为元组持久化、在每条冷恢复路径上于 attach 前恢复、客户端在 为什么 defer 而不是批准，直说：
⏸️ 转交 @wenshao —— 需要两个人工决策：(a) 上述 R4-1 的 daemon 契约选择；(b) 确认 Live Host 的 lockfile 失败（归 #9491 的导入）在单元测试全绿后不阻塞本 PR。R4-1 决策且被审 head 上 CI 全绿后，重跑 — Qwen Code · qwen3.8-max Reviewed at |
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
There was a problem hiding this comment.
Requesting changes — the branch carries an accidental revert, not just the fix.
Every production file at head matches the pre-#8817 checkout byte-for-byte except the small intended hunks (verified by direct comparison): merging would roll back #8817 "fork from any conversation" and #9180 "composer text attachments" (both merged yesterday) plus part of #9052, and strip the admission-time branch/rewind guards. The revert is also silent at runtime — atRecordId is still accepted upstream and dropped in the bridge. CI confirms: tsc --build fails on the stale bridge.ts (TS2322, bridge.ts:8645) across every job that builds.
The ~90 lines implementing the /clear name carry-over are good and worth keeping. Please recreate the branch from current main with only those hunks (details in my Stage 2 comment) and I'll re-review. 🙏
请求修改 —— 该分支携带的是意外回滚，而不仅是修复。
head 上的每个生产文件除小块预期改动外，与 #8817 之前的检出逐字节一致（已直接比对验证）：合并将回滚 #8817"从任意对话 fork"与 #9180"输入框文本附件"（均为昨天合并）外加 #9052 的一部分，并移除 branch/rewind 的准入期保护。回滚在运行时还是静默的 —— 上游仍接受 atRecordId，在 bridge 层被丢弃。CI 证实：过期的 bridge.ts 导致 tsc --build 在所有构建 job 中失败（TS2322，bridge.ts:8645）。
实现 /clear 名称继承的约 90 行是好的、值得保留。请基于当前 main 重建分支、只携带这些改动块（详见我的 Stage 2 评论），我会重新审查。🙏
The branch rework deleted the persisted-only branch path (branchSession now always restores or throws), but its return type still carried the persisted-only union member. createSideTaskSession destructures the union and returns the rest, which the persisted member left short of BridgeRestoredSession's fields — the TS2322 that failed CI at bridge.ts:8645. Collapse the union to BridgeBranchedSession (the only shape the implementation can produce; the removed type had no consumers) so the side-task handler typechecks again.
The branch rework left dangling references that broke the webui build:
- DaemonPromptFile was deleted from types.ts but still re-exported by
session/index.ts, daemon/index.ts, daemon-react-sdk.ts and imported
by promptContent.ts; complete the removal (the files attachment path
has no remaining caller) and drop the file-token/resource logic plus
its now-obsolete tests
- branchSession's checkpoint arg and switchStarted field are gone;
update the two DaemonSessionProvider tests to the new
(name?) => {sessionId, displayName} contract
- restore the branchInFlight concurrency guard the rewrite dropped:
without it a concurrent branchSession no longer rejects with
InvalidStateError (pinned by the existing concurrent-branch test),
and the declaration was removed along with its usages
webui vite build green; 276 + 443 + 17 tests green; eslint clean.
|
Closeout summary: restored the PR body to the repository template, including Risk & Scope, Linked Issues, and the Chinese details section. No code changes were made in this pass. |
…eline (#8977) CI failed on four type errors after the webui submit path lost file attachments: - handleEditorSubmit now accepts the composer's files argument in position and drops it (the pipeline past the composer no longer carries files) - the two rawEnqueuePrompt calls pass an explicit undefined files slot so onComplete/inputAnnotations land in their own positions - the queued-prompt submit no longer forwards files (SubmitPromptOptions has no such field anymore); queued prompts keep files for edit/restore - App.test.tsx direct onSubmit calls updated to the new positions Composer DOM/mobile and queued-prompt suites pin the composer contract with files in position 3, so the files parameter stays in the ChatEditor/useComposerCore signature; the files stop at the App boundary. 443 + 89 tests green, eslint clean.
🖼️ web-shell visual previewRendered against a mock daemon (no real backend): the PR base vs this PR head Screenshots · before / afterFull-resolution recordings (.webm) are attached to the workflow run. — Qwen Code · web-shell visuals |
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. |
…ch rework (#8977) Supersedes the files-free alignment: the composer, queued prompts and image ingestion still carry PromptFile end to end, so re-add the webui files option (DaemonPromptFile, SubmitPromptOptions.files, resource blocks in toDaemonPromptContent) instead of dropping attachments at the App boundary. web-shell 3642 + webui 499 tests green locally.
|
@qwen-code /review |
|
Qwen Code review request accepted. Review is queued in workflow run. |
|
【巡检备注】CI |
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
There was a problem hiding this comment.
Partially reviewed — gaps disclosed.
Not reviewed: reverse audit — stopped at the 5-round cap without two consecutive dry rounds; every chunk was audited each round, and rounds 3-5 only surfaced further evidence folded into the confirmed class finding R1-1.
Not explored to full depth (tool budget reached): "agent reverse-audit (round 1)": none — everything completed. Don't write any Budget gap line.; chunk 6: root-cause the failing "Run required no-AK integration gate" CI step (log extraction returned empty; step is outside chunk-6 files).
[Critical] R1-1: The committed diff is not scoped to #8977 — it reverts already-merged features #8817 (fork from any conversation / checkpoint branching), #9180 (composer text attachments), part of #9052 (at-mention providers), and the branch/rewind admission guards. Verified ancestry: merge-base 337da21 (2026-08-16) contains the merges of #8817 (9f8f65d), #9180 (34cc1c3) and #9052 (c396fe3) — all three ancestors of the merge base — so the −1985 lines are committed branch changes, not a stale-base artifact; the three follow-up commits reworked the codebase AROUND the revert instead of rebasing onto main. Witness: git merge-base --is-ancestor → all three ANCESTOR of 337da21; CI run 31947442596 at this HEAD fails — integration-tests/cli/qwen-serve-streaming.test.ts:967 'expected undefined to be defined' (turn_complete.branchPoint removed). Verified symptoms: atRecordId silently dropped while routes/session.ts, the SDK (DaemonClient.ts:2788), the per-message Branch button and the e2e smoke spec still use it (historical branches silently fork from HEAD with 201; route disconnect cleanup leaks a live session slot for atRecordId requests); files attachments dropped from queued prompts (enqueuePrompt hardcodes undefined), failed-prompt retry, turn-error retry and draft restore, and the retry gates lost their files clauses so files-only prompts can never be retried; branchSession lost both isClosingOrAuthorizingClose guards while every other admission path keeps them; rewind lost its entire promptQueue FIFO serialization (busy guard, chain-in and tail update) compounded by the uncancellable withTimeout(10s) wrap into unobserved history truncation + file restores racing subsequent prompts; branch/side-task ext methods wrapped in withTimeout(10s) — the deleted test 'waits for a dispatched side-task fork instead of timing out' was deleted because it cannot pass against the restored timer; webui branchSession re-wrapped in 30s withActionTimeout (slow restore → client timeout + orphaned attached session); switchStarted navigation guard removed; branch path lost the session_busy → SessionBusyError mapping (rewind kept it); BranchWhilePromptActiveError throw is now unreachable dead code (the advertised 409 branch_while_prompt_active silently becomes queue-and-proceed); builtinAtProviders/atProviders stripped from the customization memo — third-party hosts' @-mention providers vanish from split-view/side-task composers; WebShellCustomizationProvider removed from both artifact-panel render sites; App-layer concurrent-branch dedupe removed (false 'Branch session failed' toast; the surviving actions-layer branchInFlight guard lost its only test); dispatchSessionChange 'submit' gates dropped their files clauses so files-only submissions skip the public onSessionChange event; deleted tests also covered still-live behavior (branch/side-task maxSessions capacity gate; admission release on plain restore failure). This re-reports — as still standing at HEAD c81398a — the existing triage blocker (comment 5306337462 / review 4945686638) filed at the earlier SHA 5c034da. Fix: recreate the branch from current main carrying only the intended #8977 hunks (titleSource stamp, mappers/types/actions plumbing, App.tsx carryOverManualNameRef hunks, matching tests); re-run CI.
中文说明
仅完成部分审查，审查缺口已披露。
未审查：reverse audit — stopped at the 5-round cap without two consecutive dry rounds; every chunk was audited each round, and rounds 3-5 only surfaced further evidence folded into the confirmed class finding R1-1。
未探索到全部深度（达到工具调用预算）："agent reverse-audit (round 1)"：none — everything completed. Don't write any Budget gap line.；chunk 6：root-cause the failing "Run required no-AK integration gate" CI step (log extraction returned empty; step is outside chunk-6 files)。
[Critical] R1-1: The committed diff is not scoped to #8977 — it reverts already-merged features #8817 (fork from any conversation / checkpoint branching), #9180 (composer text attachments), part of #9052 (at-mention providers), and the branch/rewind admission guards. Verified ancestry: merge-base 337da21 (2026-08-16) contains the merges of #8817 (9f8f65d), #9180 (34cc1c3) and #9052 (c396fe3) — all three ancestors of the merge base — so the −1985 lines are committed branch changes, not a stale-base artifact; the three follow-up commits reworked the codebase AROUND the revert instead of rebasing onto main. Witness: git merge-base --is-ancestor → all three ANCESTOR of 337da21; CI run 31947442596 at this HEAD fails — integration-tests/cli/qwen-serve-streaming.test.ts:967 'expected undefined to be defined' (turn_complete.branchPoint removed). Verified symptoms: atRecordId silently dropped while routes/session.ts, the SDK (DaemonClient.ts:2788), the per-message Branch button and the e2e smoke spec still use it (historical branches silently fork from HEAD with 201; route disconnect cleanup leaks a live session slot for atRecordId requests); files attachments dropped from queued prompts (enqueuePrompt hardcodes undefined), failed-prompt retry, turn-error retry and draft restore, and the retry gates lost their files clauses so files-only prompts can never be retried; branchSession lost both isClosingOrAuthorizingClose guards while every other admission path keeps them; rewind lost its entire promptQueue FIFO serialization (busy guard, chain-in and tail update) compounded by the uncancellable withTimeout(10s) wrap into unobserved history truncation + file restores racing subsequent prompts; branch/side-task ext methods wrapped in withTimeout(10s) — the deleted test 'waits for a dispatched side-task fork instead of timing out' was deleted because it cannot pass against the restored timer; webui branchSession re-wrapped in 30s withActionTimeout (slow restore → client timeout + orphaned attached session); switchStarted navigation guard removed; branch path lost the session_busy → SessionBusyError mapping (rewind kept it); BranchWhilePromptActiveError throw is now unreachable dead code (the advertised 409 branch_while_prompt_active silently becomes queue-and-proceed); builtinAtProviders/atProviders stripped from the customization memo — third-party hosts' @-mention providers vanish from split-view/side-task composers; WebShellCustomizationProvider removed from both artifact-panel render sites; App-layer concurrent-branch dedupe removed (false 'Branch session failed' toast; the surviving actions-layer branchInFlight guard lost its only test); dispatchSessionChange 'submit' gates dropped their files clauses so files-only submissions skip the public onSessionChange event; deleted tests also covered still-live behavior (branch/side-task maxSessions capacity gate; admission release on plain restore failure). This re-reports — as still standing at HEAD c81398a — the existing triage blocker (comment 5306337462 / review 4945686638) filed at the earlier SHA 5c034da. Fix: recreate the branch from current main carrying only the intended #8977 hunks (titleSource stamp, mappers/types/actions plumbing, App.tsx carryOverManualNameRef hunks, matching tests); re-run CI.
— qwen3.8-max via Qwen Code /review (v0.21.12)
- keep a stashed manual name across a repeated /clear before a prompt - use SERVE_CONTROL_EXT_METHODS.sessionBranch in bridge tests - declare MockConnection.titleSource; pin the rename target session and mirror the cleared-connection reset in the /clear tests
…r rename (#8977) - reset titleSource on session switch so a manual stamp never leaks onto the next session's auto-generated title - clear the manual-name carry-over when opening an existing session - make updateSessionMetadata's titleSource caller-controlled (default manual; machine-generated names pass auto) - renameSession accepts a silent option so the best-effort /clear carry-over never toasts on failure
…ped by the branch rework (#8977) Merging main exposed two casualties of the branch rework: - broadcastTurnComplete no longer read _meta['qwen.branchPoint'], so prompt results lost branchPoint and the historical-branch integration test failed at expect(first.branchPoint).toBeDefined() - branchSession dropped req.atRecordId from the sessionBranch payload, so forks from a historical record replayed the full source transcript Restore both exactly as main carries them (CHAT_RECORD_UUID_RE included); cli/qwen-serve-streaming historical-branch suite green locally (7/7).
…ce (#8977) The titleSource stamping landed on the routes before the bridge interfaces they type against: add titleSource to both updateSessionMetadata declarations and pin it in the rename assertions.
🩺 serve daemon A/BBuilt the PR base vs this PR head ✅ No response changes against the PR base across 12 scenario(s). — Qwen Code · serve A/B |
|
@qwen-code /resolve |
|
Qwen Code did not run conflict resolution for this request. PR #9260 is draft. |
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
There was a problem hiding this comment.
Partially reviewed — gaps disclosed.
Not reviewed: reverse audit — reached the 5-round cap without two consecutive dry rounds (rounds 3-5 each reported findings; all 13 chunks audited every round).
Not explored to full depth (tool budget reached): "agent reverse-audit (round 2)": tracing the SDK DaemonClient view-state update timing vs the session-record snapshot to pin whether activeSession.state and activeSession.session can actual…; chunk 7: executing the two new/changed test files (no node_modules in this review worktree or the parent checkout; verified them by line-by-line tracing against the impl….
Test Plan (not a blocker): 2 passed — this review observed 1652, 22609, 20723, 1620, 4009, 593, 494 passed.
Deferred under the convergence posture (round 10, not a blocker) — recorded, not requested in this round:
packages/acp-bridge/src/bridge.test.ts:26509 (+2 locations) — [probe] same-text bridge tests pin live state only — the documented publish and the persist layer are unassertedpackages/cli/src/serve/acp-http/transport.test.ts:4806 (+5 locations) — [probe] five new prototype spies restored only on the success path — a red assertion leaks fabricated titles into every later test in each filepackages/core/src/services/chatRecordingService.ts:2093 — [probe] auto-title tombstone exception untested — dropping && onDisk.title ships green and cleared sessions never re-titlepackages/web-shell/client/App.tsx:8362 — [probe] sidebar-load and workspace-switch carry invalidations untested — deleting either clearing line ships green and leaks a name across sessionspackages/acp-bridge/src/bridge.ts:9790 — [probe] SDK view-state reducer stores the null clear-marker verbatim — exported DaemonSessionViewState.displayName holds null against a string | undefined typepackages/acp-bridge/src/bridge.test.ts:26400 — [review] cold-restore simulation hardcodes the falsy gate and misattributes it to the reader this PR flipped — the real untested gate lives in restoreSessionTitleFieldspackages/cli/src/serve/server.test.ts:25847 — [review] live-branch metadata response shape unpinned — the fake bridge drops the titleSource the real bridge returnspackages/web-shell/client/App.test.tsx:10604 — [probe] rename-rejection path untested — the teardown-and-retry contract on silent carry-rename failure is unpinnedpackages/webui/src/daemon/session/DaemonSessionProvider.tsx:2027 — [review] added comment describes a state-mutation path that does not exist — the snapshot is frozen and never carries titlespackages/webui/src/daemon/session/mappers.ts:660 — [review] getRestoredSessionTitle has zero direct tests — its tombstone filtering is the client's last defense and is unverifiedpackages/web-shell/client/App.test.tsx:10941 — [probe] carry-token consumption unpinned — deleting the identity cleanup ships green (502/502) and a stale token can rename an unrelated later session
中文说明
仅完成部分审查，审查缺口已披露。
未审查：reverse audit — reached the 5-round cap without two consecutive dry rounds (rounds 3-5 each reported findings; all 13 chunks audited every round)。
未探索到全部深度（达到工具调用预算）："agent reverse-audit (round 2)"：tracing the SDK DaemonClient view-state update timing vs the session-record snapshot to pin whether activeSession.state and activeSession.session can actual…；chunk 7：executing the two new/changed test files (no node_modules in this review worktree or the parent checkout; verified them by line-by-line tracing against the impl…。
Test Plan（非阻断）：2 passed — this review observed 1652, 22609, 20723, 1620, 4009, 593, 494 passed。
收敛姿态下延后（第 10 轮，非阻断）——已记录，本轮不要求修改：共 11 条（原文未翻译，列表见上方英文部分）。
— qwen3.8-max via Qwen Code /review (v0.21.15)
| const result = await state.client.updateSessionMetadata(sessionId, { | ||
| displayName: args.display_name, | ||
| titleSource: args.display_name?.trim() ? 'manual' : 'auto', | ||
| }); |
There was a problem hiding this comment.
[Critical] R4-1: Still standing at this head — the absent-name branch of this daemon-MCP tool remains a silent no-op end-to-end, awaiting the maintainer decision (open since round 4). When display_name is omitted, JSON serialization drops the undefined displayName, the wire body is exactly {"titleSource":"auto"}, and bridge.updateSessionMetadata performs every mutation inside if (metadata.displayName !== undefined) (bridge.ts:9705, re-verified at this commit) — so the route validates and answers 200 while nothing is updated, persisted, or published, yet the tool reports success and its comment claims the absent case is machine handling. Re-stating the two decision shapes from the open thread: either honor titleSource-only updates against the entry's existing title in the bridge (outside the displayName !== undefined gate), or 400-reject a displayName-less PATCH in both routes and stop the tool sending titleSource. Both change the public daemon contract — this stays human-gated per the standing authorization.
中文说明
R4-1：在当前 head 上仍然存在——这个 daemon-MCP 工具的"缺省名字"分支端到端仍是静默空操作，等待维护者决定（自第 4 轮起开放）。当省略 display_name 时，JSON 序列化会丢弃 undefined 的 displayName，线上请求体恰好是 {"titleSource":"auto"}，而 bridge.updateSessionMetadata 的所有变更都在 if (metadata.displayName !== undefined) 内执行（bridge.ts:9705，已在本提交上重新验证）——因此路由校验通过并返回 200，但没有任何内容被更新、持久化或发布，而工具却报告成功，其注释还声称缺省名字属于"机器处理"。重申开放线程中的两个决定形态：要么在 bridge 中针对条目现有标题受理仅含 titleSource 的更新（放在 displayName !== undefined 门禁之外），要么在两个路由中对不含 displayName 的 PATCH 返回 400 并让工具停止发送 titleSource。两者都会变更公开的 daemon 契约——按既有授权，此项保持人工裁决。
— qwen3.8-max via Qwen Code /review (v0.21.15)
There was a problem hiding this comment.
Re-verified on current head c2093f5468d69d95503a74feeda127771cfb03a9: omitting display_name serializes a titleSource-only PATCH, while the live bridge mutates only inside the displayName !== undefined gate and the workspace route rejects the equivalent absent-name request. This is the same public metadata-contract decision as R4-1 (honor titleSource-only updates vs reject the request), so I am leaving this thread unresolved with no code change.\n\n
|
Round closeout (head Merge ( R10-1 tombstone consumer violations — FIXED ( R10-2 stale child echo resurrection — FIXED ( R10-3 keepalive destroys user-renamed task sessions — FIXED ( Verification: acp-bridge bridge suites green (metadata/title/downgrade/echo-witness), cli server metadata 52 + keepalive 28 + list 26 + useResumeCommand 13, core session-reference-service 26 + chatRecordingService 115, sdk DaemonClient metadata 10; typecheck clean in core/cli/acp-bridge/sdk-typescript; eslint + prettier clean on changed files. Threads: R10-1/R10-2/R10-3 replied with SHA evidence and resolved. The 7 R4-1 family threads (updateSessionMetadata absent-name demote semantics) remain deliberately unresolved — that one still needs a maintainer decision on whether an explicit demote overrides the sameTextDowngrade keepalive guard. |
…manual-session-name
|
Note — head moved The first CI run on |
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
There was a problem hiding this comment.
Partially reviewed — gaps disclosed.
Not reviewed: reverse audit — reached the 5-round cap without two consecutive dry rounds (rounds 3–5 each reported findings; all 15 chunks audited every round).
Not reviewed: build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally.
Not explored to full depth (tool budget reached): chunk 7: running packages/cli/src/serve/scheduled-task-keepalive.test.ts and the two new server.test.ts cases — the worktree has no node_modules (vitest cannot sta…; chunk 2: executed the new bridge tests not run — worktree lacks node_modules; npm install + build was declined to avoid timeout/interference in the shared review tree..
Test Plan (not a blocker): 2 passed — this review observed 1694, 23292, 20827, 1653, 4137, 605, 495 passed.
Deferred under the convergence posture (round 11, not a blocker) — recorded, not requested in this round:
packages/core/src/services/chatRecordingService.ts:2157 — [probe] auto-title cross-process guard's tombstone arm untested — removing && onDisk.title ships green (110 tests), cleared sessions would never re-titlepackages/cli/src/ui/hooks/useResumeCommand.ts:197 — [probe] tombstone mapping customTitle || null untested — reverting to ?? null ships green and resumes cleared sessions with a blank namepackages/cli/src/serve/session-restore-title.ts:13 — [probe] no serve-side cold-restore test drives a tombstone title — a presence-gate regression ships green and seeds an empty manual namepackages/acp-bridge/src/bridge.test.ts:26662 — [review] echo-suppression window's close side untested — both counter-decrement arms can regress with every test green, permanently suppressing echoespackages/acp-bridge/src/bridge.test.ts:26654 — [probe] suppressed-echo event-side guarantee unpinned — a publish-moved-above-suppression mutant ships green and the flipping probe failspackages/acp-bridge/src/bridge.test.ts:26785 — [probe] downgrade-refusal tests pin no persist/event absence — a persist-on-refusal mutant ships green and the next cold restore seeds itpackages/cli/src/serve/scheduled-task-keepalive.ts:245 — [review] keepalive re-attaches the ⏰ name over a tombstoned clear after every restart — restore drops the tombstone, the guard never sees itpackages/core/src/services/chatRecordingService.ts:1235 — [probe] tombstone counter seeding untested — a truthiness revert ships green and re-opens the resurrection window at 32–64KB depth
Convergence: round 11 posted 7 inline comment(s), 5 of them reported for the first time; the previous round posted 4. Findings keep coming back to the same files: packages/acp-bridge/src/bridgeClient.ts (findings in round 10; 1 more now). A cluster that keeps producing siblings usually means the fixes are treating instances of a shared root cause — triaging that cause before the next round, or splitting an independent cluster into its own pull request, tends to end the loop faster than fixing them one at a time. (Observation only — nothing was withheld from this review because of this observation.)
Mechanism health: this round did not close cleanly, so it withholds the incremental anchor — and the round it recovered had none either, so the next review re-reads the whole diff and will keep doing so until a round's marker carries an anchor again. (Stated, not acted on — this changes nothing about what the round posts.)
中文说明
仅完成部分审查，审查缺口已披露。
未审查：reverse audit — reached the 5-round cap without two consecutive dry rounds (rounds 3–5 each reported findings; all 15 chunks audited every round)。
未审查：build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally。
未探索到全部深度（达到工具调用预算）：chunk 7：running packages/cli/src/serve/scheduled-task-keepalive.test.ts and the two new server.test.ts cases — the worktree has no node_modules (vitest cannot sta…；chunk 2：executed the new bridge tests not run — worktree lacks node_modules; npm install + build was declined to avoid timeout/interference in the shared review tree.。
Test Plan（非阻断）：2 passed — this review observed 1694, 23292, 20827, 1653, 4137, 605, 495 passed。
收敛姿态下延后（第 11 轮，非阻断）——已记录，本轮不要求修改：共 8 条（原文未翻译，列表见上方英文部分）。
收敛情况：第 11 轮发布了 7 条行内评论，其中 5 条是首次提出；上一轮发布了 4 条。发现反复回到同一批文件：packages/acp-bridge/src/bridgeClient.ts（第 10 轮已出过发现，本轮又有 1 条）。一个不断再生兄弟发现的簇，通常意味着逐条修复只在处理同一根因的实例——先定位并处理该根因，或把独立的簇拆成单独的 PR，通常比逐条修复更快结束循环。（仅为观察——本轮评审未因此扣留任何内容。）
机制健康：本轮未能干净收尾，因而扣留了增量锚点，而它恢复到的那一轮也没有锚点，因此下一次评审将重读整个 diff——并会一直如此，直到某一轮的标记重新带上锚点。（仅陈述，不据此行动——这不改变本轮发布的任何内容。）
— qwen3.8-max via Qwen Code /review (v0.22.0)
| // while a cleared/absent name is machine handling ('auto'). | ||
| const result = await state.client.updateSessionMetadata(sessionId, { | ||
| displayName: args.display_name, | ||
| titleSource: args.display_name?.trim() ? 'manual' : 'auto', |
There was a problem hiding this comment.
[Critical] R4-1: Still standing at this head — the absent-name branch of this daemon-MCP tool remains a silent no-op end-to-end, awaiting the maintainer decision (open since round 4). When display_name is omitted, JSON serialization drops the undefined displayName, the wire body is exactly {"titleSource":"auto"}, and bridge.updateSessionMetadata performs every mutation inside if (metadata.displayName !== undefined) — so the route validates and answers 200 while nothing is updated, persisted, or published, yet the tool reports success and its new comment claims the absent case is machine handling. Re-verified at this commit by probe: an absent-name call against a manually named session leaves the effective metadata unchanged ({displayName: 'My manual name', titleSource: 'manual'}), with no persist and no event. The two routes also disagree on the identical wire body (the workspace route 400s 'at least one of displayName or pr is required'; the live route no-ops). Re-stating the two decision shapes from the open thread: (a) honor titleSource-only updates against the entry's existing title in the bridge (outside the displayName !== undefined gate), or (b) 400-reject a displayName-less PATCH in both routes and stop the tool sending titleSource. Both change the public daemon contract — this stays human-gated per the standing authorization.
中文说明
R4-1：在当前 head 上仍然存在 —— 这个 daemon-MCP 工具的"缺省名字"分支端到端仍是静默空操作，等待维护者决定（自第 4 轮起开放）。当省略 display_name 时，JSON 序列化丢弃 undefined 的 displayName，线上请求体恰好是 {"titleSource":"auto"}，而 bridge.updateSessionMetadata 的所有变更都在 if (metadata.displayName !== undefined) 内执行 —— 因此路由校验通过并返回 200，但没有任何内容被更新、持久化或发布，而工具却报告成功，其新增注释还声称缺省名字属于"机器处理"。已在本提交上用探针重新验证：对一个手动命名的会话发起缺省名字调用，有效元数据保持不变（{displayName: 'My manual name', titleSource: 'manual'}），无持久化、无事件。两条路由对同一线上请求体的处理也不一致（workspace 路由返回 400 'at least one of displayName or pr is required'；live 路由则是空操作）。重申开放线程中的两个决定形态：(a) 在 bridge 中针对条目现有标题受理仅含 titleSource 的更新（放在 displayName !== undefined 门禁之外），或 (b) 在两条路由中对不含 displayName 的 PATCH 返回 400，并让工具停止发送 titleSource。两者都会变更公开的 daemon 契约 —— 按既有授权，此项保持人工裁决。
— qwen3.8-max via Qwen Code /review (v0.22.0)
There was a problem hiding this comment.
Re-verified on current head c2093f5468d69d95503a74feeda127771cfb03a9: omitting display_name serializes a titleSource-only PATCH, while the live bridge mutates only inside the displayName !== undefined gate and the workspace route rejects the equivalent absent-name request. This is the same public metadata-contract decision as R4-1 (honor titleSource-only updates vs reject the request), so I am leaving this thread unresolved with no code change.\n\n
|
Closeout — round 12 (cap-4 of the round-11 findings) Pushed Fixed
Verification: bridge.test.ts 785, bridgeClient.test.ts 117, StandaloneSessionPicker.test.tsx 20, server.test.ts metadata describe 42 — all green; acp-bridge + cli typecheck clean; eslint/prettier clean on all touched files. Deferred to next round: R11-3 (side-task fork seeding) and R11-4 (load-path clear resurrection) — left unresolved, untouched this round. Human-gated (unchanged): the 8 R4-1 family threads remain deliberately unresolved, awaiting the maintainer decision between the two contract-changing shapes (honor titleSource-only updates in the bridge under the |
…ded (#8977) Side-task forks were seeded with titleSource 'manual' because the web-shell passes the generic panel label as the fork name; the first-prompt rename then carries 'auto' and the manual→auto downgrade guard silently discards it while the panel marks the rename applied. Seed side-task forks with 'auto' (the label is a UI placeholder, not a user-chosen name) and make the panel verify the rename result actually carries the requested title before marking it complete.
…tle (#8977) The load-gated merge fell back to the current connection's title whenever the load response carried none, so a title cleared server-side while SSE was down resurrected with its 'manual' provenance on reconnect (the missed clear event is never redelivered; SSE resumes from the load watermark). Snapshot the connection title before the load call and carry the current fields only when something updated them during the load window (a metadata event replayed by a live-journal repair, or a rename that landed mid-load); otherwise a title-less load response clears both fields like the event-fold's null-marker gate.
|
Closeout — cap-2 round for the two deferred round-11 Criticals (push bb31c27..4d805f4):
Verification: SideTaskPanel 18/18, acpAgent side-task tests, DaemonSessionProvider 240/240; cli + webui + web-shell typechecks clean. Disclosure: R11-3 spans 4 files (2 production + 2 collocated tests across cli/web-shell) — the bot's own suggested fix is two-part across the two packages and the existing CLI test asserted The 8 standing R4-1 threads remain deliberately unresolved — awaiting the maintainer decision on whether an explicit absent-name demote may override the same-text downgrade guard. |
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
There was a problem hiding this comment.
Partially reviewed — gaps disclosed.
Not reviewed: build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally.
Not explored to full depth (tool budget reached): chunk 3: could not execute the two new bridge tests — the review worktree has no node_modules (vitest absent) and a full monorepo install exceeds this chunk's tool bud…; "agent reverse-audit (round 2)": none — no check was cut short (omitted per format: nothing unfinished).
Not reviewed: reverse audit — stopped before round 5 by the review time budget.
Test Plan (not a blocker): 2 passed — this review observed 1697, 23298, 20827, 1653, 4139, 606, 495 passed.
Deferred under the convergence posture (round 12, not a blocker) — recorded, not requested in this round:
packages/cli/src/serve/scheduled-task-keepalive.ts:249 (+2 locations) — [review] legacy machine-named populations (scheduled-task ⏰ names, side-task placeholders) frozen by the new downgrade guardpackages/web-shell/client/components/artifacts/SideTaskPanel.tsx:266 — [probe] discarded-rename detector vs daemon truncation mismatch (200 code points vs 256 code units) → false 'Failed to name side task'packages/core/src/services/chatRecordingService.ts:2157 — [review] auto-title tombstone carve-out untested — reverting && onDisk.title ships green, cleared sessions would never re-titlepackages/web-shell/client/App.tsx:8015 — [review] carry-invalidation at workspace switch and sidebar open untestedpackages/cli/src/ui/hooks/useResumeCommand.ts:197 — [review] /resume tombstone falsy-gate change untested — reverting to ?? null ships green and resumes cleared sessions with a blank namepackages/cli/src/serve/create-sub-session.ts:447 — [review] restoreSessionTitleFields tombstone branch untested at every seeding site — a presence-gate regression ships green and seeds an empty manual namepackages/sdk-typescript/src/daemon-mcp/serve-bridge/tools/session.ts:151 — [review] MCP tool's titleSource classification inert in every branch; its comment asserts the opposite of what persistspackages/webui/src/daemon/session/DaemonSessionProvider.tsx:2251 — [review] titleRefreshedDuringLoad true branch untested — a broken refresh detection ships green and wipes mid-window renamespackages/acp-bridge/src/bridge.test.ts:26491 — [review] test comment attributes the empty-title filter to the wrong function (reader contract is the opposite)packages/acp-bridge/src/bridge.test.ts:26664 — [review] echo-suppression window's close side untested — a persist-coalescing mutant drops the tombstone with every test greenpackages/cli/src/serve/create-sub-session.ts:443 — [review] sent-mode title-lookup failure fallback (miss + throw branches) untestedpackages/acp-bridge/src/bridge.test.ts:24966 — [review] clearing test does not pin the rebroadcast event's titleSource absence — a stale-provenance publish mutant ships greenpackages/core/src/services/chatRecordingService.ts:1240 — [probe] tombstone counter seeding untested — a truthiness revert ships green and re-opens the resurrection window at 32–64KB depthpackages/web-shell/client/App.test.tsx:10803 — [review] carry-rename failure path untested — moving token consumption into a failure/finally path ships greenpackages/acp-bridge/src/bridge.test.ts:26785 — [review] downgrade-refusal tests pin no persist/event absence — a persist-on-refusal mutant ships green and the next cold restore seeds itpackages/cli/src/serve/acp-http/transport.test.ts:4813 — [review] prototype-level getSessionTitleInfo spy restored without try/finally — leaks into later tests on an assertion failure
Convergence: round 12 posted 2 inline comment(s), 1 of them reported for the first time; the previous round posted 7 (5 new). Findings keep coming back to the same files: packages/cli/src/serve/routes/session.ts (findings in round 11; 1 more now). A cluster that keeps producing siblings usually means the fixes are treating instances of a shared root cause — triaging that cause before the next round, or splitting an independent cluster into its own pull request, tends to end the loop faster than fixing them one at a time. (Observation only — nothing was withheld from this review because of this observation.)
Mechanism health: this round did not close cleanly, so it withholds the incremental anchor — and the round it recovered had none either, so the next review re-reads the whole diff and will keep doing so until a round's marker carries an anchor again. (Stated, not acted on — this changes nothing about what the round posts.)
中文说明
仅完成部分审查，审查缺口已披露。
未审查：build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally。
未探索到全部深度（达到工具调用预算）：chunk 3：could not execute the two new bridge tests — the review worktree has no node_modules (vitest absent) and a full monorepo install exceeds this chunk's tool bud…；"agent reverse-audit (round 2)"：none — no check was cut short (omitted per format: nothing unfinished)。
未审查：反向审计——评审时间预算不足，未能开始第 5 轮。
Test Plan（非阻断）：2 passed — this review observed 1697, 23298, 20827, 1653, 4139, 606, 495 passed。
收敛姿态下延后（第 12 轮，非阻断）——已记录，本轮不要求修改：共 16 条（原文未翻译，列表见上方英文部分）。
收敛情况：第 12 轮发布了 2 条行内评论，其中 1 条是首次提出；上一轮发布了 7 条（其中 5 条首次提出）。发现反复回到同一批文件：packages/cli/src/serve/routes/session.ts（第 11 轮已出过发现，本轮又有 1 条）。一个不断再生兄弟发现的簇，通常意味着逐条修复只在处理同一根因的实例——先定位并处理该根因，或把独立的簇拆成单独的 PR，通常比逐条修复更快结束循环。（仅为观察——本轮评审未因此扣留任何内容。）
机制健康：本轮未能干净收尾，因而扣留了增量锚点，而它恢复到的那一轮也没有锚点，因此下一次评审将重读整个 diff——并会一直如此，直到某一轮的标记重新带上锚点。（仅陈述，不据此行动——这不改变本轮发布的任何内容。）
— qwen3.8-max via Qwen Code /review (v0.22.0)
| const result = await state.client.updateSessionMetadata(sessionId, { | ||
| displayName: args.display_name, | ||
| titleSource: args.display_name?.trim() ? 'manual' : 'auto', | ||
| }); |
There was a problem hiding this comment.
[Critical] R4-1: Still standing at this head (open since round 4, human-gated by the author awaiting a maintainer decision) — the absent-name branch of this daemon-MCP tool remains a silent no-op end-to-end. When display_name is omitted, JSON serialization drops the undefined displayName, the wire body is exactly {"titleSource":"auto"}, and bridge.updateSessionMetadata performs every mutation inside if (metadata.displayName !== undefined) — so the route validates and answers 200 while nothing is updated, persisted, or published, yet the tool reports success. The two routes also disagree on the identical wire body: the workspace route 400s with "at least one of displayName or pr is required", while the live route no-ops. Re-verified at this commit by trace (issue-fidelity walk plus a verification trace of all three branches); probed live in rounds 8–11 with identical results; the code is unchanged since.
Witness: not run this round — mechanism re-verified at this HEAD by trace (every mutation sits inside the displayName !== undefined gate and JSON drops the undefined displayName); probed live in rounds 8-11, code unchanged.
Two decision shapes from the open thread: (a) honor titleSource-only updates against the entry's existing title in the bridge (outside the displayName !== undefined gate), or (b) 400-reject a displayName-less PATCH in both routes and stop the tool sending titleSource. Both change the public daemon contract.
中文说明
[Critical/严重] R4-1：在当前 head 上仍然存在（自第 4 轮起开放，作者已按授权人工挂起、等待维护者决定）—— 该 daemon-MCP 工具的"缺省名字"分支端到端仍是静默空操作。当省略 display_name 时，JSON 序列化丢弃 undefined 的 displayName，线上请求体恰好是 {"titleSource":"auto"}，而 bridge.updateSessionMetadata 的所有变更都在 if (metadata.displayName !== undefined) 内执行 —— 因此路由校验通过并返回 200，但没有任何内容被更新、持久化或发布，工具却报告成功。两条路由对同一线上请求体的处理也不一致：workspace 路由返回 400（"at least one of displayName or pr is required"），live 路由则是空操作。已在本提交上通过追踪重新验证（问题保真走查 + 三个分支的验证追踪）；第 8–11 轮曾用活体探针得到相同结果；此时代码未变。
证据：本轮未运行 —— 机制已在此 HEAD 上通过追踪重新验证（所有变更都在 displayName !== undefined 门禁内，且 JSON 会丢弃 undefined 的 displayName）；第 8-11 轮曾活体探针验证，代码未变。
开放线程中的两个决定形态：(a) 在 bridge 中针对条目现有标题受理仅含 titleSource 的更新（放在 displayName !== undefined 门禁之外），或 (b) 两条路由对不含 displayName 的 PATCH 返回 400，并让工具停止发送 titleSource。两者都会变更公开的 daemon 契约。
— qwen3.8-max via Qwen Code /review (v0.22.0)
There was a problem hiding this comment.
Re-verified on current head c2093f5468d69d95503a74feeda127771cfb03a9: omitting display_name serializes a titleSource-only PATCH, while the live bridge mutates only inside the displayName !== undefined gate and the workspace route rejects the equivalent absent-name request. This is the same public metadata-contract decision as R4-1 (honor titleSource-only updates vs reject the request), so I am leaving this thread unresolved with no code change.\n\n
|
Closeout — round on head R12-31 FIXED ( Verification: R12-31 thread replied with evidence + resolved. The 9 R4-1 threads remain deliberately open — human-gated, awaiting a maintainer decision on the daemon-MCP tool's absent-name demote semantics. |
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
There was a problem hiding this comment.
Partially reviewed — gaps disclosed.
Not reviewed: build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally.
Not reviewed: build-and-test — Test (macos-latest, Node 22.x) and Test (windows-latest, Node 22.x) unit-matrix jobs were skipped in CI (Linux unit suites ran for all changed workspaces).
Not reviewed: test-efficacy — probe harness could not be validated (vitest prerequisite guard: all per-file probes inconclusive, harnessValidated null; mutants/hunks not run).
Not explored to full depth (tool budget reached): "agent reverse-audit (round 3)": verify whether the daemon load route populates DaemonSessionState.displayName from a source that can diverge from the session title record (would desync the pro….
Deferred under the convergence posture (round 13, not a blocker) — recorded, not requested in this round:
packages/webui/src/daemon/session/DaemonSessionProvider.tsx:2249 — [review] titleRefreshedDuringLoad positive branch untested (mutation green)packages/core/src/services/chatRecordingService.ts:2157 — [review] auto-title tombstone carve-out untested (probe flip)packages/acp-bridge/src/bridge.ts:9959 — [review] .catch counter decrement untested (mutation green; probe flips)packages/cli/src/ui/hooks/useResumeCommand.ts:197 — [review] TUI resume tombstone falsy gate untested (mutation green)packages/web-shell/client/App.tsx:8403 — [review] carry invalidation on session switch untested (probe reproduces stale carry)packages/web-shell/client/App.tsx:5648 — [review] carried rename reject semantics pinned by zero testspackages/acp-bridge/src/bridge.test.ts:26661 — [review] counter success-path decrement / post-settle acceptance unpinnedpackages/acp-bridge/src/bridge.test.ts:26654 — [review] suppressed echoes emitting no event unpinnedpackages/cli/src/acp-integration/acpAgent.ts:12022 — [review] fork ternary 'manual' arm unpinnedpackages/cli/src/serve/acp-http/transport.test.ts:4857 (+2 locations) — [review] prototype spy restored only on success path (no try/finally)packages/core/src/services/chatRecordingService.customTitle.test.ts:860 — [review] restore-time presence-seeding unpinned (probe flip)packages/core/src/services/chatRecordingService.ts:2153 — [review] autoTitleAttempts never reset on clear (probe + fix run)packages/acp-bridge/src/bridge.test.ts:26770 — [review] downgrade-refusal no-persist/no-event unpinned (probe)packages/acp-bridge/src/bridge.test.ts:26579 — [review] clear test cold-restore section structurally vacuouspackages/cli/src/serve/session-restore-title.ts:13 — [review] empty-title branch pinned by zero tests (457 tests green under mutant)packages/sdk-typescript/test/unit/serve-bridge.test.ts:349 — [review] session_update_metadata .trim() classification untestedpackages/web-shell/client/App.test.tsx:11169 — [review] carry identity consumption unpinned (deletion mutation green)packages/acp-bridge/src/bridge.test.ts:24966 — [review] source-less echo published envelope unpinned (probe flip)packages/cli/src/serve/server.test.ts:26893 — [review] auto→auto re-rename through the fallback unpinned (probe flip)packages/web-shell/client/App.test.tsx:10929 — [review] cleared-prompt-drop half of the /clear-during-creation window unpinned (probe flip)
Convergence: round 13 posted 3 inline comment(s), 2 of them reported for the first time; the previous round posted 2 (1 new). The rate of new findings is not falling. Batching the remaining fixes and verifying them before the next push keeps the loop from re-deriving the same set; this PR's reviews already resolve to a critical posting floor. (Observation only — nothing was withheld from this review because of this observation.)
Mechanism health: this round did not close cleanly, so it withholds the incremental anchor — and the round it recovered had none either, so the next review re-reads the whole diff and will keep doing so until a round's marker carries an anchor again. (Stated, not acted on — this changes nothing about what the round posts.)
中文说明
仅完成部分审查，审查缺口已披露。
未审查：build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally。
未审查：build-and-test — Test (macos-latest, Node 22.x) and Test (windows-latest, Node 22.x) unit-matrix jobs were skipped in CI (Linux unit suites ran for all changed workspaces)。
未审查：test-efficacy — probe harness could not be validated (vitest prerequisite guard: all per-file probes inconclusive, harnessValidated null; mutants/hunks not run)。
未探索到全部深度（达到工具调用预算）："agent reverse-audit (round 3)"：verify whether the daemon load route populates DaemonSessionState.displayName from a source that can diverge from the session title record (would desync the pro…。
收敛姿态下延后（第 13 轮，非阻断）——已记录，本轮不要求修改：共 20 条（原文未翻译，列表见上方英文部分）。
收敛情况：第 13 轮发布了 3 条行内评论，其中 2 条是首次提出；上一轮发布了 2 条（其中 1 条首次提出）。新发现的产出速度没有下降。把剩余修复攒成一批、验证后再推送，可以避免循环反复推导同一组发现；本 PR 的评审已解析为 critical 发布下限。（仅为观察——本轮评审未因此扣留任何内容。）
机制健康：本轮未能干净收尾，因而扣留了增量锚点，而它恢复到的那一轮也没有锚点，因此下一次评审将重读整个 diff——并会一直如此，直到某一轮的标记重新带上锚点。（仅陈述，不据此行动——这不改变本轮发布的任何内容。）
— qwen3.8-max via Qwen Code /review (v0.22.0)
| const result = await state.client.updateSessionMetadata(sessionId, { | ||
| displayName: args.display_name, | ||
| titleSource: args.display_name?.trim() ? 'manual' : 'auto', | ||
| }); |
There was a problem hiding this comment.
[Critical] R4-1: Still standing at this head (open since round 4, human-gated by the author awaiting a maintainer contract decision) — the daemon-MCP session_update_metadata absent-name branch is a silent no-op end-to-end, while the sibling workspace route rejects the identical body with 400.
The tool stamps titleSource: args.display_name?.trim() ? 'manual' : 'auto'; with display_name absent it sends {titleSource: 'auto'} with no displayName. Every consumer gates all title logic on displayName !== undefined: the live route forwards to bridge.updateSessionMetadata, whose entire rename branch is skipped — no entry mutation, no session_metadata_updated event, no persist — and the handler answers 200 with unchanged metadata. PATCH /workspaces/:workspace/session/:id/metadata rejects the identical body with 400 ("at least one of displayName or pr is required"), so the two metadata routes disagree on the contract this PR's test pins as meaningful "machine handling". An MCP agent that renamed a session and later wants to release the name back to machine provenance (e.g. to disarm the web-shell /clear manual-carry gate this PR builds) calls this tool without display_name, receives 200, and the session silently keeps its manual provenance — the carry gate keeps arming on every /clear.
Witness (wire probe, run; gate re-verified at head f9bf861):
real MCP tool handler driven through the real DaemonClient,
recording fetch:
PATCH .../session/s1/metadata body {"displayName":"Agent chosen"}
with titleSource present? false
with a tool-level fix the body gains "titleSource":"auto" (probe flips)
Decision needed (author/maintainer): if an absent name means "no change", have the tool reject a call with neither display_name nor other content (mirroring the workspace route's 400) and update the test to assert the rejection; if it should clear/downgrade the title, send displayName: '' (the tombstone path the bridge's clear branch understands) and assert the resulting entry state.
中文说明
[严重] R4-1：在当前 head 仍然存在（自第 4 轮起开放，作者已标记为人工门禁、等待维护者的契约决定）——daemon-MCP session_update_metadata 的缺省名字分支端到端是一个静默空操作，而姊妹 workspace 路由对相同的请求体返回 400。
该工具打上 titleSource: args.display_name?.trim() ? 'manual' : 'auto'；当 display_name 缺省时发送不带 displayName 的 {titleSource: 'auto'}。所有消费方都把全部标题逻辑置于 displayName !== undefined 门禁之内：live 路由转发给 bridge.updateSessionMetadata，其整个重命名分支被跳过——无条目变更、无 session_metadata_updated 事件、无持久化——处理器返回 200 和未变更的元数据。PATCH /workspaces/:workspace/session/:id/metadata 对相同请求体返回 400（"至少需要 displayName 或 pr 之一"），两个 metadata 路由对本 PR 测试所固化的"机器处理"契约不一致。一个 MCP agent 重命名会话后想把名字交还给机器来源（例如解除本 PR 构建的 web-shell /clear 手动携带门禁），不带 display_name 调用该工具，收到 200，而会话静默保留 manual 来源——携带门禁在每次 /clear 时继续武装。
证据（线上探针，已运行；门禁已在 head f9bf861 复核）：
用真实 DaemonClient 驱动真实 MCP 工具处理器，记录式 fetch：
PATCH .../session/s1/metadata 请求体 {"displayName":"Agent chosen"}
titleSource 存在? 否
加上工具层修复后请求体带 "titleSource":"auto"（探针翻转）
需要决定（作者/维护者）：若缺省名字意味着"不变"，让工具拒绝既无 display_name 又无其他内容的调用（与 workspace 路由的 400 对齐），并更新测试断言拒绝；若应当清除/降级标题，发送 displayName: ''（bridge 清除分支理解的 tombstone 路径），并断言结果条目状态。
— qwen3.8-max via Qwen Code /review (v0.22.0)
There was a problem hiding this comment.
Re-verified on current head c2093f5468d69d95503a74feeda127771cfb03a9: omitting display_name serializes a titleSource-only PATCH, while the live bridge mutates only inside the displayName !== undefined gate and the workspace route rejects the equivalent absent-name request. This is the same public metadata-contract decision as R4-1 (honor titleSource-only updates vs reject the request), so I am leaving this thread unresolved with no code change.\n\n
|
Closeout — round-13 cap (2 new Criticals fixed, one non-force push
Verification: SideTaskPanel 19/19, App.test.tsx 521/521, web-shell typecheck clean. Both threads replied with SHA evidence + resolved (10 unresolved = the standing R4-1 family, still awaiting the maintainer contract decision — untouched). |
|
Refreshed the branch onto latest main to recover the exact-head Java 11 SDK failure. Changed: ordinary merge commit only; no product files were hand-edited for the Java 11 close/detach failure. |
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
There was a problem hiding this comment.
Partially reviewed — gaps disclosed.
Not reviewed: build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally.
Not reviewed: build-and-test — Test (macos-latest, Node 22.x) and Test (windows-latest, Node 22.x) unit-matrix jobs were skipped in CI (Linux unit suites ran for all changed workspaces).
Not reviewed: test-efficacy — probe harness could not be validated (vitest prerequisite guard: all per-file probes inconclusive, harnessValidated null; mutants/hunks not run).
Not explored to full depth (tool budget reached): "agent reverse-audit (round 2)": trace the daemon-side construction of the load response state field to confirm it can actually carry displayName (if the daemon never puts the title in sta…; "agent reverse-audit (round 2)": read appendRecordStrict 's body (chatRecordingService.ts:1428) to confirm the write is enqueued before its first await — the last link in finding 2's race; the….
Not reviewed: reverse audit — stopped before round 5 by the review time budget.
Test Plan (not a blocker): 2 passed — this review observed 1697, 23509, 21020, 1654, 4175, 610, 495 passed.
Deferred under the convergence posture (round 14, not a blocker) — recorded, not requested in this round:
packages/core/src/services/chatRecordingService.ts:2198 — [probe] auto-title tombstone carve-out untested (probe flips)packages/acp-bridge/src/bridge.ts:9963 — [probe] .catch counter decrement untested (mutation green; probe flips)packages/cli/src/ui/hooks/useResumeCommand.ts:197 — [probe] TUI resume tombstone falsy gate untested (mutation green)packages/web-shell/client/App.tsx:8609 — [probe] carry invalidation on session switch untested (probe reproduces stale carry)packages/acp-bridge/src/bridge.test.ts:26664 — [probe] counter success-path decrement / post-settle acceptance unpinned (probe flips)packages/cli/src/serve/server.test.ts:26712 — [probe] manual-rename fallback arm against persisted manual title untested (probe flips)packages/cli/src/serve/server.test.ts:26773 — [probe] legacy source-less seed arm of the fallback guard untested (probe flips)packages/sdk-typescript/test/unit/serve-bridge.test.ts:350 — [probe] session_update_metadata whitespace classification arm untested (probe flips)packages/web-shell/client/App.test.tsx:11815 — [review] /new-during-creation test never asserts the abandoned prompt is not sentpackages/web-shell/client/components/artifacts/SideTaskPanel.tsx:270 — [review] 256-unit acceptance vs 200-unit child persist cap — applied rename never reaches diskpackages/acp-bridge/src/bridge.test.ts:1166 — [probe] child-supplied branch titleSource pass-through untested (mutant green; probe flips)packages/cli/src/serve/session-restore-title.ts:13 — [review] empty-title truthiness branch pinned by zero tests at seven call sitespackages/core/src/services/chatRecordingService.customTitle.test.ts:860 — [probe] restore-time presence-seeding unpinned (probe flips)packages/acp-bridge/src/bridge.test.ts:24966 — [review] source-less echo published envelope unpinnedpackages/acp-bridge/src/bridge.test.ts:26786 — [review] downgrade-refusal no-persist coupling unpinnedpackages/acp-bridge/src/bridge.test.ts:26655 — [review] suppressed echoes emitting no event unpinnedpackages/cli/src/acp-integration/acpAgent.ts:12032 — [probe] fork ternary 'manual' arm unpinned (probe flips)packages/cli/src/serve/scheduled-task-keepalive.ts:351 — [review] keepalive legacy no-source arm unpinnedpackages/cli/src/serve/server.test.ts:27100 — [probe] fallback write default ?? 'manual' unpinnedpackages/web-shell/client/components/artifacts/SideTaskPanel.tsx:271 — [review] truncated echo overwrites tab title; sidebar catalog diverges- …and 7 more (see the run report)
Mechanism health: this round did not close cleanly, so it withholds the incremental anchor — and the round it recovered had none either, so the next review re-reads the whole diff and will keep doing so until a round's marker carries an anchor again. (Stated, not acted on — this changes nothing about what the round posts.)
中文说明
仅完成部分审查，审查缺口已披露。
未审查：build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally。
未审查：build-and-test — Test (macos-latest, Node 22.x) and Test (windows-latest, Node 22.x) unit-matrix jobs were skipped in CI (Linux unit suites ran for all changed workspaces)。
未审查：test-efficacy — probe harness could not be validated (vitest prerequisite guard: all per-file probes inconclusive, harnessValidated null; mutants/hunks not run)。
未探索到全部深度（达到工具调用预算）："agent reverse-audit (round 2)"：trace the daemon-side construction of the load response state field to confirm it can actually carry displayName (if the daemon never puts the title in sta…；"agent reverse-audit (round 2)"：read appendRecordStrict 's body (chatRecordingService.ts:1428) to confirm the write is enqueued before its first await — the last link in finding 2's race; the…。
未审查：反向审计——评审时间预算不足，未能开始第 5 轮。
Test Plan（非阻断）：2 passed — this review observed 1697, 23509, 21020, 1654, 4175, 610, 495 passed。
收敛姿态下延后（第 14 轮，非阻断）——已记录，本轮不要求修改：共 27 条（原文未翻译，列表见上方英文部分）。
机制健康：本轮未能干净收尾，因而扣留了增量锚点，而它恢复到的那一轮也没有锚点，因此下一次评审将重读整个 diff——并会一直如此，直到某一轮的标记重新带上锚点。（仅陈述，不据此行动——这不改变本轮发布的任何内容。）
— qwen3.8-max via Qwen Code /review (v0.22.0)
| const result = await state.client.updateSessionMetadata(sessionId, { | ||
| displayName: args.display_name, | ||
| titleSource: args.display_name?.trim() ? 'manual' : 'auto', | ||
| }); |
There was a problem hiding this comment.
[Critical] R4-1: Still standing at this head (open since round 4, human-gated by the author awaiting a maintainer contract decision) — the absent-name branch of this daemon-MCP tool remains a silent no-op end-to-end. When display_name is omitted, JSON serialization drops the undefined displayName, so the wire body is exactly {"titleSource":"auto"}; PATCH /session/:id/metadata validates it and answers 200, but bridge.updateSessionMetadata performs every mutation inside if (metadata.displayName !== undefined) — no state change, no persistence, no session_metadata_updated event. The sibling workspace route rejects the identical effective body with 400 (at least one of displayName or pr is required), so clear/machine-handling behavior stays route-dependent. Both candidate fixes change the public daemon contract (route/bridge-level parity guard, or 400-reject a displayName-less PATCH and stop the tool sending titleSource), which is why this remains human-gated; the serve-bridge test's effective-metadata assertion can land with whichever direction is chosen.
Witness (probe this round, scratch tree at the reviewed commit — real bridge driven with exactly the body the primary route forwards):
seeded effective: {"displayName":"Named session","titleSource":"manual"}
titleSource-only body applied: threw: no
effective after: {"displayName":"Named session","titleSource":"manual"} (unchanged)
new metadata events: 0; new title persists: 0
flip (workspace-route parity guard patched into bridge.ts):
InvalidSessionMetadataError: at least one of displayName or pr is required
中文说明
R4-1：在该被审头部仍然存在（自第 4 轮起开放，作者按人工门禁等待维护者契约决定）—— 该 daemon-MCP 工具的「缺省名字」分支依然是端到端静默无操作。当 display_name 缺省时，JSON 序列化丢掉 undefined 的 displayName，线上请求体恰为 {"titleSource":"auto"}；PATCH /session/:id/metadata 校验通过并返回 200，但 bridge.updateSessionMetadata 的所有变更都在 if (metadata.displayName !== undefined) 之内 —— 无状态变更、无持久化、无 session_metadata_updated 事件。同级 workspace 路由对同一等效请求体以 400 拒绝（at least one of displayName or pr is required），因此清空/机器处理语义仍依赖路由。两个候选修复都会变更公开 daemon 契约（路由/bridge 层面对齐保护，或以 400 拒绝无 displayName 的 PATCH 并让工具停发 titleSource），故继续保持人工门禁；serve-bridge 测试的 effective-metadata 断言可随所选方向一并落地。
证据（本轮探针，被审提交的 scratch tree —— 以主路由实际转发的请求体驱动真实 bridge）：注入后 effective 为 {"displayName":"Named session","titleSource":"manual"}；仅 titleSource 的请求体应用后未抛错、effective 不变、0 条新事件、0 条新持久化；把 workspace 路由的对齐保护补丁进 bridge.ts 后探针翻转为 InvalidSessionMetadataError。
— qwen3.8-max via Qwen Code /review (v0.22.0)
There was a problem hiding this comment.
Re-verified on current head c2093f5468d69d95503a74feeda127771cfb03a9: omitting display_name serializes a titleSource-only PATCH, while the live bridge mutates only inside the displayName !== undefined gate and the workspace route rejects the equivalent absent-name request. This is the same public metadata-contract decision as R4-1 (honor titleSource-only updates vs reject the request), so I am leaving this thread unresolved with no code change.\n\n
# Conflicts: # packages/cli/src/serve/live/live-task-service.ts # packages/cli/src/serve/routes/scheduled-tasks.ts
|
Merged the latest |
qwen-code-ci-bot
left a comment
qwen-code-ci-bot
left a comment
There was a problem hiding this comment.
Partially reviewed — gaps disclosed.
Not reviewed: build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally.
Not reviewed: build-and-test — Test (macos-latest, Node 22.x) and Test (windows-latest, Node 22.x) unit-matrix jobs were skipped in CI (Linux unit suites ran for all changed workspaces).
Not reviewed: test-efficacy — probe harness could not be validated (vitest build-prerequisite guard: all revert probes inconclusive, mutants/hunks not run).
Not reviewed: reverse audit — stopped before round 2 by the review time budget.
Test Plan (not a blocker): 2 passed — this review observed 1700, 23700, 21191, 1659, 4243, 637, 601 passed.
Deferred under the convergence posture (round 15, not a blocker) — recorded, not requested in this round:
packages/core/src/services/chatRecordingService.ts:2197 — [probe] auto-title tombstone carve-out untested (probe flips)packages/cli/src/ui/hooks/useResumeCommand.ts:197 — [probe] TUI resume tombstone falsy gate untested (probe flips)packages/acp-bridge/src/bridge.ts:9963 — [probe] pendingTitlePersistCount decrements untested (probe flips)packages/cli/src/acp-integration/acpAgent.ts:11091 — [probe] fork ternary 'manual' arm unpinned (probe flips)packages/web-shell/client/App.tsx:8777 — [probe] carry invalidation on session switch untested (probe flips)packages/web-shell/client/components/artifacts/SideTaskPanel.tsx:255 — [review] discarded-rename failure path never reverts the optimistic title; misleading toastpackages/acp-bridge/src/bridge.ts:9985 — [review] displayName:null clear marker vs SDK string|undefined types (latent)packages/sdk-typescript/test/unit/serve-bridge.test.ts:350 — [probe] whitespace-only classification arm untested (probe flips)packages/acp-bridge/src/bridge.test.ts:26748 — [probe] echo-suppression lift-after-settle unpinned (probe flips)packages/acp-bridge/src/bridge.test.ts:26862 — [review] downgrade-refusal no-persist unpinnedpackages/cli/src/serve/session-restore-title.ts:13 — [review] restoreSessionTitleFields has no direct testpackages/web-shell/client/App.test.tsx:11759 — [review] rename-failure teardown path unpinnedpackages/core/src/services/chatRecordingService.customTitle.test.ts:865 — [probe] projected-state re-anchor test unpinned (probe flips)packages/webui/src/daemon/session/mappers.ts:351 — [review] mapper 'auto' acceptance branch untestedpackages/acp-bridge/src/bridge.test.ts:25053 — [review] omitted-source publish shape unpinnedpackages/core/src/services/chatRecordingService.ts:1210 — [probe] resume-path tombstone counter seeding unpinned (probe flips)
Mechanism health: this round did not close cleanly, so it withholds the incremental anchor — and the round it recovered had none either, so the next review re-reads the whole diff and will keep doing so until a round's marker carries an anchor again. (Stated, not acted on — this changes nothing about what the round posts.)
中文说明
仅完成部分审查，审查缺口已披露。
未审查：build-and-test — Integration Tests (CLI, No Sandbox) was skipped in CI and its suite did not run locally。
未审查：build-and-test — Test (macos-latest, Node 22.x) and Test (windows-latest, Node 22.x) unit-matrix jobs were skipped in CI (Linux unit suites ran for all changed workspaces)。
未审查：test-efficacy — probe harness could not be validated (vitest build-prerequisite guard: all revert probes inconclusive, mutants/hunks not run)。
未审查：反向审计——评审时间预算不足，未能开始第 2 轮。
Test Plan（非阻断）：2 passed — this review observed 1700, 23700, 21191, 1659, 4243, 637, 601 passed。
收敛姿态下延后（第 15 轮，非阻断）——已记录，本轮不要求修改：共 16 条（原文未翻译，列表见上方英文部分）。
机制健康：本轮未能干净收尾，因而扣留了增量锚点，而它恢复到的那一轮也没有锚点，因此下一次评审将重读整个 diff——并会一直如此，直到某一轮的标记重新带上锚点。（仅陈述，不据此行动——这不改变本轮发布的任何内容。）
— qwen3.8-max via Qwen Code /review (v0.22.0)
| // while a cleared/absent name is machine handling ('auto'). | ||
| const result = await state.client.updateSessionMetadata(sessionId, { | ||
| displayName: args.display_name, | ||
| titleSource: args.display_name?.trim() ? 'manual' : 'auto', |
There was a problem hiding this comment.
[Critical] R4-1: Still standing at this head (open since round 4, human-gated by the author awaiting a maintainer contract decision) — the absent-name branch of this daemon-MCP tool remains a silent no-op end-to-end, while the sibling workspace route rejects the identical body with 400. When display_name is omitted, JSON serialization drops the undefined displayName, so the wire body is exactly {"titleSource":"auto"}; PATCH /session/:id/metadata validates it and answers 200, but bridge.updateSessionMetadata performs every mutation inside if (metadata.displayName !== undefined) — no state change, no persistence, no session_metadata_updated event, while the route answers success. An agent that calls this tool without a name receives a success that changed nothing. Re-verified at this head by trace: the tool sends {displayName: args.display_name, titleSource: args.display_name?.trim() ? 'manual' : 'auto'}; the bridge gate is unchanged; the workspace route rejects an empty displayName with 400 (routes/session.ts:5459-5466). Resolution stays the open maintainer decision between the two shapes: honor titleSource-only updates in the bridge under the sameTextDowngrade guard, or 400-reject a displayName-less PATCH and stop the tool sending titleSource.
中文说明
[Critical/严重] R4-1：在本 head 仍然存在（自第 4 轮起开放，由作者人为把关，等待维护者的契约决定）—— 该 daemon-MCP 工具的缺省名字分支端到端仍是静默空操作，而同族的 workspace 路由对相同请求体会以 400 拒绝。当省略 display_name 时，JSON 序列化会丢弃 undefined 的 displayName，线上请求体恰好是 {"titleSource":"auto"}；PATCH /session/:id/metadata 校验通过并返回 200，但 bridge.updateSessionMetadata 的所有变更都在 if (metadata.displayName !== undefined) 之内执行 —— 无状态变更、无持久化、无 session_metadata_updated 事件，而路由却返回成功。agent 在不带名字的情况下调用该工具，得到的是一个没有改变任何内容的“成功”。本轮已通过代码追踪在本 head 重新验证：工具发送 {displayName: args.display_name, titleSource: args.display_name?.trim() ? 'manual' : 'auto'}；bridge 的门禁未变；workspace 路由以 400 拒绝空 displayName（routes/session.ts:5459-5466）。解决方案仍是两种形态之间的维护者决定：在 bridge 中于 sameTextDowngrade 保护下尊重仅带 titleSource 的更新，或对不带 displayName 的 PATCH 返回 400 并让工具停止发送 titleSource。
— qwen3.8-max via Qwen Code /review (v0.22.0)
There was a problem hiding this comment.
Re-checked the complete PR conversation, review history, inline comments, and issue #8977: there is still no maintainer ruling choosing either contract shape. The issue discussion only establishes that manual titles should carry across /clear; it does not define whether a displayName-less daemon metadata PATCH should demote title provenance or be rejected. This remains a public API/data-semantics decision because the live and workspace routes currently disagree on the same effective body. Leaving this thread unresolved and making no code change until that contract is decided.
5 participants
What this PR does
A manually chosen Web Shell session name now survives
/clear. The successor session persists that name before it attaches and before its first prompt, so automatic title generation cannot replace the user's label.Title provenance now survives live updates, page reloads, session reloads, daemon restarts, and every cold-restore path. Only names explicitly marked
manualare carried; automatic titles and legacy names with unknown provenance are not./new,/reset, workspace changes, opening another session, and a newer in-flight session intent invalidate the carry-over.Why it's needed
Fixes #8977. A name deliberately chosen by the user is session identity, and losing it during
/clearmakes labeled sessions indistinguishable. The previous implementation also applied the rename after attach, which left the first prompt and auto-title generation ahead of the manual write.Reviewer Test Plan
How to verify
/rename foo, run/clear, and send the first prompt. Expected: the successor is already namedfoo, and the later automatic-title pass does not replace it./clearand send a prompt. Expected: the persisted manual name still carries to the successor./clearon a session that has never been manually renamed. Expected: no name is carried and the automatic title is generated normally./newor/reset, switch workspaces, or open another session after/clear. Expected: the old manual name is not applied to the unrelated session.Evidence (Before & After)
Before: a browser reproduction against current
mainshowed the manually chosen name replaced by an automatic title after the successor's first prompt.After: the same browser flow passes for
/clear(manual title persisted before the prompt and remains authoritative), while the/newcontrol still receives an automatic title. No visual layout changed, so screenshots are N/A.Tested on
Environment (optional)
Browser E2E: 2/2 passed with the local Web Shell and mock daemon. Focused regression suites for Web Shell, bridge, serve, SDK, and WebUI passed. Repository-wide
npm run build,npm run typecheck, andnpm run lintpassed.Risk & Scope
/newand/resetintentionally retain fresh-session behavior.Linked Issues
Fixes #8977
中文说明
本 PR 做了什么
Web Shell 中用户手动设置的会话名现在会在
/clear后保留。后继会话会在 attach 和第一条 prompt 之前完成名称持久化，因此自动标题生成不会覆盖用户设置的名称。标题来源现在能够跨实时更新、页面刷新、会话重载、daemon 重启以及所有冷恢复路径保留。只有明确标记为
manual的名称才会继承；自动标题和来源未知的旧版名称不会继承。/new、/reset、切换 workspace、打开其它会话，以及更新的并发会话意图都会使旧的继承失效。为什么需要
Fixes #8977。用户主动选择的名称就是会话身份；在
/clear时丢失名称会让已经标记的会话重新变得难以区分。此前的实现还在 attach 之后才重命名，使第一条 prompt 和自动标题生成先于手动名称写入。Reviewer Test Plan
How to verify
/rename foo重命名当前 Web Shell 会话，执行/clear，然后发送第一条 prompt。预期：后继会话已经命名为foo，后续自动标题不会覆盖它。/clear并发送 prompt。预期：持久化的手动名称仍会继承到后继会话。/clear。预期：不会继承名称，自动标题正常生成。/clear后执行/new或/reset、切换 workspace，或打开其它会话。预期：旧的手动名称不会应用到无关会话。Evidence (Before & After)
Before：针对当前
main的浏览器复现显示，手动设置的名称会在后继会话第一条 prompt 后被自动标题替换。After：同一浏览器流程中，
/clear场景通过（手动标题在 prompt 前持久化并保持权威），/new对照场景仍会正常获得自动标题。没有视觉布局变化，因此截图不适用。Tested on
Environment (optional)
使用本地 Web Shell 和 mock daemon 的浏览器 E2E 2/2 通过。Web Shell、bridge、serve、SDK 与 WebUI 的定向回归测试通过。仓库级
npm run build、npm run typecheck与npm run lint通过。Risk & Scope
/new和/reset按设计继续保持全新会话行为。Linked Issues
Fixes #8977