Repository navigation
fix: backport 0.12.1 late batch 5 (21 fixes merged 11-08) - #33229
Conversation
#33132) The Issue Radar Investigate and Review header buttons were disabled while their session started. A disabled button leaves the tab order, so a keyboard user who pressed one lost focus to the page, and a failed start opened its error popover with focus lost. Mark the button aria-disabled and aria-busy while busy and ignore a press then; it stays disabled when blocked for another reason. Original author: Rohan Khanderia (cherry picked from commit a57d5e7)
) A preset tile's apply button held the preview image, with its alt set to the preset name, and a caption with the same name, so its accessible name was "Tuxedo Tuxedo". Each body-part swatch image likewise repeated the caption right below it. Both images are now decorative (alt=""), and the caption alone names the control. The combined "final" preview keeps its alt, since its caption says something else. The coverage tests address presets and swatches by role and label instead of image alt, and a new test pins one name per control. Original author: Rohan Khanderia (cherry picked from commit 426ef59)
#33129) A diagnostic row with a line number is named "Go to line N", and that aria-label replaces the row's content, so a screen reader announced the line and never the error itself. The message span is now linked as the row's aria-describedby, so the name still says where the row goes and the description says what is wrong. Original author: Rohan Khanderia (cherry picked from commit 8d42b86)
…33155) A sub-agent row left in retry_wait with no lease, whose run's tombstone already records a user stop, was never settled: the boot reconcile's queued pass reads retry_wait rows but acts only on the refusal and queued-stop markers, so count_pending kept the parent slot answering 409 slot_subagents_running on every turn until the store was edited by hand. The queued pass now cancels such a row when the stop belongs to the row's current attempt. An attempt is named by the generation of the claim that began it: the registered run keeps it as _taskq_attempt, its tombstone records it as taskq_attempt (carried through the delivery rewrite), and the pass asks the subagent probe about the generation of the row's latest claimed event. A stop from an earlier attempt, or one that names no attempt, leaves the row as it was. The cancel is fenced on retry_wait and the generation read. Builds on #22252 by @cruisercohen. Closes #22244 Original author: Bolin Chen (cherry picked from commit e7836f7)
A send from a little above the bottom glides there smoothly. Two follow paths misread that glide's own mid-flight position. The pre-paint re-pin read it as a scroll-up and released follow; the tail re-baseline read it as a layout clamp and moved the reference the glide's landing is matched against. Either way the glide stopped at the old bottom and the reply streamed on below the reader, and an app switch then restored that spot instead of the live end. Both now stand down while a smooth pin is in flight, as the post-paint pin already does, so the glide's arrival re-targets the live bottom. Closes #29818 Original author: Bolin Chen (cherry picked from commit 8554a00)
UX Review (Opus 5.5) — ✅ PASS (a BLOCK with no Trigger and no Outcome does not block)UX-level review of UX-Verdict: PASS cannot evaluate: missing screenshots of the visible changes in this backport. The PR attaches no image, so the blind read was not performed; "the main PRs show it" is not evidence for this revision. Evidence gaps
Copies of each main PR's own screenshots, attached to this PR's description, would close these gaps. The blind read must then be re-run. [UX-REVIEWED] c849f9c Description read[DESCRIPTION-READ] 33c1df5ca3f6fe6bf0d5eac16869def771697faa096580ed693b71ad32570b95 That sha256 names everything this verdict read: the PR description, plus the 3 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
Design Review (Opus 5.5) — ✅ PASS (a BLOCK with no Trigger and no Outcome does not block)Design-level review of Design-Verdict: PASS cannot evaluate: missing screenshot of the tunnel tile with its new inline error notice, and of the cron-tab Stop-refused state, rendered on BlockersThe reshaped dashboard surfaces have no rendered evidence for this revision Notes
[DESIGN-REVIEWED] c849f9c Description read[DESCRIPTION-READ] 0e0b7d7ccf06374b4613365013fd28b33f4c5a8607d741067e1a13792d25e1c0 That sha256 names everything this verdict read: the PR description, plus the 1 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
First Principles Review (Opus 5.5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Every item is a fix already merged on Notes
What this change shipsInventory (10 items) — 10 justifiedIntent: ship 20 bug fixes already merged on
[FIRST-PRINCIPLES-REVIEWED] c849f9c Description read[DESCRIPTION-READ] 398f1f38f1cf8c4e91d6d69099f1aac69c76aaa64ba9eb17c46286a665626df1 That sha256 names the PR description this verdict read. Recompute it with |
GPT 6.1 Review — ✅ no blocking findingsGPT 6.1 completed its review of This comment is updated in place on each push. Review detailsNo findings. False positive or not applicable? A repository writer can comment: |
Opus 5.5 Review — ✅ no blocking findingsReviewed Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
The knowledge chip on a chat message printed its token count as a number followed by a bare 'tokens' word. English read '1 tokens', and no language could apply its own plural forms or put the noun before the number. The count is now one plural message, written for every shipped language. Original author: Rohan Khanderia (cherry picked from commit d1a9a81)
…#33184) A sub-agent completion that waited out the 1200 s delivery cap reset the parent session unconditionally, so a cron, channel or webhook parent still busy in its own long turn was killed mid-turn. By the time the timeout arm runs, the cancelled delivery route has released any lease it held, so a turn that still holds the session is the parent's own. The arm now resets only when no turn holds the session (sessions.turn_holder) and logs the reset it declined. A stuck injection is still reset, including when its release has just handed the lease to a queued sibling delivery, which then gets a fresh provider. Builds on #26275. Original author: Bolin Chen (cherry picked from commit f216ae2)
`kirocrew cloud destroy` warned that the uploaded source object could not be removed whenever the launch shipped no source (an installed package), printed a cleanup command that fails, and then said all AWS resources were deleted. Only a source upload creates the bucket, so an owner-pinned delete answering NoSuchBucket now counts as removed. When the object really could not be removed, the last line names only the stack. A stack with no InstanceId output (still being created, or a failed create) makes destroy and status print "(id unknown)" instead of a blank. Closes #32101 Original author: Bolin Chen (cherry picked from commit ee677c3)
The Session Breakdown tree's occupancy gauge is decorative
(aria-hidden) and carried its figure ("6% of window") only in a title,
which reaches a pointer and nothing else. A keyboard, touch or screen
reader user never learned how full a sub-agent's window got. Keep the
gauge hidden and add the same figure as visually hidden text, so it
joins the node row's spoken name.
Original author: Rohan Khanderia
(cherry picked from commit b2a4ba3)
#33170) `kirocrew cloud launch` went silent for the 5-10 minute install wait. The on-box log tail waited for an instance id read from the stack's InstanceId output, which only exists once the whole stack completes, so no `[box]` line ever printed. The id now comes from the Instance resource's physical id in the stack events. A quiet stretch of 30 s prints the elapsed time, and every cloud UI line flushes, so output piped to a file or `tee` streams as well. Part of #32099 Original author: Bolin Chen (cherry picked from commit 1af545f)
The Files and Library panes of the AWS Control drive end in a Load more button that took native disabled while the next page was fetched. The button stays on screen during the fetch and after it when more pages remain, so disabling the pressed button blurred it to the body and a keyboard user lost their place at the bottom of the list. Both panes now share one LoadMoreButton held with aria-disabled. A ref latches the press, so a second press that beats the re-render is refused, and the fetch keeps the default cancelRefetch so a press during a background refetch still loads the next page. The Files button is keyed by folder, so the latch of a folder left mid-fetch does not hold the parent listing's button. Fixes #33114 Original author: Rohan Khanderia (cherry picked from commit c0d78cd)
…3169) An update that fails part-way on a held file can leave the app's data/ partial and the only complete copy in .<name>-data-tmp. The next update deleted that copy as stale and reported success, so whatever was only in it was lost. update_app now refuses that state before anything moves, names both paths and writes an audit record, the same call the built-in adoption already makes. With no data/ the copy is still put back. Closes #32674 Original author: Bolin Chen (cherry picked from commit aca65b4)
…#33070) The tunnel status API returns state and error as separate fields, so a connected tunnel can still carry an error. The badge ignored it and showed plain green Connected. Show a warning label and put the error in the tooltip; connected without an error and the other states are unchanged. Fixes #33066 Original author: Joe Guo (cherry picked from commit 686ed3b)
#33176) A new chat with no folder project gets dashboard.default_project. _birth_default_project resolves that path with os.path.realpath on a worker thread and then asked the boolean is_sensitive_path, which resolves it again on the resolver pool and answers True while the pool is in a stall cooldown. So a stall read as a sensitive match and the chat silently opened in the workspace default. It now asks is_sensitive_canonical_path about the path it resolved, which answers inline off the event loop with the same gate list. A configured directory that really is sensitive is still skipped. The chat_handlers facade binds the new name for its owners, and the security spec lists this caller beside the other canonical-gate users. Closes #32702 Original author: Bolin Chen (cherry picked from commit 21b5e16)
…sk cases (#33048) The worker prompt and work_report schema described needs_human as 'only a person can unblock you', so a worker tagged plain design choices with it and the goal-conductor skill then put the worker's options in front of the person. Now needs_human covers only the five ask cases (credentials or permission, spend, irreversible or deleting someone else's work, a costly uninferable preference, a choice the person asked to make); the skill treats it as the worker's claim, checks it, and otherwise takes the recommendation with action=decide. The session_send gotcha keeps the Normal-approval fact. Worker spec digests re-pinned. Original author: Joe Guo (cherry picked from commit 89b7d3b)
…32839) Regenerate, edit-resend and rewind refused with kiro_prerequisite_required whenever kiro-cli was missing or signed out, even on a slot whose turn runs on Claude, Codex or another harness that never reads kiro-cli's sign-in. They now ask the kiro-cli gate only when the backend the session manager's applied provider factory picks for the slot's session (member DMs and crewmate pins included) signs in from kiro-cli's store. The applied config is read rather than config.json, which runs ahead of the factory until the config watcher applies a write. On such a foreign harness the one sign-in fact Crew holds is the slot's last outcome, so edit-resend and rewind, which cut history before their turn runs, refuse with harness_sign_in_required while the last turn ended signed out. Restoring the replaced tail after a failed foreign turn stays open on the issue. Builds on #10414 by @amansk. Part of #14512 Original author: Bolin Chen (cherry picked from commit 91ee8a0)
…#33174) On Windows a read of a file another process holds open fails with PermissionError. The listing routes already retry the Kiro-global mcp.json read and answer a retryable 503 mcp_config_busy, but the toggle, toggle-tool, toggle-all, remove and DELETE routes read it with a plain read_text, mostly on the event loop where no retry runs, so a held file escaped as a bare 500. The agent-config sync behind them read kirocrew.json the same way. Read the Kiro-global file through the retrying read on a worker thread in all five routes and answer a hold that outlasts the retry with the same 503 before anything is written. The two agent-config syncs read through the retrying read too, and a held file still raises out of them: kiro-cli loads that file, so skipping the sync would answer success while the server stays mounted. The four routes that sync answer it with the 503. Closes #32722 Original author: Bolin Chen (cherry picked from commit d57b80f)
kirocrew logout printed only "HTTP <code>" when /api/logout answered with an error, dropping the body's reason. A 500 "could not persist session revocation; logout not completed" read the same as a refused secret, so the operator could not tell that nothing was revoked. Move cron_trigger's bounded, sanitising reader of a gateway error body to loopback_http.http_error_reason, beside deadline_expired, and use it in both cron trigger (output unchanged) and logout. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> Original author: Yu-Chen Lai (cherry picked from commit 3f8c2a4)
A persistent cron job runs on its tab's cron:<id> session with no task on the tab, and the tab read that session's turn lease only for channel keys. So while a run was going, the tab showed idle with no Stop. The tab now reads the lease for a cron link as it does for a channel link, and the reconcile pass's lease-flip check covers every tab, so the row is pushed when the run starts and ends. Closes #29705 Original author: Bolin Chen (cherry picked from commit 30bff54)
A dashboard user who is not the owner may not cancel a cron run, so their Stop on a running cron tab leaves the run alone. The server answered that press with "not running", which the dashboard reads as "nothing is running": it settled the pane idle and hid Stop while the cron kept going. The press is now refused with the cron cancel route's own owner_only code. The turn keeps its running state and Stop stays. The main chat shows the refusal on its action-error notice, as the side pane already did, with its own words: only the dashboard owner can stop the run, so pressing Stop again does not help. Pause all counts a refused stop as not stopped, so Resume all sends that tab no Continue. Original author: Ray Xu (cherry picked from commit 23f532c)
8e534ad to
c849f9c
Compare
Problem / Motivation
Goal: Bring 20 fixes merged on
mainon 2026-11-08 (#32441 and 19 merged between 13:42 and 14:24 UTC) ontorelease/0.12.0for 0.12.1, one commit each, in a PR of their own.Why it matters
Each fix changes behaviour a 0.12.0 user can hit today. An app update no longer runs while the app's data sits in two places, so it cannot delete the only complete copy. A busy parent keeps running when a sub-agent completion delivery times out, a parked task row is cancelled when its attempt is stopped, and task runs list only this data home's runs. A scheduled cron run shows as running in its tab, and a non-owner's Stop there is refused instead of hiding Stop while the run goes on. History reruns ask for kiro-cli sign-in only when the turn runs on it. MCP toggles answer 503 on a held
mcp.json,dashboard.default_projectsurvives a resolver stall, the tunnel warns on an error while connected, and logout shows the gateway's reason. Cloud launch streams its install log and destroy output matches what it removed. The rest fix lost keyboard focus, screen-reader text, the knowledge chip token count, chat follow during reprice and the conductor's needs_human check.Not a goal
main; nothing is changed only here.CHANGELOG.mdchange.What changed (motivation → approach → change)
Each main squash commit was cherry-picked onto
release/0.12.0@80d6211963, in merge order onmain(the #32441 + #33124 pair last, together), one commit each, with the(cherry picked from commit …)trailer. The repository's leak guard requires the public identity, so every commit is authored as Ray Xu and names its original author on anOriginal author:line.a57d5e77bb99fe7c0a3e426ef59d7ad4b08c65b38d42b860e4b907041d63e7836f795ed141c255e08554a00b17b697020899d1a9a812d7e3da1870b5f216ae2886f99c57695cee677c3c19f63888f1fbb2a4ba3e4cb70e0f334b1af545f1159375941808c0d78cd3319c7a23cc19aca65b4c8746ccd6d7e1686ed3b2f0d098cdf9e721b5e16602208d70786289b7d3b05d9802cb20b691ee8a053f5284849e47d57b80f6f649cdc51df33f8c2a4c64e5535d684130bff5460e7e8e84d68523f532c4adc849f9c9a9Hand-resolved hunks:
src/kiro_crew/dashboard/chat_regenerate.py(edit-resend, after_subagents_attached_response) andsrc/kiro_crew/dashboard/chat_rewind.py(rewind, after theslot_closingcheck): only the newif not kiro_turn:sign-in refusal is kept. 0.12 has no crew-pause guard next to it, so that context is not brought in.website/src/pages/chat/page/busyTurnControls.ts: the newshowActionErroroption is added;onSlashRefusedkeeps 0.12's stage list (no'fork'stage there). Itsdashboard/handlers/crew_pause.pychange and the matchingtest/test_crew_pause.pytest are left out: 0.12 has no Pause all.test/test_agent_refactor_spec_bytes.py: the eightkirocrew-worker.jsondigest pins, the only pins fix(crew): conductor checks a worker's needs_human against the five ask cases #33048 changes, are set to this branch's bytes.test/test_cli_gateway_client.py: import block takesimport asyncioonly (0.12's file does not useio).docs/system-specs/modules/app-kit-platform.mdanddocs/system-specs/modules/learn-cron-dashboard.md, keeping release's wording with each PR's own sentences added.One file under
.github/changes:.github/sync-io-in-async-fs-baseline.txt, a test baseline, carried as-is by the #33174 pick (49cdc51df3). No workflow, action or script changes.Left out of this batch:
acp/kas_permissions.pyandagent_sdk/kas_tool_view.py; needs the tools-settings alias work (canonical_tools_settings,tools_settings_aliases) and the trailing-*deny twins. Too long a chain.taskrunner_parts/run_registry.py:114, the legacy registry adoption does synchronous file IO and an fsync write on the event loop at startup.mainhas the same code; it waits for amainfix and 0.12.2.Backwards compatibility
None: each fix keeps its main PR's compatibility notes; nothing here changes stored data or config beyond what those PRs already shipped on
main.Tests
Proof: not provable -- other: a backport of main commits whose proofs ran on their own PRs
release/*, so the related set ran here, one pytest process at a time with no xdist, withTMPDIRoutside~/.kiro/crew: the 23 Python test files the batch changes plustest/test_taskrunner_atomic_persistence.py, 1,347 passed.check-i18n-strings.mjsOK (220 untranslated, baseline 332),i18n-check.mjswithI18N_BASE_REF=80d6211963OK, andvitest run src/i18n/70 files, 1,194 passed.Manual verification
N/A - each fix was verified on its own main PR. The 0.12.1 RC soak is the end-to-end check.
Screenshots / video
Why no screenshot: the frontend picks carry the same rendered change as their main PRs, which show it.
Related Issues
Backport to the 0.12 line for 0.12.1 of #33132, #33127, #33129, #33155, #32499, #33134, #33184, #33183, #33130, #33170, #33117, #33169, #33070, #33176, #33048, #32839, #33174, #27134, #32441, #33124.
Sibling sweep
Siblings: none -- each source PR carried its own sibling sweep; this PR only carries those fixes.
Pattern harvest
Not generalizable: this PR only carries fixes already on
main. The source PRs have their own harvest.Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)Contribution License Agreement
Contributions are accepted under the project's Apache-2.0 license; see the Licensing section of CONTRIBUTING.md.