Repository navigation
fix: backport 0.12.1 late batch 2 (42 fixes merged after 11-07 22:00Z) - #33025
Conversation
Sensitive Change ReviewSensitive files changed:
How to clear this check:
Only the author's comment counts this way. A reviewer can instead |
ee53ec1 to
ae92990
Compare
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: |
First Principles Review (Opus 5.5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS This PR only backports fixes already merged on What this change shipsInventory (10 items) — 10 justifiedIntent: give 0.12.0 users 39 fixes already merged on
[FIRST-PRINCIPLES-REVIEWED] 5e213e2 Description read[DESCRIPTION-READ] 888817117d1173bc27062b22a9762f5a188799a954617f1aa590b680ec86fd9f That sha256 names the PR description this verdict read. Recompute it with |
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. The blind read was not performed because this backport changes user-visible dashboard controls and copy, and its description has no images and it commits none. Evidence gaps
Notes
[UX-REVIEWED] 5e213e2 Description read[DESCRIPTION-READ] 0d2e92280ff8d8baf61153a8a39b37e3cc8c739028b6b6e5ac254503a50b478c 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 screenshots of the new link-preview failure card and the registered-theme Overlays toggle as they render on this branch. Both surfaces are new under BlockersNew user-visible surfaces with no rendered evidence of this revision Notes
[DESIGN-REVIEWED] 482e23b Description read[DESCRIPTION-READ] b7125db7628510915c6dd05ceecf75019d2773301f7603931fb4a65076d312c1 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 |
Security Scope Review -- 🔴 could not complete (fails closed)No scope verdict for Security scope review -- adjudicated candidates
No candidate flipped from allowed to refused. [SCOPE-REVIEWED] 5e213e2 |
ae92990 to
200d2dd
Compare
200d2dd to
482e23b
Compare
482e23b to
7e92025
Compare
The Slack step rebuilt .env from its KEY=VALUE pairs, so each save deleted every comment and blank line, re-spaced unrelated keys, and read an indented '# KEY=value' comment as a key named '# KEY'. It now saves through the dashboard writer's locked line rewrite, which replaces only the Slack app token, bot token and owner ID lines (every spelling of each) and keeps all other lines byte for byte. The reader strips each line before the comment test, as the boot parser does. Fixes #32675 (cherry picked from commit ec11b37) Original author: @CrysisDeu
…up (#32579) GatewayManager arms its watchdog only on a start that spawned or adopted a daemon, and _init_mcp_gateway had no branch for a start that returned False. Any failed start (a daemon owned by another live gateway, a draining incumbent, a slow or failed spawn, the election rounds spent) left the gateway without a broker until it restarted, with every stubbed server on per-session exec. On a failed start the gateway keeps the manager in _mcp_gateway_pending and starts mcp-gateway-start-retry, which awaits start() on the same manager after each L4 ladder delay (1s doubling to a 60s cap, jittered, no attempt cap) and publishes the broker through the helper a boot start uses, including the dashboard handle that was wired at boot with no broker. Reusing the manager keeps its lifetime stand-down budget and election gates across attempts. _stop_mcp_broker cancels the retry first and shuts the pending manager down as well, so a daemon an attempt spawned cannot outlive the gateway, and the sharing toggle stops a pending retry before it builds a new manager. The lifecycle spec said the watchdog retried after a refused start; it names the gateway's retry instead. (cherry picked from commit 76b1664) Original author: @bolichen97
…2468) A Kiro's Chromium that cannot load a shared library exits 127 at once, but its stderr went to /dev/null and the start only watched the Harness daemon, so the person waited about 60 s for "the Harness daemon did not start in time" with no hint of the cause. On macOS and Linux the browser's stderr now goes to a pipe the gateway drains for the browser's life, keeping its newest 4 KiB in memory. The daemon wait also watches the browser: one that exits fails the start at once with its exit code and its last stderr lines, redacted, which for a missing library is the loader's own line naming it. Closes #30312 (cherry picked from commit 908add8) Original author: @bolichen97 Hand-resolved: added the TYPE_CHECKING import of Callable that main carries from an earlier commit, so the new published_stderr_tail annotation resolves on release.
…log (#32571) The gateway-failure dialog opened as a modal child of the window it is about. On Windows and Linux a modal child disables its parent, so while the dialog waited (for example "port 5476 already in use") the splash's Close window button, and the window's own captions, did nothing. Open it as a plain child on Windows and Linux: it still stays above its window, and the window can be closed to the tray while it waits. macOS keeps the modal sheet. The desktop-app doc says so. (cherry picked from commit 2caa69b) Original author: @bolichen97
…#32914) The MD Notebook connect screen's Clone / Attach button set native `disabled` while busy. A disabled button leaves the tab order, so a keyboard user who pressed it lost focus to the page for the whole clone. It now uses `aria-disabled` plus an in-handler guard while busy, and stays natively disabled only for an empty field. (cherry picked from commit 1aea792) Original author: @RohanK6
Switching between regenerated replies swapped the text but left the turn stats footer showing the newest reply's model, credits and time. A variant now keeps its reply's turn stats, and a switch puts them back on the row its turn stamped, so each reply shows its own figures. Builds on #22749 by @LuckVd. Closes #22724 (cherry picked from commit 652d578) Original author: @bolichen97
…#32905) A session/new model list is entitlement-filtered, and a session born while an entitlement lookup races a token refresh can advertise a degraded subset. refresh_advertised_models replaced the cross-session cache with it and persisted it to provider_models.json, so the degrade seeded every later session and the cold picker until a healthy session rewrote it. Both session/new callers now pass confirm_shrink=True: a list that drops a cached id is held in memory instead of applied, the cache keeps its ids and takes only the list's additions, and reading the same list again replaces the cache, so a real downgrade still narrows it one session later. The kiro --list-models catalog keeps plain replace. Builds on #28617 by @cruisercohen. Fixes #22344 (cherry picked from commit 0983357) Original author: @bolichen97
…2527) (cherry picked from commit 2f60cf0) Original author: @bolichen97
…32906) Under LANG=C with UTF-8 mode off, the first emoji print raised UnicodeEncodeError and killed kirocrew setup, service install and service uninstall. ensure_utf8_console() now switches a POSIX stdout/stderr that cannot encode those glyphs to errors="replace", keeping its ASCII encoding. Closes #29741 Co-authored-by: Cary Wu <41397851+cwu98@users.noreply.github.com> (cherry picked from commit d9a472f) Original author: @bolichen97
…32612) Settings > Browser read "Not connected · Connect" when the engine was installed but Kiro's Chromium, the browser it would launch, was not on disk. Connect could only fail with "install it in Settings", and the card then offered only Retry. The card's state looked at the engine alone and never at chromium.downloaded. An install is now needed when the engine is missing, or when the browser is Kiro's Chromium and it is not downloaded. The card offers Install for that case, and a failed install there is the card's to show rather than the Engine row's under More. Closes #30385 (cherry picked from commit 752c0b9) Original author: @bolichen97
) In Settings, Security, an expanded posture row with more than 25 items shows a Show N more button that is removed once pressed. Keyboard focus then fell to the page body. Focus now moves to the first row the press revealed, so a keyboard or screen reader user continues where the new rows begin. (cherry picked from commit 7de2532) Original author: @RohanK6
The spend chart's Try again rendered only while its query held an error, which React Query clears when a query with nothing cached refetches, so the chart swapped to its skeleton and the pressed button unmounted. The harness notice's Try again stayed mounted but was natively disabled while it read. Either way keyboard focus fell to the page. Hold the chart's last error message through a first-load retry so the notice and button stay up, mark both busy buttons with aria-disabled plus a click guard, and label them Retrying while they read so the retry is visibly running. (cherry picked from commit e26b1aa) Original author: @RohanK6
On Windows, CPython's subprocess.run kills only the direct child on TimeoutExpired and then calls communicate() with no timeout, so a launcher's child (Git for Windows' cmd\git.exe, a .cmd or exe shim) holding the pipes keeps the call open until it exits. Add platform_compat.run_killing_tree_on_timeout, which on Windows kills the tree with kill_popen_tree, drains under a 5 s bound and re-raises the same TimeoutExpired. Route run_limited, no_index_diff, api_file_diff's git calls and doctor's kiro-cli --version and node -v probes through it. POSIX is unchanged. Fixes #32755 Fixes #32742 Fixes #32703 (cherry picked from commit 5d4fbc0) Original author: @CrysisDeu (picked by hand: conflict only in docs/system-specs/modules/cli.md docs/system-specs/modules/learn-cron-dashboard.md)
Deleting a crewmate tore down its config row, crew log, avatar, goal check-ins, worktrees and private template copy, but never its pinned thread's session_map.json entry. A bad dashboard:member-<slug> record then outlived the crewmate, and nothing else could clear it. The shared removal teardown now destroys the thread's keys through the gateway's session manager, which also ends a live process for the thread: the V1 key, the current store's key and earlier store generations on the exact .memory- boundary. A slug another live crew still derives keeps its entries, and a failure never fails the delete. The dashboard delete and both prunes pass the manager; the CLI delete has none and leaves the map alone. Builds on #28613 by @cruisercohen. Co-authored-by: Matt Cohen <1832296+cruisercohen@users.noreply.github.com> (cherry picked from commit bed3c8a) Original author: @bolichen97
…#32934) POST /api/spawn built its own execution context for a parent session with no execution record and always passed it as _execution_context, so the manager's fallback to the parent's live selection never ran. A channel conversation has no record until its first turn is saved, so a sub-agent spawned in that turn ran kirocrew and lost the parent agent's prompt, tools, hooks and MCP servers, while the same request from a dashboard chat inherited the parent's agent. The fallback moves into execution_context.recordless_parent_execution, which resolve_spawn_execution and the route both call. When the parent has no record and the request names no agent, the route reads the parent's live selection (SessionManager.get_agent_selection) and admits the child from it. A parentless request and a parent on the default template keep kirocrew. An unusable selection is refused with 409 member_identity_unavailable, as spawn_async refuses it, rather than run as kirocrew. The availableAgents policy for such a parent is unchanged. Test doubles that send a recordless parent give a concrete get_agent_selection tuple, as testing-conventions.md requires. Closes #32798 (cherry picked from commit 0a60009) Original author: @bolichen97
A theme registered through the branding seam mounts its overlays (the Bikini Bottom mascot, for one) with no way to hide them, because the Experience section's Overlays switch only exists for installed packs. Give such a theme that same switch, on the same per-theme record, on by default, and skip its overlays in the shell while it is off. Closes #20542 (cherry picked from commit 2ebb1d1) Original author: @bolichen97 (picked by hand: conflict only in docs/decisions/README.md)
…32929) The pending-skills panel's not-found notices reused "dismissed" for a candidate that left the queue in another window or through the agent, so the same word named the user's own Dismiss button and something that happened elsewhere. The five out-of-band values now say "removed": the two settled variants (candidate_no_longer_pending_dismissed, approve_failed_not_found_dismissed_named) and the three hedges (approve_failed_not_found, approve_failed_not_found_named, candidate_no_longer_pending), in all 12 catalogs plus en-XA. Key names are unchanged. SkillsTabValidationReason pins the new copy on every not-found approve and dismiss path, and the approve-validation capture script waits on the new settled sentence. Builds on #22679 by @dodgeblaster. Co-authored-by: Gary Jennings <garysjennings@gmail.com> (cherry picked from commit 36fb266) Original author: @bolichen97 (picked by hand: conflict only in website/src/test/SkillsTabValidationReason.test.tsx; took main's version, which release now matches up to this change)
…9354) A streaming reply auto-scrolls the transcript. When it moved the link under the pointer, the layer hid every unpinned card on scroll, so a card the user was reading closed by itself. A card that is already open now stays where it is on such a scroll. The pointer-out the browser fires for the scroll is ignored; the first real pointer move outside the card and the link sends the usual leave, so the app's own leave grace still decides when it closes. A show still in its hover delay is still cancelled by a scroll that moves its link. Escape, blur, another provider's hover and pinned cards are unchanged. (cherry picked from commit f8e2574) Original author: @buluoray
…2505) (cherry picked from commit 1dde2fb) Original author: @bolichen97
A file tree row is a button whose name is its text, so a modified file read as "app.py M": the badge letter is sighted shorthand with no meaning to a screen reader and no way to translate it. Hide the letter from assistive tech and give the row a visually hidden status word (modified, untracked, added, deleted, renamed, or "git status <code>" for any other code) from the catalog, in every locale. (cherry picked from commit 93a9c7f) Original author: @RohanK6
…2507) A file over knowledge.max_ingest_file_mb in a watched folder is an expected skip: the ingestion pipeline logs it as a WARNING naming the config key and raises FileTooLargeError. The folder scan caught that in its generic handler and logged it again with logger.exception, so every scan added an ERROR line and a traceback for a file the user only has to raise the cap for. The scan now skips the error log for FileTooLargeError and records the row exactly as before: status 'failed' with the raise-the-cap message, so the Knowledge page still shows the file with its reason and Retry. Closes #30347 (cherry picked from commit 0c34a42) Original author: @bolichen97
#32513) A directive tool's result reaches the model with its internal [[KIROCREW_SESSION_DIRECTIVE]] marker line, so a reply that repeats the result carried that line into the chat bubble and the stored transcript. The marker is applied from the tool row, never from the reply. The segment flush now drops each reply line that is only a marker reading as a directive, or only the refusal tag, and keeps every other line; sentinel bytes inside a sentence or in code stay as written. (cherry picked from commit 1091827) Original author: @bolichen97
…#31731) On a host with no kiro-cli, KAS takes Kiro Crew's own engine on the automatic route, and the gateway downloads the pinned engine at the first chat. kirocrew doctor printed a missing pin there as a warning but also put it in its issue list, so a healthy fresh install ended with "Fix these issues" and exit 1. On the automatic route a missing pin now keeps its warning row and download hint, and is an issue only when the gateway will not install it on this host (an unsupported platform or glibc, or a version override it does not install unattended); the row then says so instead of promising a download. An operator who turned agent.kas_bundle on, an incomplete pinned directory and an unreadable engines directory still count as issues. (cherry picked from commit d549bec) Original author: @bolichen97 (picked by hand: conflict only in docs/system-specs/modules/cli.md)
…ssing (#32938) A request naming ?workspace= whose directory was absent fell through to the default roots and returned another directory's files with an empty root. Answer it like a missing ?project= root instead: 404 workspace_not_found. Fixes #32846 (cherry picked from commit cad0ff4) Original author: @CrysisDeu
7e92025 to
5e213e2
Compare
Problem / Motivation
Goal: Bring 39 fixes merged on
mainafter 2026-11-07 22:00 UTC ontorelease/0.12.0for 0.12.1, one commit each, in a PR of its own.Why it matters
Each fix changes behaviour a 0.12.0 user can hit today: wrong or missing UI text and focus, refused or stuck actions, lost dictation, cron, sub-agent, task-runner and channel bugs, KAS sign-in and certificate problems, crash-dump and diagnostics gaps, and Windows publish retries. They are the user-facing
fix:rows merged onmainin the last day, freeze 2026-11-08 16:00Z.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@ec9e31b334, in merge order onmain, 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.This batch is built on
release/0.12.0after the other batch (#33024) merged.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-resolved hunks:
src/kiro_crew/agent_sdk/__init__.py: added theTYPE_CHECKINGimport ofCallablethat main carries from an earlier commit, so the new annotation resolves (flake8 F821 otherwise).docs/system-specs/modules/security.md: release's text kept.test/test_security_gate_liveness.py: release's version kept, with_PACKAGE_LINE_BUDGETraised by 4, from 34,697 to 34,701, for this change's Bearer-branch lines (the test passes at that budget).test/test_crew_fired_spec_hooks.py,test/test_hooks_composition_contract.py: release's versions kept, this change's new subagent session-id test added, and_BASE_SHAPE_DIGESTset for the one changedpermission_pre_tool_blocksignature (checked: it is the only shape line that differs from release).docs/system-specs/modules/learn-cron-dashboard.md: release's text kept.docs/system-specs/modules/cli.md,learn-cron-dashboard.md: release's text kept.docs/decisions/README.md: release's index kept.website/src/test/SkillsTabValidationReason.test.tsx: main's version taken; release's copy of the file now matches main's up to this change.docs/system-specs/modules/cli.md: release's text kept.The left-out list for the whole set is in #33024. Left out of this batch as well:
update_schedule.pymarks the crash dump clean before the final shutdown drain, so a SIGKILL during that drain makes the next start delete the dump as a clean exit. Kept onmainonly. #31159, which went in as its prerequisite, stays: it is a fix on its own.taskrunner.pytakes a Resume's agent only for aplannedrun, so a paused run saved by 0.12.0 (no agent field) resumes on the default agent instead of the one selected. Kept onmainonly.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: every Python test file the batch changes plus the two hook test files (64 files), 4,927 passed, 1 skipped and 1 failed. The failure istest_acp_harness_contract.py::test_a_missing_self_served_binary_names_its_install_command[copilot], which fails the same way on a cleanrelease/0.12.0on this host.tsc -p tsconfig.app.json --noEmitpasses.check_file_line_capfails only ondashboard/chat_api/resume.py, which is already 2,008 lines on release.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 #32816, #32579, #32468, #32367, #32872, #32848, #32873, #32879, #32020, #32832, #32890, #31495, #31159, #32611, #32458, #32638, #32608, #32571, #32914, #32790, #32905, #32527, #32906, #32612, #32877, #32880, #32828, #32924, #32934, #32134, #32929, #29890, #29354, #32505, #32916, #32507, #32513, #31731, #32938.
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.