Repository navigation
fix: backport 0.12.1 late batch 1 (47 fixes merged after 11-07 22:00Z) - #33024
Conversation
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 PR attaches and commits no images, so the blind read was not performed. Meanwhile the diff changes dozens of user-visible strings, controls and states across the dashboard. Evidence gaps
Notes
[UX-REVIEWED] 4d961fd Description read[DESCRIPTION-READ] dfc4ff16266023363ddd75231fd390bb45b39323faece0a0dcf1f6d4c61996ed 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) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks: this mechanically cherry-picks 47 fixes that were already reviewed and merged on Notes
[DESIGN-REVIEWED] 4d961fd Description read[DESCRIPTION-READ] 82f3c2bab133bfd03d80fde53594054c24d8c92afc64ad7949ea3f6783af1250 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 Each item is a cherry-pick of a fix already merged on Notes
What this change shipsInventory (10 items) — 9 justifiedIntent: give 0.12.0 users 47 fixes already merged on
[FIRST-PRINCIPLES-REVIEWED] 4d961fd Description read[DESCRIPTION-READ] 50584f1c1d98db4c5e0fbdc26fcbe37d5e32355ce195b3c30d99bf0dfdb9c2eb That sha256 names the PR description this verdict read. Recompute it with |
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: |
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: |
…on-board widgets (#32595) (cherry picked from commit 326033c) Original author: @iamwhatever
The Workflows author view disabled its Run button while the script validated and ran. A disabled button leaves the tab order, so a keyboard user who pressed Run lost focus to the page. Mark it aria-disabled and aria-busy instead, and claim the press in a ref so a second press in the same tick still starts only one run. (cherry picked from commit 18ec128) Original author: @RohanK6
The Overview health line, the Usage card's no-usage note and the
WakaTime line each glued catalog fragments around a value in English
order. Hindi and Korean put the uptime last, Japanese put the provider
name after the sentence, and German read the status "Up" as the
direction ("Nach oben"). Render each as one catalog message with
the value in place, reusing the Usage tab's existing no-usage sentence.
(cherry picked from commit c13765b)
Original author: @RohanK6
Create team, Save, the name field and the delete confirm took native disabled while the request they started was out. Disabling the focused control blurs it to the body, outside the dialog, so a keyboard user lost their place before the request answered. Hold them with aria-disabled (the name field read-only) and refuse a second send in the handler instead. Fixes #32407 (cherry picked from commit 6a81d5c) Original author: @RohanK6
) Co-authored-by: w <w@k.local> (cherry picked from commit ea26f31) Original author: @bolichen97
…2340) A goal loop whose last cycle could not run (stopped_reason structural_terminal or session_start_failures) rendered "Paused · the last cycle could not run" in the same warn-toned status line as a user pause, so a failure read like an ordinary pause with no alert. That line now renders through ErrorNotice (alert semantics, no hand-off beside the unsaved goal fields); every other state keeps the status line. Closes #21637 (cherry picked from commit bc6579b) Original author: @bolichen97
…#32711) A stock Btn has a transparent fill and a faint --border edge. On the 12% ok/warn tint of a banner that edge falls to about 1.0:1 against the background, so the Dev Fleet Undo, the job-log Cancel, the artifact slug-collision Open buttons and the MCP Apply/Discard pair looked like plain text until hovered. Give each an opaque --bg-elevated fill and a --border-strong edge, as #31924 did for Open Approvals. Hover and the danger text stay as they were. The wider sweep of the same shape is tracked in #32636. Closes #31903 (cherry picked from commit a4e1271) Original author: @bolichen97
…xt (#30799) * fix(config): describe Mint Timeout's below-10 fallback in its help text The help said "Clamped to [10, 120]", but InstancesConfig.__post_init__ discards a value below 10 and uses the transport default (SSH 30s, SSM 90s), and the loader reads a non-numeric value as 30s. Say what the code does for each input. Fixes #30396 * fix(config): say a non-numeric Mint Timeout uses the transport default validate_config_data drops a non-number before the section builder runs, so the 30s _safe_float fallback is unreachable on KiroCrewConfig.load(). (cherry picked from commit 0b4e7fe) Original author: @tim-finnigan
) The Schedule page agent roster Retry used a native disabled while the roster reloaded, which takes the focused button out of the tab order and drops a keyboard user to the page. Make it inert with aria-disabled and a guard instead, and pin focus and the ignored second press. (cherry picked from commit bfc0ef1) Original author: @RohanK6
The Retry under a failed Prompts or Workflow library list read lost keyboard focus the moment it was pressed. It rendered only while the query held an error, which React Query clears when a query with no data refetches, so the pressed button unmounted for the whole retry, and it was natively disabled while busy. Keep it mounted with queryReadFailed and mark the busy state with aria-disabled plus a click guard, as the memory cards do. (cherry picked from commit d3665ee) Original author: @RohanK6
Every field in InstalledApp.from_dict is coerced with a no-raise wrapper so a malformed installed.json degrades to defaults rather than crashing a reader. schemaVersion was the one exception: int() raises on a non-numeric string, a list or null, so a corrupted or hand-edited record crashed from_dict with an uncaught ValueError/TypeError, unlike every other bad field. _read_installed catches only JSONDecodeError, UnicodeDecodeError and OSError, so that raise escaped to most callers. Read schemaVersion through a shared _read_schema_version helper that mirrors _read_enabled: an unusable value reads as 1 with a warning, forcing the legacy-migration pass, so the field degrades like the others instead of raising. Fixes #32542 (cherry picked from commit 75aed02) Original author: @RohanK6
…2775) The confirm-hide button, the stored-context banner note and a token's agent line were each built from catalog pieces joined in code with an English space or ': '. Japanese, Korean and Chinese read the button in English word order, and Japanese and Chinese got an ASCII colon or space between full-width text. Each line is now one catalog message with placeholders, in all shipped catalogs. English reads the same. (cherry picked from commit fe2d88c) Original author: @RohanK6
…32650) Research Lab kept campaigns.db beside app.json in <data home>/apps/auto-research/. An update or an adoption onto the app's own repository keeps only data/ and .app_secret of that directory, so moving the app out of core would delete every campaign row with the retired tree (gap 1 in the app repository's PORTING.md). The database now lives in data/. On every gateway start, from the startup hook that runs whether or not the app is enabled and before any connection opens, a database kept beside app.json is copied in once through SQLite's online backup (rows still in the -wal come along), written owner-only under a staging name, checked with PRAGMA quick_check and published with a hard link, so it never replaces a database already in data/. Once published the old file and its sidecars are removed. A link at the old name is an operator's redirection and stays in use. Until a copy is published db_path keeps answering the old file, so no row is hidden; a lock someone else holds on the old file makes the copy give up after a few seconds instead of holding up the start, and a failed copy is retried at the next start. Part of #26967 (cherry picked from commit b42d198) Original author: @bolichen97
A host deny (or a parent's spawn_steer) steers text into a running sub-agent turn, and kiro-cli has the model acknowledge it with a [STEERING steer-<id>: ...] line. The run's end-of-turn cleanup stripped the [OPTIONS: ...] line but not that acknowledgement, so the completion event and result.txt (which spawn_status reads) both handed it to the parent as part of the answer. Strip complete acknowledgements once at the end of the turn, through a new messaging.driver.strip_steering_markers that reuses the channel layer's steer-marker parser, fed in bounded slices so its cost stays linear. sanitize_channel_replay_text now calls the same function. Closes #29807 (cherry picked from commit 991e33f) Original author: @bolichen97
…32543) The CLI built its HistoryConsolidator without a vector store. Without one a pass drops the lessons and structured facts it extracts, so the CLI pass drained the tail, printed done, and kept neither. The gateway's pass, given the store, kept them. Hand the CLI pass the default VectorMemoryStore and memory.migrated, as the gateway does, with embeddings wired like a named store's, and close the store when the command ends. The CLI spec row says what the pass now stores. Closes #30339 (cherry picked from commit 8730b41) Original author: @bolichen97
#32231) Connect saved the repository link exactly as it was pasted, so anything pasted after the repository name, a token included, was written to the app's saved settings in plain text. It now saves only the repository the link names, so nothing else from a pasted link is kept. (cherry picked from commit 5941a9c) Original author: @bolichen97
With no gateway reachable, kirocrew app enable/disable said the change takes effect at the next gateway start. A gateway that is running anyway (the CLI cannot reach one on Windows or from a sandboxed shell) re-reads the app's crons and hooks from the files within a minute, so the enabled cron was listed and fired at once. Only the backend process waits for a gateway start. The message, the CLI spec and the App Kit reference say so, and a test pins the minute to the two poll intervals behind it. Closes #30328 (cherry picked from commit bf4fc9c) Original author: @bolichen97
…2298) The PATCH route read the record before the manager lock and skipped the teardown when every named transport field matched that read. Another edit plus a connect landing between the read and the write let a request that set a field back to its old value skip the teardown, so the record named one machine while the live forward served another. reconfigure() takes an optional needs_teardown predicate and runs it under the lock first. False writes in that same hold and leaves the tunnel and its self-heal alone; True takes the existing barrier and teardown. The route passes it for every edit naming a transport field; a disable still always tears down. Closes #20991 (cherry picked from commit 69431d5) Original author: @bolichen97
The watchdog learned the loop thread's id from its first heartbeat, which lands one interval after start. A stall in that window had no thread to read major faults from, so a paging loop exited on the plain budget. start() runs on the loop thread, so record the id there. (cherry picked from commit a3f4e3a) Original author: @bolichen97
A package installer writes one copy of a shared agent per dependent package (<Package>-<name>.json plus a local-<Package>-<name>.json twin), all declaring the same name. Every crew binding writer (crew create, both crew-update paths, the locked rebind and the default-agent write) runs _foreign_private_copy_owner, which refused any name matched by more than one spec file with 409 lineage_unverifiable. The crew editor re-sends kiro_agent on every save, so a crew bound to such an agent could not even be edited. Lineage is keyed by the declared name, and that name's strict sidecar read has already found no record by the time the file scan runs. When every claimant resolves under the target's own name, none of them is a private copy whichever one a session runs, which is the verdict the spawn gate already gives the same twins. _AmbiguousTemplateName now carries the name each match resolves under, and the ownership read falls through only when that set is exactly the target. A claimant resolving under another name, such as a <target>.json declaring a fork-backed name, keeps the fail-closed refusal. Fork, publish, reset and template create still refuse, since they must pick one file. The existing ambiguity tests move to that mixed-name shape so the sentence, masking and file listing stay pinned; new tests cover the rebind, the generic editor save, the direct ownership read and a recorded private copy with twins. crew-mode.md says which claimants the refusal covers and that the session start picks the file through agent.spec_pins. (cherry picked from commit 336b847) Original author: @bolichen97
…rd (#32821) A Power whose one server shares its name gets the namespaced id power-<name>-<name> (power-figma-figma), which the MCP sign-in card printed as-is, in code font, in the sentence and on the button. Show that name once, capitalized, in normal text. Every other id keeps its full id and the code font, and the relay still gets the raw id. Builds on #28532 by munahad. Closes #28524 Co-authored-by: munahad <344993864+munahad@users.noreply.github.com> (cherry picked from commit 7951d1f) Original author: @bolichen97
…litter (#32829) (cherry picked from commit db353b7) Original author: @CrysisDeu
…32632) A skill that sets the steering key inclusion: always stays on demand, because skills pin only on always: true. The listing ignored the key silently, so an author who shares one SKILL.md with a tool that reads inclusion had no signal. The listing now logs one warning per saved file naming the skill and the always: true key, and the skills doc lists inclusion as a key skills do not read. Closes #30773 (cherry picked from commit f5b9bdc) Original author: @bolichen97
#32827) The daily SEL prune and four tmp + rename writers (Bot staging, recording meta, avatar staging, adoption archive) published with a bare os.replace. On Windows any reader holding the destination fails that rename on the first attempt. They now rename through atomic_write.replace_with_retry, which stays a plain rename on POSIX and on the event loop. The prune caught a failed rewrite in the same except arm as the rotation lock, so a held rename was logged as a lock skip and never reached the heartbeat's prune_failed counter. A rewrite failure is now raised after the lock is released; a lock that cannot be taken still skips as before. The adoption archive's rename sat inside the download's OSError mapping and was reported as network_unreachable. A failed local rename is now the transient stage_io_failed (disk_full for ENOSPC), retried with the same backoff and shown with the dashboard's "local" remedy copy. Fixes #32684 Fixes #32683 (cherry picked from commit b6419ee) Original author: @CrysisDeu
A gateway that exec'd into a live target kept KIROCREW_LIVE_EXECED in its environment, so every process it spawned inherited it. The foreground Make live restart is one of those: its detached `kirocrew restart` brought up a replacement that saw the marker, skipped the pointer and booted the old checkout again, so the second Make live (and Switch back) only staged the change. The exec'd image now pops the marker as it returns, so a gateway it later starts reads the pointer like any other boot. Closes #30376 (cherry picked from commit 7458a23) Original author: @bolichen97
…32592) A session directive such as monitor_start called through Tool Search under a resource-v2 tool loader arrives as {"tool_id": "tool::kirocrew-core::<tool>", "arguments": {...}}. The consumer only unwrapped the envelope when tool_id was exactly kirocrew-core::<tool>, so it hashed the whole wrapper, missed the record the MCP server parked and the directive was never applied. event_input_digest now accepts both exact spellings for the directive tool the call already resolved to; any other server, tool or prefix still hashes the whole envelope. The agent-host contract names the second spelling. Closes #30665 (cherry picked from commit c991635) Original author: @bolichen97
…ro (#32815) Fixes #32723 (cherry picked from commit 48c4454) Original author: @CrysisDeu
…pen (#32618) (cherry picked from commit 73377ad) Original author: @bolichen97
#32614) On a KAS session, a call that Tool Search deferred carries the agent's purpose inside the envelope's arguments. The pet's approval bubble read the purpose from the raw input's top level only, so it showed no purpose for those calls. It now prefers the tool_purpose the gateway already puts on the permission row, which is read through the envelope, and keeps the raw-input read as the fallback. (cherry picked from commit e3e6c82) Original author: @bolichen97
) A plan imported from YAML with no heading comment and no path fell through auto_name to its first non-empty line, which in every workflow spec is the top-level `agents:` key. The plan was listed as "agents:" and exported as agents.yaml. Once the YAML is decomposed, the plan is named again with its first step's title standing in for that first-line fallback; a saved workflow's own name, a heading comment and a spec path still win. Closes #30364 (cherry picked from commit 8103f62) Original author: @bolichen97
Two Hooks page names glued translated text together with a hard-coded
English em dash: the unsent-hook badge ("never fires — <reason>") and
the ON switch of a trigger that never fires ("Disable hook — <note>").
Japanese and Chinese screen readers heard a foreign separator
mid-sentence, and no locale could re-punctuate either name.
Move each name into one catalog key with a placeholder,
pages.hooksPage.stored_only_reason and
pages.hooksPage.disable_hook_with_note, so each locale owns its
punctuation. English stays byte-identical.
(cherry picked from commit b767105)
Original author: @RohanK6
A batch transcript that settles while Chat is unmounted is held by the transcript inbox and handed over when Chat mounts again. The composer's session effect also redelivered held results synchronously on mount, inside the composer's own mount effects and before ChatPage's draft restore, which then overwrote the transcript: the composer came back empty and the saved draft stayed empty. The session effect now redelivers only when the session changes; on mount the inbox's deferred hand-over delivers after the host's mount effects. The inbox reports when a held result is being delivered, so the "Dictation added" cue shows on a return as well as on a slot switch. Closes #29776 (cherry picked from commit 1718c9c) Original author: @bolichen97
) With both Slack tokens saved and KIROCREW_OWNER_ID empty, the gateway keeps Slack off and the settings page only said "Needs setup": the config reads as not configured, and the panel hid every hint in that state. The panel now says Slack stays off until an owner is set and points at the Owner Slack member ID field, in every shipped language. Closes #30491 (cherry picked from commit eb0ecb9) Original author: @bolichen97
Crew's own KAS bundle runs Node directly, and Node trusts only its built-in roots unless NODE_USE_SYSTEM_CA is set. Behind a TLS-inspecting proxy whose CA lives only in the OS store, every engine call failed on the bundle route. kiro-cli already defaults the variable to 1 when it spawns KAS, so the relay route worked. The bundle spawn now defaults NODE_USE_SYSTEM_CA=1. An operator's own value, 0 included, is kept, as on kiro-cli. The relay is left to kiro-cli. Fixes #32561 (cherry picked from commit 9b5bbfb) Original author: @iamwhatever
#32792) Under wecom.allow_all_users, configured_targets() lists every warm peer, so a stranger who messaged the bot once could be the single available direct target that sole_direct_target reads as the channel owner. Mark those peers learned on ConfiguredChannelTarget and have sole_direct_target skip learned targets, so an owner-DM path names only a configured userid or owner_id. The dashboard picker still lists them. (cherry picked from commit 1782ff8) Original author: @RohanK6
…32373) A loop a restart caught mid-cycle comes back paused as interrupted_cycle, which the goal popover rendered as a bare "Paused". A refused turn whose claim release the store would not take stopped re-arming but stayed active with nothing armed, so the popover showed a running loop with only a log line saying why. The popover now words interrupted_cycle in every locale, and that failed release stops the loop under claim_unsaved in the same write that records the claim undelivered. The popover shows that failed save as an error notice in place of the status line, like a cycle that could not run. Part of #22831 (cherry picked from commit 8e251ac) Original author: @bolichen97
7eec2d6 to
4d961fd
Compare
Problem / Motivation
Goal: Bring 45 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@a0b61daa99, 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.#32373 sits at the end of this batch, after its prerequisite #32340, which is in it. #29890 (a fix) goes in ahead of #29354, which needs it.
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-resolved hunks:
src/kiro_crew/docs/skills.md: release's text kept; main's sentence not carried (doc only).Left out of both batches (code conflicts, with the prerequisite each needs):
website/src/i18n/en.context.json.website/src/store/chat/activity.ts.dashboard/chat_utils_slots.py.acp/harness/kas_cloud.py.ArtifactsPageGallery.tsx.crew_log/emit_runtime.py.dashboard/handlers/usage.py.apps/routes_install.py,routes_lifecycle.py,AppDetailPage.tsx; needs 4 earlier commits, two of them features.dashboard/chat_fork.py,state_parts/slot_session.py; needs 4 earlier commits, including features #9750 and #17165.crew_pause.py,chat_api/continue_turn.py; needs 4 earlier crew-mode commits, one a feature.acp/client_transport.py,runtime_wire.py,session_handle_turn.py; needs 5 earlier commits, mostly features.messaging/dispatch.pyand the channel turn engines.cli_server.py; needs 5 earlier commits, two of them features.website/src/store/chatSlice.ts.knowledge/SourcesList.tsx.dashboard/chat_turn/queue_drain.py; needs 7 earlier commits.chat_turn/queue_drain.py.acp/kas_agents.pyand the tools-settings readers.chat_runner.py,ChatInput.tsx,ContextShelf.tsx; needs 20+ earlier commits.mainonly, by decision; not backported to the 0.12 line.dashboard/handlers/usage_sessions.py; needs #26951, which stays onmainonly.test_dashboard_chat_turn_wiring.py::test_first_slash_preserves_pending_replay_for_next_prompt(a test it changes) fails on release: the replay mark is set twice.Not picked by the bar (not a user-facing fix): #32710, #32398, #32236, #32313, #32482, #32590, #32558, #32383, #32627, #32452, #32843, #30652, #32551, #32486 (CI and release scripts, the source-build frontend staging step #32486, prepare-pr and pod/Dev Fleet tooling, a log-only change, perf). Every
feat:,test:,docs:,ci:andrefactor:PR merged in the window was skipped by type.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 on the rebased head, one pytest process at a time with no xdist: every Python test file the batch changes plustest_usage.pyandtest_usage_daily_credits.py(44 files), 2,303 passed.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 #32595, #32553, #32540, #32585, #32538, #32504, #32340, #32711, #32747, #30799, #32557, #32773, #32546, #32775, #32650, #32483, #32543, #32514, #32774, #32806, #32525, #32517, #32429, #32231, #32523, #32298, #32596, #32588, #32821, #32829, #32822, #32632, #32827, #32586, #32592, #32815, #32618, #32614, #32582, #32793, #32474, #32577, #32587, #32792, #32373.
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.