Repository navigation
fix: backport 0.12.1 late batch 3 (26 fixes merged 11-08 08:00-09:23Z) - #33093
Conversation
…ntries (#31558) browse_ui computed the one-word auto entries over the whole Settings area before it removed the entries that a settings.<tab> sub-area holds. The top-level settings listing then reported omitted_one_word_auto for controls it never listed, and each tab listing counted the same controls again. The settings listing now drops sub-area entries from the one-word list as well, so omitted_one_word_auto counts only the entries this listing left out. Closes #31552 Original author: rohnkum (cherry picked from commit 6bd5b30)
) The Skills page reads the description and always flag of kiro-cli skills with _parse_skill_description, which decoded the file as plain utf-8. A SKILL.md saved as UTF-8 with a byte-order mark kept U+FEFF in front of the opening --- fence, so the frontmatter parser found no fence. The row then showed an empty description and always=False. The file is now decoded with utf-8-sig, which drops a leading BOM, as the skills loader does for the same files. Closes #31553 Original author: rohnkum (cherry picked from commit 5bbc82e)
… artifact (#28661) `_handle_to_artifact` read `report_artifact_slug` before authoring the report, then `_publish_report_artifact` created or updated the artifact and persisted the slug. Two first exports for the same campaign both read a NULL slug, both created an artifact, and the later persist left the other one as an orphan duplicate. The handler now holds a per-campaign asyncio lock, `_report_export_lock`, across the slug read, the create-or-update and the persist, and re-reads the slug once the lock is held. The second export then updates the first one's artifact. The lock entry is dropped when no export holds or waits for it, and the store and DB work still runs in threads. Closes #ISSUE Original author: rohnkum (cherry picked from commit ca3b5e7)
…href (#28541) `static_response()` built the injected `<base href>` by concatenating the served file's folder, as file-system text, onto the `/<project-id>/` prefix. A browser parses that value as a URL, so a folder named `section#one` gave `/p1/section#one/`, where `#one/` is a fragment and `asset.png` resolved to `/p1/section`. `?`, `%`, spaces and `"` broke the base the same way. The no-entry diagnostic page linked the HTML files it found with the same unquoted concatenation. Each folder path is now quoted with `urllib.parse.quote(..., safe="/")` before it is joined to the prefix, in both places, and `rewrite_html()` HTML-escapes the finished href. The `/<project-id>/` prefix is left as the caller built it, so it is not quoted twice. Closes #28537 Original author: rohnkum (cherry picked from commit 906be01)
…32939) A reply that repeats an MCP app tool's result carried the [kirocrew-mcp-app:<id>] marker into the saved assistant row. The segment flush and its hold check now cut each well-formed marker (and the one space after it) with mcp_apps_render.strip_quoted_markers. Fixes #32809 Original author: Zezhen Xu (cherry picked from commit fce8bc2)
…8657) `publish_sync.unpublish()` and `publish_sync.delete_for_artifact()` are async, but they called the synchronous `provider.reachable_for()` on the gateway's event loop. For Personal Drive that call resolves the publication's profile through `_profile_for`, which runs an `sts get-caller-identity` AWS CLI subprocess with a 30 second timeout. While it ran, every other request, SSE stream and heartbeat on the loop waited. Both call sites now run the probe with `asyncio.to_thread`, as the module already does for the artifact store and as `serving_notice()` does for the same reason. The provider interface is unchanged. A new test records the thread each probe runs on and asserts it is not the loop thread. Closes #ISSUE Original author: rohnkum (cherry picked from commit d7a4537)
…al edit (#31524) The artifact list route caches each artifact's loaded content and its prefix snippet under (version, updated_at). A live file pointer reads its source file on every get, so an edit made outside the store changes the content without moving either value. Content search kept matching the old text, missed the new text, and the snippet kept showing the old text until the gateway restarted. The cache key for a live file pointer now also carries the source file's modification time and size, so an external edit misses the cache and the file is read again. Artifacts that keep their own copy keep the old key. Closes #31522 Original author: rohnkum (cherry picked from commit e81ef77)
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: |
c0edb32 to
2d6e8d3
Compare
First Principles Review (Opus 5.5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS This PR only cherry-picks 24 fixes already merged on Notes
What this change shipsInventory (10 items) — 9 justifiedIntent: give 0.12.1 users 24 user-facing fixes already on
[FIRST-PRINCIPLES-REVIEWED] 058171a Description read[DESCRIPTION-READ] fb426d4ec9671b8a8bbcb5555acd0c1da881f506dd5c66239c1bbee1bc3922ba 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 of the new Crew Board "refresh failed" notice and the restored MCP Servers list view. The blind read was not performed because this revision supplies no image. Evidence gaps
[UX-REVIEWED] 058171a Description read[DESCRIPTION-READ] 1de56c960b00518dc9bfba629442b0cd48624acdd311d86ef876428a2b03fb7b 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: these are 24 cherry-picks of fixes already merged on Notes
[DESIGN-REVIEWED] 058171a Description read[DESCRIPTION-READ] 9ad4d7f6c5a2bb4f69cad557d57d2cb9f24a47f0ca0e6562e2acfdd371ba0108 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 |
…ancelled (#28224) RunnerAdmission ran accept_async, admit's claim and claim_only_async through _db, which awaits a writer-thread job. A caller cancel cannot stop a job the writer has already started: the write commits and the result is dropped. The accepted row stayed queued (the orphan sweep skips this incarnation's rows), a claimed step row stayed admitted under this incarnation's lease (the cancel arm had no generation to end it with), and a run's container row stayed starting with no handle, all until the next boot. These kinds have no dispatcher, so nothing else settles them. The three now go through _db_owned. It shields the job, and a caller cancelled while the job runs queues _end_row_on_cancel for the row the job names, from the job's done-callback, so the cancel arm still awaits nothing. A job the writer had not started yet does nothing, so a cancel that wins the race still writes no row. Closes #28220 Original author: rohnkum (cherry picked from commit ab89ed1)
… a BOM (#31591) The prompt listing and the prompt detail read decoded each prompt with plain utf-8. A file saved "UTF-8 with BOM" kept a leading U+FEFF, which hid the column-0 frontmatter fence. Both then showed the first body heading, or nothing, instead of the declared description. _extract_sop_description, _gated_sop_description and the description decode in _api_user_prompt_detail now use utf-8-sig. It is still a strict decode, so a file that is not UTF-8 still has no description. The detail read's content copy and edit base are unchanged. Closes #31585 Original author: rohnkum (cherry picked from commit b39b1cd)
…ilt from (#28667) `_pr_ai_fingerprint` and `_build_pr_ai_prompt` derived the same inputs two different ways. The fingerprint capped each raw comment body at 1200 characters, while the prompt stripped the body first and then capped it, so an edit past raw offset 1200 in a comment with leading whitespace reached the model without changing the digest. The prompt also listed the first eight failing checks in provider order, while the fingerprint hashed all checks sorted, so a reordered check list changed the prompt but not the digest. In both cases `read_pr_ai_cache` served the stale summary. The fix adds `_pr_ai_comment_text`, which strips, caps and marks truncation once, and `_pr_ai_failing_names`, which sorts failing names before the cap of eight. The prompt and the fingerprint both use them. Unchanged input keeps the same digest, and the frozen fingerprints in the contract test are unchanged. Closes #ISSUE Original author: rohnkum (cherry picked from commit 265f25d)
…n two letter cases (#31721) On a case-insensitive volume, realpath keeps the letter case it was given. Two steering roots that differ only in case therefore produced two different dedup keys for the same file, and collect_folder_steering emitted every document under that root twice. The collector now also records each emitted document's device and inode, and skips a candidate whose file identity was already emitted. The reader refuses hardlinks, so two distinct documents never share an identity. Closes #31705 Original author: rohnkum (cherry picked from commit 6980b66)
…uns (#33012) The Meetings list's Sync calendar button set native `disabled` while the sync ran. A disabled button leaves the tab order, so a keyboard user who pressed it lost focus to the page for the whole sync. It now uses `aria-disabled` with the same dimmed look, and the click claims the sync in a ref, so a second press is ignored even before the pending state re-renders. Original author: Rohan Khanderia (cherry picked from commit eed0a06)
…32992) The Expand button on an answered clarifier in the Research Lab grill tree was disabled while it fetched follow-up questions. A disabled button leaves the tab order, so a keyboard user who pressed it lost focus to the page. Mark it aria-disabled and aria-busy while it runs, keep it disabled at max depth, and claim the press per node in a ref so a second press before the busy state renders is refused. Original author: Rohan Khanderia (cherry picked from commit ece4f40)
* fix(crew-board): keep the board when a background poll fails The Crew board polls every 10 s. When one poll failed, the page swapped the whole board for the "Could not read this board" notice, even though React Query still held the last good board. Every row unmounted, so open event lists and a Stop failure notice vanished after a gateway blip. A failed refresh now keeps the board and shows an inline notice above it saying it may be out of date, with the backend's message and the agent hand-off. A first read that fails still shows the error, and a 404 still shows the no-ledger state, since that is the ledger's absence (a purge removes it) rather than a blip. * chore: screenshots for #32900 (dropped before merge) Original author: Rohan Khanderia (cherry picked from commit 43ac0cd)
… name (#31664) build_bundle named the report temp <report>.<_RUN_ID>.tmp. _RUN_ID is fixed once per process, so every build in one process used the same temp name for a given --out. The temp is claimed with an exclusive create, so a temp left behind by an earlier build made every later build fail with FileExistsError on that name. The temp name now adds a fresh random token per build after the run id. A leftover from an earlier build no longer collides, and it is left in place because this build did not write it. The rollback still removes the temp this build wrote, because it uses the same name. Closes #29040 Original author: rohnkum (cherry picked from commit daa91a1)
…ecord again (#31592) A slot's pre-projection ledger document is carried into the crew log as one entry. Each field was clamped, but the entry as a whole was never checked against the crew log's entry ceiling, and the artifact count was not capped. A document with long non-ASCII fields or many artifacts built an entry the writer refused. The carry then released its claim and asked the caller to try again, and every retry built the same entry, so the slot refused all updates. The carry payload now keeps the newest artifacts up to the artifact cap. It then drops the oldest artifacts, and then the rejected approach, until the entry fits one crew log line. The carried event counts what was dropped. The read preview uses the same builder, so it shows the same record. Closes #31583 Original author: rohnkum (cherry picked from commit aaccdfd)
…33014) The blocked card's Fix my access button shows and hides the sign-in steps, but it carried no aria-expanded and named no region, so a screen reader heard only a label that changes. It now reports aria-expanded and points aria-controls at the steps it opens, the same way the History menu button does. Original author: Rohan Khanderia (cherry picked from commit b031b35)
… filter (#29407) * fix(memory): match stored backslashes and quotes in the memory editor filter The memory editor filter in memory_edit._rows searched a fact's value in its JSON-encoded form. A backslash was searched as two backslashes and a double quote as an escaped quote. Typing a fact's text exactly as stored found nothing. The filter now searches the decoded value. Strings are used as stored, dict keys and leaves are joined by spaces, and other scalars keep their JSON form. The filter now agrees with the find-and-replace edit, which already works on the decoded value. Closes #29403 * fix(memory): walk nested values without recursion in the memory editor filter _search_text recursed once per nesting level. A semantic value stored a few hundred levels deep, which validate_semantic accepts, raised RecursionError. The ValueError/TypeError handler in _rows did not catch it, so every list, filter, preview and bulk apply on the memory page failed. _search_text now walks the value with an explicit stack, as _contains_memory_search_text does, and _rows also falls back to the row text on RecursionError. Original author: rohnkum (cherry picked from commit 31dc75a)
…32867) A Global memory failure at gateway startup that is not a damaged memory.db (for example the preferences.md seeding write failing) told the user to "Inspect or cancel this store's pending restore". The restore pass has already run by then and reports its own failures, so no pending restore exists and the advice sent people looking for one. That failure now names what the failing step opened (the memory files directory, or memory.db for its open) and asks the user to fix the quoted error and restart. A damaged memory.db keeps its own advice, and restore-pass failures keep the restore advice. Builds on #28561 by @cruisercohen. Fixes #23681 Original author: Bolin Chen (cherry picked from commit 0d68e97)
) The Context Breakdown trend heading ("Last 12 turns") and the fold line ("1 of them listed above") rendered their numbers through a raw {{count}}, which i18next writes in Latin digits whatever the language. The chart's own axis labels beside them go through fmtNumber, so under Bengali the heading read "12" next to axis labels written "১২". The three keys now render {{n}}, formatted with fmtNumber like every other number on the panel; {{count}} is still passed to pick the plural form. Original author: Rohan Khanderia (cherry picked from commit 7290777)
Pause and Start building went natively disabled while their own stop or handoff request ran, which takes the pressed button out of the tab order, so a keyboard user landed on the page body. Give the app's pill Btn a busy state that reads aria-disabled, ignores clicks and keeps the disabled look while staying focusable, and use it for both. A second press still sends nothing, so the double-handoff guard holds. Original author: Rohan Khanderia (cherry picked from commit 285971e)
A macOS launch refused for a writable bundle always said the app "is installed in a directory this user can write", whichever probed location failed, and named none of them. An operator whose install directory was locked down but whose bundle executable was user-owned could not tell where to look. _writable_component now returns the first path that would let this user change what runs (or None), naming a replaceable executable as the file itself, and resolve_target picks the install-directory message when that path is the bundle's parent and a component message otherwise, each naming the path. Closes #23891 Original author: Bolin Chen (cherry picked from commit 6d07df1)
Coming back to Customize from a chat opened Connections on Services, with the MCP Servers filter empty and every tool list closed. The sidebar's Customize entry opens a bare /capabilities, and the sub-tab lived only in ?view= while the filter and open rows were McpTab state. The sub-tab, the filter and the open tool lists are now kept in sessionStorage for the browser tab, and a resumed MCP Servers is written back into the URL. A view=mcp-servers link still opens MCP Servers. Builds on #25967 by @alparamonov. Closes #25961 Original author: Bolin Chen (cherry picked from commit 1823c3e)
2d6e8d3 to
058171a
Compare
Problem / Motivation
Goal: Bring 24 fixes merged on
mainon 2026-11-08 between 08:00 and 09:23 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: lost keyboard focus and untranslated digits, MCP Servers and Crew Board views that lose their place, skill and prompt files with a byte-order mark read wrong, oversized session ledgers that stop recording, a drive probe on the event loop, and app and preview bugs. They are the user-facing
fix:rows merged onmainthis morning, before the 2026-11-08 16:00Z freeze.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@dc384c6114, 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.1cf402902e27d7b2f0246bd5b3099e38f424935c5bbc82ea8fe7632ab767ca3b5e75de72b38b9731906be0116379d33dc916fce8bc2f2144802dd0dcd7a45378bed855b4db49e81ef77d666b0d00b79bab89ed1f39a31603a702b39b1cd78ffef35f84cf265f25d0cbf3f829b73f6980b663f593fa90b182eed0a065c0b23ebd3f71ece4f407e299b412f5a243ac0cd051516147f3dedaa91a1e0577d74146a4aaccdfd4fbf5578ea5d7b031b3537d0bd054c2f431dc75aebdb8075cba8d0d68e97cc24fdbf1e6a1729077734b46d61591e7285971e1e0ff81ce6fb96d07df11cf53367d4e741823c3e48cc0edb32b66Hand-resolved hunks:
temp-screenshots/32900/are left out; they are PR evidence, not part of the fix.docs/feature-map/README.md: release's rows kept, with this change's three phrases added to the Connections row.Left out of this batch:
ok: true/ "not running" fromturn_stop.py, so the composer drops the running state while the cron keeps going. Kept onmainonly.GalleryPanel.tsxdiscards a pending pet lookup instead of deferring it, so a failed save followed by a retry saves the previous pet. Kept onmainonly.slack/handler_runtime/approvals.py. Its prerequisite is #30692 (enterprise permission_mode rule, 51 files), which itself conflicts in source. Too long a chain.config.loader.read_endorsed_config_text, which 0.12 does not have.release/0.12.0.en-XA.json,en.manual.json).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: every Python test file the batch changes (17 files), 1,416 passed and 9 skipped. With this host'sTMPDIRunder~/.kiro/crew,test_design_tweak_backend.pyanswers 403 on this branch, on a cleanrelease/0.12.0, and onmainat fix(design-tweak): URL-quote folder names in a static preview's base href #28541's own merge commit: the preview's Kiro Crew internal-path guard refuses the test's temporary project. WithTMPDIRelsewhere the file gives 169 passed, 1 skipped on bothmainand this branch.CrewCompanionGallery.coverage.test.tsx, 183 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 #32988, #31558, #31559, #28661, #28541, #32939, #28657, #31524, #28224, #31591, #28667, #31721, #33012, #32992, #32913, #31664, #31592, #33014, #29407, #32867, #33013, #33001, #31363, #32347.
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.