Repository navigation
fix: 0.12.1 late merges 1 (22 fixes merged 00:15-01:29Z, KAS #33435) - #33586
Conversation
dafb01e to
1a83475
Compare
… 1 (#31782) kirocrew memory restore raised FileNotFoundError when the store had no backup yet, or when --from named a missing file. _memory_cmd only turns ValueError and SQLite errors into its one-line Error: refusal, so the operator saw a full traceback instead. The restore branch now raises ValueError when no backup is found, and converts the FileNotFoundError from restore_from_backup into a ValueError. Both cases print one Error: line on stderr and exit 1. Closes #31767 Original author: rohnkum (cherry picked from commit 4541fe1)
The App Store listing never offered Update for an app installed from a versionless row of an index-tier registry (any user-added registry that is not owner-tier) unless the user pressed refresh. Only refresh_registries read the branch tip of such an app; the listing's own hourly index fetch resolved the row to the installed checkout, so update_status kept reading the lapsed source manifest cache. The listing's index fetch now hands a freshly fetched index to installed_app_tips.refresh_installed_tips in the same gather as the owner-tier store-art prewarm, before the rows resolve. A row whose source manifest cache is still fresh is skipped, so the listing clones each tip at most once per manifest TTL; the explicit refresh expires that cache first and still reads every qualifying row. The row selection, which now reads cache files, runs off the event loop. Original author: Bolin Chen (cherry picked from commit d73680e)
…3494) A catalog or index-tier app install fetches credential-free in the strict sandbox, which hides ~/.ssh, so ssh reads no ssh config and no agent. An app repo on a host reached only through a ~/.ssh/config route (ProxyCommand, ProxyJump, Host alias) then failed with ssh's "Could not resolve hostname" and a network and proxy hint that does not apply to ssh. When ssh's own unresolved-hostname line is in the output of that credential-free install clone, pinned-commit fetch or fast-forward pull, the install now returns git_clone_failed_strict_ssh_route with a fixed sentence: the strict sandbox hides ~/.ssh, so the ssh config and agent were not used, and the operator can clone the repository and run `kirocrew app install <dir>`. The sentence is a constant, so no git output or host name reaches it. Owner-credentialed clones, HTTPS remotes and auth refusals keep their current messages. Original author: Bolin Chen (cherry picked from commit 20adeea)
The Sync button on an installed app's detail page took native disabled while its run was out, and a successful run then reloaded the page through the full-page loader, which unmounted the button. Either way a keyboard user lost their place. Hold Sync with aria-disabled while its own run is out, refuse a press while any action runs, leave it natively disabled during other actions, and refresh the app in place after a successful Sync so the pressed button stays mounted. Original author: Rohan Khanderia (cherry picked from commit 293acc3)
On Windows every ssh:// registry index clone and app install failed when git used Win32-OpenSSH. The child env allowlist (_SAFE_ENV_KEYS, shared by minimal_env and anonymous_git_env) dropped ProgramData, and Win32-OpenSSH exits 255 with no output when it is unset, so the log showed only git's generic "Could not read from remote repository". ProgramData is a location hint, not a credential, so it joins the other Windows location hints in the allowlist. The plain index clone failure warning now also names git's exit code, so a silent ssh failure is visible in the log. Original author: Bolin Chen (cherry picked from commit 59fed20)
#33491) A repo on gitlab.com, or on a self-managed GitLab host the dashboard lists in dashboard.gitlab_hosts, was refused with "expected a repo URL on one of: github.com", which points the user at github_hosts, a list that can never make a GitLab repo reviewable. The refusal now says the host is a GitLab one, and names the host instead of echoing the typed text in quote marks. The add-repo field also glued an SCP-style remote (the git@host:group/repo.git form git remote -v prints) onto https://github.com/, and the backend found no host in it. The page now sends a remote as typed, and parse_repo_ref reads it as its https URL before the same exact host check, with plain string parsing. The GitLab list only picks the wording; nothing new is admitted. Both route callers run the parse off the loop because a refusal can now read config.json. The same change is on the app's mirror repository. Original author: Bolin Chen (cherry picked from commit 19529ce)
…opy (#31750) A link of 24 characters or fewer got no random tail after https://example.invalid/, so every short link became that same prefix. Two different short links merged into one, and a map keyed by two of them was refused as two keys that could not be told apart, which skipped the dashboard page check. Every invented link now gets at least eight random characters after the prefix, so distinct short links stay distinct and a map keyed by them is copied. Closes #31737 Original author: rohnkum (cherry picked from commit 62851a3)
…31556) The adaptive controller counted every finished run in the completion_rate denominator, including cancels, permission denials, invalid params, context-length errors and turn limits. Five operator cancels in a minute fired the completion_rate signal, and with one slow start in the same window the two signals corroborated and halved the exec cap. build_sample now divides successful runs by successful plus attributable runs, so non_congestion outcomes never feed the signal, as the adaptive-concurrency spec states. The spec's signal table now names the denominator. Closes #31549 Original author: rohnkum (cherry picked from commit 34b59bb)
`kirocrew memory restore --store ''` treated the empty name as absent. The `or DEFAULT_MEMORY_STORE` fallback swallowed it, so the verb staged a restore of the default store from its newest backup and exited 0. A shell that expands an unset variable produces exactly this call. With --cancel-pending it cancelled the default store's staged restore instead. The restore branch now defaults the store only when the flag is absent, as `_memory_verb` already does for export and import. An empty name raises a ValueError, which `_memory_cmd` prints as one error line with exit 1, and nothing is staged or cancelled. Closes #31743 Original author: rohnkum (cherry picked from commit 9185ade)
#33426) A native disabled attribute removes the row action button from the tab order, so a keyboard user who presses Cancel or Delete loses focus to the page body until the request ends. Mark the button aria-disabled and aria-busy instead, matching the busy-button pattern used elsewhere, and keep the dimmed styling. runAction already ignores a press while the run is busy, so a second press is a no-op. Co-authored-by: kirocrew-agent <agent@kirocrew.local> Original author: Raymond Chen (cherry picked from commit 306d133)
…k time (#28324) due_watch_items() defers an item whose last check is more recent than the tier's floor (economy 30 minutes, balanced 10, active 5), but it read lastCheckedAt, a field nothing writes. A check records lastChecked (watchlist_file's historyEntry update), so every checked item looked never checked and the floor never deferred anything. It now reads lastChecked. The floor tests built their items with lastCheckedAt by hand; they now use the real field, and a new test records a check through apply_watchlist_update and the watchlist file, the way a real check does. Closes #28320 Original author: rohnkum (cherry picked from commit 7eef1a8)
…h 400 (#28286) GitLab has no "request changes" review and Azure DevOps records neither verdict, and their clients refuse the verb they lack. The review route read the head and called the client anyway, and _pr_action_error mapped the refusal to a 502 provider_error, so a known capability gap read as an upstream outage. A bulk approve on Azure DevOps failed every row the same way after a head read each. Both routes now check the verb against the KEY's own client's PR_REVIEW_EVENTS first, as the merge route already does with PR_MERGE_METHODS, and answer 400 unsupported_review_event before anything is read. Closes #28283 Original author: rohnkum (cherry picked from commit 3cff821)
…ays (#28492) The update_plan description says to pass null to clear a metadata field, but the incremental schema typed mood and narrative as string, planner_notes as object and needs_replan as boolean, so validation refused null. Past the schema, apply_update skipped a None value and left the field unchanged. The four keys now take null, and apply_update removes a key passed as None. An omitted key is still left unchanged, and full_replace is unchanged. Closes #28491 Original author: rohnkum (cherry picked from commit e2ad9e7)
… the store (#33487) A pending-deploy store item that was not an object, or whose created_at_epoch was not a number, raised from the expiry filter. The pending list, every new preview, and every confirm or dismiss then failed until the file was edited by hand. Such an item can never be confirmed, so the filter now drops it the way it drops an expired one. Original author: Rohan Khanderia (cherry picked from commit 31ff778)
#33429) expand_spec_args built its placeholder source by reading each name under its own spelling, so on Windows (HOME/USER unset, USERPROFILE/USERNAME set) a POSIX-written ${HOME} or ${USER} in a stdio MCP server's command, args or env stayed literal and the server launched on a path that does not exist. Fill a missing HOME/USER from its Windows twin before expanding; a set HOME/USER still wins and no name outside _SPEC_ARG_VARS is read. One helper feeds command, args, env and the probe, so all four surfaces get it. Closes #32728 Co-authored-by: kirocrew-agent <agent@kirocrew.local> Original author: Raymond Chen (cherry picked from commit 961571c)
…3506) The skill review and the refine path read a pinned: value of true, 1 or yes as a pin and carry it forward, but the lifecycle that archives idle auto skills accepted only true. A hand-edited pinned: yes or pinned: 1 skill was archived while every other reader called it pinned. All four readers now go through one frontmatter_pinned helper, so they agree. Every value that pins a skill today still does. Original author: Rohan Khanderia (cherry picked from commit e52b555)
…#33370) When the System Sessions tab cannot read session usage and has no earlier rows, it shows a 'Can't read session usage' panel with Retry. Any refetch, a Retry press or the 5s poll, turned the query back to pending, which swapped the panel for a loading skeleton and took the focused button with it. The panel now stays up from the first failure until data arrives, Retry reads Retrying... and stays focusable (aria-disabled, ignoring a second press), and the failure text renders through ErrorNotice with its Ask-the-agent hand-off. Original author: Rohan Khanderia (cherry picked from commit 0b61495)
* fix: drop a session card's source-link chips at /clear The chips are derived from the transcript, and /clear keeps the transcript, so a cleared session's card kept showing the pull requests and issues of the conversation the person had dropped. The derivation now stops at the newest user row that typed /clear. * docs: say the /clear chip boundary is the typed row, not a clear that ran A queued or refused /clear also drops the earlier chips, unlike the plan pill, which resets only on a clear that ran. Raised by the Design and First Principles reviews on 1d59e5b. --------- Co-authored-by: EC2 Default User <ec2-user@ip-172-31-10-71.ec2.internal> Original author: Jie Lan (cherry picked from commit daf575d)
) A Teams turn edits its "🔧 <tool>…" progress message into the first chunk of the answer. When a tool approval card was posted after that message, the edit put the answer above the card, so the newest thing in the chat was the card and the reply sat hidden over it. Once a card follows the progress message, settle the message to the done line and post the answer as a new message below the card. Original author: Rohan Khanderia (cherry picked from commit 851ba86)
…3003) When a plain pip install is below the policy minimum version, the gateway's version-compliance log line gives the in-place upgrade and the restart after it. An agent can read that log line, and both steps are the user's to run: the upgrade replaces the package the running gateway is loaded from, and the restart is refused to agents. Both branches of the line (the signed-wheel command and the verification-failure note) say to run them from your own terminal, and a decision entry records the wording. Closes #25156 Original author: Bolin Chen (cherry picked from commit f2faabf)
…3435) The install report decided its closing advice from agent.kas_bundle alone, so on a host with no kiro-cli (the flag off but the route resolving to the bundle automatically) it told the user the bundle route was off and to turn it on, while kirocrew doctor reported the bundle in use. Read the same predicate doctor reads (kas_bundle.kas_route): print the turn-it-on hint only when the route is genuinely the relay, and otherwise say Crew will use the engine. Co-authored-by: kirocrew-agent <agent@kirocrew.local> Original author: Raymond Chen (cherry picked from commit b0a62b6)
1a83475 to
8731686
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: |
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 changed dashboard states. The blind read was not performed because the description attaches no image and the PR commits none. Evidence gaps
Notes
[UX-REVIEWED] 8731686 Description read[DESCRIPTION-READ] 8ea6ff70d46e1e6e94a56580930a8e15c408f23469d91718eb35c7466c42d919 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 |
First Principles Review (Opus 5.5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Every item is a fix that already merged on Notes
What this change shipsInventory (10 items) — 9 justifiedIntent: ship 22 fixes that already merged on
[FIRST-PRINCIPLES-REVIEWED] 8731686 Description read[DESCRIPTION-READ] cf7b19274cf8947389cdca4ff1719a92097c4bd37048dd165db19ada8deda3b1 That sha256 names the PR description this verdict read. 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: there is no screenshot of the Sessions read-failure panel as this backport renders it. The panel now shows an inline ErrorNotice with an "Ask the agent" button in place of the subtitle, and the evidence list for this revision has no images. BlockersThe reshaped Sessions error panel has no rendered evidence for this revision Notes
[DESIGN-REVIEWED] 8731686 Description read[DESCRIPTION-READ] 550db12423c5168a799097ecbd0a614de1f396505d32d6301ee35be6134ac9c7 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 |
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: |
Problem / Motivation
Goal: Bring 22 fixes merged on
mainon 2026-11-09 (between 00:15 and 01:29 UTC) ontorelease/0.12.0for 0.12.1, one commit each, in one "late merges" batch.Why it matters
Each fix changes behaviour a 0.12.0 user can hit today. On KAS, the installer no longer tells a user to turn on a bundle route that already resolves by itself (user-reported, #31907).
kirocrew memory restorewith no backup, or with an empty--store, prints one error line instead of a traceback. App Store installs show Update for branch-tracking apps, explain a strict-sandbox ssh clone failure, and keepProgramDataso Windows ssh clones work. Stdio MCP specs resolve${HOME}/${USER}on Windows. Teams posts the answer below an approval card. Several dashboard buttons keep keyboard focus while busy, and a session card drops stale source-link chips at/clear.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@9f7010e187(rebased fromf6d1874badafter #33529 and #33549), one commit each inmainmerge order, 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.4541fe1ce35db699ab6cd73680e13aa0fce66dd120adeeaef83f46877c5f293acc3fb14c03442c8959fed2070dda5bec65b419529cec97ac178820d962851a3d52c3b004a6c134b59bbc75b39979d5ad9185ade2c4eabadf47bb306d1331e7f4790d80fe7eef1a8ffcee1b6c9fe23cff821d4c2647b4fd58e2ad9e74eb9fca330a4a31ff77843f88a4b87920961571c82b5c3b28fe95e52b555e8d068de2eff60b614952a6a59585f371daf575d81e3767483d61c2b47ac3430c91346945851ba86e411bc506260cf2faabf17041ade93e68b0a62b6e06873168629a#31782 is picked before #31756 because both edit the same restore branch of
src/kiro_crew/cli_commands_memory.py.Hand-resolved hunks:
src/kiro_crew/cli_commands_memory.py(theif source is None:branch): comment-only conflict. 0.12 carries a longer comment above the oldraise FileNotFoundError; this takes the PR side (theValueErrorraise and theFileNotFoundError→ValueErrorconversion), as onmain.src/kiro_crew/env.py(docstring ofexpand_spec_args): docstring-only conflict. Takes the PR side, which adds the WindowsUSERPROFILE/USERNAMEparagraph; the code hunk applied cleanly.docs/decisions/README.md(decision index): one index-row conflict. Adds only this PR's row, after the measure-button row fix(release): backport twenty-one user-facing fixes to 0.12.0 line (s3b3) #33549 brought in.main's neighbouring meetings-sketch row is not brought in, because its decision file does not exist on 0.12.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 25 test files these picks change ran here, one file per pytest process, in a freshuv venv -p 3.12withPYTHONNOUSERSITE=1, no xdist andTMPDIRoutside~/.kiro/crew: 1,126 passed, 0 failed on this head (873168629a), re-run after the rebase;test_memory_backup.pyalso ran (26 files, 1,200 passed).tsc -b --forceexits 0.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 #31782, #33495, #33494, #33366, #33488, #33491, #31750, #31556, #31756, #33426, #28324, #28286, #28492, #33487, #33429, #33506, #33370, #26731, #33507, #33515, #33003, #33435.
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.