Repository navigation
fix: backport 0.12.1 fast batch v2b-1 (20 user-facing fixes) - #32959
Conversation
…1098) A remote crew's SSH tunnel stayed in ERROR after a short network outage until the operator clicked reconnect. When the rebuild and the re-mint both failed before an ssh child was running, no exit callback fired and the recovery returned without scheduling another attempt. After the attempt cap it only ran a diagnosis. Self-heal now schedules the next attempt after each of those failures, and after the cap it retries every five minutes. A tunnel that is still down one hour after its first failed recovery stops retrying and keeps an error that says it gave up. A healthy probe, a reconnect or a disconnect clears the clock. Closes #31088 (cherry picked from commit 9762e13)
* fix: read an unreadable cgroup usage as unknown posture * fix: keep measured tight or critical memory beside unreadable usage --------- Co-author on main: Jacob Largent (cherry picked from commit 5275d7b)
…#31267) _persist_channel_config, the allowlist and tracking list helpers and the Slack config modal each write under config.json slack. When the slack entry, slack.channels or one channel held a non-object value, or an allowlist or tracking list was not a list, setdefault returned it and the next index, call or append raised a bare TypeError or AttributeError. Each mutate now returns None for the wrong shape, so nothing is written, and the writer raises a ValueError naming the entry. !channel answers it in the thread and the channels-modal selects log it. Every button keeps its order, and the error ends the click the way an unreadable config does. The config modal logs a warning and changes nothing. (cherry picked from commit 603115b)
On Windows, saving a file from the dashboard (file editor, steering update, SKILL.md save) replaced it with a copy that took its folder's permissions. A file the user had narrowed below its folder came back wider after one save. The save now keeps the file's own permissions, and refuses the save, leaving the original untouched, if it cannot. Part of #13035 (cherry picked from commit f11c19b)
update_status.published_index answered every app by name from the first source that listed it. An app installed from a second registry that publishes the same name was compared against the first one's version, while its Update installs from the source the install recorded (_pinned_registry_entry). The app's banner and the automatic update then saw a release the Update would not install, or missed the one it would. An installed registry app that recorded sourceUrl and sourceRegistry is now answered from the cached row with that URL in that registry, and gets no answer when no cache holds that row. Other names keep the first-claimant rule, except that a bundled seed row outranks a catalog row for a different repository, as _resolve_registry_row resolves it. (cherry picked from commit a4ed25d)
On Windows, a process whose parent exited first is out of reach of the parent-PID walk that drains a runtime's tree, and a root that exits on its own only had its own registry lines retired. Its MCP helpers (and any browser they started) kept running and holding memory. - apply_job_limits hands the Job handle it creates to the admitted tree whose pinned root it bounds; after the exact-handle drain completes the tree calls TerminateJobObject and closes it. KILL_ON_JOB_CLOSE stays unset, and a spawn with no Job tears down by the drain alone. - A confirmed observed exit on Windows runs the same drain as the kill path, which also retires the root's tracking. Closes #24821 (cherry picked from commit da2595b)
…31843) The Discord and Telegram dispatchers named a new generation's predecessor only from the allocation's capture of the session mapping. With the mapping entry gone (a restart that lost it) the next crew log cited nothing, and every create passed no determination, so a slot's first store and an unnamed predecessor wrote the same silence. Read the slot's newest crew log from the store before a claim with no registered session, keep the capture ahead of it, let the store decide while the live session owes a history replay, and pass the determination through to session/opened, as the dashboard runner does. The owner-DM surface pin now requires the store tier behind the capture. Closes #14127 (cherry picked from commit b6111e3)
…while held (#21377) * fix(autonudge): read a legacy loop's pull request between cycles and while held An ungated (gate=false) prompt loop naming one pull request now reads the pull request's state every five minutes while it sleeps and runs its cycle early, once per head, when the head conflicts with its base (BEHIND does not wake it: under up-to-date protection every sync is behind again). The read never suppresses a cycle and never ends the loop on a merged or closed reading, so the gate=false opt-out keeps its promise. A loop held for an unanswered tool approval reads the same state on every held tick: a pull request that merged or closed during the hold settles the loop as monitor_terminal with the usual expiry notice. The reconciler now rescues a held loop whose timer died, since a held loop keeps its own timer. * test(autonudge): stub the cycle sleep seam in the ledger-extension timer tests Main's _nosleep fixture now replaces firing's asyncio namespace; the timer's cycle wait lives in subject_reads._nap, so the fixture stubs that seam too. --------- Co-author on main: Kyle Seaman (cherry picked from commit 5902e29)
…ard (#25343) Reuse useKeyboardLift (already used by the bottom terminal panel) on the composer root instead of adding a second keyboard-lift hook. The hook now measures the box's own bottom edge against the visible band, so a composer above a docked terminal rises only by the part the keyboard covers. Callers spend the lift as bottom padding inside the measured box, so its bottom edge stays put (no feedback when re-measured, including a fullscreen terminal) and the composer dock's measured height grows, keeping the newest transcript lines visible above the lifted composer. The terminal now measures its padded wrapper instead of the panel. Refs #25240 Co-authored-by: kirocrew-agent <agent@kirocrew.local> (cherry picked from commit 69814f4)
…31120) A registry whose index rows carry no version leaves the version to each app's app.json. For an installed app the listing reads that file from the installed checkout and caches it apart, and only the owner-tier prewarm writes the source manifest cache that update_status reads as what the branch publishes. An app installed from an index-tier registry of this kind was never shown, or automatically given, a newer release. refresh_registries now reads those tips after its manifest expiry. For each versionless row of a fresh index-tier index that names an app installed from that row, one credential-free throwaway clone of the branch tip supplies app.json, written as the source manifest cache. At most 20 rows per registry, 4 at a time, within 30 seconds. (cherry picked from commit 85936d6)
) (cherry picked from commit b85bbfe)
…31487) A gateway started with --approval yolo still stopped dashboard chats on an approval card: only the background approval path read the flag, and dashboard chats read the process-wide safety override, which nothing armed from it. At startup the dashboard takes the declared grant from the flag (audit source cli), the same grant agent.dangerouslySkipPermissions uses: it is refused when an approval_modes policy denies yolo and bounded by the yolo_duration scope and yolo_max_seconds. Picking another mode in the dashboard ends it for dashboard chats. Closes #29697 (cherry picked from commit f2d2537)
#32114) When a cron or sub-agent run is force-stopped after its root process has exited, the kill could not walk the tree from the root, so the children of a recorded child (an MCP server's own children, say) ran on while the run was recorded and audited as reaped. Behind a gone root the kill now walks from each recorded child that still reads back as the run's, before the group signal and the sweep re-parent its descendants out of reach, and names each unrecorded descendant still running afterwards as a kill failure. They are reported, never signalled. The cron reaper and cancel audit rows carry the failure on kill_failed, as the sub-agent reaper's row does. Closes #13544 (cherry picked from commit 0c15056)
#32392) The heartbeat prepends its HEARTBEAT_KEEP reminder to every task, and a Slack-thread auto-nudge prepends the cycle header (and sometimes a ledger snapshot) to the loop's message. Neither caller passed skill_trigger_text, so skill triggers matched Kiro Crew's own words and loaded skills nobody asked for. Pass the task and the loop's message, as the other unattended callers already do; a monitor wake is unchanged. Builds on #25617. Co-authored-by: Leon <leon.lai.yc@gmail.com> (cherry picked from commit bac6d9a)
The installer scripts already install a pinned, sha256-checked Session Manager plugin, but the gateway's own installer still downloaded the package from AWS's plugin/latest path and installed it as root with no checksum, and the copy-paste command the dashboard shows did the same (its dnf/yum line installed straight from the URL). cloud/ssm.py ships in the wheel without packaging/pins/aws-tools.env, so it carries a copy of the plugin version and the six macOS/Linux package checksums, kept equal to the env file by test_installer_tool_pins.py. Both install paths now fetch the pinned release and refuse a package whose sha256 does not match before any sudo step. Closes #24313 (cherry picked from commit 37aad40)
…#32318) The whisper model and ffmpeg decoder fetch (stt.models.stream_pinned_payload) and the update engine's manifest and wheel fetch (wheel_engine_transfer's _read_bounded) opened their connection with urllib.request.urlopen, whose default redirect handler follows a hop to any host, plaintext included. Both now open it through asset_downloader.build_opener, the one place that answers which redirects a pinned transfer follows. The stt urls are source constants or the operator's environment override and their upstreams redirect to a download host of their own, so they take the cross-host, https-only policy. The release CDN serves the manifest and wheel itself, so that transfer keeps the same-host policy, except while an operator's KIROCREW_CDN_BASE mirror is set. Each loop keeps its own bounds (pre-write size ceiling, cancel, total deadline, mkstemp staging), which download_to does not carry, so only the connection moves. Part of #22230 (cherry picked from commit 9ea1ed8)
A single Stop of a queued run whose store cancel did not land was told to the caller anyway, and its one re-post was never read, so a contended store could leave the row queued for the next start to run. The stop still drops the window entry and registers its stopped record at once, which the claimer refuses. The unlanded cancel is now settled through the refused-row settle's own retry, now shared by both: a queued-stop marker goes to disk for a persistent run, then the cancel is retried until the store answers. The record's report and cancel()'s answer wait for that, released early once the marker is confirmed and the store holds the row for cancelled. The boot reconcile settles a queued or admitted row carrying the marker cancelled, the tombstone pruner keeps it until then, and the shutdown drain no longer deletes it when it cancels the stop's undelivered report. Stop all and the parent-end teardown use the same settle and still publish without waiting. Part of #21339 (cherry picked from commit f8d6525)
…2113) A run spawned by another sub-agent has subagent:<id> as its parent, a key no tab shows, so its panel frames named a slot nobody reads and the run never appeared in any Subagents panel, live or after a reconnect. Admission now stamps SubagentInfo.root_session_key from the spawning run's record, and the live per-run frames and the reconnect replay carry that root chat's slot. The queued depth, the wave-start marker, the injection-failure transcript card and subagent_status stay on the parent's key. The client drops or ends the approval card a nested spawn raised on its parent's key, so the run is counted once. Closes #13116 (cherry picked from commit 1a1d650)
) On Linux an orphaned pod HOME could not be reclaimed: `pod down` and `pod prune` both failed on `Unit ... not loaded`, because stop_pod returned on any failed `systemctl stop` before it deleted anything. And `pod ls` listed a HOME as an orphan whenever its unit was not active, so a gateway started outside the unit was shown with a delete command while it served, and `pod down` deleted it on a host where the template unit is installed. stop_pod now reads the unit state before stopping. When systemd was not running the pod and a HOME is there, a failed stop means nothing to stop, and the delete goes ahead only when the new pod/home_liveness.py sees nothing using the HOME: no gateway holds its gateway.lock and no readable process has its cwd or an open file under it. orphan_homes applies the same check, so pod ls, pod prune and Dev Fleet list only HOMEs the delete would take. Builds on #12105 by @dimwael. (cherry picked from commit 46e26c0)
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 ChatEmbed busy Send button's new "Steer" state. The PR supplies no images, so no blind read ran and nobody has looked at the changed controls. Evidence gaps
Notes
[UX-REVIEWED] 15222c2 Description read[DESCRIPTION-READ] a246070710ffb609d4930358715c37e0e10a2d9f13441e7066b1130f1e40671d 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 This PR copies fixes that are already merged on What this change shipsInventory (10 items) — 10 justifiedIntent: give 0.12.0 users the fixes that already landed on
[FIRST-PRINCIPLES-REVIEWED] 15222c2 Description read[DESCRIPTION-READ] 223067c423663d3e387d946980b8912e429c997c702e180f265952cc6300828b 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: missing a screenshot or recording of this revision's ChatEmbed busy-Send control and of the iPad composer lift. Manual verification is also N/A for code that only runs on Windows and iPad, which CI does not run on BlockersNo rendered or manual evidence that this revision works where CI cannot check it Notes
[DESIGN-REVIEWED] 15222c2 Description read[DESCRIPTION-READ] c515b6c78d62f208e5cfa27e349fb4b6b9322cdf6a83721dab4a8f0e715a370c 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 |
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: |
Problem / Motivation
Goal: Bring twenty fixes for problems a 0.12.0 user can hit, merged on
mainafterv0.12.0, ontorelease/0.12.0, one commit each, in a PR of its own.Why it matters
Each fix changes behaviour a 0.12.0 user can run into today. Every pick applied to
release/0.12.0cleanly, so the code each one replaces is on release.Not a goal
main; nothing is changed only here.CHANGELOG.mdchange.What changed (motivation → approach → change)
Each main squash commit was cherry-picked with
-xontorelease/0.12.0@5dfe33aae5, in merge order onmain, one commit each. Every pick applied cleanly, and every commit keeps its original author.9762e131a323e3eacf5f5275d7bc0ae76aea48c6603115b9847ee779a126d0c6eddaaf12de711935f11c19b95cc19d62a075a4ed25d310c2ef68efc4da2595bcfd2657c1f2c9b6111e3333d730253a345902e29db3a733f7abce69814f4b60641e29424985936d6be4504c9ebfd0b85bbfeb63e3f117b60bf2d2537ff9e4d09ca47c0c15056eb92b854b12a9bac6d9aaca77ab853e8937aad408c2d55e124f3e9ea1ed889033c96f242ff8d65252668a828165891a1d65045ef187e5056646e26c009815222c2f8bPicks whose co-author trailers carried internal addresses have each trailer rewritten as a plain "Co-author on main: " line, because the commit guard refuses the address.
Backwards compatibility
Compatible: each change is the one already merged on
main, and its compatibility note applies unchanged.Removes nothing: no setting, key or caller is removed.
Tests
Proof: not provable -- other: a backport of main commits whose proofs ran on their own PRs
release/*, so the direct test modules were run here: the 66 Python test files these picks touch gave 4,230 passed and 36 skipped.tsc -p tsconfig.app.json --noEmitpasses; the 15 vitest files these picks touch gave 210 passed;lint:i18npasses at the same count as clean release.mypy src/kiro_crewpass.i18n:checkwithout a base reports the same findings class as a cleanrelease/0.12.0on this host, and fewer of them.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 #31098, #31095, #31267, #26156, #31695, #31118, #31831, #31843, #21377, #25343, #31120, #31978, #31487, #32114, #32392, #32380, #32318, #32312, #32113, #31828.
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.