Repository navigation
fix: backport 48 fixes to 0.12.1 (batch b1 of 4) - #33974
Conversation
(cherry picked from commit 7b00f4a)
(cherry picked from commit 08c052c)
…30810) prove.py reverts each hunk with `git apply -R -` and restores it with `git apply -`, feeding the patch on stdin with text=True. On Windows a text-mode pipe writes every \n as \r\n, so the patch no longer matched LF files and every revert failed with "patch does not apply". Both calls now go through _git_apply, which encodes the patch with utf-8/surrogateescape and passes bytes. The _git capture that produces those diffs had the mirror problem on every platform: text mode reads with universal newlines, so a CRLF file's diff lost its CRs and could not be reverted either. _git now captures bytes and decodes without newline translation. Fixes #29345 (cherry picked from commit 39ccfec)
…30860) The agent prompt recommended $KIROCREW_SCRATCH and said it is reclaimed once the session's processes are gone, but never said those processes are themselves bounded by the agent runtime. So `nohup ... &` read as a supported way to outlive a turn, and an agent told a user a training run "runs by itself in the background" until a runtime recycle stopped it hours later with nobody told. The prompt now says a process launched from the shell ends with the runtime and forbids the "runs by itself" promise. Work that must outlive the turn goes to a one-shot cron command job when it fits that job's one-hour timeout, and otherwise to a command the user starts outside Kiro Crew. The cron guide explains the same limit with a working example, and the test ties the quoted cap to the cron_add schema. (cherry picked from commit 61a9c63)
The signing service refuses a request body over 102,400 bytes. The universal app lists every nested Mach-O, and the Feishu SDK libraries took that body from 96,725 to 136,880 bytes, so every universal macOS signing failed with "request entity too large". A manifest under 100,000 bytes is still one request, unchanged. A bigger one is signed in the order the service documents for nested code: the regular files go out in parts under 90,000 bytes each against the same upload, their signed bytes are copied into the app, and a final request signs the bundle entries and the app over that tree. Cutting the list and merging the outputs alone would leave the app seal covering one part only. Co-authored-by: w <w@k.local> (cherry picked from commit 1543450)
The About page and the update popup already show a pending update's notes when the update feed carries them, but the macOS, Linux and Windows publish lanes write their channel files by hand and never set releaseNotes, so the notes block stayed empty on every platform. Each lane now writes releaseNotes from the version's CHANGELOG.md section, picked the way Settings > Releases picks a build's row and printed as one JSON string, which is also a one-line YAML scalar. A version with no section, a nightly in practice, gets empty notes. Closes #6521 (cherry picked from commit c38eec4)
…tside the checkout (#30841) A test golden path's pytest basetemp (pytest-of-<user>) landed under the worktree root because child_env points TMPDIR/TEMP/TMP there. A real checkout's own rootdir conftest snapshots its root and fails a clean run that leaves a new entry, so pytest exited 1 and the gate judged an unchanged clone broken. Point each test row's pytest scratch base outside the worktree instead; real pass/fail verdicts are unchanged. Co-authored-by: kirocrew-agent <agent@kirocrew.local> (cherry picked from commit 360960f)
…g request (#31401) The first real run of the split universal signing was refused at part 1: "Electron Framework.framework: code object is not signed at all". Every signing task, a part included, ends with the service's app pass, which signs the frameworks under Contents/Frameworks. Part 1 listed only backend files, so the executables inside those frameworks (Electron's Helpers, Squirrel's ShipIt) were still unsigned when the pass reached them. Every request now lists the shell entries (helper apps, frameworks and the executables inside them), and only the backend files under Contents/Resources are cut across the parts. The final request signs the shell over the merged tree; a part's signed shell files are not copied. sign.sh also waits for every part before it fails, so one run reports each part's answer. A manifest under 100,000 bytes is still one request, unchanged. Co-authored-by: w <w@k.local> (cherry picked from commit f490ce1)
…line (#31712) After Report a problem builds its bundle it says "Diagnostics ready — {{redactions}} secret(s) redacted across {{files}} file(s)." Both counts sat in one single-form string, so one secret read "1 secret(s)" and German, Spanish, French and Portuguese copied the shortcuts. Read each count from its own plural part, with the verb or preposition that must agree inside it, and keep the outer sentence one key so each language still orders the parts. Numbers are locale-formatted. (cherry picked from commit 486806e)
The JWT spelling matches on shape alone, so a host label holding eyJ (honeyJar.atlassian.net) was cut in half by the log record filter, refused as a credential by the decisions gate, and mangled by the dashboard's sanitizer. The chat scrubber already counted a hit only when its first segment decodes to a JSON object. The log filter and the sanitizer now apply that same check, the gate leaves the JWT spelling to the scrubber, and the check lives in one stdlib-only module the log filter can import at bootstrap. A JSON object Python cannot parse still counts, so the check fails closed. Builds on #16573 by NicholasRBowers. (cherry picked from commit 5635855)
…31203) PUT /api/crons/{id}/secrets converted expected_ts with float() on both the approve and the deny path. A JSON integer has no width limit, and float() of a very long one raises OverflowError, so the request answered 500. Refuse it with the route's own 400 invalid_secret_env before anything is audited or written; the pending request stays. (cherry picked from commit 7315727)
…32218) Settings > Computer Use read Accessibility and Screen Recording by running `kirocrew computer doctor --json` in a child with a 5s budget. The two TCC calls take milliseconds, but the child first imports the whole CLI (about 1,700 modules) and, with the feature on, projects the chat backend's agent spec. That measured 2.9 to 4.6s warm and over 11s on a cold cache, so on a busy host the probe timed out and both rows read "Could not check" while both grants were held. The handler now spawns `python -P -m kiro_crew.computer_use.permission_probe` under the gateway's own interpreter. The module asks the same `probe_permissions()` the doctor asks and prints one JSON object, importing only the driver seam (about 0.5s warm). It stays out of process, so a native fault still costs one row and not the gateway. A test fails if the module loads `kiro_crew.cli` or `kiro_crew.agent`. Closes #32193 (cherry picked from commit 1ef44e7)
…as (#32363) When /etc/systemd/system/kirocrew.service is an alias of another, stopped unit and our per-user unit is installed and stopped, `kirocrew restart` skipped the stopped-service refusal and started a detached gateway outside the user unit. The refusal now falls back to the per-user unit when the system file is not ours, so it refuses and names the user unit instead. Closes #25979 (cherry picked from commit 67a7bc8)
A Strands key field's Save and the remove confirm's Remove went natively disabled while their write ran, dropping a keyboard user's focus to the page, and a refused write left it there. A landed write unmounted the pressed button, so focus fell again. Both now use aria-disabled with a guard; a refused remove keeps its confirmation open; and once the field shows its new state focus moves to Replace after a save, to the empty input after a remove, or back to a refused dangling Remove, unless the user already moved it elsewhere. (cherry picked from commit 52d6e05)
…32398) Where pysqlite3 is installed, the stdlib sqlite3 is a second SQLite library in the gateway. Its open of a database the gateway already holds through _sqlite_compat resets the -shm under that connection, whose next read fails with SIGBUS or "database disk image is malformed". Pod sign-in staging snapshotted kiro-cli's data.sqlite3, which the gateway also reads through _sqlite_compat, with the stdlib driver; it now uses _sqlite_compat too, as does the test that stands in for kiro-cli. A new test lists every module that references the stdlib connect with the database each opens, so the next such open has to say what it opens. (cherry picked from commit 666f962)
The gateway writes its pid record off its readiness path, seconds after the credential pod up waited for. The mint proves port ownership from that record, so a mint right after readiness could not prove it, withheld the token, and pod up exited 0 with an empty token. For a pod it started and mints for, pod up's readiness now also waits, inside the same budget, for the mint's own ownership check to pass. When the budget ends first the pod is still reported up and the mint explains the withheld token as before. (cherry picked from commit 2eeceb4)
#32627) The page check removed its throwaway browser profile with one ignore_errors rmtree right after the group kill. A helper the browser starts in its own session (its crash handler) is not killed with the group and can add an entry while the tree is removed, so the directory was left in the temp dir with nothing said. The removal now passes over the profile again until it is gone, bounded by the reap limit, and logs a warning when it still cannot. Closes #30934 (cherry picked from commit d802999)
…32486) On an overlay filesystem a real static/dist that lives in a lower image layer cannot be renamed aside: the rename fails with EXDEV, so both the link and the copy fallback of `kirocrew frontend stage` failed and the old build stayed live. On EXDEV the directory is now copied aside, removed and the link moved in; a removal that fails or is interrupted part-way is undone from the copy, so the old build keeps serving. Closes #29779 (cherry picked from commit ddd42be)
…tart refuse it (#32814) On Windows the gateway.lock pid stamp is unreadable under the msvcrt byte lock, so lock_holder answers LockProbeError(held=True) and _config_dir_owner returned None: the desktop-supervised refusal from #25705 never fired. On that one answer, name the holder from the gateway's fresh run/gateway-<port>.pid record instead. Fixes #32714 Co-authored-by: Zezhen Xu <zezhexu@amazon.com> (cherry picked from commit aa08aa6)
…32484) A dashboard session that owes a history replay re-armed the replay lease on every exit but a clean landed end_turn. A replay-bearing turn that streamed output and then failed (a throttle 5xx after the first chunk, a timeout, any error after output) therefore made every recovery turn rebuild the whole session context, although the live session already held it. Such an exit now notes the replay delivered to the provider's context incarnation, using the same kept-the-prompt rule the first-turn history debt settles on. The lease stays armed so close_all keeps the prior SID, later prompts on that native conversation are not built as a session start, and the next landed non-slash end_turn promotes the fresh SID. A respawned process is a new incarnation: a turn admitted just before the respawn commits nothing, and the next one sends the replay again. Closes #29535 (cherry picked from commit 12a1a7d)
…32205) A loop-stall crash dump holding only its startup header was read as a clean exit, so startup swept it and rotation sacrificed it first. The header is written at boot, though, so a gateway that was killed or crashed without the stall alarm firing leaves the very same file, and that file was the only trace of the death. An orderly end now leaves an empty <dump>.clean marker beside its dump, never touching the dump's bytes: from the gateway's shutdown right after it disarms the watchdog, and from a pre-exec stop before an in-place restart, whose exec skips that shutdown. A dump's first header line carries a token saying its build writes that marker, so a dump an earlier build wrote still reads as a clean exit after an upgrade. The sweep deletes a header-only dump of a dead owner, with its marker, only when it reads as a clean exit. One that does not is kept as an unexplained exit: rotation sacrifices it after clean exits and before unreadable files and dumps with stacks, kirocrew doctor shows an information-only "unclean:" row with the count and the newest, and the diagnostics bundle leaves markers out and gives these headers only the crash-dump slots nothing else fills. The dump reader opens without blocking, so a FIFO named like a dump is refused instead of hanging. Closes #14724 (cherry picked from commit 0956e65)
(cherry picked from commit 86a4fef)
…32551) Pod boot copies kiro-cli's data folder into the pod so the pod is signed in. kiro-cli 3 also unpacks its engine there, under kas/, and the copy never finished it: the folder is bigger than the 512-file cap, and the walk held every folder's descriptors until the end, so under a pod's 1,024-descriptor soft limit it ran out and skipped the rest. kiro-cli then trusted the partial engine and failed with ERR_MODULE_NOT_FOUND. kas/ at the store root is no longer copied; kiro-cli unpacks a complete engine itself on first start. Each folder's descriptors are closed before the next one opens, in both staging twins. Closes #30357 (cherry picked from commit 8ad14ca)
kirocrew chat sent the typed text to the provider bare. Every mirror rules a custom agent's prompt and its skill:// resources withheld on the premise that both reach the harness in the context message ContextBuilder.build_message assembles, which the dashboard, channels, cron and kirocrew run send. So on claude, codex, opencode, goose and the other mirrored harnesses, `kirocrew chat --agent X` ran the harness's own default persona with none of X's skills. On those harnesses the CLI now builds the first message of each session with build_message, as those surfaces do: the -m message, the REPL's first turn, and the next turn after a compaction restart, a typed /clear or /compact, or a changed context_incarnation, since each of those sessions holds none of it. kiro-cli and KAS load the agent themselves, so a CLI session on either still sends the typed text bare, as before. A backend process that exits between turns shows up only once the send that respawned it changed the incarnation: the process's own liveness would be a shared runtime's answer, not this session's. Later turns are sent as typed. A turn that opens with a slash command is sent as typed too, because a harness runs a command only when it opens the prompt, and the context waits for the next ordinary turn. A build that raises prints one line and sends the typed text; the next turn tries again. build_message decides per harness what the message holds. It reads the store resolve_agent_bindings resolved: a member bound to a named store gets neither memory nor lessons, since the CLI does not bind that store and Global memory is not the member's to read. The builder lives in a new cli_chat_context module, imported inside _chat so other CLI commands do not load the context stack. The provider-shutdown tests in test_cli_chat_permission.py stub the tool gate with a real _ToolGate over a default HookManager, as production always builds one: _chat hands the gate's hooks to the context builder before the backend starts. (cherry picked from commit eba4f3b)
…33150) The in-place update restart wrote the crash dump's clean-exit marker before its final, unbounded callback drain. A gateway killed with SIGKILL while that drain hung kept the marker, so the next boot read its header-only dump as a clean exit and deleted it, and the doctor's unclean: row missed that death. The marker is now written after that drain. A zero-wait look right after the write catches work that arrived while it was being written: it withdraws the marker and drains that work first, so a received message is still never left behind the exec, and no channel-wide ingress quiesce is needed. Closes #32782 (cherry picked from commit 36deb7c)
The default workspace root is one folder per OS user, so every data home with no workspace of its own runs its task runner in the same taskrunner_main folder and loads and rewrites the same runs.json. Another home's runs showed up in the task list after a restart. In a work dir under that shared root, a KIROCREW_HOME override now keeps its registry in runs-<home digest>.json. The default home, and every work dir outside the shared root (a pod's own workspace, a configured taskrunner.workspace_dir), keep runs.json. Closes #29713 (cherry picked from commit a1f286a)
…#33188) close_slot waited (for the slot's guarded history write, the app close hook, the AutoNudge lock) and then acted on the slot's key: it retired the auto-nudge loop by key and popped the key, asking only at its last pre-pop check whether the slot still held that key. When another close popped the slot and a send recreated the key during a wait, the close stopped the replacement's loop or monitor, or popped the replacement, and its rollback could not put either back. After each of those waits the close now checks that the slot still owns its name and refuses with 409 slot_replaced before the next step that acts by key, releasing only its own close fence. The tab DELETE route answers that refusal as a 409, and session_close answers it as its own target_replaced. Closes #32903 (cherry picked from commit 2eaea3d)
A gateway under systemd, launchd, Docker or a detached spawn wrote its startup dashboard link, ?token= included, to a kept log on every boot, and again from the slow-browser auto-open fallback and from kirocrew restart. Anyone reading that log inside the link's five minute window could redeem it for a 20 hour dashboard session. dashboard.urls.login_link_may_print() is true only when stdout is a terminal, and all three prints ask it. Off a terminal the gateway prints each URL without its query string plus a kirocrew token hint, and restart mints no link. A terminal's output is unchanged, and browser auto-open still receives the link. (cherry picked from commit c9b9b1d)
The runtime liveness check reads its own root pid for the handle-count probe and marks the site pid-owner-ok, like the RSS probe beside it. The pinned owner-marked count still said 8, so the Static Ratchet Gate fails on main. Pin it at 9. (cherry picked from commit 044a860) Backport note: release/0.12.0 pins 7 owner-marked sites (it lacks main's code_review_sage review_pool.py marker), so the #32376 pick's handle-count probe makes 8, not 9. The pin is 8 here.
Sensitive Change ReviewSensitive files changed:
|
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks: each hunk is a cherry-pick of a fix that was already reviewed on main, and none of the dashboard changes adds or reshapes a surface. They change wording, plurals, the folder-chip translation and Strands key focus handling. Notes
[DESIGN-REVIEWED] 4bcbd58 Description read[DESCRIPTION-READ] 37e156cedc1fd36e282b3556e812780ec7ac70b60a9d2563dd11285af78ce6f8 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 |
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 no image, so the blind read was not performed, and three dashboard surfaces this backport changes went unseen. Evidence gaps
Notes
[UX-REVIEWED] 4bcbd58 Description read[DESCRIPTION-READ] b05f9f0786bb4e3e68a145921dea037191395194aab61443e48a3c70e577fc2b 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 reviewed main merge cherry-picked onto the release branch, so no new shape is decided here. One premise can't be read: the intent file is cut off at 8000 bytes before the two release-only commits are listed. Notes
What this change shipsInventory (10 items) — 9 justifiedIntent: give 0.12 users the bug fixes that already merged to main, as a patch release. This is a FIX.
[FIRST-PRINCIPLES-REVIEWED] 4bcbd58 Description read[DESCRIPTION-READ] ffa232d7e3c9e1ecba0e54d630039d630a43e768d15e31647d63f5fd97edf45b That sha256 names the PR description this verdict read. Recompute it with |
Sensitive change review
|
Security Scope Review -- ✅ PASSWhich legitimate operations does Verdict basis: adjudicated, zero confirmed regressions. Security scope review -- adjudicated candidates
No candidate flipped from allowed to refused. Scope-Verdict: PASS Reason: This diff does not touch the shell deny composite ( TIGHTENING: skill-script validator sensitive-path regex now accepts ADJUDICATED: 25 candidates, 0 confirmed newly-refused PLATFORM COVERAGE: I wrote posix and windows rows; no UNADJUDICATED:
[SCOPE-REVIEWED] 4bcbd58 |
GPT 6.1 Review — ✅ human override acceptedHuman judgment by @bolichen97 overrides the GPT 6.1 finding for [GPT-OVERRIDE] 4bcbd58 This comment is updated in place on each push. The model was not re-run because an authorized human decision supersedes it. False positive or not applicable? A repository writer can comment: |
|
/ai-review override gpt 4bcbd58 pre-existing: agent: verbatim cherry-pick of main a1f286a (#32457), same patch-id; origin/main still runs this one-time shared-registry adoption on the loop (run_registry.py:115). The adjudicator rated it LOW harm: a self-clearing first-boot stall, not a security, data-loss or crash defect. Moving it off the loop is a main fix first, then a backport, not a change to this pick. |
Human judgment recorded@bolichen97 recorded a pre-existing release (class
This decision applies only to this commit. A new push requires a new judgment. |
|
/ai-review override fable 4bcbd58 not-applicable: agent: Opus discovery cannot finish one pass over this 879,120-byte, 233-file diff: both runs ended with no [OPUS-DISCOVERY] marker and no finding (https://github.com/kirodotdev/KiroCrew/actions/runs/42656690652). 48 of the 50 commits are verbatim cherry-pick -x of merged main commits (e.g. a1f286a #32457, 5635855 #31997); GPT, Design, UX, First Principles and Security Scope reviewed this whole head, the 2 release-only fix-ups included. |
Human judgment recorded@bolichen97 recorded a not-applicable release (class
This decision applies only to this commit. A new push requires a new judgment. |
Opus 5.5 Review — ✅ human override acceptedReviewed Human judgment by @bolichen97 overrides the Opus 5.5 finding for [OPUS-OVERRIDE] 4bcbd58 Verdict recorded from an authorized human decision for commit False positive or not applicable? A repository writer can comment: |
Problem / Motivation
Goal: Ship 47 user-facing fixes that merged to main after the 0.12 cut, plus the one test pin they need, to 0.12.1 users (hotfix batch b1 of 4).
These fixes are on main and not on
release/0.12.0. Each one is a bug that 0.12 users can hit today. All 48 main commits cherry-pick onto release with no conflict once the earlier picks are in; two release-only commits make two of them build on release's own code, listed below.Why it matters
0.12 users keep hitting these bugs until a patch release carries the fixes. They cover the CLI (
bench,pod,restart,chat), macOS signing and desktop update notes, the dashboard (chat replay, slot close, usage, artifacts, Computer Use, the Strands key focus, Report a problem), crash dumps, task runs and crons, Windows path and credential handling, KAS tool globs, subagent admission and background runtimes, and a dashboard login link, token included, that every service-managed gateway wrote to its log.Not a goal
git cherry-pick -xof the main merge commit, or one of the two release-only commits listed below, each of which keeps release's own code and applies only what one pick needs.What changed (motivation → approach → change)
The 0.12.1 hotfix list was re-reviewed PR by PR against
v0.12.0: a bug in released code, not on release yet, and not an app PR. This batch is slices hf1 to hf3 of that list, applied in main merge order, onecherry-pick -xof each main merge commit, onrelease/0.12.0at34898f9795, after the release owner's late-4 (#33742) merged. The whole list was assembled and tested as one staging branch first (see Tests), and each batch is a contiguous slice of it, so every merged batch leaves release at a state that branch already passed through.7b00f4af96kirocrew benchprinted its errors and refusals on stdout, mixed in with its resultsb81f3076f808c052cdc6kirocrew pod --helpnamed systemd units and a journal on every host, and offeredpod restarton Windows, which refuses it39ccfec70eprove.pyrevert failed with "patch does not apply", because the patch went through a text-mode pipe (fixes #29345)61a9c6393bnohup ... &as a way to outlive a turn4188c0f348154345004ec38eec4166360960f2b7f490ce1095486806e1385635855f0eeyJ(honeyJar.atlassian.net) was cut in half in logs, refused by the decisions gate and mangled by the dashboard sanitizer as if it were a JWT73157272abexpected_tswas too large for a float answered 500 instead of 4001cacf51b35banner_metarules1ef44e73d067a7bc86f5kirocrew restartstarted a detached gateway beside a stopped user unit when a systemkirocrew.servicealias of another unit hid it (fixes #25979)52d6e05345666f96222d2eeceb4b07kirocrew pod upcould exit 0 with an empty token, because it minted before the gateway wrote its pid recordd8029998fdddd42be19bkirocrew frontend stagefailed with EXDEV and the old build stayed live (fixes #29779)aa08aa6001kirocrew stopandrestartdid not refuse the desktop app's supervised gateway (fixes #32714)12a1a7d50f0956e6534b86a4fef09b8ad14ca919eba4f3b053kirocrew chatsent the typed text bare, without the session context (custom agent prompt, skill resources) every other surface sends36deb7c842a1f286a4e02eaea3d7ed8509138a7ckirocrew pod downreported a failed stop for a pod whose gateway took longer than 15 s to exit (fixes #33270)2bc7c8f190c5f6643172infornanas a time6264d9ef0f0781f8567eb1bbe69f6418de2125b2/opt/homebrew/binand/usr/local/bin, which resolve to folders on the current drive that any local account can create3e4a72f85d33af78eeef\\?\extended-length prefix did not count as inside the provider CLI workspacead9a25a7e2ssh_auth_sock) reached a pooled MCP backend instead of being scrubbedd5728edc39e1f14081accaf79f4e4604e5fa68e8allowedToolsentry with?went to KAS, which reads?as a literal where kiro-cli reads it as any one character21c57e69778f29b1dcf7c9b9b1df40044a860a15Hand-resolved hunks. One, in a release-only commit. The staging branch carries this commit after #32411 (batch b3), so its context held #32411's lines:
src/kiro_crew/dashboard/chat_turn/prompt_assembly.py_turn_settle_debts, in "settle a failed replay turn on release's kept-the-prompt rule"settle_drained_subagent_noticesrequeue block (they arrive with #32411 in batch b3); keep only this commit's_first_turn_history_deliveredhoist and the replay call changeRelease-only commits. Written for release, not picked, each needed by a pick in this batch. They sit after the hf2 picks, in front of hf3.
_kiro_cli_kept_the_prompt, which main moved intoturn_stats.pyin feat #17165 and release does not have, so a replay-bearing turn that did not land raised NameError (flake8 F821). This settles it on release's own kept-the-prompt rulequick_prompts.user_quick_promptfrom feat #31463 (saved prompts as slash commands), which is not on release, so importingkiro_crew.cli_chat_contextraised ImportError. This reads a CLI slash command againstQUICK_PROMPTSonlyPrerequisites not picked:
_kiro_cli_kept_the_promptintoturn_stats.py): a feature; the first release-only commit above instead.App content. #33272 carries a Dev Fleet backend hunk (2 of its 11 files) that reads the new pod stop wait; the core fix is the reviewed change, so its app hunk stays.
File line cap. This batch grows
src/kiro_crew/env.pyfrom 2,016 to 2,029 lines (#33562). Release is already over 2,000 lines onacp/kas_agents.py,dashboard/chat_api/resume.pyandenv.pywithout a baseline row, and no gate on a release PR runscheck_file_line_cap.py. Splitting a file here would be a refactor main never shipped in this form, so the batch ships as is.Dropped: #30872 (pluralize the Dev Fleet force-remove worktree count). It is an app PR (Dev Fleet strings only), so it is out of scope; it and its release-only revert stay on the staging branch, where they cancel, and neither goes to release.
Backwards compatibility
Compatible: each commit is the main commit unchanged, or, for the two release-only commits, only what one pick needs applied to release's own code. Each main PR states its own compatibility.
Removes nothing: no field, API, key or config is removed beyond what each main PR already shipped and documented.
Tests
Run on the pretest branch at
db6479eb12on a Linux DevSpace. Its tree,0af3c8270f, is this branch's tree byte for byte: the same 50 commits in the same order, with the same one hand resolution (the resolvedprompt_assembly.pyblob isd49a5a1e07on both).pytest -n 6 --dist loadgroup --max-worker-restart=2 --timeout=120 --no-cov, with an isolatedTMPDIR,--basetempandKIROCREW_HOME, on the 83 backend test files the picks touch plus all 22test/*ratchet*.pyfiles (104 files after removing duplicates, 5,200 items): 5,165 passed, 35 skipped, 0 failed.test_flake_pattern_ratchet.py,test_wait_discipline_ratchet.py), measured againstrelease/0.12.0at34898f9795, once by base SHA and once in PR mode (GITHUB_BASE_REF=release/0.12.0): 117 passed both times.npm ci, thenvitest runon the 7 website test files the picks touch, withI18N_BASE_REF=34898f9795: 7 files, 171 tests passed.release/0.12.0or comes from the host), and the website suite (vitest 65,459 passed, electron 3,265 passed).Proof: not provable -- other: a cherry-pick batch of already-merged main fixes, each proven in its own main PR
Manual verification
N/A: each fix was verified in its own main PR, and this batch carries them unchanged except for the two release-only commits above. The staging build that carries every batch was installed and run on Amazon Linux 2, Amazon Linux 2023 and macOS (install, gateway boot, dashboard, chat, MCP gateway, a scheduled run, restart, stop) with an isolated data home, and the collected bug reports were re-run against it with no regression.
Screenshots / video
Screenshots come from each pick's own main PR; the picked change is the same. Of the four picks that change a dashboard file, #31712, #32218 and #32552 carried no screenshot on main: #31712 changes only the noun and verb forms of one line, #32218's two
website/files change only comments, and #32552 changes only keyboard focus.#33334 (from its main PR): the artifact toolbar of an artifact in no folder. English is unchanged; German shows
Ordnerinstead offolder.Related Issues
0.12.1 hotfix batch b1 of 4 (slices hf1 to hf3): #30153, #30256, #30240, #30810, #30860, #31083, #31289, #31223, #30841, #31401, #31712, #31997, #31203, #29957, #32218, #32363, #32552, #32398, #32482, #32627, #32486, #32814, #32484, #32205, #32583, #32551, #32932, #33150, #32457, #33188, #33272, #33334, #33508, #33534, #33536, #33571, #33562, #33566, #33564, #33555, #33484, #33443, #33635, #33544, #32376, #33663, #33579, #33695.
On the release line this carries the fixes for #29345, #6521, #31707, #32193, #25979, #30934, #29779, #32714, #29535, #14724, #30357, #32782, #29713, #32903, #33270, #33307, #33473, #33291, #30716 and #32069, through the picks named in the table.
Sibling sweep
Siblings: none -- a backport batch; each pick's own PR carried its sweep
Pattern harvest
Not generalizable: a release backport batch, not a new defect fix.
Checklist
cherry-pick -xof a main commit with a Conventional Commits title, or a release-only commit listed aboveContribution License Agreement
Contributions are accepted under the project's Apache-2.0 license; see the Licensing section of CONTRIBUTING.md.