Skip to content

fix: v0.12.2 hotfix batch 1 (82 fixes from main) - #34375

Merged
bolichen97 merged 84 commits into
release/0.12.0from
hotfix/0.12.2-batch1
Nov 10, 2026
Merged

bolichen97 merged 84 commits into
release/0.12.0from
hotfix/0.12.2-batch1

Conversation

@bolichen97

Copy link
Copy Markdown
Contributor

Problem / Motivation

Goal: Ship the first batch of the 0.12.2 hotfix list to 0.12 users: 82 user-facing fixes from main, plus the partial prerequisite one of them needs, on release/0.12.0.

These fixes are on main and not on release/0.12.0 (v0.12.1). Each one fixes a bug in code that 0.12 users run today. 59 cherry-pick with no conflict; 23 needed a hand-port onto release's own code; one prerequisite (#33667) is carried in part for #33729 and #33809.

Why it matters

0.12 users keep hitting these bugs until a patch release carries the fixes. This batch covers security and Windows file access, chat and the dashboard, KAS and agent specs, sessions and restarts, channels, schedules, desktop and the sandbox.

Not a goal

  • No new behaviour on release: each commit is a git cherry-pick -x of the main merge commit; a hand-port keeps release's own code, applies only that fix's change, and says what it left out in a Backport note: line of its commit message.
  • No prerequisite feature picked to make a conflict go away (the one partial carry is listed below).
  • No app PRs, and no CI-only, docs-only or test-only commits, except the one release-only commit listed below that a pick needs to pass.
  • The fixes merged to main after this list was cut come in a later batch.

What changed (motivation → approach → change)

The 0.12.2 candidate list was reviewed PR by PR against release/0.12.0 (109 fix commits merged to main after the 0.12.1 cutoff): a bug in released code, not on release yet, and not an app PR. This batch applies the 82 included fixes in main merge order, one cherry-pick -x of each main merge commit, on release/0.12.0 at b9a5b555da. The whole batch was assembled and tested as one staging branch first (see Tests); this branch is that staging branch rebased onto the two release commits that landed since, and git range-diff shows every patch unchanged.

# PR Main merge commit Fix
1 #33858 3f9e35351a report a GitLab MR as conflicting behind another blocker
2 #33839 1d1a3633c5 read one pull request spelled in two cases as one subject
3 #33837 a4e191816b skip a non-string comment parent_id in the forward filter
4 #33283 52a2ebb3c2 four launch approval and settings card fixes
5 #33841 ae43f1bc3c write a slot savepoint when the unit grew during the pass
6 #33840 b4356a6fcd read past a transcript row with no role or content
7 #33860 9617c69161 keep a GitHub rate limit retryable when the REST status read fails
8 #33712 cd481171e7 report a remote update unverified when its old version is unknown
9 #33771 710d5e9458 refuse a stage or offer from a channel the user left
10 #33785 0c0dd8267f give the judge verdict a text reason
11 #31086 eef45f9ba5 reopen a Bot whose memory.db was deleted
12 #33892 26c8a9733a tighten a file-search scope check on windows
13 #33788 2b5cfe395e word every arm refusal and judge the click by its offer
14 #33895 21e79015c3 fold the work rebuild's longer unit list in its own cell
15 #31521 7c55e42b2c discard a resumed partial that the server refuses with 416
16 #31755 12ad3e80b7 parse a HEARTBEAT.md saved with a UTF-8 BOM like one without
17 #31780 c5758a3c9d fall back when og:title or og:description is blank
18 #33925 d0e911b617 skip the drive-root /tmp kiro-log probe on windows
19 #33476 3100756cac report a busy gateway when artifact or a spawn poll times out
20 #33951 9b409f3314 refuse from-chat plans and refine while task runner is disabled
21 #33946 3150511252 name a skill saved with a bom by its frontmatter
22 #33947 04655300a4 probe one Windows launcher once, whatever its spelling
23 #33856 0d44748a0e call a GitHub PR with over 100 passing checks review-ready
24 #33263 3b995c8e53 pluralize the crew board's failed attempts count
25 #33944 eacf826036 fail a cancelled step in a taskrunner parallel group
26 #33945 11da4e1069 skip a stored run id the sequence cannot count
27 #33942 521b278717 keep a user mcp server whose name has a colon in the default spec
28 #28136 2eda983ddf keep a held key result's goal link through both Undos
29 #30512 6b19bc920c count distinct Bedrock KB ids toward the kb_ids cap
30 #33652 5a22d7e1f3 bound gateway-launch.log reads and rotate it on fresh spawn
31 #30558 278b887dd4 mark unresolved fields as not current on a scripts-off page
32 #28391 883bb6071d name the text-field rule on the Back/Forward rows
33 #33961 970e91ba08 retry the install rename through a brief windows hold
34 #33960 0476068510 say an unrecorded write was undone, not stored
35 #31590 c3f87a3683 keep the streak when the local date steps back a day
36 #33959 586f239bc4 collapse a home path at any escaping depth in bundles
37 #28138 79bff1f741 count a settled strReplace's real change, not the whole file
38 #33333 39aecc7296 keep an svg artifact's stylesheet off the page around it
39 #33940 7ba1662453 skip a record nested past the stack limit as damage
40 #33760 558497d049 three gateway supervision and shell version fixes
41 #34074 8134b0f71e stop overtaken debounce timers in two file searches
42 #34000 df1eeb64bb ask the ACL question of each middle link hop's directory on Windows
43 #29906 398d49b74c render Slack-formatted send_message bodies legibly in the bell
44 #33877 2c6f61e882 stop a deeply nested crons.json value blocking every save
45 #33755 47d9e55d7b make list, status and destroy agree on a removed crew
46 #33759 4e9db0dfb0 let the desktop app confirm vscode file links with a drive letter
47 #29892 d2c64b2139 follow active tool discovery instructions
48 #34084 ad34e36825 let one waiter through per embed rate-limit token
49 #33857 044a1b7804 count only the newest Azure DevOps status per check
50 #33938 bf02448a81 read an ANSI .env in UTF-8 mode instead of failing boot
51 #33972 648c52db89 tolerate a non-string --skill-name-filter in agent discovery
52 #33590 a2a0a62f61 say when a company SSO sign-in has no Kiro profile
53 #34099 092be23315 keep line breaks in the dictation preview
54 #33948 16d40aafbe report a launch job file with a wrong-typed field as unreadable
55 #34085 cb51e30fd3 drop a session store scan that a move overlapped
56 #34153 0ed959d3c0 drop a quoted directive marker line from channel replies
57 #33968 db4ef699b0 judge catalog walk paths without the resolver pool
58 #34005 2c678c350f sweep staging trees an ended engine install left behind
59 #33767 08e371d261 label only new bare URLs in the chat nav panel
60 #33667 94dfacdb2d keep focus on the remaining issue radar buttons while busy
61 #30885 45f5c583a4 fail a step whose agent turn was cancelled mid-run
62 #33279 90be52d879 three handshake, provider and model registry fixes
63 #33830 b08eb54466 list a reset's unrecorded unit last on the ledger read
64 #33847 8cc73e991c keep Crew's tools for a claude session that asks for auto beside a foreign settings file
65 #30683 3987f3aa14 probe the spawned kiro-cli for v3 and report the kiro pick engine
66 #33729 c0a645527e keep knowledge source buttons focusable while they save
67 #20041 a76084d602 auth/credential failures do not count toward auto-pause
68 #33870 62428f3358 count a runtime's reparented session members against the RSS ceiling
69 #33809 bd88d41d71 keep Capabilities and Memory read buttons focusable while busy
70 #33900 c1c0912814 offer only resumable found stacks and never preselect one
71 #33687 92a2e9f9e3 flag chat trust lost on restart, recover failed chunks
72 #33927 877b4900af reap a session crumb whose start is past float range
73 #33871 e3db305f82 let the update check's fetch finish and reap its killed tmp packs
74 #34069 028e0e3383 tell a bot chat outside its own thread to do the work itself
75 #32550 6992769c8b keep the query on the /connections redirect
76 #33935 88b7d61276 end a recycled session's leftover processes and report them mid-turn
77 #34086 3605300026 remove the fetch scratch dir when a check is cancelled
78 #33227 1943bc883c say Windows policy blocked the agent binary instead of a generic error
79 #34071 77cb0eaa0e make boot.allow_terminal refuse the dashboard terminal
80 #34144 4bbbc50200 expose edition MCP servers' tools on the team lead
81 #33670 72f6877a59 voice reply, restart close and rss recycle fixes
82 #32179 9d68dc6e22 load workspace steering once, and name use_aws and inheritance-off
83 #33963 73e3275d98 keep a folder file the scan could not see

Release-only commit (needed for a pick's test to pass on release):

  • 0077a017d9 fix: build the GitLab fake fetch inline in the conflict test (backport assembly)

Hand-ports with a Backport note: (what each left out on release):

Backwards compatibility

No behaviour change beyond each listed fix. No config key, CLI flag or API is removed or renamed.

Tests

  • The staging branch (5d32d238a5, same patches) passed the whole backend suite as the 8 release candidate shards on a DevSpace with no failure caused by a pick: each remaining failure also fails on release/0.12.0 or comes from the host (no IPv6 loopback, low /tmp, sandboxed rg), plus one timing flake that passed 3 reruns. Website and Electron suites pass.
  • flake8, isort, mypy --platform linux, tsc -p tsconfig.app.json, gen:ui --check, black and comment-history: clean on the staging tip. The file line cap is advisory here: 4 of the 5 files over it were already over on release.
  • Every test file the picks change: 6313 passed; the one failure needs the voice extra and fails the same way on release.

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 hand-ports above. The staging build was installed and run on Amazon Linux 2, Amazon Linux 2023 and macOS (install, gateway boot, dashboard, a real chat turn, MCP gateway, a scheduled run, restart, stop) with an isolated data home, and 80 collected bug reports were re-run against it: 14 are now fixed, the 20 that 0.12.1 fixed stay fixed, and none regressed.

Screenshots / video

Screenshots come from each pick's own main PR; a hand-port changes how a fix sits on release's code, not what it draws.

Related Issues

0.12.2 hotfix, batch 1:
#33858, #33839, #33837, #33283, #33841, #33840, #33860, #33712, #33771, #33785, #31086, #33892, #33788, #33895, #31521, #31755, #31780, #33925, #33476, #33951, #33946, #33947, #33856, #33263, #33944, #33945, #33942, #28136, #30512, #33652, #30558, #28391, #33961, #33960, #31590, #33959, #28138, #33333, #33940, #33760, #34074, #34000, #29906, #33877, #33755, #33759, #29892, #34084, #33857, #33938, #33972, #33590, #34099, #33948, #34085, #34153, #33968, #34005, #33767, #33667, #30885, #33279, #33830, #33847, #30683, #33729, #20041, #33870, #33809, #33900, #33687, #33927, #33871, #34069, #32550, #33935, #34086, #33227, #34071, #34144, #33670, #32179, #33963.

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

  • Each commit is a cherry-pick -x of a main commit with a Conventional Commits title, or a release-only commit listed above
  • Existing tests pass and new tests added for new functionality
  • Self-review completed; code follows project style guidelines
  • Documentation updated (if applicable)
  • No secrets, credentials, or internal references in the diff

Contribution License Agreement

Contributions are accepted under the project's Apache-2.0 license; see the Licensing section of CONTRIBUTING.md.

RohanK6 and others added 30 commits November 9, 2026 22:49
#33839)

GitHub owner and repository names are case-insensitive, and the bare
number resolver already folds case. The URL and shorthand readers did
not, so an instruction naming #42 beside the URL
#42 read as two pull
requests, refused to gate, and left the watch on its plain timer.
Compare slugs case-folded in both readers. A different number still
refuses.

(cherry picked from commit 1d1a363)
…lter (#33837)

comments.json is decoded verbatim, so a hand edit can leave a comment
whose parent_id is a list or an object. When that comment's thread_id
names no comment, the forward filter falls back to walking parents and
tests the parent_id against a dict, which raised TypeError. The
comments route then failed with exclude_resolved=true for the whole
artifact.

A parent_id that is not a string now ends the walk, the same answer a
missing parent already gets, so the other threads still filter.

(cherry picked from commit a4e1918)
- An approved MCP launch is admitted before the per-server launch cap,
  so 64 other agents declaring the same server no longer refuse it.
- A long launch (one --include-tools per tool) is shown in full and can
  be approved; config set stub_servers names any unapproved server.
- Approve update pins the launch resolved now, so a launch that moved
  is approved without a gateway restart instead of answering 409. A
  press that pinned an older launch shows the newer one, and says so
  through ErrorNotice with its agent hand-off: on the page for the row,
  inside that server's block for the Review dialog.
- A setting.change card path accepts a Bot name with a hyphen or a
  capital, so it no longer fails with "invalid characters".

(cherry picked from commit 52a2ebb)
…ass (#33841)

_persist_slot_fold skipped the write whenever the newest unit grew during
the pass (memo.reached != sampled.last_seq). A live conductor appends about
once a second and a long fold takes minutes, so the biggest cells never got
a savepoint and every restart or eviction refolded them whole.

On a unit with a cut counter the count sampled before the pass and re-read
at the write brackets every record the pass read, so the savepoint is now
written at reached. A unit with no counter keeps the old skip (a digest
covers fewer records than the pass folded).

A cold fold also writes a savepoint every 4,096 entries on the way, beside
the unit it is reading, under the same cut-count evidence, so a restart in
the middle of a long fold resumes near where it stopped. The oversize-cell
skip is unchanged.

Part of #33827.

Co-authored-by: Raymond Chen <chenmingwei23@users.noreply.github.com>
(cherry picked from commit ae43f1b)
)

The session replay and the fallback history recall pick their rows
with `_quota_tail`, which indexed `m["role"]` and `m["content"]`. The
transcript reader returns every non-metadata JSON object as a row, so
one hand-edited or torn row without either key raised KeyError out of
every context build for that chat.

The row is now read with `.get`: a row with no role is skipped like
any non-turn row, and a row with no content replays as empty text.

(cherry picked from commit b4356a6)
…3771)

The desktop freshness gate refused a build from a channel the install
no longer follows only when the switch landed during its own wait: every
check re-points the feed's channel first, and nothing recorded where the
staged bytes or the offer came from. A switch made before the click,
followed by a feed that fails or hangs, let fail-open install or
download the old channel's build.

The offer and the stage now carry the channel whose feed served them,
and one guard compares that, and the channel the judging check was sent
to, with the selected channel. A build the selected channel also serves
is taken over by that channel's check and still installs. The refusal
is its own verdict, so the quit path names the channel change instead
of a withdrawal.

A Retry after an installer failure re-arms the before-quit listener, so
a quit pressed during its wait is held as it is on the first click. The
Electron suites tick FRESHNESS_CHECK_TIMEOUT_MS instead of a copy of it,
and one test pins its documented 8s.

(cherry picked from commit 710d5e9)
The eval CLI records a judge verdict, then redacts its reason before
printing it. The redactors take text only, so a reply such as
{"score": 5, "reason": null} raised there, and the CLI's error path
then marked the turn failed. A passing verdict was lost because of
its reason field.

judge_turn now always returns the reason as text: an empty string for
a missing or null reason, str() of anything else.

(cherry picked from commit 0c0dd82)
A user who deleted a Bot's memory.db to wipe its memory could never
open that Bot's chat again. The Bots page said "Bot 'x' changed during
private memory assignment", which was not true: the store check's real
reason (database missing) was swallowed and reported as a change.

When the store is the same at request entry and under the lock:

- if its database file is gone, the config binds the store to that Bot
  alone by name and immutable id, and any leftover WAL/SHM/journal file
  is a regular file: close this process's cached handles, take the
  store's exclusive use lock, rename those side files aside (never
  delete them), create a new empty member database with the same
  identity and owner-only permissions, log it once, and open the chat.
  A store another process still holds open is refused instead;
- any other failure (identity mismatch, a link or non-file at the
  database or a side-file path) keeps refusing, and the refusal names
  the store check's reason instead of claiming a change.

The new-chat precheck admits that same repairable case, so starting a
chat with the Bot heals it too. A store that really changed during the
wait still gets "changed". The security and memory specs record this
one exception and its guards.

(cherry picked from commit eef45f9)
On Windows an operator's file-search scope rule missed a search root
written with a namespace prefix, a spelling Windows opens as the same
folder, so a search the operator meant to refuse ran. The root is now
folded the way the sensitive-path check already folds it, so the rule
also covers those spellings. Other platforms are unchanged.

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit 26c8a97)
…33788)

A managed-venv arm re-checks the release feed first, so its no-verdict
refusal became the normal answer whenever that check could not reach
the feed: the panel showed the gateway's English "run a check first",
in every language, right after a check had run.

A failed re-check now refuses with its own code, arm_check_failed, and
the panel words it and every other arm refusal code from the catalog,
keeping the offer and its button up unless the refusal says the offer
is gone; the failed re-check's line names that button as the retry.
The notice keeps the request's report, so asking the agent
still hands over the status, code and body the catalog line no longer
shows. The click sends the version its button offered, which the
gateway uses to tell an upgrade click from a channel-move click; an
upgrade the cache held when the request arrived still refuses a move,
and callers that send none keep the cached judgement. zh-CN's
withdrawn-offer line uses the release-channel term the rest of the
panel uses.

(cherry picked from commit 2b5cfe3)
…ll (#33895)

The work-ledger rebuild reads the board with also_slots, so its unit
list is the board's own units plus its workers'. Every other reader of
the board's work cell lists the own units only, and the warm cell is
keyed by slot and fold, not by list. A rebuild therefore left the cell
holding a list the next ordinary read could not continue, so that read
resumed from disk, and it wrote its savepoint beside the list's newest
unit, a worker's unit, over the worker's own work.json.

read_slot_state now folds a list that the supplement lengthened as a
plain fold, with no warm memo and no savepoint, and answers
SlotFoldNotReady for it on the request path. Same value: a warm cell
answers what a cold fold answers.

Co-authored-by: Raymond Chen <chenmingwei23@users.noreply.github.com>
(cherry picked from commit 21e7901)
…16 (#31521)

download_to kept the staging file on every transport failure when resume
was on. A payload with no declared size, such as a feature-video poster,
can leave a partial that already holds every byte. The next resume then
asks for a range at the end of the file, a CDN answers 416, and the same
range is sent on every later attempt, so the poster never installs.

A 416 answer to a ranged request now discards the partial, so the next
attempt starts from zero and installs the verified file.

Closes #31520

(cherry picked from commit 7c55e42)
…without (#31755)

A HEARTBEAT.md saved as "UTF-8 with BOM" was decoded as plain UTF-8, so
the first line kept U+FEFF in front of it. A first-line heading then no
longer started with "#" and ran as a task, after which the rewrite
deleted it. A first-line task kept its "- [ ] " marker in the task text.

The task read now decodes as utf-8-sig. The rewrite strips the mark
before parsing and writes it back in front, so the file's other bytes
stay as they were.

Closes #31740

(cherry picked from commit 12ad3e8)
…31780)

extract_meta chose og:title over <title>, and og:description over the
meta description, by testing the raw attribute value. A whitespace-only
value such as content=" " is truthy, so it won the fallback test and
was then stripped to an empty string. The preview lost a title and
description the page did serve, and title_complete went False.

The og values are now stripped before the fallback test, so a blank og
value counts as absent, the same as a missing one.

Closes #31763

(cherry picked from commit c5758a3)
…33925)

On Windows, Path("/tmp") is drive-relative and names \tmp at the root
of the current drive, a folder kiro-cli does not use there and any
local account can create. The support bundle and the kiro_cli_logs
tool picked kiro-cli logs from it whenever %TEMP% held none, so a
file someone else put there shipped as the user's own kiro-chat.log.

Add the fixed /tmp candidate only off Windows. Windows keeps its
TEMP/TMP candidates, which is where kiro-cli writes its logs there.

Closes #33294

(cherry picked from commit d0e911b)
…ut (#33476)

`kirocrew artifact` turned every failure into its raw exception text, so a
gateway that was up but slow to answer read as a bare "Error: timed out",
unlike `status`, `token` and `spawn list`, which say the gateway is busy.
It now classifies the failure through loopback_http.deadline_expired: a
read says busy and asks for a retry, and a save, update or delete says
busy and points at `kirocrew artifact list`, since the write may land
after the client gives up. Any other failure keeps its own text.

`spawn run`'s result poll had the same shape and said "lost connection to
gateway" on a timeout; it now says the gateway is busy and points at
`kirocrew spawn list`.

Closes #29744

(cherry picked from commit 3100756)
…33951)

Switching Task Runner off in Library refused start, plan, execute and
retry, but POST /api/taskrunner/from-chat still created a planned run
and POST /api/taskrunner/refine still started a model session, with no
page to watch or stop either from. Both routes now answer the same 409
app_disabled refusal before anything is created.

Closes #32883

(cherry picked from commit 9b409f3)
_skill_name decoded the SKILL.md head as plain UTF-8, so a file saved
as UTF-8 with BOM kept U+FEFF before its fence. The frontmatter was
not found and the skill was named after its folder, so the KAS skill
rule left out the name KAS actually uses.

Decode the head as utf-8-sig, as every other SKILL.md reader does. A
file without a BOM decodes exactly as before.

Fixes #33911

(cherry picked from commit 3150511)
)

MCP discovery probes names that start the same process one after
another, keyed by command plus args. On Windows the key compared the raw
command string, so npx and npx.cmd, a case variant, or a full path to
the same launcher got separate keys and separate locks, and both names
were probed at once: the overlap the key exists to prevent.

Key the command on Windows by its folded basename with the PATHEXT
suffix removed, the same folding the module already applies when it
compares resolved commands. POSIX keeps the command as written.

Fixes #33915

(cherry picked from commit 0465530)
A Crew board row with failures showed the label and the number as two
pieces glued in JS: "failed attempts 1" in English, with the number
stuck after the words in every language and no singular form. Russian
read "неудачных попыток 1" for any count.

Pass count for the plural form and n as the locale-formatted number,
give the key the forms each language needs, and register it as plural.

(cherry picked from commit 3b995c8)
_execute_tasks runs a parallel group with gather(return_exceptions=True)
and decided each step with `isinstance(result, Exception) or not result`.
gather captures BaseException, so a cancelled step arrives as a
CancelledError, which is not an Exception and is truthy -- the step was
read as a success and the run continued past a step that never finished.
The sequential branch has no such hole: it awaits the step directly.

A step passed only when it returned True; anything else (a BaseException
including CancelledError, or a falsy bool) is a failed step and routes to
_try_replan. Adds a regression test that fails on the current code.

Fixes #33924.

(cherry picked from commit eacf826)
On start the workflow service reads the highest wf_NNNNNN among the
runs it reloads, so new ids continue past it. It accepted any id whose
tail passed str.isdigit, so a stored wf_ followed by a superscript two
raised from int() and the service failed to start, and an id past the
allocator's 64-bit range became a floor the allocator refuses, so
every new run failed. Only an ASCII-digit tail within that range now
counts; the allocator never issues any other, so nothing can collide.

(cherry picked from commit 11da4e1)
…8136)

Removing a key result and a goal it served, then undoing both, brought the key result back without that goal. _strip_goal and restore_goal walked only the live key results, never the one held in removed_key_result, so the goal's removal did not record the held key result and its Undo could not relink it. Both now include the held key result, in either removal order.

Closes #28130

(cherry picked from commit 2eda983)
…30512)

_parse_kb_ids compared the number of raw comma pieces with the cap of 10
before it dropped blank pieces and repeated ids. Ten distinct ids with a
trailing comma, or with one id repeated, were refused as "more than 10
knowledge bases". The lenient search-time parse cut the raw pieces first,
so a real id after a blank or a repeat could be dropped.

The cap now applies to the normalized, deduplicated list. Strict parsing
refuses only when an eleventh distinct id appears, and the lenient parse
stops once it holds ten ids. The docstring and the knowledge spec say so.

Closes #30509

(cherry picked from commit 6b19bc9)
…pawn (#33652)

The two gateway failure dialogs read the whole gateway-launch.log into one
string to show its last 60 lines, and the log only ever appends with no cap.
On a long-lived install the file reaches hundreds of MB, so the synchronous
read in the main process stalls for seconds and spikes RSS the moment the
gateway fails, and past V8's string limit the read throws and the dialog shows
an empty log pane.

Add readLogTail, which reads only the file's last 256 KiB, and route both
reader sites through it. Add rotateLaunchLog, a size-gated one-generation
rename in the shape native-logging.js already uses for chromium.log. Rotate in
the supervisor's fresh-spawn path, just before the new gateway opens its log
fd, so a shell that reuses a surviving gateway never rotates the log that
gateway is still writing to (off Windows through an inherited descriptor, on
Windows through the path).

Closes #32719

Co-authored-by: kirocrew-agent <agent@kirocrew.local>
(cherry picked from commit 5a22d7e)
CrysisDeu and others added 17 commits November 9, 2026 22:49
…e RSS ceiling (#33870)

A job a timed-out shell command leaves running is reparented to init and
dropped out of the watchdog's parent walk, though it keeps the runtime's
session id. The RSS ceiling's /proc map now also lists each process under
its session leader. Measurement only: kill paths keep the plain map.

Refs #33843

Co-authored-by: Zezhen Xu <zezhexu@amazon.com>
(cherry picked from commit 62428f3)
…busy (#33809)

* fix(a11y): keep Capabilities and Memory read buttons focusable while busy

The default-template Retry, Knowledge graph Retry, webhook roster Retry,
episodic Search/Clear/tag-filter/Preview and the Recall and preview-page
Retry buttons used a native disabled while their read ran, which takes the
focused button out of the tab order and drops a keyboard user to the page
body until the read lands. Swap native disabled for aria-disabled + aria-busy
with a click/handler guard, keeping the same dimmed look. The retries that
render only on failure gate on queryReadFailed so they stay mounted through
their own retry instead of unmounting while error is briefly cleared.

Closes #33087

* fix(a11y): reuse Btn/SendBtn busy; keep Retry mounted through its retry

Reuse the Btn busy prop (aria-disabled+aria-busy, dimming, click guard, stays
in the tab order) on the Retry/Clear/Recall buttons instead of writing those
out by hand, and give SendBtn the same busy prop for the episodic Search and
inspector Preview. Also from review:
- KnowledgeGraph computed the failure state after the isLoading early return,
  so a retry with nothing cached hit the loading branch and unmounted the
  focused Retry -- compute it first and gate the loading return on
  isLoading && !failed.
- MemoryRecordsEditor's preview-page Retry was gated on previewPage.error,
  which clears on retry, so retain a failed flag across the pending phase.
- Show a loading line (retrying / loading_graph) in the webhook roster and
  knowledge-graph rows while the error text is briefly cleared, so the Retry
  is not left context-less.

(cherry picked from commit bd88d41)
…#33900)

With no usable saved launch, `cloud launch` listed every stack named
`kirocrew-*` and preselected the first. In a shared account, Enter, a
non-TTY run or `--yes` resumed a stack whose instance was gone, one
another run was still creating, or one Kiro Crew never made.

A found stack is now offered only when it is neither failed nor in
progress and a live managed instance carries its tag, read with one
filtered describe-instances call. Every other stack is named with the
`cloud status --tag` that checks it. "Create a new installation" is
the default, and `--yes` keeps no found stack, since nothing yet tells
this machine's stack with a lost record from another person's.

Part of #32071

Backport note: index-only conflict in docs/decisions/README.md:
release does not carry the 2026-11-04 Ultra Mode pause entry or the
2026-11-06 work-ledger entry main has above the insertion point, and
keeps its conductors row elsewhere, so only this PR's row is added,
after release's 2026-11-03 KAS sign-in row. The decision entry, both
specs, the ec2.py and wizard.py change and the tests apply as on main.

(cherry picked from commit c1c0912)
#33687)

Two dashboard chat failures that the owner never sees.

- Trust lost on restart: a chat's Trust and Read modes live in memory
  only, so any gateway restart (Restart, an applied update, kirocrew
  restart, a crash relaunch) dropped them silently and an armed loop
  stalled on approvals. The flush and shutdown save now record which
  open chats hold a mode; the next boot posts one bell note and one
  system row for each chat that came back without it. The record is
  display-only and never restores a mode. App-owned chats, chats a
  live auto-approve grant reaches, and crewmate chats whose Profile
  seeds the mode back are not told.
- Failed lazy chunk: a chunk that failed to load in the chat pane (the
  unrestored-tabs notice, the crew window, the composer's goal and
  monitor control or its environment chip) blanked the pane or the
  whole chat page, and Try Again replayed the cached rejection. A
  shared lazyRetry helper now re-imports on each Try Again, offers
  Reload page, and keeps each failure in that part's own place, so the
  sessions sidebar and the rest of the chat stay usable. Ask the agent
  on a failed crew window closes the window, so the chat it opens is
  not left under the window's cover.

Backport note: lazy-chunk recovery half only. The trust-lost-on-restart
half is dropped: chat_trust_restart.py imports
kiro_crew.approval_mode_policy, which is not on release, so its
save_transaction.py, dashboard_persistence.py and session_restore.py
hooks, test_chat_trust_restart_notice.py, the composition-contract count
bump and the trust paragraphs in session.md and app-notifications.md
stay out. The 11 locale conflicts sit next to main-only cache_hit_rate
keys and take only automation_load_failed and environment_load_failed.
learn-cron-dashboard.md takes the lazyRetry routing text from main onto
the release paragraph. The website code and tests apply as on main.

(cherry picked from commit 92a2e9f)
…33927)

The crashed-session backfill reads each open-session crumb with
`_read_crumb`, which returned `float(started_at)` for any number. An
int past float range raised OverflowError, which ended the whole scan
at that crumb. The crumb was never removed, so every later boot stopped
at it again and the crumbs after it were never reaped or reported.

A start time that is not a finite number now marks the crumb malformed,
the same as one that is not JSON, so it is reaped and the scan goes on.

Backport note: release's _read_crumb returns a 4-tuple with no harness
(main's harness field is unreleased), so the finite-start check returns
release's 4-tuple malformed shape. The test applies as on main.

(cherry picked from commit 877b490)
… tmp packs (#33871)

The source checkout update check fetched under a 30s total deadline. A
killed fetch keeps none of its download, so a checkout far enough behind
that its fetch needs more than 30s failed the same way on every check and
never recovered. Each killed try also left a tmp_pack_* in the checkout's
object store, and no reaper ran on this path.

Raise the check's deadline to 300s, the largest bound git_tmp_packs
vouches for, and reap stale tmp_pack_* after every fetch through the
shared isolated-fetch helper, timed-out ones included.

Fixes #33828

Co-authored-by: Zezhen Xu <zezhexu@amazon.com>
(cherry picked from commit e3db305)
…4069)

A chat that runs a bot on its conductor outside the bot's own thread
(the agent picker, a chat folder's default agent, an agent channel) has
no worker and no session tools, yet its Delegation text sent it to a
worker hand-off and to blocks only the bot's own thread gets. A session
without a member key gets a Delegation text that says to do the work in
the chat; the bot's own thread keeps its text.

The user docs and the crew-mode spec say which chat hands work to the
worker, and a decision entry records the maintainer decision.

Closes #28546

Backport note: index-only conflict in docs/decisions/README.md:
release's index ends at its 2026-11-08 bot-chat skills row, without
the main-only rows above this PR's, so only this PR's row is
appended there. The decision entry, both docs, the agent_spec.py and
sections.py change and the test apply as on main.

(cherry picked from commit 028e0e3)
/connections?view=mcp-servers landed on Services: the route sent every
old link to a fixed /capabilities?tab=mcp and dropped its query. A
ConnectionsRedirect beside the other query-keeping redirects appends the
incoming query after tab=mcp, so the old link opens MCP Servers.

Closes #30392

(cherry picked from commit 6992769)
…them mid-turn (#33935)

A runtime's leftover processes (its session's members outside its live
tree, such as a job a timed-out shell command left running, set -m jobs
included) are ended after an RSS recycle once the runtime is gone. During
a turn they are logged with pid, size and name and the chat gets one
notice; ending them then is opt-in behind
session.watchdog_end_leftovers_mid_turn (default off). macOS counts the
same session members through KERN_PROC_ALL filtered by getsid.

Fixes #33843

Co-authored-by: Zezhen Xu <zezhexu@amazon.com>

Backport note: release lacks platform_compat.terminate_pids_pinned,
which the new session_pid_rss.py leftover kill calls; main added it in
#31975 (266e126). This folds in only that helper from it: the
function in platform_compat_kill.py with its IS_LINUX and
get_process_start_id imports, its TYPE_CHECKING forwarded name in
platform_compat.py, its _MOVED entry in
test_platform_compat_refactor_facade.py and its platform-compat.md row.
None of the #31975 tool-stall changes (acp/liveness*,
session_handle_watchdog and the rest) come along. #33870, picked before
this, is the other prereq. Everything else applies as on main.

(cherry picked from commit 88b7d61)
…34086)

The update fetch builds a throwaway git dir under the crew scratch root
in an off-loop step. A cancel that landed while that step ran, such as
a gateway shutdown during the 12-hourly check, dropped the result while
the thread went on to build the dir, and an error after the dir was
made also left it. The dir has no scratch owner record, and the sweep
never deletes such a dir, so each one stayed for good. The prepare step
now removes its own dir on any error, and on a cancel the caller and
the worker settle under one lock which of them removes it, so the
cancel never waits on the worker.

(cherry picked from commit 3605300)
…neric error (#33227)

* fix(kas): say Windows policy blocked the agent binary instead of a generic error

A spawn still refused with winerror 5 or 1260 (AppLocker, EDR or group policy) after main's one retry raises a non-transient AcpSpawnBlocked, and the chat card names the binary in the reader's language. Only the winerror code decides; no message text is read. Other spawn errors are unchanged.

Part of #33225

* ci: retrigger after GitHub Actions outage

---------

Co-authored-by: Joe Guo <zejiangg@amazon.com>

Backport note: the _terminal_error_meta docstring in recovery.py takes
the windows_spawn_blocked sentence on the release text, without the
AmbiguousAgentSpecError paragraph from main (unreleased). The pinned
token counts move by one from the release values:
test_transient_notice_parity.py goes to 79 backend and 79 frontend
tokens (legacy stays 80) and 92 copy keys, and gatewayNotice.test.tsx to
79. The spawn_blocked code, the locale keys and the other tests apply as
on main.

(cherry picked from commit 1943bc8)
…al (#34071)

A security policy that states boot.allow_terminal as anything but a JSON
true now refuses the dashboard terminal with the same answer
dashboard.terminal.enabled: false gives. A policy that leaves the key out
has no opinion, and translated enterprise policies no longer state it.

Fixes #30367

Backport note: release has no enterprise_policy_fetch
TRANSLATION_REVISION, so the t3 bump is dropped and an older build's
cached translation is handled by the content trim alone
(_without_legacy_terminal_refusal, whose docstring no longer names the
revision); test_terminal_policy_gate.py drops the one revision test, and
the offline-cache test in test_enterprise_policy_fetch.py rewrites the
cache under release's etag. enterprise_policy_translate.BOOT drops
allow_terminal and keeps release's sandbox wording. service_boot.py
registers only the terminal teardown hook, without main's neighbouring
credential-redaction hook (#30976, not on release).

(cherry picked from commit 77cb0ea)
Ultra Mode switches a chat to kirocrew-team-lead, whose spec is written
from build_agent_config. That adds each edition server to mcpServers but
not to tools, and kiro-cli exposes only servers named in tools, so the
lead mounted the edition servers and could call none of their tools.

The installer now runs _register_edition_refs, as the crewmate worker
base does. Refs go in tools only, never allowedTools, so every call
still prompts. A standalone build has no edition servers, so its lead
spec is unchanged.

Backport note: one doc conflict in crew-mode.md: the new team-lead row
lands after release's crew_mode_agent row, and the neighbouring rows keep
release's wording (Crew Mode, default kirocrew-conductor, mate thread),
without main's later renames. The code applies as on main.

(cherry picked from commit 4bbbc50)
- Slack voice replies are spoken on the default transport path. With
  messaging.use_transport on (the default), handle_message_transport
  now speaks a finished answer through _reply_by_voice, the voice-memo
  signal survives the busy-session queue, and the owner's !voice on
  opts a thread in instead of answering that replies stay text.
- Every in-process restart (Restart to Update, the dashboard restart
  and the Slack /kirocrew restart) closes sessions through
  close_sessions_for_restart, which leaves a close still running after
  20 s (5 s for Slack) and goes on to the exec. A background task that
  outlived its cancel used to hold the restart forever on the old
  version; _restart_gateway also logs each step it starts.
- The RSS watchdog holds a session after its third rebound in a row
  instead of recycling it after every turn, logs it once and tells the
  chat which setting (session.watchdog_rss_max_mb) is below the
  session's footprint. The hold lifts a day after the last recycle,
  and a held tree that doubles its footprint is still recycled.

Backport note: slack/events.py takes the close_sessions_for_restart
TYPE_CHECKING import without the code_channels slash import next to it
on main (#21429, unreleased). The commands.py row of
test_slack_post_message_call_sites.py goes from 78 to 77, since release
already counted one call fewer than main. The session.md and
voice-streaming.md conflicts are context only: the close_all row and the
Slack reply row take the change from main, while the newer
drain_active_turns wording and the WhatsApp reply row beside them on
main stay out. Everything else applies as on main.

(cherry picked from commit 72f6877)
…nce-off (#32179)

On KAS, Crew's relative .kiro/steering resource entry is dropped when KAS's
native load covers the whole tree: KAS loads it itself and honours manual and
fileMatch, so they no longer load every turn or twice. The entry is kept as
before when the tree may hold an inclusion: auto file, a link, or anything
unreadable; auto handling is #32153. A session with Kiro's default resources
turned off is told AGENTS.md and steering still load. An agent that lists
use_aws is told AWS calls run in the shell, and
toolsSettings.use_aws.deniedServices travels as shell denies that leave the
shell grant answered.

Fixes #32138

Co-authored-by: Joe Guo <zejiangg@amazon.com>

Backport note: two context-only conflicts. kas_agents.py: release has
no kas_skill_scope import and no batch_guard step after
authored_policy, so only the with_use_aws_denies import and call are
added, in the same place relative to release's lines.
agent-spec-fields.md: main's permissions, toolsSettings and hooks rows
carry main-only text, so release's rows are kept and only this PR's
two toolsSettings edits (the use_aws.deniedServices sentences and
"the rest of `use_aws`") are applied to release's row. Release reads
toolsSettings as written (no canonical_tools_settings), so the
use_aws_shell_denies docstring says only the `use_aws` spelling is
read. kas_steering.py, harness/kas.py, steering-and-hooks.md and the
tests apply as on main.

(cherry picked from commit 9d68dc6)
os.walk skips a directory it cannot list, and the folder walk skipped
a file it could not stat. Either can be transient: a permission
change, or a network mount that drops out mid-walk. The scan then
read every tracked file in that gap as deleted and removed its
indexed items, which only a fresh, paid re-ingest brings back.

Record the directories the walk could not list and the files it
could not stat, and keep their rows and items. A file that is
really gone is still removed.

A file whose stat answers not-found (or not-a-directory) is gone, so it is still removed:
it was deleted after the listing, or it is a symlink whose target is
gone. An AsciiDoc includer whose target the walk could not see keeps
its indexed text too.

Backport note: conflicts only where main's hunks sit next to the
ignore-pattern keep of #31410 (unreleased). folder_watcher.py adds the
unlisted and unlisted_real sets to release's step 1 walk call (which
has no tracked or pattern_ignored arguments) and the not-unseen check to
release's removed list; _walk gains only the two new keywords.
knowledge.md adds the two new sentences to release's bullets, without
main's ignore_patterns paragraph. The test applies as on main.

(cherry picked from commit 73e3275)
…t assembly)

The #33858 backport added a test that calls _gitlab_fetch, a helper
that only main has (it came with #22488, a feature not on release), so
the test raised NameError and flake8 failed with F821. Build the fake
fetch the way every other GitLab test on release does.
@bolichen97
bolichen97 requested review from a team and KiroCrewCaptain as code owners November 10, 2026 06:51
@bolichen97
bolichen97 requested review from Zedmor and removed request for a team November 10, 2026 06:51
@bolichen97
bolichen97 merged commit 7f0d8cf into release/0.12.0 Nov 10, 2026
13 of 22 checks passed
@bolichen97
bolichen97 deleted the hotfix/0.12.2-batch1 branch November 10, 2026 06:51
@github-actions github-actions Bot added the readiness: checking Automated validation is still running label Nov 10, 2026
@github-actions

Copy link
Copy Markdown
Contributor

Sensitive Change Review

Sensitive files changed:

  • src/kiro_crew/hook_runtime/search_targets.py

  • No reasoning template for 0077a01 yet.

How to clear this check:

  1. The PR author posts the template below as a PR comment, with
    every field filled (N/A needs a reason). Keep the
    Head commit line: it must name the current head SHA.
  2. The check turns green within a minute, with no push.
  3. A new push resets this check: post it again for the new head.

Only the author's comment counts this way. A reviewer can instead
paste the filled template into an "Approve" review on this head.
Merging still needs a reviewer's approval.

## Sensitive change review
- Head commit: 0077a017d9a830eb093a5a6305d4c790c29e78a7
- What rule changed:
- Before -> after (what was allowed/blocked, now):
- Worst case if wrong (breaks commands / leaks secret / locks user out):
- Evidence checked (which sandbox result proves it):
- No regression:
  - Tools still work (which tools/commands were run, result):
  - Backward compatible (old config / old data / old callers still work, how checked):
  - Existing tests (which suites passed on this head):
- How to undo:

@github-actions github-actions Bot removed the readiness: checking Automated validation is still running label Nov 10, 2026
@github-actions

Copy link
Copy Markdown
Contributor

First Principles Review (Opus 5.5) — ✅ PASS

Premise-level review of 0077a017d9a830eb093a5a6305d4c790c29e78a7 — why this exists and whether the shipped surface is the smallest honest version. Updated in place on each push. A BLOCK verdict blocks PR readiness; PASS does not.

First-Principles-Verdict: PASS

This is a backport of fixes that were already reviewed on main. No item adds a capability or a default on its own, and the one release-only commit only fixes a test so a pick passes.

What this change ships

Inventory (5 items) — 5 justified

Intent: Give 0.12 users the bug fixes main already has, without waiting for 0.13. This is a FIX.

  1. 59 fixes cherry-picked unchanged from main, each a bug in released 0.12 code — justified
  2. 23 hand-ported fixes, keeping release's code and noting what was left out in a Backport note: line — justified
  3. Part of fix(a11y): keep focus on the remaining issue radar buttons while busy #33667 carried as the prerequisite for fix(a11y): keep knowledge source buttons focusable while they save #33729 and fix(a11y): keep Capabilities and Memory read buttons focusable while busy #33809, as the description declares — justified
  4. The GitLab conflict test builds its fake fetch inline, because _gitlab_fetch exists only on main (the test failed with NameError and flake8 F821) — justified
  5. Two UX evidence screenshots under temp-screenshots/p110/, following the existing temp-screenshots/ convention — justified

[FIRST-PRINCIPLES-REVIEWED] 0077a01

Description read

[DESCRIPTION-READ] f4b3a1b947aa27fefe59d9e935f65b0694bad634b60df298d23bdba8bdf839c3

That sha256 names the PR description this verdict read. Recompute it with .github/scripts/pr-description-capture.sh; if it differs, the description changed after this verdict was formed and any finding drawn from the description is unproven.

@github-actions

Copy link
Copy Markdown
Contributor

Security Scope Review -- 🔴 could not complete (fails closed)

No scope verdict for 0077a017d9a830eb093a5a6305d4c790c29e78a7 (the candidate stage failed). This lane fails CLOSED: an unmeasured tightening must not read as "nothing newly refused". See the job logs, then re-run -- or record /ai-review override scope 0077a017d9a830eb093a5a6305d4c790c29e78a7 <class>: <reason> if a human has judged the scope by hand.

@github-actions

Copy link
Copy Markdown
Contributor

GPT 6.1 Review — ✅ no blocking findings (all downgraded on adjudication)

GPT 6.1 flagged blocking issues on 0077a017d9a830eb093a5a6305d4c790c29e78a7; Opus 5.5 adjudication downgraded every one of them. Downgraded: the remedy each one needs is disproportionate to the harm it prevents. Adjudication downgraded all 1 blocking finding(s): the remedy each one requires is disproportionate to the harm it prevents.

This comment is updated in place on each push.

Review details

BLOCKING -- website/src/hooks/useChatNavigation.ts:95 -- An empty batch prevents later URLs from receiving labels

const misses = linksToResolve.filter(l => !labelStore.has(labelKey(l))).slice(0, NAV_LINK_BATCH_MAX)
Trigger: Open a transcript with 20 bare URLs when the background resolver times out, then append a separate message containing another URL. The endpoint returns HTTP 200 with 20 empty summaries; useChatNavigation stores none, so the first 20 misses and query key remain unchanged.
Outcome: React Query reuses the infinitely fresh empty response. The newly added URL is never submitted for labeling, even after the resolver recovers.
Anchor: defect
Fix: Track empty responses as settled fallbacks so batching advances past attempted links; handle retries independently of batch advancement.

[BLOCK-MERGE-DOWNGRADED] 0077a01
[GPT-REVIEWED] 0077a01

Adjudication (Opus 5.5) — is blocking on each finding proportionate?

F1 — LOW (disproportionate-remedy). Confirmed conditions:

  • The handler turns a resolver failure into HTTP 200 with empty strings padded to the batch length (src/kiro_crew/dashboard/*chat_nav*.py:270-288).

  • The empty strings are never stored (useChatNavigation.ts:105-109), so the same first 20 misses come back (:95), the key stays the same (:96), and with staleTime: Infinity (:113) the empty reply is reused.

  • This only sticks when there are at least 20 bare-URL misses. With fewer, a newly added URL lengthens the miss list, which changes the key and resends the batch.

  • What the user sees: link labels stay missing, a cosmetic gap that does not lose data. Labels stay off for this page until a reload. A reload clears both the in-memory query cache and the tab-local store (:53-59).

  • Recovery: none until a reload, so this is not fully self-correcting. That is still low harm, because the only consequence is a missing label.

  • Cost of the real fix: a second "attempted but empty" state, plus a retry/backoff policy that has to be designed separately from how batches advance. That is new state and new branches only to recover labels after an outage, which costs more than this rare cosmetic gap.

    [ADJUDICATION] 0077a01 total=1 uphold=0 downgrade=1
    DOWNGRADE F1 website/src/hooks/useChatNavigation.ts:95 reason=disproportionate-remedy
    [GPT-ADJUDICATED] 0077a01

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 0077a017d9a830eb093a5a6305d4c790c29e78a7 false-positive: <reason, 20 to 500 characters>

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.