Skip to content

fix: v0.12.2 hotfix batch 2 (65 fixes from main plus release tooling) - #34586

Merged
bolichen97 merged 68 commits into
release/0.12.0from
hotfix/0.12.2-batch2
Nov 10, 2026
Merged

bolichen97 merged 68 commits into
release/0.12.0from
hotfix/0.12.2-batch2

Conversation

@bolichen97

Copy link
Copy Markdown
Contributor

Problem / Motivation

Goal: Ship the second batch of the 0.12.2 hotfix list to 0.12 users: 65 user-facing fixes from main plus 2 release-tooling commits, on release/0.12.0.

These fixes are on main and not on release/0.12.0 (v0.12.1 plus batch 1, #34375). Each one fixes a bug in code that 0.12 users run today. 43 cherry-pick with no conflict; 24 needed a hand-port onto release's own code. Two are release tooling: #34362 gives the universal macOS desktop build 90 minutes, and #30916 keeps the backend trees out of the universal merge, which is what made that build slow on this branch.

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.
  • Fixes merged to main after this list was cut come in a later batch.

What changed (motivation → approach → change)

The fixes merged to main after the second 0.12.2 review (84 fix commits) were reviewed PR by PR against release/0.12.0 with the same rule: a bug in released code, not on release yet, and not an app PR. This batch applies the 67 picks in main merge order, one cherry-pick -x of each main merge commit, on release/0.12.0 at 7f0d8cf88e. The whole batch was assembled and tested as one staging branch first (see Tests); this branch is that tested branch.

# PR Main merge commit Fix
1 #30916 4dc3ac42bf keep backend-dist out of the universal macOS merge
2 #34001 0465e3ab66 name the real exit status when a start dies at stdout EOF
3 #34230 12a449fe3a open no new HTTP session for a send after close
4 #33977 bb5a4c04bd stop an old approval answer from approving a later tool
5 #34199 ae6cbd69ae refuse a pod name with a trailing newline
6 #34198 4b91dc7240 refuse unresolvable workspace dirs with a 400, not a 500
7 #34281 d53a6746ab keep approval cards answerable when a stop fails
8 #34275 8cd53380b7 keep focus on an armed delete button while it runs
9 #33733 0741d899f5 name the cause when an SSM crew's dashboard is silent
10 #34282 2d36a98336 keep a streaming turn running when its send fails late
11 #34283 f764987074 report an unfinished kiro sign-in and retry it in one click
12 #34280 0075cd3fb5 relocate every frozen pip launcher in windows scripts\
13 #34285 5355e0cce9 read model_managed strictly where the answer feeds a write
14 #34023 3467dd6ca5 let the cursor menu pick an item past nine by number
15 #33975 44d165faef import a .kirobot whose JSON part starts with a byte-order mark
16 #34306 dfa624b834 name a launch's denied action and point to iam-policy
17 #34309 9e39adffed keep a later turn's renderer when an earlier turn ends
18 #34307 ba7a0c8a3d never suggest a bot's home folder for a new chat
19 #34255 2648f907c9 keep the steering entry when front matter may spell auto with an escape
20 #34203 46d6f36aab strand no shared claim when its rollback is cancelled
21 #30750 648aa71a5f count the chat folders on the discovery pool so the sidebar list does not wait on process walks
22 #33967 432a94f644 write a session savepoint when the unit grew during the pass, and seal it on close
23 #30939 b8f144f363 carry the gateway instance env on Crew server stubs
24 #34288 44c02c5aed tell the conductor which of its worker's mcp servers failed
25 #34059 9dd2d99350 point spawn_sub_agents results at the configured home
26 #23442 96fda7e7ed bound the workspace cli.json read and keep it off the event loop (#23396)
27 #34326 3154bfd0f1 re-measure a ref's point after scroll-into-view settles
28 #34248 3533f0b0a2 size the account Refresh button to its own label
29 #34366 d1ed22de5b mask credentials in long capability messages
30 #34365 223d3a7aa1 answer no item id for a create the undo removed
31 #34229 0583ffa3bd identify fresh when the gateway resume host stays unreachable
32 #34321 87100f2005 keep the review draft when Slack refuses its approve post
33 #33212 14d19477cb send agentSpawn hook context again after a compaction
34 #34076 0835976cb0 let an oauth_endpoints.json entry clear a token-safe client_id
35 #34080 062d92c6af report a refused systemd stop instead of success
36 #34356 e9fb171420 keep workspace settings writes out of the global kiro settings
37 #34134 f84483f1b1 drop provenance records from a copied gateway entry
38 #34381 a0f9261a08 settle every polly reply, however long, with voice_complete
39 #34331 7c9d44a0c6 refuse a windows junction wherever a directory symlink is refused
40 #34286 b799fc2688 kill a starting repair or npm child's group when cancelled
41 #34383 7ce599b241 report a crew's origin refusal, not a bad credential
42 #34357 7a6d75e5ab tell crystallize the candidate layout; stop blaming scripts
43 #34362 c0fe7e75a6 give the universal macOS desktop build 90 minutes
44 #34167 113992594f keep terminal and voice sockets through a loop stall
45 #34173 d503d56422 keep stateless cron runs out of the session search window
46 #34090 8f6fe6dbfe withhold a model pin the session's model select does not list (#29723)
47 #33875 4697bd820d join an in-flight resume prefetch, not a second agent
48 #33810 81ecb8b6c8 forward remote crew pane ports, write config owner-only
49 #18517 9c5e565308 run a requeued steer on its own row instead of a duplicate
50 #33802 24041871e2 move kiro-cli's broken KAS engine copy aside once instead of crash-looping
51 #34002 5fa893fd24 keep a slack approval click on the chat that asked
52 #33867 60e867f396 four bot editor, agent picker and roster fixes
53 #34020 67f5836d83 tighten a channel member permission check
54 #34224 1b1f265ca6 count every tip click and name the tip in tips.feedback
55 #34225 eb4a973033 answer a worker steer send before its ledger record is written
56 #34216 9f1c79dc97 refuse approval_mode=auto from app tokens on cron create and update
57 #34141 9442f2ac1e run the local token owner check on its own pool
58 #34308 f9788c5388 report a systemctl or journalctl timeout instead of crashing
59 #34322 650a18f2dd keep a later turn's renderer when an earlier turn ends
60 #34344 a1e64da623 land a sent kiro-cli context only on a kiro-cli slot
61 #34337 99133a6f25 let a stable release reach the insider index and image
62 #34034 56ed09453b let Outlook's own webLink through the link guard
63 #34377 4b56b32871 remove a stopped data home's skill views when its gateway stops
64 #34370 fbbb07dcd3 refuse saved task-plan workflows while task runner is disabled
65 #34340 715d4aa307 explain the windows cloud launcher block once
66 #34320 31924d946f keep a closed tab closed when its late save lands
67 #34298 9d1ce77a1b keep the steer acknowledgement out of webhook and task results

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

  • 21fa270a10 fix: drop main-only _sends_prompt in turn gates (backport assembly)

The clean pick #33212 read a local that exists only on main; as picked, every chat turn after a session's first raised NameError. This one-line fix keeps release's own condition.

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 (this branch's tip, 21fa270a10) passed the whole backend suite as the 8 release candidate shards on a DevSpace: 252,175 passed, 30 failed, none caused by a pick (each also fails on the batch-1 base and on release/0.12.0: host artifacts, the line-cap advisory and one app test). Website vitest 65,732 passed, Electron 3,335 passed.
  • 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: the files over it were already over 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: 13 of 13 checks on each. The collected bug reports were re-run against it: 21 more are now fixed, the ones 0.12.1 and batch 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 2:
#30916, #34001, #34230, #33977, #34199, #34198, #34281, #34275, #33733, #34282, #34283, #34280, #34285, #34023, #33975, #34306, #34309, #34307, #34255, #34203, #30750, #33967, #30939, #34288, #34059, #23442, #34326, #34248, #34366, #34365, #34229, #34321, #33212, #34076, #34080, #34356, #34134, #34381, #34331, #34286, #34383, #34357, #34362, #34167, #34173, #34090, #33875, #33810, #18517, #33802, #34002, #33867, #34020, #34224, #34225, #34216, #34141, #34308, #34322, #34344, #34337, #34034, #34377, #34370, #34340, #34320, #34298.

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.

bolichen97 and others added 30 commits November 10, 2026 02:35
Co-authored-by: w <w@k.local>
(cherry picked from commit 4dc3ac4)
…34001)

The reader loop's EOF branch read returncode before the child was
reaped, so the reason every pending request failed with said
"process exited (rc=None)" while the reap a moment later measured the
real code. The chat's stop row printed that text as its last error.

Wait for the exit first, under the same _KILL_REAP_TIMEOUT the
registry retirement right after it uses. A status still unread when
the wait ends reads "<not reaped>", the label the death log already
uses, never a bare None.

Closes #32740

(cherry picked from commit 0465e3a)
In a Webex conversation, an approval answer that a failed turn never
used could approve a tool in a later turn of the same conversation,
with no prompt shown for it. A turn's unused answers are now discarded
when it ends, so a later request always asks again. A test pins it and
the messaging spec states the rule.

(cherry picked from commit bb5a4c0)
The pod name pattern ended in `$`, which also matches just before a
trailing newline, so validate_name accepted "name\n" and the peer-port
scan counted a "name\n.env" file as a pod. Such a name can never be a
systemd instance, and pod prune's own guard exists to skip names like
it rather than shell a bogus systemctl stop that fails every future
prune. Anchoring with `\Z` closes it for both readers of the pattern.

(cherry picked from commit ae6cbd6)
* fix: refuse unresolvable workspace dirs with a 400, not a 500

The dashboard workspace create and update handlers let Path.resolve()
and expanduser() exceptions escape on a dir string the host cannot
resolve (an embedded NUL, a ~user naming no account). The requested
dir now answers 400 invalid_path with no write.

A stored entry that cannot resolve was also resolved eagerly before it
reached _ws_dir_in_use, outside that helper's skip, so it blocked every
create or update of an unrelated workspace. The handlers now skip such
entries, the way _ws_dir_in_use documents.

The resolves these edits touch now run through asyncio.to_thread, so
the sync-io-in-async fs baseline for workspaces.py drops from 12 to 5.

Fixes #32845

* test: pin the gone-account case on every host

(cherry picked from commit 4b91dc7)
The composer's Stop marked every pending approval card rejected the
moment it was pressed, before /stop answered. When the stop then
failed (an HTTP error, or a 2xx reply with ok:false), the turn kept
running and the server kept waiting on the card, but the card had no
buttons left. Coordinator (registry) cards were swept as well.

The sweep now runs inside requestStop, only after a stop the press
sent has landed, only while that chat is on screen and no newer turn
has started on it, and it leaves coordinator cards to the registry,
as the new-turn sweep already does.

Closes #33468

(cherry picked from commit d53a674)
The four arm-then-confirm buttons (a hook's Delete, a job's Delete,
a job's Revoke all secrets and the leaked-runtime Reclaim) took a
native disabled and the label "..." while their request ran. A
disabled button cannot hold focus, so the keyboard press that
confirmed the delete dropped focus to the page body, and a failed
delete left the reader nowhere; the dots were untranslated and read
as punctuation.

While pending each button is now aria-disabled and aria-busy, keeps
its translated label as its name, and shows a spinner over it, the
same shape as a job's Run and Cancel. useArmedDelete already ignores
a press on a pending row, so no second request goes out.

Closes #33233

(cherry picked from commit 8cd5338)
…#33733)

Diagnose on an SSM crew answered "is the remote gateway running?" for
every failure past the first rung, so a user whose AWS profile was
refused, whose Remote user did not exist, whose gateway listened on
another port or who had not installed Kiro Crew checked the one thing
that was fine.

The second rung's one send-command now also curls each port a gateway
run-marker names and checks for the launcher the token mint would run.
Diagnose reports ssm_access_denied, ssm_user_unknown, remote_wrong_port
or remote_not_installed, each naming the setting to change (Remote
user, Remote port, Remote kirocrew path, or the three SSM permissions)
with no ARN or account id. Anything it cannot tell apart stays
remote_down. No new AWS call or permission.

Part of #32117

(cherry picked from commit 0741d89)
…4282)

A send whose POST fails with a connection reset or a refusal after the
server already started the turn ended the local turn anyway, so the
Thinking indicator vanished under a reply that kept streaming and the
composer offered a fresh send instead of steer or stop. A send queued
behind a running turn did the same when it failed. The web chat
composer and the Request a Feature send now end the local turn only
while that send still owns its optimistic running mark (the check the
send-race and composer-locked arms already made) and, in the composer,
only while no turn is streaming on the slot.

Closes #33469

(cherry picked from commit 2d36a98)
)

A closed or blocked sign-in tab left the browser-wait screen on a
spinner whose only change after 30 s was a hint blaming the network,
and both "Start over" and "Use a different sign-in" went back to the
provider list, so a retry took two clicks.

After 30 s the wait screen now says the sign-in has not finished and
offers Start over. Start over, there and on the failure screen,
cancels the attempt, waits for the gateway to confirm nothing landed,
and begins again with the same provider on the same transport
(Company SSO keeps its start URL and region). "Use a different
sign-in" stays the way back to the provider list, and the code screen
says when its code expires, from the begin answer's expires_at.

Closes #32062

(cherry picked from commit f764987)
#34280)

pip writes each Scripts\<name>.exe with the build machine's absolute
python.exe in its #! line. Only kirocrew.exe was replaced, so the other
console scripts (httpx, pwcpp, tqdm, f2py, ...) exit 1 with no output on
a user's host, and env.py puts Scripts\ on every agent's PATH.

Both Windows bundle writers now run packaging/relocate_windows_scripts.py
after the kirocrew shim. For each launcher whose #! names an absolute
path it writes the zip's __main__.py as <name>-script.py and a
relocatable <name>.cmd that runs it under %~dp0..\python.exe with the
same environment scrub, then deletes the .exe. Native uv.exe, uvw.exe
and uvx.exe have no launcher zip and stay as they are; a frozen launcher
with no payload fails the build.

Fixes #32692

(cherry picked from commit 0075cd3)
get_model_managed folds an unreadable sidecar into None, the same value
as "nothing recorded". Three callers still used that lenient answer to
decide a write, so one transient read error (a Windows sharing
violation from a second process) silently changed model tracking:

- lift_and_strip_bookkeeping lifted a stale spec model_managed over
  the owner's recorded False, and the refresh then re-synced the model
  to the shipped default. It now uses one strict, locked set-if-unset
  (claim_model_managed_if_unset takes the value), and the cc_model lift
  reads strict too. An unreadable sidecar raises, as a lasting hold
  already did in set_model_managed.
- _install_dashboard_author_agent skipped carrying the owner's pin. It
  now reads strict and keeps the spec's model on an unreadable sidecar.
- api_agent_fork and _record_publish_model left the copy with no flag
  and answered 200. They now read strict, so the existing
  bookkeeping_failed path unwinds the copy.

Fixes #32673

(cherry picked from commit 5355e0c)
)

The cloud wizard's cursor menu acted on the first digit pressed, so
typing 13 in a 13-item list picked item 1 while its hint said
"press 1-13". The menu now keeps the digits typed so far: a number that
one more digit could still extend moves the cursor and waits for that
digit or Enter, and any other digit picks at once, so a menu of nine
items or fewer behaves as before.

Part of #32071

(cherry picked from commit 3467dd6)
…33975)

A .kirobot whose kirobot.json, agent.json or other JSON part was saved
by a Windows editor or PowerShell 5.1 starts with the UTF-8 byte-order
mark. The reader decoded it strictly and handed json.loads text whose
first character is U+FEFF, which it refuses, so the import was refused
as "not UTF-8 JSON" although the file is UTF-8.

Parse each JSON part through loads_user_json, the repository's reader
for hand-edited JSON, which drops one leading mark and is otherwise
json.loads. Non-UTF-8 bytes are still refused by the strict decode, and
the file's sha256 is still taken over the raw bytes. The RFC's file
rule names the dropped mark.

Closes #32708

(cherry picked from commit 44d165f)
…4306)

A failed cloud launch now puts "grant `<action>` (see `kirocrew cloud
iam-policy`) and retry" ahead of the raw AWS error when a permission
was denied, whether the denial is in the AWS CLI's stderr or only in a
failed stack event such as iam:PassRole. The WaitCondition's reason is
the instance's setup log and is not read for it. With the action
known, `kirocrew cloud launch` prints its grant line instead of the
--keep-on-failure tip, and the dashboard, which keeps a launch error's
first 400 characters, shows the remedy before the role and resource
ARNs.

On a build with host policies, a denied iam:AttachRolePolicy leads
with the second policy from `kirocrew cloud iam-policy --host-policies`
instead, since the first policy already grants the one attach it
allows.

The CLI's stderr is no longer cut at 200 characters, which split the
stack name in the describe-stack-events command it prints. The shared
aws.checked error, the stack lookup and the permissions-boundary read
lead with their remedy the same way.

Closes #32227

(cherry picked from commit dfa624b)
…#34309)

A Telegram turn publishes its renderer under the session key before it
acquires the session, so a second message admitted in that window
publishes its own renderer under the same key. When the first turn
ended, its finally popped the key unconditionally and took the second
turn's renderer away, so that turn's mid-turn steers lost their chip.

Remove the entry only when it is this turn's own renderer, as the
Discord dispatcher already does.

(cherry picked from commit 9e39adf)
)

The folder suggester picks from every folder that is not hidden, so a
Bot's home folder, and the folders inside it, could be offered. Only
that Bot files a chat there, so the card's Move is refused with
mate_home_locked; the dashboard rolls the move back and clears the card
with no message, and the session stays unfiled.

The suggester now leaves out every folder in a Bot's home, using the
same subtree set the folder writers use. When those are the only
folders, no suggestion is made and no card is shown.

(cherry picked from commit ba7a0c8)
…th an escape (#34255)

Co-authored-by: Joe Guo <zejiangg@amazon.com>
(cherry picked from commit 2648f90)
…ed (#34203)

A cancelled warm_mint_all rolls its claims back through
_release_shared_claims, which disposed each row and only then removed
it from the mint table. A second cancel landing during that rollback,
such as a dropped premint request followed by gateway shutdown, cut
the loop at a dispose: that row and every later one stayed minting
with no caller holding their tokens. Nothing withdraws such a row, and
it keeps the shared mint process from ever retiring. The release now
takes every matching row out of the table with no await, then shields
the disposal of their holdings.

(cherry picked from commit 46d6f36)
…idebar list does not wait on process walks (#30750)

(cherry picked from commit 648aa71)
…e pass, and seal it on close (#33967)

A session fold sampled the unit's seq and segment fingerprint before its
pass, and checkpoint.save refused the write when the files moved at all,
even by an append. A pass over a unit that kept growing therefore wrote
no savepoint, and the next restart folded the whole unit again.

On a unit with a cut counter the witness now names what the pass folded
and the fingerprint the write compares against is the file as the pass
left it; held_still still re-reads the cut count after the pass, so a cut
during it retries cold. A unit with no counter keeps the old rule (a
digest covers only what it hashed).

On session/closed the eager worker now seals the warm bundle before it
forgets it (seal_session_folds): one write at the closer whatever the
256-entry threshold says, under the session's fold lock, with a cut
witness re-read at the write. A unit with no counter is left as it is.

Part of #33827.
Refs #33845.

Co-authored-by: Raymond Chen <chenmingwei23@users.noreply.github.com>
(cherry picked from commit 432a94f)
codex starts every stdio MCP server from an allowlist plus the element's own
env. A pooled broker stub for a Crew server thus carried only its session
token. When the broker was unreachable, the stub's fallback started the real
server with no KIROCREW_BOUND_PORT and no home override, so the server found
its gateway by fallback and could authenticate against another instance.

control_plane_instance_env() in providers/mirrors/identity.py returns the
instance half of the identity env: the managed home override and the bound
port, with no session key or token. control_plane_identity_env() now builds
on it. codex_projection() merges it into each stub whose name is one of
Crew's servers. A session-bound Crew stub that is_session_bound_crew_stub()
admits stays mounted exactly as given.

(cherry picked from commit b8f144f)
)

A Bot's worker runs its own MCP servers in a hidden session. When one
failed to start or its sign-in expired, the conductor's hand-off note
still listed it among the worker's servers and the worker report said
nothing, so the Bot asked the user to add a server the worker already
had instead of saying what to fix.

Both now read the worker slot's stored MCP session report: the
hand-off note leaves an unusable server off the worker's list and
names it with how it failed, and each worker report carries the same
line. Names only; the failure text never reaches the model.

Part of #32778

(cherry picked from commit 44c02c5)
…34059)

With KIROCREW_HOME set to a symlink, the "read the full result at ..."
pointer that spawn_sub_agents gives for a long child result named the
link's target. The pointer was built inside the managed MCP process,
which runs on the already-resolved data home, so the configured
spelling never reached it.

GET /api/spawn/{id} now carries a finished run's result_path, spelled by
the gateway (agent_dir_for_display), and spawn_sub_agents uses that path
instead of deriving one. The MCP spawn facade stops serving
agent_dir_for_display, which nothing on that side may use any more.

Closes #31466

(cherry picked from commit 9dd2d99)
…#34326)

A click by ref measured the target in the same call that scrolled it into
view, before the page's scroll handlers ran. A page that collapses a header
on scroll moved the target after that measure and the click landed below it.

Scroll once, then re-measure without scrolling after a short settle until
two reads agree. A re-scroll would re-centre the element and hide the shift.

(cherry picked from commit 3154bfd)
)

The Kiro Account modal's Refresh button reserved room for "Refreshing…"
even while idle, so "Refresh" sat against the left edge of the pill with
a 27px gap after it. Both labels shared one grid cell, which froze the
width so the button could not resize mid-animation -- the resize that
left a stale copy of the button painted over it on mobile WebKit.

The residue came from animating that resize, not from the resize: the
class list carries transition-colors and no longer transition-all. So
the button holds one label, swapped with the state, and is the width of
the word it currently shows: 85px idle with no slack, 112px while it
reads "Refreshing…". The accessible name is that same visible text, so
no aria-label is needed to say which state the button is in.

(cherry picked from commit 3533f0b)
A long message from an edition capability manager could reach the
dashboard with part of a credential in clear text. The capability
replies that mask their message now mask such a credential in full,
whatever the message's length, and keep the same 500-character limit.

Closes #33086

(cherry picked from commit d1ed22d)
…4365)

When the crew log does not confirm a work_ledger_record create, the
route undoes the write, removes the item it created, and answers 503
crew_log_unrecorded. The answer still carried that item's id, so it
named an item that no longer exists, and the packaged doc told readers
not to repeat a create whose answer carries an id, while the message
says to retry it.

The refusal now answers item_id null when the undo removed the created
item. A write about an existing item keeps its id, and so does a create
whose undo did not complete, since that item may still be on the board.
The work-ledger refusal row tells the caller from the message whether
repeating the write is safe, and the RFC sentence says the same.

Part of #33878

(cherry picked from commit 223d3a7)
chenmingwei23 and others added 19 commits November 10, 2026 11:48
…f crash-looping (#33802)

A home directory copied to another machine can leave kiro-cli's unpacked KAS
engine (<data dir>/kas/<version>-<sha>/) without its node_modules. kiro-cli only
checks that the directory exists, so every `kiro-cli acp --agent-engine v3` exits
with "failed to canonicalize KAS server path: .../acp-server.js" and never
re-extracts. Kiro Crew showed a generic process death and the warm pool retried
every 30 s forever.

The spawn entry now reads that line from a failed start's stderr. When the path
sits in kiro-cli's own data dir, <version> is a real directory and the server
entry does not resolve, it moves <version> to <data dir>/<version>.kas-broken-<UTC>
(outside kas/, which kiro-cli sweeps) and starts once more. Otherwise, or when the
second start fails the same way, the provider raises AcpKasExtractBroken, one
non-retryable message naming the directory and the mv command, and the warm pool
stops respawning until the directory is moved. agent.kas_repair_kiro_cli_extract
(default true) turns the move off. A start that succeeds runs none of this.

kas_extract_repair.repair is recorded in test/link_screen_sites.py as a
site that refuses every link: a symlinked or junctioned kas/<version> is
never moved, and os.rename on a real directory follows no link.

Closes #33762

Co-authored-by: Raymond Chen <chenmingwei23@users.noreply.github.com>

Backport note: release has four fail-closed agent switches, not main's
five: agent.conductor_own_child_verbs is unreleased. loader_pipeline.py
adds only kas_repair_kiro_cli_extract to release's list, and its comment
names it after kas_persist_mcp_sign_in and counts five;
test_config_load_pipeline_field_defaults.py adds only that key's row.
acp_runtime_start.py keeps release's kas_cli_settings import line (main
also imports kiro_cli_default_model there for #34155, unreleased) and
adds the AcpKasExtractBroken import. Every other file applies as on
main.

(cherry picked from commit 2404187)
A press on a linked chat's Slack approval card answers that chat's own
pending request and nothing else, so a second chat waiting at the same
moment keeps its request open until its own card is answered.

Closes #33464

Backport note: test_slack_linked_approval.py conflicts on
test_trust_under_a_policy_deny_approves_the_one_tool, which #30692
(unreleased) added and release does not have, so that test's hunk is
dropped. learn-cron-dashboard.md takes main's sentence edit onto
release's chat.py paragraph. Production files and the other tests apply
as on main.

(cherry picked from commit 5fa893f)
Bot editor Abilities pane: GET .../capabilities and GET .../worker-kit
now give up on a stalled file read after 20 s with 503
capabilities_unavailable, and the dashboard bounds both reads at 30 s,
so a hung mount shows the pane's load error instead of Loading forever.

Bot editor resources: the Abilities list now shows every resource the
worker loads, file:// documents an add copied in included, under a new
Resources group with Remove for added rows. The Custom agents detail
PATCH refuses mcpServers and resources (400 definition_key_read_only)
instead of answering 200 and dropping them.

Chat agent picker: a crewmate on the shared memory store is withheld
only when it is the template under another name. One with its own
workspace, pins, description, rules, goal, briefing or harness choice
stays pickable; the catalog sends an own_setup flag for what the row
does not carry.

Members roster: the GET /api/members read retries through a short
network drop and re-reads every 30 s while failed, so a one or two
second drop no longer leaves "Couldn't load your starred Bots" on the
dock, and a shown notice clears once a read lands. A permission
refusal (a guest session's 403 owner_only) is never re-read, for this
read and the other two that re-read while failed. The render stub
answers GET /api/members with an empty roster, so the render bench
measures a working gateway rather than a failing roster read.

Backport note: conflicts were textual only. agent_catalog.py keeps both
config imports. useAgents.ts takes main's header comment (release's
body already has no default-agent exemption). en-XA.json and
en.manual.json add band_resources beside release's band_remove_stale.
The spec, feature-map and configuration.md rows take main's added text
on release's wording (release's mate naming and feature_previews row).
The two new load-error tests match release's "Bot's" casing of
crewEditorNext.load_error.

(cherry picked from commit 60e867f)
On a backend where Kiro Crew runs an agent's own hooks itself, the
agent's guard now gets its say on a channel member's calls before any
approval, as it already does on the other gated turns.

Part of #32812

Backport note: channel.py's import block on release lacks main's
unreleased channel_account_check and channel_shell_base imports, so
only the channel_spec_hooks import is added. agent-spec-fields.md's
hooks row on release lacks main's Dismiss-link sentences; only the
channel-agent sentence is reworded as on main.

(cherry picked from commit 67f5836)
…34224)

* fix(tips): count every tip click and name the tip in tips.feedback

* fix(tips): read curated ids through the tips loader and retry a failed ack

shipped_curated_tip_ids now builds its set from tips._load_curated_tips,
so the packaged curated file has one reader. A failed ack request clears
the once-per-tip marker, so the next click sends it again.

---------

Backport note: telemetry-otlp-export.md and learn-cron-dashboard.md on
release lack main's Feature/Activity/Entity telemetry rework, so on
release the tips.feedback row's feature column is reworded to read a
shipped curated tip id or doc, and the Tips paragraph gains main's
Learn-more ack sentence. test_recommendation_metrics.py uses the
release curated tip zero-token-cron (same cron-and-scheduling.md doc
link) in place of schedule-repeated-work, which only main ships
(#30192).

Co-authored-by: Raymond Chen <chenmingwei23@users.noreply.github.com>
(cherry picked from commit 1b1f265)
…tten (#34225)

* fix(chat): answer a worker steer send before its ledger record is written

POST /api/chat on a Crew Mode worker and the side panel's steer route awaited the person-steer ledger record, which takes the board lock and waits for the crew-log writer's commit. A busy writer held the send past the client's timeout, so the composer showed 'Delivery not confirmed' for a message that was delivered, and a resend duplicated it.

Both routes now answer ledger_pending=true and record the steer in the background through one helper. Records for one worker take a per-worker lock in send order; the queue drain and a read steer use the same helper. A failed record is logged and sent to the owner's tabs as a crew_steer_unrecorded frame, which sets the same note a receipt's ledger_error set.

* Keep a steer's unrecorded note when its failure frame beats the receipt

Both steer callers now clear the old note on submit instead of on a
ledger_pending receipt, so a background record failure that reaches the
tab before the HTTP receipt is no longer erased. The stuck-writer test
waits with finished() instead of a short wait_for.

---------

Backport note: release has no dashboard/chat_api/crew_steer.py (main
split it out of chat_send.py in #33545), and adding the record helpers
to release's chat_send.py would take it past the 2,000-line cap. So
CREW_STEER_UNRECORDED_FRAME, _STEER_RECORD_LOCKS,
record_crew_steer_later, _release_steer_record and _record_crew_steer
go, as on main, into a new chat_api/crew_steer.py owner holding only
them; chat_handlers.py composes it and re-exports those five names,
and the composition contract test lists it in _OWNER_MODULES (release
has no _SYNC_ONLY_OWNERS). crew_steer_consumed stays in chat_send.py
and now calls record_crew_steer_later, a facade global, so
test_crew_mode_steer_lineage.py's _held fixture calls it through
chat_handlers, as main's does. learn-cron-dashboard.md names the new
owner.

Co-authored-by: Raymond Chen <chenmingwei23@users.noreply.github.com>
(cherry picked from commit eb4a973)
…nd update (#34216)

Only the owner may set a cron job to auto-approve its tool calls. POST /api/crons and PATCH /api/crons/{id} now return 403 (code app_auto_approval_refused) when an app token asks for approval_mode=auto; the refusal is audited under the app's name. An app PATCH that changes what a job does (message, agent, model, schedule, channel) also turns the owner's auto off in the same write, as the cron_update tool does. Owner writes are unchanged.

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

Backport note: api_cron_update takes only the _refuse_app_auto_approval
check; the new_chat_per_run block beside it on main is not on release.

(cherry picked from commit 9f1c79d)
Backport note: the relay viewer half is dropped, since
dashboard/handlers/relay_viewer.py is the post-0.12.0 relay viewer
feature and is absent on release; with it go its test in
test_local_auth_executor.py and the relay viewer route in the new
executors.py docstring bullet. executors.py keeps release module
docstring (rewritten on main) and adds only the local_auth_executor
bullet; the function, pool global, worker cap and shutdown hook apply as
on main.

(cherry picked from commit 9442f2a)
#34308)

systemctl() let subprocess.TimeoutExpired escape, but the pod verbs'
dispatch turns only PodError into its one-line refusal, so a user
manager that did not answer within the call's timeout ended `pod ls`,
`status` or `up` in a traceback. It now raises PodError (never
PodBackendAbsent, since no answer is no proof that no pod runs).

recent_journal() had the same escape, and `pod up` reads it on a failed
boot before stopping the unit it started, so a stuck journalctl also
skipped that teardown and left a crash-looping unit running. It now
reports the unreadable journal in the tail it returns.

(cherry picked from commit f9788c5)
…34322)

A Teams turn publishes its renderer under the session key before the
shared pipeline acquires the session, so a second message admitted in
that window publishes its own renderer under the same key. When the
first turn ended with no chips pending, its finally popped the key
unconditionally and took the second turn's renderer away, so that
turn's cards and chips resolved against nothing.

Remove the entry only when it is this turn's own renderer, as the
Discord and Telegram dispatchers do.

Backport note: keeps release's longer comment above the check and adds
main's identity-check lines under it.

(cherry picked from commit 650a18f)
A kiro-cli chat sent to a peer whose new slot starts on KAS (the
default) or another backend reported a full resume: the importer
installed the kiro-cli pair and joined it with provider acp on any
slot, and that slot cleared the join at its first turn
(detect_provider_switch), reopening from the transcript with the pair
left orphaned.

The import now asks the member-aware backend gate the provider factory
uses whether the slot starts on the backend the context came from (KAS
for a layer_b_kas session, kiro-cli for a layer_b pair). Anywhere else
it installs nothing, answers prefix and marks the tab transcript only,
as the KAS direction already did.

Closes #32884

Backport note: session_transfer_layers.py keeps release
select_provider_backend call, which has no session_backend_pick override
(post-0.12.0); only its result is renamed to starts_on, so
_slot_starts_on returns starts_on == backend as on main.

(cherry picked from commit a1e64da)
After a stable release the insider pip index and the ghcr :insider tag
kept serving the last release candidate, older than stable: the index
step wrote only the publishing channel's index, and the docker lane
moved only the publishing channel's alias.

A stable run now also lists its wheel in feed/insider/simple/. That
needs no feed guard, because an index only adds a version and pip
installs the newest one listed. The docker lane moves :insider to a
stable release's digest as its last alias write, behind the insider
feed guard run against the repo's tags, so an insider line already
ahead keeps its alias.

Fixes #32646

Backport note: publish-cli.yml on release has no lts PROJECT variable,
so publish_index is ported with release's literal kirocrew project
name. Release has none of main's stable-to-insider feed mirroring
(#32710 and the desktop lanes), so docs/build/release.md gains a short
bullet for the insider index and image only (main's edits the longer
bullet that describes that mirroring), the publish-cli.yml table row
gains only the insider-index sentence, and the publish-docker.yml
comment and the test module docstring no longer point at the
"Mirror stable feed to insider" steps.

(cherry picked from commit 99133a6)
Backport note: _PACKAGE_LINE_BUDGET is release's 35,048 raised by the same
67 lines (35,115; the package measures 35,114 on this branch).

(cherry picked from commit 56ed094)
…34377)

A data home's native skill views live in the shared ~/.kiro/agents and
only that home removes them. A shared runtime's kill never released the
skill projections it held, and the gateway leaves through os._exit, so
the views a stopping gateway ran on waited for that home's next boot
drain; a home stopped and then deleted never has one.

A completed, authorized runtime kill now releases its spawn and adopted
projections, as the direct client's teardown already does, and the
gateway teardown awaits one bounded prune of this home's unused views
once every session has ended. Ownership stays per home: no home removes
another home's views.

Closes #34149

Backport note: one docstring conflict in
acp/skill_projection_parts/alias_reclaim.py: release module docstring
still names the legacy side-file migration in the boot drain (removed on
main by #32139), so it is kept and only this change adds the stop prune
to its list. The code applies as on main.

(cherry picked from commit 4b56b32)
…34370)

With Task Runner switched off in Library, its own start, plan, execute,
retry, from-chat and refine routes already answer 409 app_disabled, but a
saved task-plan workflow still reached Task Runner through the Workflows
run route and /workflow, and was planned and executed there. Such a run
has no page to retry a failed step or resume a pause on.

The task-plan branch of the saved-workflow start now reads the same
Library switch before anything is planned and returns the same refusal;
the run route maps it to 409 app_disabled. Saved Python workflows do not
read the switch. A decision entry records the maintainer ruling.

Closes #33879

Backport note: the spec row and the handler test land beside release's
rows and tests; main's DELETE route row and definition update test
next to them are not on release.

(cherry picked from commit fbbb07d)
* fix(settings): explain the windows cloud launcher block once

On a Windows gateway the dashboard's cloud launcher answers 400
posix_host_required before any AWS call, but the Set-up tab still drew
the AWS checklist: the raw POSIX error, a credentials row saying they
were not checked, Ask the agent, and a disabled Launch, with no guide.

The tab now shows one notice instead. It says the launcher needs a
Linux or macOS gateway and the profile was not checked, and names the
two ways forward: the CLI launch wizard in PowerShell, or a gateway
inside WSL, which needs admin rights and a restart. A gateway known to
be Windows skips the preflight entirely. The Windows guide gains the
WSL steps the notice links to.

Fixes #32150

* fix(settings): route the refused preflight through the error notice

When the preflight refusal, not the platform signal, is how the Set-up
tab learns the gateway is Windows, the notice's sentence now renders
through ErrorNotice with the rejection's report and agent hand-off.

The desktop installer adds nothing to PATH, so the CLI sentence and the
Windows guide now point desktop installs at the bundled shim's full
path, and the guide says installing the AWS CLI usually needs admin
rights. The notice now says the AWS profile is not the problem and that
the Windows install keeps working. The fr, ru, hi and bn strings use the
informal register their style guides require. The feature map lists the
notice and the settings coverage gate waives it, as it holds no setting.

(cherry picked from commit 715d4aa)
A turn that outlives the close's cancel wait still saves its slot when
it ends. A full save rebuilds the close fields from the slot, which
reads open, so that late write erased the close the slot's own archive
had just committed and the dismissed tab came back after a restart.

The archive save now marks its slot when its write commits, under the
transcript lock. A later full save of a marked slot whose close fence
still stands writes back the close the line carries, closed_id
included, and keeps a linked or channel transcript's mtime as the
close does. Every other save is unchanged: a reopen is a new slot, a
close that puts its slot back releases the fence, and the hand-over
drain runs only when the archive did not commit.

Closes #8361

Backport note: slot_lifecycle.py conflicts because main wraps the
single-tab archive save in _timed_close_stage (unreleased). Release's
direct save_slot_off_loop call gets only the
after_commit_under_lock=slot.note_close_recorded line; the bulk cleanup
hunk and every other file apply as on main.

(cherry picked from commit 31924d9)
…34298)

A host deny steers its reason into a running webhook, task-runner or
task-refine turn, and kiro-cli has the model acknowledge it with a
[STEERING steer-<id>: ...] line. These collectors kept the turn text
raw, so that line reached the webhook's notification and owner DM,
the task step's recorded result, and the refined task spec.

Each now drops complete acknowledgements from the finished turn text
with messaging.driver.strip_steering_markers, the call the sub-agent
result already makes.

Closes #32476

Backport note: test/test_optimizer_refine_deny_notice.py is not on
release (#30076 moved the refine deny tests there, unreleased); on
release they live in test/test_handlers_auto_improvement_deny_notice.py,
so main's new refine test is added there unchanged, after
test_refine_a_failing_steer_still_rejects, and uses that file's own
doubles. learn-cron-dashboard.md adds main's new line after release's
Webhook Hooks paragraph. Production files, taskrunner.md and the other
tests apply as on main.

(cherry picked from commit 9d1ce77)
The #33212 pick (532f31f) applied clean, but its agentSpawn re-send
branch reads _sends_prompt, a local that main's _turn_dispatch_gates
binds from turn._agent_started. Release has no agent-started turn:
every turn that reaches this gate sends its prompt, which is the case
main's _sends_prompt is True for. As picked, every non-new dashboard
turn raised NameError at that elif (flake8 F821, mypy name-defined,
13 NameError failures in test/test_kas_spec_hooks_turn_loop.py).
Drop the always-true operand so the branch behaves as main's does for
a turn that sends its prompt.
@bolichen97
bolichen97 requested review from a team and KiroCrewCaptain as code owners November 10, 2026 15:16
@bolichen97
bolichen97 requested review from buluoray and removed request for a team November 10, 2026 15:16
@bolichen97
bolichen97 merged commit f0994b5 into release/0.12.0 Nov 10, 2026
16 of 23 checks passed
@bolichen97
bolichen97 deleted the hotfix/0.12.2-batch2 branch November 10, 2026 15:16
@github-actions

Copy link
Copy Markdown
Contributor

Sensitive Change Review

Sensitive files changed:

  • src/kiro_crew/security/exfil.py

  • src/kiro_crew/security/exfil_oauth.py

  • src/kiro_crew/security/redaction.py

  • src/kiro_crew/security/redaction_entropy.py

  • src/kiro_crew/security_posture_sinks_gateway.py

  • No reasoning template for 21fa270 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: 21fa270a10ab3bb2543787ea8c30805b0c094451
- 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

Copy link
Copy Markdown
Contributor

First Principles Review (Opus 5.5) — ✅ PASS

Premise-level review of 21fa270a10ab3bb2543787ea8c30805b0c094451 — 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

Every item is a cherry-pick -x of a fix already merged and reviewed on main, so its provenance is the main merge commit. The two assembly commits only fix NameErrors the picks introduced, which unwedges the build.

What this change ships

Inventory (10 items) — 10 justified

Intent: give 0.12 users the bug fixes main already carries, without waiting for 0.13. This is a FIX (a backport batch).

  1. Security and Windows file-access fixes from main (junction refusal, credential masking, approval_mode=auto refused for app tokens) — justified
  2. Chat and dashboard fixes (renderer kept, focus kept, closed tab stays closed, Refresh sizing) — justified
  3. KAS and agent-spec fixes (broken engine copy moved aside once, agentSpawn context re-sent after compaction) — justified
  4. Session and restart fixes (savepoint sealing, resume prefetch joined, systemd stop or timeout reported) — justified
  5. Channel fixes (Slack approval stays on its chat, review draft kept, member permission check) — justified
  6. Schedule fixes (stateless cron runs kept out of search, task-plan workflows refused while the runner is off) — justified
  7. Desktop and sandbox fixes (pane port forwarding, owner-only config, workspace settings kept out of global) — justified
  8. The universal macOS desktop build gets 90 minutes, and backend-dist stays out of the universal merge — justified
  9. The _sends_prompt operand dropped from the fix(kas): send agentSpawn hook context again after a compaction #33212 pick, because it is a main-only local that raised NameError on release — justified
  10. The GitLab conflict test builds its fake fetch inline, because _gitlab_fetch exists only on main — justified

[FIRST-PRINCIPLES-REVIEWED] 21fa270

Description read

[DESCRIPTION-READ] 75b76230a7b6cb105a349506abf4bfc39db39eb4c28302663f38beb66115b2e6

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 21fa270a10ab3bb2543787ea8c30805b0c094451 (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 21fa270a10ab3bb2543787ea8c30805b0c094451 <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

GPT 6.1 completed its review of 21fa270a10ab3bb2543787ea8c30805b0c094451 and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] 21fa270

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 21fa270a10ab3bb2543787ea8c30805b0c094451 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.