Skip to content

fix: backport 0.12.1 fast batch v2b-3 (20 fixes, 4 docs/tests conflicts resolved) - #32991

Merged
buluoray merged 20 commits into
release/0.12.0from
crew/slack/0.12.1-fast-v2b-3
Nov 8, 2026
Merged

buluoray merged 20 commits into
release/0.12.0from
crew/slack/0.12.1-fast-v2b-3

Conversation

@buluoray

@buluoray buluoray commented Nov 8, 2026

Copy link
Copy Markdown
Contributor

Problem / Motivation

Goal: Bring twenty fixes for problems a 0.12.0 user can hit, merged on main after v0.12.0, onto release/0.12.0, one commit each, in a PR of its own.

Why it matters

Each fix changes behaviour a 0.12.0 user can run into today. Their code applies to release/0.12.0 cleanly; where a pick conflicted, the conflict was only in a doc, a test, translator notes or a generated index.

PR Area What a 0.12.0 user sees today
#30296 memory Memory restore crashes with RecursionError on a deeply nested journal
#30771 security dd to a raw disk device is not refused
#30848 snapshot Snapshot CLI fails on a read-only store
#31052 messaging Channel token check fails for users behind the channel's proxy
#30992 apps An app's window fails to open when Crew runs as a Linux service
#31081 chat long chats stay responsive; settled turns unmount folded tool pills
#31018 instances remote crew pane gives one ssh line forwarding every needed port
#30774 security git publish check no longer bypassed through -C path expansion
#31829 spec builder teardown no longer frees a slot a newer run just took
#31621 app sdk agent questions stay open for an app embed in its own window
#31789 file picker a newer search is no longer cancelled by an older debounce
#31985 knowledge Oversized Knowledge upload is refused with a message, not silently
#31774 subagent Windows subagent finish no longer blocks the gateway event loop
#30516 UI Settings actions confirm in-app instead of native browser dialog
#32090 kas built-in kiro-cli agents refused with a reason and a suggestion
#31970 dashboard switched reply variant kept after reloading long transcripts
#31601 mcp template MCP servers honour waitForReady, so tools load on time
#31949 dashboard palette Skills tab shows the active chat agent's skills
#32402 memory Same memory proposal filed dozens of times
#32246 browser Browser ops on Laya fail with needs_ref on busy pages

Not a goal

  • No fix that is not already merged on main; nothing is changed only here.
  • No version bump, no tag, no CHANGELOG.md change.
  • Rows whose pick conflicts in product code.

What changed (motivation → approach → change)

Each main squash commit was cherry-picked with -x onto release/0.12.0 @ 19bd79b4a8, in merge order on main, one commit each. Sixteen applied cleanly and keep their original author. Four conflicted only outside product code and were resolved by hand:

PR How the conflict was resolved
#30296 test_member_memory_backup.py: both new imports kept (release errno, pick gc)
#31018 en.context.json (translator notes): per-key merge; the pick's pane_port_unreachable note taken, release's plugin_registry_name note kept
#30516 SkillsTabValidationReason.test.tsx: the pick's in-app-confirm test shape kept, with release's list-failure assertion and per-alert loop inside it (7 related files, 258 tests pass)
#31949 docs/system-specs/modules/learn-cron-dashboard.md: word-level 3-way merge, both sides kept

A hand-resolved commit is authored by the batch owner, because the commit guard refuses any other author identity, and names the original author in a Co-authored-by: trailer (or a plain "Co-author on main" line where their address is internal).

# PR main commit release commit pick
1 #30296 refuse a too deeply nested restore journal or snapshot manifest 0b2b1b5837 eef2edc01f hand-resolved
2 #30771 refuse one more destructive disk write 7328a22fa6 3446b4492b clean
3 #30848 stage a permission-read-only store's database writable 9694b2482a c0b733210b clean
4 #31052 verify channel tokens through the channel's proxy 81d35e178f f7018832c2 clean
5 #30992 open an app's window under the installed linux service 0f93202938 6303b2cfdd clean
6 #31081 stop settled turns keeping every folded tool pill mounted 323c4f65ec ccd2aafb1a clean
7 #31018 give one ssh line that forwards every remote crew pane port 2327f144f9 bcd3b51d2e hand-resolved
8 #30774 tighten a git publish check 4b5126a6e8 55d8d9e506 clean
9 #31829 keep a slot minted under the key a teardown freed 923189f006 4cbc2d9fb7 clean
10 #31621 keep an agent ask open for an app embed in its own window 90621cc9d3 294157794a clean
11 #31789 stop an overtaken picker debounce cancelling a search a04855dde0 a0671a4ce6 clean
12 #31985 refuse an over-limit upload before sending it 4c8274fc16 99250ad05a clean
13 #31774 build the settle's result path off the event loop 9a9b540321 45bb214c27 clean
14 #30516 confirm settings actions in-app instead of natively c6d7c365e3 8a11573b1d hand-resolved
15 #32090 refuse kiro-cli's built-in agents with a reason and a pick c6da224208 80bc423eb0 clean
16 #31970 keep a switched variant on bounded transcript reads fd51cbcd85 d91e6866b8 clean
17 #31601 template MCP additions carry waitForReady and merge that server first fc1110cbdc 0bcf5c56f5 clean
18 #31949 scope the palette skills tab to the active chat's agent 959bc047fd 03c749a703 hand-resolved
19 #32402 keep one memory proposal when every session re-proposes it b42bb2d889 a8fd47d53d clean
20 #32246 keep decision choices within laya's option limits 136d7a2218 8f23b5585f clean

Backwards compatibility

Compatible: each change is the one already merged on main, and its compatibility note applies unchanged.

Removes nothing: no setting, key or caller is removed.

Tests

Proof: not provable -- other: a backport of main commits whose proofs ran on their own PRs

  • CI does not run the backend suite on PRs into release/*, so the direct test modules were run here: the 20 Python test files these picks touch gave 1,196 passed, 4 skipped and 2 xfailed.
  • Frontend: tsc -p tsconfig.app.json --noEmit passes; the 19 vitest files these picks touch gave 469 passed; gen-ui-index --check and lint:i18n pass.
  • flake8, isort, the black gate, mypy src/kiro_crew and scripts/docs-lint.sh pass.

Manual verification

N/A - each fix was verified on its own main PR. The 0.12.1 RC soak is the end-to-end check.

Screenshots / video

Why no screenshot: the frontend picks carry the same rendered change as their main PRs, which show it.

Related Issues

Backport to the 0.12 line for 0.12.1 of #30296, #30771, #30848, #31052, #30992, #31081, #31018, #30774, #31829, #31621, #31789, #31985, #31774, #30516, #32090, #31970, #31601, #31949, #32402, #32246.

Sibling sweep

Siblings: none -- each source PR carried its own sibling sweep; this PR only carries those fixes.

Pattern harvest

Not generalizable: this PR only carries fixes already on main. The source PRs have their own harvest.

Checklist

  • At most two commits (one is the norm), with a Conventional Commits title (feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)
  • 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.

buluoray and others added 20 commits November 8, 2026 06:51
…anifest (#30296)

(cherry picked from commit 0b2b1b5)

Co-authored-by: Rohan Khanderia <51836404+RohanK6@users.noreply.github.com>
A command that writes a raw disk with no input file was allowed while
the deny list claimed to cover it. It is refused now; ordinary file
writes, searches and commit messages that quote it stay allowed.

(cherry picked from commit 7328a22)
…#30848)

A store made read-only by file permissions (chmod a-w), not by a read-only
mount, failed kirocrew snapshot with "attempt to write a readonly database"
while the same layout on a read-only mount gave rc 0.

The source read already handled the read-only case: a closed WAL database that
SQLite cannot open mode=ro is read immutable. The failure was on the
DESTINATION. The staging tree walk copies each file with the SOURCE file's
mode, so the byte copy of a 0o444 store database is itself 0o444, and the
read-write open that writes the consistent copy over it is refused. On a
read-only mount the files keep an ordinary mode, so that copy stays writable.

Restore owner write on the staged copy before the consistent copy overwrites
it. The copy is this process's own file in its private staging tree; the live
store is never touched. Covered for WAL and delete journal modes.

Co-authored-by: kirocrew-agent <agent@kirocrew.local>
(cherry picked from commit 9694b24)
…1052)

On a host that reaches the internet only through HTTPS_PROXY, saving a
new Discord, Telegram or Webex bot token reported the service as
unreachable and stored the token unverified, while the running channel
connected fine. The token checks opened a plain aiohttp session, which
ignores the proxy environment, and dialed directly.

Each check now passes the proxy its own channel client uses: the Discord
and Webex clients' _resolve_proxy, and for Telegram the client's
forward-proxy rule, which skips the proxy for a loopback API base so the
token never leaves the machine. The messaging spec names the rule for
each save route.

Part of #20185

(cherry picked from commit 81d35e1)
…0992)

An app's Open action decided "remote" from the gateway's own
environment only. The installed Linux service is a system unit, so its
environment never names a display server, and every Open on a desktop
host answered "Kiro Crew is running remotely" instead of launching.

When the gateway names no display, ask the desktop session the user's
systemd manager names, the same checked lookup the file manager already
uses, and launch when it is live. The launcher gets that session's
display, X cookie and desktop name; the session never supplies a bus
address, so the child's bus reach is unchanged.

(cherry picked from commit 0f93202)
…31081)

With "Show Thinking Inline" off (the default), a finished turn folds its
working steps behind "Worked through N steps", but every folded tool
pill stayed mounted: hidden at zero height, still in the DOM and still
subscribed to the store. A turn with 100 tool calls kept about 2,000
hidden elements, and the transcript keeps several such turns mounted,
so the chat page's DOM grew with every tool-heavy turn.

A settled fold now drops the content of its folded tool rows in every
fold mode, as the fold already did with "Show Thinking Inline" on. The
row element stays, so opening the fold mounts the pills again; rows
that can still go live (running, an MCP App, a pill the user opened)
keep their content as before.

Part of #23281

(cherry picked from commit 323c4f6)
…e port (#31018)

When the browser cannot reach a remote crew pane's port, the pane error
now prints one ssh -L line with every crew pane port, so one paste fixes
every pane instead of one failure per crew. The dashboard port stays out:
the browser sees only its local end, so a same-number forward for it
could aim at the wrong gateway port.

Closes #31012

(cherry picked from commit 2327f14)

Co-author on main: Bolin Chen
Refuse one more git push spelling whose destination cannot be
determined before it runs, the case reported in the linked issue.
Quoting the value keeps the command allowed.

Closes #18520

(cherry picked from commit 4b5126a)
…31829)

A Spec Builder delete whose transcript archive fails, and Create's
orphan recovery when a worker outlives its cancel, put the popped
slot back under its key unconditionally. Between the pop and that
put-back the key is free, so any opener can mint a new slot under
it, and the put-back then overwrote that live slot.

Put the slot back only while the key is still free, as close_slot
and the idle sweep already do.

(cherry picked from commit 923189f)
…ow (#31621)

An app's ChatEmbed reads the agent's ask_question card by polling
GET /api/chat/slots/{slot}/question; it is sent no question_card
frame. open_agent_ask counted only dashboard sockets, so with no
dashboard tab connected (an app in its own window) the ask was
withdrawn at once and the agent fell back to plain text before the
embed could poll it.

Each read of that route that passes its existing checks now stamps
the slot, and an ask that reaches no socket is kept while that stamp
is at most 20 s old (twice the embed's slowest poll). Writer and
reader use the coordinator's one clock, and the stamp lives on the
slot, so it goes away with it.

Closes #25379

(cherry picked from commit 90621cc)
…31789)

A file-picker debounce timer armed for an earlier query (the '' a pick
leaves behind) could fire after a newer query had committed but before
the effect cleanup cleared it. Publishing that stale query moved the
search key off the newer query's in-flight fetch, cancelling it, so the
old root's placeholder rows showed as settled and a click relativized
against the wrong project. The timer now publishes only the query that
is still current.

(cherry picked from commit a04855d)
Co-authored-by: w <w@k.local>
(cherry picked from commit 4c8274f)
)

taskq_settle built result_ref by calling agent_dir_for_display, which
resolves the run directory with two realpath calls. That ran on the
event loop for every finished subagent, even though the finish itself
is posted to the task-store writer thread. On Windows realpath opens
every path part, so a busy disk stalled the loop for tens of seconds
and the loop-stall watchdog ended the gateway.

The path is now built inside the job that runs the finish, so the
posted path does it on the writer thread. The row_settled path, which
never writes the row, no longer builds it at all.

Closes #31768

(cherry picked from commit 9a9b540)
…30516)

Importing a config in Replace mode, deleting or dismissing a skill,
dismissing all pending skills, and deleting a steering file asked
through the browser-native confirm(). In the desktop app that is an OS
dialog that blocks the renderer, and on Windows the window can stop
taking keyboard input after it closes until it is re-activated (#15976).

The Portability, Skills and Steering tabs now ask through the in-app
useConfirm dialog. The title is a short action label and the unchanged
question is the dialog body, because Modal truncates a title to one
line. Each action captures its target before it waits and acts only on
a true answer: the chosen import file, mode and memory choice, the skill
key, the candidate and its reviewed digest, the candidate queue shown
when the question opened, and the steering file with its slot and
project.

Steering deletion keeps invalidating every steering listing and keeps
evicting every cached copy of a deleted user file. For a workspace file
it now evicts only the captured project's copies, so another project's
file with the same name keeps its cache. The query client sets
staleTime: Infinity, so an entry left in the cache is not refetched on
its own. Switching the Skills tab to the budget view, including through
browser history, or the owning candidate row disappearing, cancels its
pending question.

Remove the three files from the native-dialog allowlist.

Part of #15976

(cherry picked from commit c6d7c36)

Co-authored-by: abirkel <34585063+abirkel@users.noreply.github.com>
…32090)

kiro_default, kiro_planner and kiro_guide are built into kiro-cli, so no
spec file defines them, and a KAS session for one was refused as a
missing agent definition with no way forward. KAS has no safe match:
its own closest modes decide some tool calls without asking Crew. The
KAS harness now refuses them with a reason that names the agent as a
kiro-cli built-in and names Kiro Crew's default agent to pick instead.
An app that declares the same name keeps its app-aware refusal.

Fixes #32006

Co-author on main: Joe Guo
(cherry picked from commit c6da224)
…31970)

A bounded slot-detail read served the disk copy of every row it matched
to the live window by id, so an in-place edit the save had not flushed
yet (a variant switch whose save failed and waits for the next flush)
repainted the text the user had just switched away from. The no-limit
read serves the window and showed the right text, so the two disagreed.

The id walk now serves the window's copy of a row it matched, when the
id appears once in the disk region; a repeated caller-supplied id keeps
the disk copies. Rows disk holds that the window does not (another
writer's append) stay where disk put them, so row counts, total and
next_before do not change.

Part of #7526

(cherry picked from commit fd51cbc)
…first (#31601)

* test: red-first: template MCP additions carry waitForReady and merge that server first

* fix: template MCP additions carry waitForReady and merge that server first

---------

Co-authored-by: Raymond Chen <raymondchen625@gmail.com>
(cherry picked from commit fc1110c)
The command palette's Skills tab listed the unscoped catalog while a
$token only expands for a skill in the chat agent's own skill://
mapping, so the palette could insert a token that is sent as literal
text. With a chat open it now asks /api/skills with that chat's
session key and agent (the default template for a chat with none), as
the $ picker does, leaves out a project skill the operator has not
trusted, and keys its result cache on that scope so another chat's
list is never served. The spec sentence that still called $-token
expansion global is corrected.

Closes #6027

(cherry picked from commit 959bc04)

Co-author on main: Bolin Chen
On a crewmate's private store, consolidation saves a changed fact as a
proposal for the owner. propose_conflict reused a pending row only when
the source matched, but consolidation proposes as consolidation:<session>
with a per-span source_ref and a fresh confidence, so every session filed
the same proposal again and one record collected dozens of them.

A pending proposal is now reused when it makes the same change at the
same base revision and operation, with provenance (source, confidence,
source_ref, observed_at and the row's updated_at clock) set aside.

Part of #29663

(cherry picked from commit b42bb2d)
Laya's local server refuses a Choice of more than 100 options and a
request whose options add up to more than 512, and the browser point
sent up to 200 per Choice, so on Laya a target-only op ended as
needs_ref and the reflex loop got no answer on a busy page.

Cap the pick Choice and each reflex target head at 100 options for
every provider, as skills.select already does. The reflex cap counts
options rather than elements, since a dropdown row adds up to 50, and
a dropdown is offered whole or not at all, so a reflex request carries
at most 465 options.

Closes #30973

(cherry picked from commit 136d7a2)
@buluoray
buluoray requested review from a team and KiroCrewCaptain as code owners November 8, 2026 07:12
@buluoray
buluoray requested review from Zedmor and removed request for a team November 8, 2026 07:12
@github-actions

github-actions Bot commented Nov 8, 2026

Copy link
Copy Markdown
Contributor

First Principles Review (Opus 5.5) — ✅ PASS

Premise-level review of 8f23b5585fa706ae89aa793c07d640283b01eb3b — 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 twenty fixes already merged on main. Whether each fix should exist was settled in its own source PR. The files touched line up with the twenty listed PRs, and nothing new was added only for the release branch.

What this change ships

Inventory (10 items) — 10 justified

Intent: give 0.12.x users twenty fixes that are already on main, by applying them to release/0.12.0 for 0.12.1. This is a FIX (backport).

  1. Memory/snapshot restore refuses a journal or manifest nested too deeply, instead of crashing with RecursionError (fix(memory): refuse a too deeply nested restore journal or snapshot manifest #30296). The snapshot CLI works on a read-only store (fix(snapshot): read a permission-read-only store like a read-only mount #30848) — justified
  2. The shell gate refuses dd to a raw disk, and the git publish check can no longer be bypassed with -C (fix: refuse one more destructive disk write #30771, fix: tighten a git publish check #30774) — justified
  3. Channel token checks go through the channel's proxy (fix(messaging): verify channel tokens through the channel's proxy #31052) — justified
  4. App windows open under the Linux service, and agent asks stay open for an app embedded in its own window (fix(apps): open an app's window under the installed linux service #30992, fix(app-sdk): keep an agent ask open for an app embed in its own window #31621) — justified
  5. Long chats stay responsive, and a switched reply variant survives a reload (fix(chat): stop settled turns keeping every folded tool pill mounted #31081, fix(dashboard): keep a switched variant on bounded transcript reads #31970) — justified
  6. The remote crew pane gives one ssh line that forwards every port it needs (fix(instances): give one ssh line that forwards every remote crew pane port #31018) — justified
  7. Spec-builder teardown, the picker debounce and the Windows subagent settle no longer race (fix(spec-builder): keep a slot minted under the key a teardown freed #31829, fix(frontend): stop an overtaken picker debounce cancelling a search #31789, fix(subagent): build the settle's result path off the event loop #31774) — justified
  8. An over-limit Knowledge upload is refused with a message (fix(knowledge): refuse an over-limit upload before sending it #31985). Settings actions confirm in-app instead of in a native browser dialog (fix(overview): confirm settings actions in-app instead of natively #30516) — justified
  9. kiro-cli built-in agents are refused with a reason. Template MCP servers honour waitForReady. The palette Skills tab is scoped to the active chat's agent (fix(kas): refuse kiro-cli's built-in agents with a reason and a pick #32090, fix: template MCP additions carry waitForReady and merge that server first #31601, fix: scope the palette skills tab to the active chat's agent #31949) — justified
  10. A duplicate memory proposal is filed only once, and Laya decision choices stay within its option limits (fix: keep one memory proposal when every session re-proposes it #32402, fix(browser): keep decision choices within laya's option limits #32246) — justified

[FIRST-PRINCIPLES-REVIEWED] 8f23b55

Description read

[DESCRIPTION-READ] 339d0d5be6a3262feaf98923dcd37795383209bd5050a20d7b1e289cd33c6311

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

github-actions Bot commented Nov 8, 2026

Copy link
Copy Markdown
Contributor

UX Review (Opus 5.5) — ✅ PASS (a BLOCK with no Trigger and no Outcome does not block)

UX-level review of 8f23b5585fa706ae89aa793c07d640283b01eb3b — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS does not.

UX-Verdict: PASS

cannot evaluate: missing screenshots of the user-visible changes. No image was supplied, so the blind read was not performed, and "the main PRs show it" is not evidence this lane can admit.

Evidence gaps

All five could come from re-attaching the source PRs' images to this description, which is the only revision this lane reviews.

Notes

  • pane_port_unreachable hedges ("may be restarting; retry") about state the pane already knows: it never loaded. Consider naming the next step instead. This is outside a backport's scope; it belongs on main.

[UX-REVIEWED] 8f23b55

Description read

[DESCRIPTION-READ] 20b3fac01ecb753c8164f01f7203b46f18d1497ec051f6ce779f5eb090acef10

That sha256 names everything this verdict read: the PR description, plus the 3 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with .github/scripts/pr-description-capture.sh over the same inputs; if it differs, the description or that evidence changed after this verdict was formed and any finding drawn from either is unproven.

@github-actions

github-actions Bot commented Nov 8, 2026

Copy link
Copy Markdown
Contributor

Design Review (Opus 5.5) — ✅ PASS (a BLOCK with no Trigger and no Outcome does not block)

Design-level review of 8f23b5585fa706ae89aa793c07d640283b01eb3b — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS does not.

Design-Verdict: PASS

cannot evaluate: missing a screenshot of the in-app confirm dialogs and the Knowledge upload refusal as they render on release/0.12.0. The description sends reviewers to the main PRs, which show another branch.

Blockers

No rendered evidence of this revision for the reshaped Settings and Knowledge surfaces
Trigger: the diff swaps native confirm() for an in-app dialog on website/src/pages/overview/SkillsTab.tsx ("+ const { confirm, cancel, confirmDialog } = useConfirm()", "+ if (viewBudget) return <>{confirmDialog}<SkillContextBudget …/></>"), and does the same in SteeringTab.tsx and PortabilityTab.tsx. It also adds an upload-limit refusal to SourcesList.tsx and scopes the CommandPalette skills tab to the active chat's agent. The evidence file lists 0 images and 0 recordings, and the description says "the frontend picks carry the same rendered change as their main PRs, which show it."
Outcome: screenshots on the main PRs show main's tree, not release/0.12.0. The SkillsTab tests were merged by hand during the cherry-pick, and the release branch's surrounding components differ from main's. So nobody has seen how the dialog, its mount points (loading state, budget view) or the refusal message actually render on the branch that ships 0.12.1.
Fix: attach screenshots taken from a build of this branch: the Skills delete confirm, one Steering or Portability confirm, the Knowledge over-limit refusal, and the palette Skills tab.
Clears when: github.com/user-attachments images of those surfaces, captured from this head, appear in the description, or the same images are committed under temp-screenshots/.

Notes

  • "Manual verification: N/A" holds here only because fix(subagent): build the settle's result path off the event loop #31774's change (_finish moved into store.run) runs the same code on every OS, not a Windows-only branch. Say so in the description instead of pointing at the RC soak.
  • The ## Backwards compatibility section has both a Compatible: line that names no reader and a Removes nothing: line. Keep only the Removes nothing: line.

[DESIGN-REVIEWED] 8f23b55

Description read

[DESCRIPTION-READ] e87d1976c945e7a1d850bbfc8f917181a8a09afe9fd001f6347cb8bc6239c374

That sha256 names everything this verdict read: the PR description, plus the 1 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with .github/scripts/pr-description-capture.sh over the same inputs; if it differs, the description or that evidence changed after this verdict was formed and any finding drawn from either is unproven.

@github-actions

github-actions Bot commented Nov 8, 2026

Copy link
Copy Markdown
Contributor

GPT 6.1 Review — ✅ no blocking findings

GPT 6.1 completed its review of 8f23b5585fa706ae89aa793c07d640283b01eb3b and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] 8f23b55

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 8f23b5585fa706ae89aa793c07d640283b01eb3b false-positive: <reason, 20 to 500 characters>

@github-actions

github-actions Bot commented Nov 8, 2026

Copy link
Copy Markdown
Contributor

Security Scope Review -- ✅ PASS

Which legitimate operations does 8f23b5585fa706ae89aa793c07d640283b01eb3b newly refuse? Updated in place on each push. A script-confirmed regression blocks; the model proposes candidates and never decides a refusal.

Verdict basis: adjudicated, zero confirmed regressions.

Security scope review -- adjudicated candidates

  • macos-latest (platform posix): 21 classified, 0 newly refused, 20 still allowed.
  • ubuntu-latest (platform posix): 21 classified, 0 newly refused, 20 still allowed.
  • windows-latest (platform windows): 6 classified, 0 newly refused, 6 still allowed.

No candidate flipped from allowed to refused.

Scope-Verdict: PASS

Reason: Both tightenings are in the shell deny composite, so the classifier can check them. No script-confirmed regression was handed to this review. One overlap is deliberate and is not written as a row: on posix, an ordinary push with an unquoted variable before the subcommand (a worker writing git -C $WORKTREE push -u origin feat/x) is now refused as git-publish-target-unverifiable. The diff accepts that loss as the cost of the word-splitting threat.

TIGHTENING: Two mechanisms got stricter.
(1) The git-publish argv floor. A word skipped between git and push that carries an unquoted $ now leads the returned push tokens, so the caller's $ check refuses the push as unverifiable. The threat is a value that word-splits into a push of a protected branch: "with d='x push origin main', git -C $d push origin feature can push main". The hunk is args: list[str] = [word for word in raw_tokens[start + 1 : i] if not platform_compat.IS_WINDOWS and any(step.active and step.char == "$" ....
(2) A new argv rule in the rule catalog, local-destructive-dd-of-device. It refuses dd when an output operand starts with a raw-disk prefix ("of=/dev/" + name for name in ("sd", "xvd", "hd", "vd", "nvme", "mmcblk", "loop", "dm-", "md", "dasd", "disk/")). The threat is dd wiping a disk from stdin when no if= is given.

ADJUDICATED: 27 candidates, 0 confirmed newly-refused

PLATFORM COVERAGE:

  • posix: These rows probe the edges of the new skipped-word $ scan:

    • a quoted $ variable and a quoted $(...) substitution;
    • a $ inside quotes glued to a flag (--git-dir="$GIT_DIR", -c core.sshCommand="... $HOME ...");
    • an env-prefix $ before git;
    • a $ in a separate cd segment;
    • tilde and relative paths;
    • unquoted $ before non-push subcommands (fetch, log).

    For dd, the rows use /dev/ outputs that sit beside the raw-disk prefixes (/dev/null, /dev/stdout, /dev/shm, /dev/fd/3), raw disks used only as stdin (< /dev/sda, < /dev/nvme0n1), a regular file under sudo (/swapfile), disk.img, and text-only mentions (echo, rg, a git commit -m). There is also a git add row, because the cheap _could_name("dd") pre-check now sends every git add through the argv reader.

  • windows: The diff assumes the Windows shell is PowerShell and does not word-split. The rows cover $env:WORKTREE, a quoted $env:USERPROFILE\... path, a drive-letter path, and (Get-Location).Path before push. Two more rows cover a backslash git add path and a dd.exe name with a $env:TEMP output.

  • Not probed: I wrote no row for a posix-style shell (Git Bash/MSYS) running on a Windows host. There the IS_WINDOWS exemption applies even though that shell does word-split. A row can't express which shell runs the command, so this stays open.

  • any: No rows. Every candidate names a path, separator or executable.

UNADJUDICATED: The deliberate overlap was left out of the candidates on purpose. Writing it as a row would admit the threat shape:

  • Guard: _git_push_args in src/kiro_crew/security/argv_floor_git.py, reached through deny_gate.is_denied → the git-publish floor → the caller's $ check → git-publish-target-unverifiable.

  • Inputs I would have submitted as legitimate:

    • git -C $WORKTREE push -u origin feat/example
    • git -C $(git rev-parse --show-toplevel) push origin feat/example
    • git --git-dir=$GIT_DIR push origin feat/example

    These are posix-only, and an agent writes them routinely. They are refused by design. The only fix for the user is to quote the variable, which the refusal message should say. A narrower rule could refuse only when the expansion can't be resolved and splits into a word equal to a git subcommand. That isn't checkable statically, so the trade-off is the author's to make. I'm noting it, not blocking on it.

[SCOPE-REVIEWED] 8f23b55

@github-actions

github-actions Bot commented Nov 8, 2026

Copy link
Copy Markdown
Contributor

Opus 5.5 Review — ✅ no blocking findings

Reviewed 8f23b5585fa706ae89aa793c07d640283b01eb3b — this comment is updated in place on each push.

Review details

No findings.

[OPUS-REVIEWED] 8f23b55

Verdict parsed from the review's SHA-scoped output markers for commit 8f23b5585fa706ae89aa793c07d640283b01eb3b.

False positive or not applicable? A repository writer can comment:
/ai-review override fable 8f23b5585fa706ae89aa793c07d640283b01eb3b false-positive: <reason, 20 to 500 characters>

@buluoray
buluoray merged commit 2fc09a0 into release/0.12.0 Nov 8, 2026
23 of 28 checks passed
@buluoray
buluoray deleted the crew/slack/0.12.1-fast-v2b-3 branch November 8, 2026 07:23
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.

5 participants