Skip to content

fix: backport 40 user-facing fixes to 0.12.1 (batch s3b1) - #33574

Merged
buluoray merged 40 commits into
release/0.12.0from
crew/slack/0.12.1-s3b1
Nov 9, 2026
Merged

buluoray merged 40 commits into
release/0.12.0from
crew/slack/0.12.1-s3b1

Conversation

@buluoray

@buluoray buluoray commented Nov 9, 2026 •

Copy link
Copy Markdown
Contributor

Problem / Motivation

Goal: Ship 40 user-facing fixes that are on main but not on release/0.12.0 to 0.12.1 users.

These fixes merged to main after the 0.12 cut. Each one is a bug today's 0.12 users can hit, and each cherry-picks onto release/0.12.0 with no conflict.

Why it matters

0.12 users keep hitting these bugs until a patch release carries the fixes. They cover the dashboard, chat, Bots, cron, artifacts, the installer, doctor and the desktop app.

Not a goal

  • No fix that needs an earlier PR first or that conflicts on release; those stay for a later batch.
  • No change written for release: every commit is a git cherry-pick -x of the main merge commit, unchanged.
  • No CI, test-only, docs-only or dev-tooling commits.

What changed (motivation → approach → change)

Slice 1 of the merged-to-main, not-on-release list was triaged row by row: on release already, user-facing, code exists on release, and a trial pick. The 40 rows that are user-facing, not on release and pick cleanly are applied here in main merge order, one cherry-pick -x each, on release/0.12.0 at f6d1874bad, after #33530 (s3b4) merged.

# PR What users hit
1 #30033 installer named the wrong restart command when re-run under a gateway
2 #30120 @ picker could not find files in .kiro/specs and steering
3 #29592 shared crew showed duplicate invites and People rows per person
4 #29470 composer height wrong after the window width changed
5 #30288 keyboard focus lost when arming or cancelling a remote update
6 #29992 batch write guard gaps in ACP file writes
7 #30439 agent prompt promised AWS CLI calls are never blocked
8 #30444 doctor did not show or warn on an old claude-agent-acp
9 #30542 skill install hint pointed at a button that does not exist
10 #30106 claude backend did not load workspace steering or project AGENTS.md
11 #30624 false claude-not-found warning when the SDK ships its binary
12 #30445 files app search matched too loosely
13 #30711 one non-UTF-8 lessons.jsonl line broke every lesson
14 #30708 cron preview listed runs on skipped dates
15 #25959 restart ran against a stopped installed service
16 #30751 an Infinity number in artifact meta.json was not treated as bad
17 #30724 inline thinking block shown as plain text, not in the reasoning row
18 #30861 live browser card missing in a Bot's chat
19 #30879 resume-all row did not name the interrupted sessions
20 #30839 a *.localhost preview name was refused as an outside address
21 #28488 frontend sanitizer leaked the second of two adjacent credentials
22 #20094 macOS app did not come to front when the main window was shown
23 #31258 a failed artifact_move still created its target folders
24 #21175 empty capped project tree looked like an error
25 #31136 wrong-shaped bodies on five upload routes gave a 500, not 400
26 #31050 a config value clamped on load was not logged
27 #31263 knowledge PATCH with a null title or type was not a no-op
28 #31027 doctor passed when the kiro-cli backend had no kiro-cli
29 #26761 a resident-resume entry took a queued subagent lane slot
30 #31090 reply strip did not show the history the model route sent
31 #31416 a failed embedding runtime showed as running on the system page
32 #31185 Windows crash dumps not found in Crashpad's reports folder
33 #31318 cron preview failed for scripts using session or crew-log calls
34 #31419 desktop pets hit-tested wrong under dashboard zoom
35 #31498 infinite tag column order gave a 500
36 #31497 cron history row with non-string text gave a 500
37 #31491 artifact with a list or object source failed without a validation error
38 #31536 computer use silently missing when the chat backend cannot load it
39 #31511 macOS governance probe missed dashboard windows
40 #31506 notebook queued-for-indexing notice not pluralized

Dropped: none. All 40 picked cleanly in order, and none was taken out for a test failure.

Another batch from slice 2 is being built in parallel; whichever merges second may need a rebase.

Backwards compatibility

Compatible: each commit is the main commit unchanged, already reviewed and merged there.

Removes nothing: no pick takes a capability away on its own beyond what its main PR already shipped and documented.

Tests

Run on this branch at 1d2f066f6a, one process at a time, backend in a clean Python 3.12 venv (PYTHONNOUSERSITE=1, one file per pytest run):

  • pytest -n0 on the 45 test files the picks touch: 2,111 passed, 5 failed. The same 5 fail identically on plain release/0.12.0 with no picks, so they are not caused by this batch (checked on both 7a89cc18d6 and f6d1874bad):
    • test/test_cli.py::TestDoctor::test_doctor_with_kiro
    • test/test_cli.py::TestDoctor::test_doctor_windows_missing_stt_engine_and_ffmpeg_is_non_fatal
    • test/test_cli.py::TestDoctor::test_doctor_slack_workspace_allowed_ok
    • test/test_folder_steering_provider_dedup.py::test_an_opted_out_kiro_chat_receives_declared_trees_once
    • test/test_folder_steering_provider_dedup.py::test_opted_out_plain_kiro_without_folder_steering_keeps_managed_declaration
  • vitest run on the 79 website test files the picks touch or that test a touched file: 1,240 passed.
  • tsc --noEmit -p tsconfig.app.json: 0 errors.

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; this batch only carries them.

Screenshots / video

Screenshots come from each pick's own main PR; every pick is applied here unchanged (cherry-pick -x, no edits), so the rendered change is the same. Picks whose main PR carried no screenshot are not shown.

#29592 (from its main PR; the picked change is identical)

#29592

#30724 (from its main PR; the picked change is identical)

#30724

#30861 (from its main PR; the picked change is identical)

#30861

#30879 (from its main PR; the picked change is identical)

#30879

#21175 (from its main PR; the picked change is identical)

#21175

Related Issues

0.12.1 backport batch from slice 1 of the merged-to-main, not-on-release list.

Checklist

  • Each commit is a cherry-pick -x of a main commit with a Conventional Commits title
  • Touched tests run; the only failures also fail on the base branch
  • Self-review completed; code follows project style guidelines
  • Documentation updated (if applicable)
  • No secrets, credentials, or internal references in the diff

Contribution License Agreement

@buluoray
buluoray requested review from a team and KiroCrewCaptain as code owners November 9, 2026 01:38
@buluoray
buluoray requested review from tim-finnigan and removed request for a team November 9, 2026 01:38
@github-actions

github-actions Bot commented Nov 9, 2026 •

Copy link
Copy Markdown
Contributor

Sensitive Change Review

Sensitive files changed:

  • src/kiro_crew/deny_guidance.py

  • No reasoning template for 1d2f066 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: 1d2f066f6ac9a6fbc5d29f5a70859304bc033258
- 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 added the readiness: checking Automated validation is still running label Nov 9, 2026
@github-actions

github-actions Bot commented Nov 9, 2026 •

Copy link
Copy Markdown
Contributor

Design Review (Opus 5.5) — ✅ PASS

Design-level review of 1d2f066f6ac9a6fbc5d29f5a70859304bc033258 — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS does not.

Design-Verdict: PASS

Nothing blocks: this is 40 unchanged cherry-pick -x copies of fixes already merged on main, and every reader of a symbol whose signature changed (_search_rg, _load_steering_resources, agent_spec_mcp_refs) is updated in this tree.

Notes

[DESIGN-REVIEWED] 1d2f066

Description read

[DESCRIPTION-READ] 7e46eb5962692ce98fcb619f24328eb40e5c524496dcba686d0ad00a09181979

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 9, 2026 •

Copy link
Copy Markdown
Contributor

First Principles Review (Opus 5.5) — ✅ PASS

Premise-level review of 1d2f066f6ac9a6fbc5d29f5a70859304bc033258 — 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

Nothing blocks. Every pick is an unchanged cherry-pick -x of a fix already merged on main. The only new surface is two batch-write-guard settings that default off.

Notes

  • Item 6 — undeclared: the row says "batch write guard gaps in ACP file writes", but the hunk also adds two settings, "batch_write_guard": {"type": "boolean", "default": False and batch_write_guard_commands, in config/loader_config.py. Nobody gets them without turning them on.

What this change ships

Inventory (10 items, of 40 picks) — 9 justified

Intent: bring 40 fixes from main to 0.12.1 users so they stop hitting known bugs. This is a FIX (a backport).

  1. Claude backend now loads workspace steering and the project AGENTS.md (fix(context): load claude workspace steering and project AGENTS.md #30106) — justified
  2. The @ picker finds files under .kiro/specs and steering (fix(dashboard): let the @ picker find files in .kiro/specs and steering #30120) — justified
  3. Shared crew shows one invite row and one People row per person (fix(shared-crew): one invite entry and one People row per person #29592) — justified
  4. Batch-write-guard toggle and command list appear in Settings, default off (fix(acp): finish the batch write guard follow-ups (#27138) #29992) — undeclared: new settings surface
  5. Agent prompt and skill hints point at Customize → Skills → Add Skill (fix: point skill install hints at the real add skill button #30542) — justified
  6. Doctor warns when claude-agent-acp is old or kiro-cli is missing (fix(doctor): show the claude-agent-acp version and warn when it is old #30444, fix(doctor): fail when the selected kiro-cli backend has no kiro-cli #31027) — justified
  7. Cron preview skips dates that would be skipped, and handles session or crew-log calls (fix(cron): keep interval and one-shot previews off skipped dates #30708, fix: let cron preview run scripts using session and crew-log calls #31318) — justified
  8. Upload, tag, cron-history and artifact routes give a 400 instead of a 500 for bad input (fix: answer 400 for wrong-shaped bodies on five upload routes #31136, fix(dashboard): refuse an infinite tag column order instead of a 500 #31498, fix(cron): skip a history row with non-string text instead of a 500 #31497, fix(artifacts): refuse a list or object source with a validation error #31491) — justified
  9. A config value clamped on load is now logged once per disk read (fix(config): log a stored value its bounds clamp on load #31050) — justified
  10. Frontend sanitizer redacts both of two adjacent credentials (fix(security): redact both of two credentials written back to back in the frontend sanitizer #28488) — justified

[FIRST-PRINCIPLES-REVIEWED] 1d2f066

Description read

[DESCRIPTION-READ] 5b31c718595f590f9f4577e619d539cf16db9b698419ef4d5774e05d2f20c502

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 9, 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 1d2f066f6ac9a6fbc5d29f5a70859304bc033258 — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS does not.

UX-Verdict: PASS

cannot evaluate: missing screenshots of the reworked People / Invite surface, the resume-all session list, and the other new or changed UI. The three supplied images show only the localhost preview card and the capped Files tree.

Evidence gaps

Notes

  • In shot-01, the blind reader would not click the green >_ dashboard.browser_local_origins chip ("I'd be afraid it runs something"). It also doesn't match the "Settings → Browser → Local preview" path in the text above it. Replace it with a link labelled with that path.
  • In shot-01, the shield with a tick reads as "OK" on a refusal message. Use a neutral or blocked icon instead.
  • Shot-02's capped-tree copy ("Nothing is wrong… too many files and folders… Type a file name") was understood correctly. It's an improvement over shot-03.

[UX-REVIEWED] 1d2f066

Description read

[DESCRIPTION-READ] cc3f13ab532a1cba643b05c05fe35c9ad0a9458fe2a3e4ea0debb3fca3c869d9

That sha256 names everything this verdict read: the PR description, plus the 6 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 9, 2026 •

Copy link
Copy Markdown
Contributor

Security Scope Review -- ✅ PASS

Which legitimate operations does 1d2f066f6ac9a6fbc5d29f5a70859304bc033258 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): 15 classified, 0 newly refused, 14 still allowed.
  • ubuntu-latest (platform posix): 15 classified, 0 newly refused, 14 still allowed.
  • windows-latest (platform windows): 9 classified, 0 newly refused, 9 still allowed.

No candidate flipped from allowed to refused.

Scope-Verdict: PASS

Reason: The patch doesn't change any shell deny tier (hooks.py, security/* and SHELL_DENY_TIERS are untouched). The only tightenings are guards outside the deny composite: the file explorer's rg provenance check and newline-query refusal, and a kirocrew restart refusal when an installed service is stopped. The classifier can't give a verdict on those, so they are named under UNADJUDICATED below rather than counted as clean.

TIGHTENING: the file explorer search now only runs an rg that passes validate_provider_executable, and it refuses queries that contain a line break. The threat is a $PATH-planted rg in an agent-writable tree: src/kiro_crew/apps/builtins/file_explorer/server.py + "``$PATH`` can reach a tree the agent writes, so the first ``rg`` on it goes through the same provenance check the dashboard's file search uses." / + return validate_provider_executable(found).

ADJUDICATED: 24 candidates, 0 confirmed newly-refused

PLATFORM COVERAGE: The posix rows cover the rg argv with the query on stdin (--file -), a cargo-installed rg path, which rg, the systemd --user service named in the restart refusal, project and global steering paths with / separators, and the AWS remediation commands. The windows rows cover the same areas in Windows spelling: rg.exe with a drive-letter root and \ separators, where.exe, kirocrew.exe, type .kiro\steering\..., and the PowerShell $env:USERPROFILE listing. The restart refusal (service_controller.installed_unit_path) has a macOS launchd form, and no candidate was written for it. That spelling is unprobed.

UNADJUDICATED:

  • rg provenance check. Call path: FileExplorerHandler /search and /health, then _rg_executable(), then shutil.which("rg"), then github_runner.validate_provider_executable. The base docstring of that check says it refuses outright when the gateway runs as root on POSIX. It also refuses binaries owned by another unprivileged account, world-writable paths, and anything under agent_writable_roots. When it refuses, search quietly falls back to the slower Python engine, so nothing is denied outright. Legitimate inputs I would have submitted:

    • a gateway running as root in a container with a distro /usr/bin/rg
    • an rg vendored inside the project (for example node_modules/@vscode/ripgrep/bin/rg placed on PATH by a dev shell)
    • a Windows rg.exe under a Scoop or winget shim directory whose ACL grants write to another principal

    Search results stay correct in all three, but the reported engine changes from rg to python.

  • Newline-query refusal. Call path: _search_using, which returns [] for any query containing \n or \r. Legitimate input: a user pastes a two-line snippet into the file explorer search box. It used to be passed to rg as one argv pattern; now it silently returns no results.

  • kirocrew restart refusal. Call path: cli_server_restart._restart, then service_controller.installed_unit_path() / unit_is_ours, then lock_holder(...), which exits 1 with reason=service_installed_not_running. Legitimate input: an operator runs kirocrew restart after stopping the service, expecting the gateway to come back. They now have to start it through systemd or launchd instead. That looks like the intended behaviour, but nothing in the diff records it as a chosen loss. I couldn't write a launchd candidate for this.

  • Display redaction rewrite. website/src/utils/sanitize.ts CredentialScan is a dashboard-side display filter and can't be expressed as a shell command, so the classifier can't check it. Where it may over-redact: an http://user:pass@ reader that now matches to the first @ anywhere after the colon, so ordinary chat text containing a URL followed later by an email address could be redacted.

[SCOPE-REVIEWED] 1d2f066

@github-actions

github-actions Bot commented Nov 9, 2026 •

Copy link
Copy Markdown
Contributor

GPT 6.1 Review — ✅ no blocking findings

GPT 6.1 completed its review of 1d2f066f6ac9a6fbc5d29f5a70859304bc033258 and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] 1d2f066

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 1d2f066f6ac9a6fbc5d29f5a70859304bc033258 false-positive: <reason, 20 to 500 characters>

@github-actions

github-actions Bot commented Nov 9, 2026

Copy link
Copy Markdown
Contributor

Opus 5.5 Review — ✅ no blocking findings

Reviewed 7c0166deaa53f229ba95757588b4a9098d28d4e1 — this comment is updated in place on each push.

Review details

No findings.

[OPUS-REVIEWED] 7c0166d

Verdict parsed from the review's SHA-scoped output markers for commit 7c0166deaa53f229ba95757588b4a9098d28d4e1.

False positive or not applicable? A repository writer can comment:
/ai-review override fable 7c0166deaa53f229ba95757588b4a9098d28d4e1 false-positive: <reason, 20 to 500 characters>

mbriones98 and others added 18 commits November 9, 2026 01:52
…#30033)

Re-running cli.sh writes the new version into the tree a running
gateway was loaded from (the managed venv is rebuilt in place, pipx
reinstalls into its own venv, and a first managed build retires the
legacy <data home>/venv), so that gateway runs a mix of old and new
code until something restarts it. Nothing told the user.

On a re-run that replaces an existing install, the installer now looks
for a running gateway with read-only probes (systemctl show for the
system and user kirocrew units, counted only when the manager reports
the canonical Id kirocrew.service so an alias of another unit is never
named; launchctl print for the launchd agent; and the pid gateway.lock
records, checked with ps) and ends with a warning naming the restart
command for that setup. It never starts, stops or restarts anything.
Restarting the service automatically from the installer is left for a
maintainer to decide.

Fixes #29736

Co-authored-by: Matt Briones <mdbrio@amazon.com>
(cherry picked from commit e4af0fd)
…ng (#30120)

The @-mention file search offers dot directories but never enters
them, so a spec's tasks.md or a steering file under .kiro could not be
attached with @. Both search paths, the FileIndex walk and the
per-request walk fallback, now take their descent decision from one
helper, file_index._dot_dir_walk_plan, driven by _OPEN_DOT_DIRS.

Below the search root, a .kiro directory is entered, but only its
specs and steering children are offered and walked, and the files
directly in .kiro are not offered. Every other dot directory (.github,
.vscode, .ssh, .aws, .kiro/settings, a nested dot dir) stays offered
but closed, _SKIP_DIRS still removes .git, and the per-entry
sensitive-path check is unchanged. The match is on whole path
segments. A symlinked .kiro/specs is offered but not followed, since
neither walk follows links.

The narrow allowlist is the approach pangiann proposed on the issue.

Closes #28703

Co-authored-by: pangiann <37578272+pangiann@users.noreply.github.com>
(cherry picked from commit 4f9b857)
)

People kept two invite entry points (the share card's Invite… dialog and an
inline card in personal mode) and showed a joined invitee twice, once as the
person and once as a Waiting row for the pending invite sent to the same email.

The Invite… button now opens the one dialog from both layouts, the inline card
is gone, and a pending email invite whose address matches a listed login is
not shown, so each person is one row and the count matches the rows. Step
prose, ⓘ tips and the dead catalog keys behind them are removed from all 12
catalogs, and the spec follows.

(cherry picked from commit 8495b59)
…th (#29470)

The width observer re-measured only when the width changed from the
last width it saw. A measurement taken by an effect during a layout
that is briefly narrower, and restored before the next rendering
step, is invisible to it, so an empty box measured at a near-zero
width keeps the 140px cap. The memo now records the width each
measurement used, and the observer re-measures whenever the current
width differs from it.

(cherry picked from commit 47b1fc8)
* fix(acp): finish the batch write guard follow-ups

Three follow-ups to the batch write guard (batch_guard.py), none changing
what the guard allows or blocks:

1. Audit labels. A consumer that writes its own SEL row when approve_tool
   returns False (the dashboard chat path and the Slack ladder) labelled a
   batch-guard refusal rejected_transport_floor, the same as a security-floor
   refusal. Each such row now reads the kept reason back through
   take_batch_refusal and the new batch_guard.refusal_outcome helper, so it
   records rejected_batch_guard (mechanism=batch_write_guard) when the guard
   refused and keeps rejected_transport_floor only when the floor did.

2. Settings surface. hooks.batch_write_guard and
   hooks.batch_write_guard_commands are declared on the hooks config field so
   the schema (and the Settings UI) carries them, both default off, and listed
   in configuration.md and the config spec. config-baseline.json regenerated.

3. GraphQL classification gap. Documented in the acp-client spec: a gh api
   graphql query body supplied through a shell command substitution or a shell
   variable reaches the guard as that unexpanded token and classifies as a
   read, so a mutation hidden there is not guarded. Left as a documented gap on
   purpose; closing it in code would make the guard refuse more than its tables
   name.

Closes #27138

* refactor(acp): one owner for taking a client's batch-guard refusal

Give the "ask the client for its kept batch-guard reason, take it once, keep
only a string" step a single owner, exposed to application code through
kiro_crew.agent_sdk (take_batch_guard_refusal / batch_guard_refusal_outcome, so
the dashboard, Slack and tool_permission consumers do not reach into the ACP
layer). The dashboard and Slack audit consumers and tool_permission.batch_refusal
all go through it instead of each repeating the getattr/take/isinstance steps.
No behaviour change.

The dashboard consumer-label repro lives in its own test file so it does not
perturb the sibling suite's facade-patch scope. Also drop issue-number history
narration from the guard helper unit test.

---------

Co-authored-by: kirocrew-agent <agent@kirocrew.local>
(cherry picked from commit 8580b50)
The credential-read refusal, the blocked-by-policy skill, the user
doc and `kirocrew doctor` each told the reader that AWS CLI calls are
not blocked. That is only true of the rule that refused the read: the
builtin catalog refuses some aws calls itself, and an enterprise
overlay can deny aws outright, so on such a host the agent followed
the advice into a second refusal and the operator was told something
false about their own host.

Each sentence now says the rule that refused the read does not block
AWS CLI calls, which no overlay can falsify; two lines that only
repeated the promise drop it.

Closes #7394

(cherry picked from commit 4f89e32)
#30444)

The claude-acp row said only that the adapter was installed, so an
adapter far below the version the installers pin looked the same as a
current one. The row now prints the version from the adapter's npm
package.json, never by running it: old releases ignore --version and
wait on stdin. Below the floor it adds a warning with the update
command. It is a warning only and does not change the exit code.

(cherry picked from commit 16a2cf4)
The skills doc, the agent prompt, the skill_fetch tool text and the
dashboard's install refusal sent users to Settings -> Skills -> Discover
to install a registry skill. That entry does not exist: installing is
the Add Skill button on Customize -> Skills. The doc's create-skill
step named an Overview page that is gone too; it is Customize ->
Skills -> Create New Skill.

(cherry picked from commit c1fa45b)
…30106)

The Claude Code backend does not read an agent spec's resources, so
the gateway loads kirocrew.json's file:// entries itself. It globbed
every entry against $HOME, so a claude chat bound to a project never
received <project>/.kiro/steering or <project>/AGENTS.md, which
kiro-cli loads for the same chat; and because the folder-steering
dedup counts the project tree as delivered on claude, a folder that
declared it dropped it too. An absolute entry made pathlib raise, and
the blanket except then emptied every entry, the global rule included.

Anchor each entry the way kiro-cli does (relative at the session's
project, else $HOME; ~/ at $HOME; absolute at its glob-free prefix),
add kiro-cli's inherited defaults (global steering; the project's
steering and AGENTS.md when a project is bound), admit each hit
against its own entry's anchor, read a document two entries match
once, and skip only the entry that cannot be globbed. The session's
project reaches the loader through steering_resources_part, so
session start and post-compaction re-injection stay on one gate.

(cherry picked from commit fcf9765)
…nary (#30624)

The Claude Code adapter's SDK installs its native binary as an
optional dependency. With it present a session starts and runs with
no claude on PATH, yet every spawn logged that the adapter would fail.
Look for the SDK's per-platform package next to the adapter, the way
Node resolves it, and warn only when neither exists.

(cherry picked from commit dc4d008)
Bring the Files app content search up to the standard the dashboard
file search already meets.

Part of #10852

(cherry picked from commit 600ae6b)
)

Co-authored-by: Leon <leon.lai.yc@gmail.com>
(cherry picked from commit 4bfb626)
Co-authored-by: Zejiang Guo <zejiangg@amazon.com>
(cherry picked from commit fbe1035)
…ber (#30751)

json.loads accepts Infinity, and int() on it raises OverflowError.
The meta.json decoder's numeric fallbacks (the publication version
map, last synced version and wrapper revision, and the fork's upstream
version) only caught TypeError and ValueError, and ArtifactStore.list
only skips a record on those too. So one hand-edited Infinity in a
single artifact made the whole Library list raise.

Those counters now fall back to 0 on OverflowError, as they do for any
other unreadable number, and an Infinity version is reported as the
ValueError a word version already is, so the list skips just that one
record.

(cherry picked from commit 8c2e758)
A model with no separate reasoning stream writes its scratchpad into
the reply text inside <thinking> tags. The messaging channels already
drop it, but the dashboard bubble printed the tag and the whole
scratchpad as prose. AssistantMessage now splits a block that opens a
line outside a code fence into the same collapsed row the streamed
reasoning uses, and hides a still-open block while the reply streams.
The answer keeps its own whitespace, and Copy, Quote, Read aloud,
follow-up chips and find-in-chat read the answer without the
scratchpad.

(cherry picked from commit 8fb88a6)
* fix(chat): show the live browser card in a Bot's chat

* fix(chat): keep a Bot's browser card across a parked send

(cherry picked from commit f7d574f)
RohanK6 and others added 22 commits November 9, 2026 01:52
…0879)

The Resume-all row said only how many sessions were interrupted, and
up to five a single press restarts them with no confirmation, so the
person could not tell which sessions it would restart. The row now
lists their titles under the count whenever a press resumes without
asking. Above that bound the existing confirmation still stands
between the press and the restart.

(cherry picked from commit 66a8552)
…an outside address (#30839)

* fix(preview): say a *.localhost name is refused as a localhost name, not an outside address

A *.localhost alias (foo.localhost) is a loopback name but is deliberately not
one of the three spellings the preview allowlist can admit, so it stays refused.
The panel routed it to the outside-URL card, which told the user it "is not a
local address" -- wrong for a loopback name. Add a refusal card that fits a
loopback name (points at Settings -> Browser -> Local preview) and route a
potentially-trustworthy refused host to it. 0.0.0.0 (unspecified, not loopback)
keeps the outside-address wording. The allow/deny decision is unchanged.

* docs(preview): add screenshot of the new *.localhost refusal card

Visual evidence for PR #30839: the real WebPreviewPanel rendering the new
loopback refusal card for http://foo.localhost:5173/ (no frame, no Allow
button). Committed under temp-screenshots/ (gitignored, force-added) because
this gh version has no --attach; maintainer prunes it from the tip post-merge.

---------

Co-authored-by: kirocrew-agent <agent@kirocrew.local>
(cherry picked from commit 41a96d2)
… the frontend sanitizer (#28488)

* fix(security): redact both of two credentials written back to back in the frontend sanitizer

* fix(security): read each frontend credential kind from its own prefix

Replace the sticky regex copies, the source-text pattern classification and
the probe budget in sanitizeCredentials with one scan that visits every start
and hands it to a reader per credential kind. Each reader is keyed on the
kind's exact prefix, consumes the characters its CRED_PATTERNS entry takes,
and returns where the credential ends; overlapping spans become one
placeholder. A credential glued after another is read from its own prefix,
and a span no longer redacts past where its last credential ends.

Refs #28366

(cherry picked from commit 6c9b032)
…ndow on macOS (#20094)

Relaunching the running app (Spotlight, a launcher re-exec), the tray's
Show item, and clicking the update notification all reach the app through
showMainWindow({ focus: true }) while another app is frontmost. On macOS
14+ win.focus() orders the window in without activating the app, so the
window does not come to the front and the other app keeps the keyboard.

showMainWindow now follows a focused show with app.focus({ steal: true })
on macOS, the same activation the summon hotkey and the gateway-error
reveal already use. A show without focus, and every other platform, are
unchanged.

The new test evaluates showMainWindow as written against a fake window and
app, and lists every focused call site so a new caller has to be a user
gesture.

Part of #19547

(cherry picked from commit 38522b7)
)

PATCH /api/artifacts/{slug}/folder resolved a `folder` path with
create_missing=True before it checked the slug, so a move for an
unknown or malformed slug answered 404/400 but left every new path
segment saved in the folder tree. The handler now reads the artifact
first, inside the same error mapping as the move, so a refused slug
returns before any folder is created.

Builds on #26869 by @rohnkum.

(cherry picked from commit 43c20f9)
#21175)

When the project tree's payload is truncated with no row at all, the
panel shows only one sentence in the tree's place. "Workspace too large
to list: limit of 10,000 items reached" read as a failure, and "items"
names nothing the user can see.

The sentence now says nothing is wrong first and counts files and
folders, keeping the limit as a parenthetical. It is translated in all
twelve catalogs, en-XA is regenerated, and the spec and the frame-32
capture assertion quote the new wording.

Closes #21025

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
(cherry picked from commit e6e1038)
* fix: answer 400 for wrong-shaped bodies on five upload routes

* test: keep route test file black-clean and facade manifests complete

---------

Co-authored-by: Jacob Largent <largjaco@dev-dsk-largjaco-kiroc-2a-dc066d04.us-west-2.amazon.com>
(cherry picked from commit d59ae0d)
A config.json value outside its allowed range was clamped by the
bounds a section builder passes to SectionReader.read with no log
line, while a wrong-typed value in the same file was logged.

SectionReader.read now logs a warning naming the class, key, stored
value and value used when the bounds change a stored value. The
warning repeats once per disk read of the config files, like the
validation warnings, so a log attached after the first load still
records it.

Closes #29827

(cherry picked from commit a5a9f6a)
#31263)

PATCH /api/knowledge/items/{id} wrote every allowed field through, so
a null title or item_type reached two NOT NULL columns and the update
failed with a 500. Treat a null there as no change, like an absent
field; a body with nothing else left still gets the existing 400.

Fixes #31257

(cherry picked from commit 781672b)
…31027)

With agent.acp_backend (or the Bot backend's Kiro CLI floor) on the
Kiro CLI backend and no kiro-cli resolvable, kirocrew doctor printed
the Dependencies row as a skip and recorded no issue, so it could end
ready while every session on that role failed to spawn.

The row now prints a failing line and records a kiro-cli issue when a
selected role is the Kiro CLI backend, on the KAS bundle route too,
where a Bot role on Kiro CLI still spawns kiro-cli. Relay KAS is not
counted: with no kiro-cli its spawn takes Crew's own engine. The role
selection is read once, before the Dependencies section, and reused
by the Connectivity rows. The doctor test helper that pins the Kiro
CLI backend pins a resolvable kiro-cli too, so those tests no longer
depend on the runner having one installed.

Closes #30075

(cherry picked from commit 4a71d19)
…slot (#26761)

When a run yields its slot and asks for it back while the cap is full,
request_resume puts a resume entry in the window. The lane snapshot
counted that entry as queued and the run's live info as waiting, so
GET /api/spawn/lanes showed one run twice ({queued: 1, waiting: 1}).

Skip resident-resume entries when counting queued, the same way the
queue-depth count already does. The run is still counted once, as
waiting.

Closes #26759

(cherry picked from commit 782c499)
The model-route reply strip renders the outcome row, which never carried
history_chars and truncated: routed_model() put them only on the call
row, and build_outcome() did not copy them, so the strip always read 0
for context sent and earlier turns truncated. Carry both counts on the
routed answer and copy them into the outcome row, and list them on the
outcome row in the decisions spec.

Closes #29243

(cherry picked from commit 162d407)
…31416)

GET /api/system reported ollama_running as true whenever the embedding
model file was present, even after the vendored llama.cpp import had
failed, which never recovers in-process. The Services tab therefore
showed the embedding runtime as running while memory was keyword-only.

Read the same runtime-load check the Memory tab's embedding status uses
(import failed on a platform that ships the libs) and report false in
that state, so both health surfaces agree.

Closes #21725

(cherry picked from commit 1c0fa40)
…31185)

On Windows Crashpad keeps every dump in Crashpad\reports\ and records
pending/completed as states in its metadata file, not as folders. The
desktop crash collector listed only pending/ and completed/, so every
Windows scan reported candidates=0: no crash was ever recorded, shown
in the crash notice, or (for a proven-foreign dump) deleted.

List reports/ beside the other two. A missing folder is already
skipped, so one list serves macOS, Linux and Windows.

A Windows install already baselined by a build that never read
reports/ would otherwise see its pre-collector dumps as new crashes,
so the activation cutoff now writes off older artifacts on every scan,
not only on the baseline run.

Closes #27763

(cherry picked from commit 571f23c)
…31318)

kirocrew cron preview ran a script against a stand-in context that
had only notify(), call_tool() and close(). A script calling
list_session_folders(), create_session_folder(), open_session(),
send_to_session(), set_session_mode() or crew_log_append() worked as a
cron but stopped the preview with AttributeError, and call_tool()
named its third parameter differently from production, so a keyword
call failed too.

The stand-in now has every public ScriptContext method with
production's parameters. Each session and crew-log call prints its
arguments, redacted, instead of reaching the gateway, and returns a
stand-in of the shape production returns. A test compares the two
classes' public methods, parameter names, kinds and defaults, so the
next gap fails a test.

Closes #25956

(cherry picked from commit 670268d)
…oom (#31419)

The Mochi and Crew Companion overlays load their pet.html from the
dashboard's origin, and Chromium keeps one persisted zoom level per
host, so a dashboard zoom other than 100% zooms the overlays too. Each
renderer reports its pet's hitbox in its own CSS pixels, but the main
process compared that box with the cursor in screen DIP. At 50% the
drawn pet never made its overlay interactive, an empty spot beside it
swallowed clicks instead, and because the zoom is persisted a relaunch
kept the pet click-through.

Map the cursor into the overlay page's pixels with its zoom factor
before the hit test, and do the same in the drag poll, the drop and the
mid-drag activation, so a drag keeps the pet under the pointer at any
zoom.

(cherry picked from commit 5e5b15e)
…31498)

json.loads reads a number too large for a float, such as 1e400, as
infinity, and int() on it raises OverflowError. The tag column
normalizer only caught TypeError and ValueError, so PATCH
/api/chat/tag-columns/{id} with that order answered 500.

An infinite order is now refused as invalid_column_payload, and the
stored column is left as it was. An order int() cannot read for any
other reason is still ignored, as it is today.

(cherry picked from commit ef7655e)
…31497)

The cron history routes redact each run's summary, error and trace
with string scanners. A history line whose text field held a number,
list or object raised there, so one such row turned the whole history
page into a 500. The three readers now skip that row, as they skip a
line that is not JSON; the rewrite paths still keep it on disk.

(cherry picked from commit 5722529)
#31491)

The store checked an artifact's source with a set lookup, so a list or
object in POST /api/artifacts raised TypeError (unhashable) and the
route answered 500. Check the type first and raise the validation
error the create route already maps to 400. Every string is checked
as before.

Fixes #31489

(cherry picked from commit b2490cb)
…#31536)

A user who turns Computer Use on and chats on a backend that will not
mount kirocrew-computer sees the toggle ON and a doctor report that says
"Enabled: yes", while the agent has no computer-use tools. On codex the
projection withholds Crew's identity-bound servers from the spec half of
the session array, so a default codex chat never gets them.

kirocrew computer doctor now asks the configured backend's projection of
the default spec, through agent_spec_mcp_refs, whether
@kirocrew-computer resolves there. When it does not, the report carries
withheld_on_backend and the text names a remedy. Codex is the one
backend whose projection admits the MCP gateway's stub for the server,
so there the line names routing kirocrew-computer in MCP Management and
restarting, or the Kiro CLI backend, and it goes away once the running
gateway routes the server (gateway_routed_names: the saved routing and
a stub in the overlay). Every other backend is pointed at the Kiro CLI
backend only: the single-binary harnesses withhold that stub as well. A
healthy report is unchanged.

agent_spec_mcp_refs takes an optional backends filter so the doctor
asks about one backend, and the projection's own withhold warnings are
kept out of the gateway log for that one question.

Refs #28436, #29557
Closes #30160

(cherry picked from commit bac144d)
…31511)

When a governing sender's webContents is destroyed, the update
governance listener only recomputes the flag on macOS if another window
is still open. It asked BrowserWindow.getAllWindows(), but dashboard
windows are BaseWindows with a WebContentsView, so with only dashboard
windows open the probe saw none and governance stayed set: the feed
lane stayed stood down after the governing window closed, even with
another dashboard window still open.

Read BaseWindow.getAllWindows() instead. On macOS 26.7.2 with the
repo's Electron 44.9.1, the closing window has already left that list
when its view's webContents fires destroyed, so the last-window
carve-out holds.

Closes #26918

(cherry picked from commit e758445)
Adding a Markdown notebook vault to Knowledge reports "Added to
Knowledge — {{n}} file(s) queued for indexing." The key had one form
and the call passed only n, so one file read "1 file(s)", German,
Spanish, French, Portuguese and Russian copied the shortcut, and a
large count showed with no digit grouping.

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

(cherry picked from commit becf26f)
@buluoray
buluoray force-pushed the crew/slack/0.12.1-s3b1 branch from 7c0166d to 1d2f066 Compare November 9, 2026 02:05
@github-actions github-actions Bot added readiness: action required A blocking check or review needs attention and removed readiness: checking Automated validation is still running labels Nov 9, 2026
@buluoray
buluoray merged commit 8e30b9d into release/0.12.0 Nov 9, 2026
38 of 70 checks passed
@buluoray
buluoray deleted the crew/slack/0.12.1-s3b1 branch November 9, 2026 05:19
@github-actions github-actions Bot removed the readiness: action required A blocking check or review needs attention label Nov 9, 2026
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.