Skip to content

fix: backport 0.12.1 wide batch 3 (15 user-facing fixes) - #32836

Merged
buluoray merged 15 commits into
release/0.12.0from
crew/slack/0.12.1-wide-3
Nov 8, 2026
Merged

buluoray merged 15 commits into
release/0.12.0from
crew/slack/0.12.1-wide-3

Conversation

@buluoray

@buluoray buluoray commented Nov 8, 2026

Copy link
Copy Markdown
Contributor

Problem / Motivation

Goal: Bring fifteen 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. Every pick applied to release/0.12.0 cleanly, so the code each one replaces is on release; the right-hand column names where.

PR Reported by What a 0.12.0 user sees today Faulty code on release
#30113 bolichen97 (#30111) On claude, manual, auto and fileMatch steering files are injected every turn, front matter included. claude steering load in context
#30105 bolichen97 (#30094) On claude, an MCP permission event carries no trusted server identity, so @server grants never apply. acp claude permission path
#30045 RohanK6 (#28951) A cloud launch job file with bad JSON does not hold off a second launch. cloud launch job read
#30163 bolichen97 An uninstalled app's in-process routes keep serving its data. apps route registry
#30022 CrysisDeu (#29811) Delete in the Semantic memory browser removes a lesson at once, with no confirm. VectorMemoryCard
#30214 RohanK6 (#30211) A non-UTF-8 tool-alias record breaks the agent spec commit for connections. connections tool-alias read
#30185 CrysisDeu (#29820) PPTX Maker's "Install engine" banner clears before the engine is on disk, so the first deck turn fails. PptxMakerPage, pptx-maker engine status
#30102 CrysisDeu (#29709) Toggling a custom MCP server creates the shared ~/.kiro/settings/mcp.json. MCP toggle writer
#30083 rubencu (#30026) Under the installed Linux service, "Show in file manager" is hidden on a GNOME desktop. dashboard file-manager probe
#30082 bolichen97 (#30052) On claude, incognito and temporary chats leave their full native transcripts on disk. claude session teardown in acp
#30112 bolichen97 (#30109) On claude, the / menu lists kiro-cli commands, and /review is sent as prose. dashboard slash-command list
#30079 bolichen97 (#30056) An auto_deny_tools @server/tool rule never denies a claude MCP call. hooks deny match
#30087 bolichen97 (#30051) A claude 529 Overloaded fails the turn instead of retrying. acp/transport_errors.py
#30221 RohanK6 (#30217) A hand-edited meta.json tags value 500s the Library search or hides the artifact. artifact meta read
#30223 RohanK6 (#30222) One non-UTF-8 byte in a cron history file breaks every history read and stops new runs being recorded. cron run-history read

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; they go in a later batch.

What changed (motivation → approach → change)

Each main squash commit was cherry-picked with -x onto release/0.12.0 @ c318c7d72a, in merge order on main, one commit each. Every pick applied cleanly, and every commit keeps its original author.

# PR main commit release commit
1 #30113 honour steering inclusion front matter in the claude load 994fde4866 66f4ec966d
2 #30105 read claude's adapter-stamped mcp server as trusted identity 8d9233d259 d599e5a8ae
3 #30045 hold off a second launch while a job file holds bad JSON a7399c56ec a07cf49994
4 #30163 stop an app's in-process routes when it is uninstalled 4f48d48e04 ff2c9d895b
5 #30022 ask before deleting a semantic memory entry 53fbe49e74 b975aa35f3
6 #30214 read a non-UTF-8 tool-alias record as claiming nothing b2111c2cab 4817e4ab3d
7 #30185 keep the install banner up until the engine prompt is on disk cff13846b7 29fcc82e50
8 #30102 keep the toggle off the shared mcp.json for store-only servers 258996b131 5ebe594efa
9 #30083 restore the file manager under the installed Linux service 1b8a19568b 160b1ecb92
10 #30082 delete claude native transcripts for disposable sessions 89192ab04d e33e6fe818
11 #30112 list claude's commands in the / menu on acp_backend claude bc42d1128f 213dcada4e
12 #30079 bind @server/tool auto_deny_tools to an mcp__ title d7337c9aea fcfda743d5
13 #30087 retry a claude 529 overloaded failure as transient e253b71ec6 de27519ea2
14 #30221 read hand-edited meta.json tags as strings fc921d4567 f52bc9b873
15 #30223 read run history as bytes so a bad byte costs one line 59e91f239b 8f2216b6e4

#30022's and #30102's internal co-author trailers are rewritten as a plain "Co-author on main" line, because the commit guard refuses the address.

#30107, the next row in merge order, conflicts in test/fixtures/acp_frames/README.md, so it waits for the docs-conflict batch.

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. #30082 deletes claude's native transcript only for an incognito or temporary session, which promises not to keep one. #30079 makes an existing auto_deny_tools @server/tool rule deny the claude MCP call it already names.

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 related set was run here. Every touched test file (24 files) gave 1,756 passed. The test files that import a touched source module (120 files) gave 6,048 passed and 7 failed.
  • The 7 failures are copilot-backend cases in test_acp_agent_env.py, test_acp_copilot_backend.py, test_acp_harness_contract.py and test_acp_launch_goldens.py. The same 7 fail on a clean release/0.12.0 @ c318c7d72a on this host.
  • Frontend: tsc -p tsconfig.app.json --noEmit passes; vitest PptxMakerPage and VectorMemoryCardCoverage, 154 passed; gen-ui-index --check and check-i18n-strings pass.
  • flake8, isort, black gate, subprocess-encoding gate, mypy src/kiro_crew and check_harness_parity.py pass. check_file_line_cap fails only on dashboard/chat_api/resume.py, which is already over the cap on release and is not touched here.

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 (#30022, #30185) 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 #30113, #30105, #30045, #30163, #30022, #30214, #30185, #30102, #30083, #30082, #30112, #30079, #30087, #30221 and #30223. Source issues: #30111, #30094, #28951, #29811, #30211, #29820, #29709, #30026, #30052, #30109, #30056, #30051, #30217, #30222.

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.

bolichen97 and others added 15 commits November 8, 2026 01:39
…ad (#30113)

The Claude Code backend's [Steering resources] load appended every
globbed steering body raw, so a document declaring inclusion manual,
auto or fileMatch reached every Claude turn and every document carried
its front-matter header into the prompt. The folder-steering
collection already withholds those modes and strips the header.

Promote that rule to folder_steering.every_session_body and run both
loads through it, so the two cannot drift. Update steering-and-hooks.md
and context-management.md row 12.

Closes #30111

(cherry picked from commit 994fde4)
#30105)

claude-agent-acp names an MCP call's server on its permission request
(_meta.claudeCode.mcpServer) and the tool in the fused toolName, but
build_permission_event never read it, so a claude MCP call carried no
(server, tool) identity and every gate tier keyed on one was off:
operator @server grants, the Reads-mode own-tool proof, the MCP
document-body exemption and the heartbeat core reads.

On a claude session the pair is now adopted on top of the tool_call
cache hit, only when the server's source is "dynamic", the server is
one Crew placed on the session array, and toolName is exactly
mcp__<server>__<tool> as the tool_call frame stamped it. Any other
frame keeps the empty identity, so the user's own ~/.claude servers
stay title-only and are never refused as foreign.

Closes #30094

(cherry picked from commit 8d9233d)
…30045)

Co-authored-by: w <w@k.local>
(cherry picked from commit a7399c5)
…0163)

POST /api/apps/<name>/uninstall returned 200 while the app's
backend.hooks routes stayed in the route registry and kept serving,
including the app's data, until the hook reconciler's next tick. The
uninstall handler runs its own backend stop and deregistration and
never called on_app_disable, the in-process half that disable reaches
through teardown_app_runtime.

teardown.stop_app_hooks_for_uninstall now runs that half for every app
before the backend stop: the disable hook is told, routes leave the
registry, on_shutdown runs from the loaded manifest when one is
retained, and job runs, the SDK entry, hook health and the loaded-hook
record are dropped. Residual execution is reported, never an abort.

Part of #26967

(cherry picked from commit 4f48d48)
The Delete button on a row of the Vector Memory card's Semantic table
removed the entry, lesson.* rows included, on the first click. It now
opens the shared confirm dialog, which names the key, says the delete
cannot be undone, and restates the action on its button. The entry is
deleted only on that press.

Fixes #29811

Co-author on main: Matt Briones.
(cherry picked from commit 53fbe49)
…ing (#30214)

load_claimed reads the Connections tool-alias record during agent spec
commit and, by its invariant 4, reads every failure as claiming nothing,
so a lost or broken record never costs a user alias. It caught OS errors
and bad JSON but not UnicodeDecodeError, so a record whose bytes are not
UTF-8 raised out of the spec commit instead.

Such a record now reads as empty, like bad JSON: it claims nothing and
strips no alias.

(cherry picked from commit b2111c2)
…s on disk (#30185)

* fix(pptx-maker): disable Start a deck until the engine is ready

The deck-start buttons were enabled regardless of engine state, so a user
could open a deck chat before the engine was installed. The first turn then
projected an engine agent whose prompt file was not on disk yet and failed.
Gate the buttons on the engine status the banner already reads, so a deck
cannot start until the engine is ready.

Closes #29820

* test(pptx-maker): add review screenshots of the Start a deck gate

Review-only evidence for the engine-readiness gate: the deck-start buttons
disabled while the engine is not ready, enabled once ready, and the error
notice shown when the engine status cannot be read. Under temp-screenshots/
(gitignored, force-added); removed from the tip after merge.

---------

Co-authored-by: kirocrew-agent <agent@kirocrew.local>
(cherry picked from commit cff1384)
…ers (#30102)

POST /api/mcp/toggle always wrote the flag into the Kiro global
~/.kiro/settings/mcp.json, adding a {"disabled": true} stub for any
server that file did not hold. A server added through Add Custom lives
only in Kiro Crew's own store, so switching it off created the shared
file the IDE owns, for a server the IDE never knew about.

When the global file has no entry for the row, the toggle now looks in
the Kiro Crew store and, when the store holds a mapping entry for it,
sets the flag there through _set_kirocrew_entry (the store's one writer, in
an existing-entry-only mode) in one read-modify-write off the event loop,
leaving the global file as it was (absent or unchanged). An enable
also lifts a store mute when the global file holds the row too, such
as a stub a per-tool disable left. A name several store keys alias to
is refused with 409, and a store that does not parse with 500
config_corrupt, before any write. The global file is read off the loop
with comments allowed and refused for comments only when the toggle
rewrites it. Servers held by the global file, or only by another
scope, keep the existing behaviour.

Fixes #29709

## Pattern harvest

Not generalizable: the toggle is the one writer that picked its target
scope without asking which scope holds the server; apply already does.

Co-author on main: Matt Briones.
(cherry picked from commit 258996b)
…rvice (#30083)

The installed Linux service is a system unit, so the gateway's own
environment never names a display server. Since #23019 that hid "Show in
file manager" and "Open with default app" on every Linux desktop running
the service, including GNOME, where xdg-open used to reach Nautilus over
the session bus.

When the gateway's own environment names no display, ask the user's
systemd manager (`systemctl --user show-environment`, 2 s timeout, answer
kept for 30 s), run with the environment `systemctl_user_env` gives
every `systemctl --user` spawn. Copy only DISPLAY, WAYLAND_DISPLAY,
XAUTHORITY, XDG_CURRENT_DESKTOP and KDE_SESSION_VERSION, each passing a
plain character check, and count a display only while its socket exists,
checked again on every use of a kept answer. The launcher gets them on
top of the same session-bus pointers. A reachable session bus alone is
not a desktop, so a headless host stays hidden and refused. The open
action asks before its stat-and-launch transaction, so no spawn lands
inside it.

Closes #30026

(cherry picked from commit 1b8a195)
…30082)

The claude CLI behind claude-agent-acp keeps its own record of every
session: projects/<project>/<sid>.jsonl, projects/<project>/<sid>/ and
file-history/<sid>/ under its config root. Crew only ever deleted the
kiro-cli replay log, so an incognito or temporary claude chat, a pruned
claude subagent run and a released claude conversation all left the
full conversation on disk, and the prune kept retrying forever.

acp/claude_transcripts.py deletes those three paths: root
CLAUDE_CONFIG_DIR else ~/.claude, the project name spelled from the
resolved cwd. A non-UUID sid, an empty cwd, an over-long project name
or any symlinked component deletes nothing and reports not done.
cleanup_session calls it on is_claude_backend off the loop,
_cleanup_session_files_sync routes PROVIDER_LABEL_CLAUDE to it with the
recorded cwd, and spawn_release reads that cwd from the trusted cleanup
record.

Closes #30052

(cherry picked from commit 89192ab)
…laude (#30112)

* fix(chat): forward claude code's own slash commands on acp_backend claude

On the Claude harness (provider "acp", acp_backend "claude") an arbitrary
leading-slash word such as /review, /init, /security-review, a
.claude/commands file or a skill was treated as prose: Crew's context was
prepended and the adapter never saw a command. The forward was granted
only by the dormant agent.provider == "claude_code" label, which the
public schema never admits.

Add ACP_BACKENDS_HARNESS_OWNED_SLASH_COMMANDS = {claude}, read through
SessionCapabilities.runs_own_slash_commands. The slash gate, the
app-grant send refusal and the Slack auto-link skip now share one
predicate, _forwards_harness_command, which asks the backend the
/todos gate resolves (live provider first, else the member-aware
selection). Kiro, KAS and every other harness keep their path.

Closes #30090

* fix(dashboard): list claude's commands in the / menu on acp_backend claude

GET /api/slash-commands offered Claude Code's command list only under
the dormant agent.provider == "claude_code" label, which the public
schema never admits, so on acp_backend claude the / menu listed
kiro-cli's /hooks, /code, /experiment, /logdump, /changelog and /tools
and left out /init, /review and /security-review.

The handler now takes the claude list where the configured backend's
runs_own_slash_commands capability holds (or the old label), adds
/todos (served natively there, still dropped where it is blocked) and
keeps the dashboard's own /agent, /goal and /prompts beside /side and
/workflow. The provider _slash_commands read is removed: nothing in the
tree ever writes it.

Closes #30109

(cherry picked from commit bc42d11)
An operator's auto_deny_tools entry in the identity spelling
(@server/tool, @server, "Running: @server/*") only bound when the call
carried a _meta.kiro server name. The claude adapter sends none, so a
claude MCP call, titled mcp__<server>__<tool>, skipped the rule and
went to approval, where a Trust or YOLO session ran it.

The mcp-auto-deny tier now reads a server-less mcp__ title through
mcp_title_to_ref, the last-__ split governance already uses, and
matches only the operator's auto_deny_tools globs against it. The
reading is deny-only: no grant tier and no deny target sees it.

Closes #30056

(cherry picked from commit d7337c9)
claude-agent-acp forwards Anthropic's overload answer as "Internal
error: API Error: 529 {...overloaded_error...}" and, once Claude Code's
own retries are spent, "API Error: Repeated 529 Overloaded errors".
Neither matched the 5xx family: the status pattern only read a code
after "HTTP" or "status", and no named token covered the overload. Both
frames were classified terminal, so the retry ladder a kiro-cli 5xx gets
never ran on the claude harness.

Read a status after the seam's "API Error:" prefix and add an
"overloaded error" named token. The patterns are shared by the
classifier, classify_provider_error and the formatter, so all three
agree. Size, auth and usage-limit verdicts stay ahead of the 5xx family.

Closes #30051

(cherry picked from commit e253b71)
The store only writes tags as a list of strings, but a hand-edited
meta.json can hold anything. A null or numeric tag made the Library
content search raise in ' '.join over the tags, so the search answered
500; a bare-string tags split into one tag per character; and a tags
that is a number dropped the whole record from every listing.

decode_meta now reads tags tolerantly, like its other fields: a bare
string is one tag, a number its text, and null, bool, object and list
entries are dropped.

(cherry picked from commit fc921d4)
…0223)

The cron history store decoded each whole JSONL file as UTF-8. One byte
that is not UTF-8 (a torn or foreign write) raised UnicodeDecodeError,
which is not an OSError, so it escaped the job history, the run history
page and the run detail read, and the trim after every append, so no
later run was recorded. Files are now read as bytes and split before
each line goes through json_line.parse_json_object_line, which decodes
strictly and skips a line it cannot use; the store's own line parser
is gone. The trim and the delete rewrite keep each line byte for byte,
opened with O_BINARY and joined with the os.linesep ending append_line
writes.

(cherry picked from commit 59e91f2)
@buluoray
buluoray requested review from a team and KiroCrewCaptain as code owners November 8, 2026 02:00
@buluoray
buluoray requested review from Zedmor and removed request for a team November 8, 2026 02:00
@github-actions github-actions Bot added the readiness: action required A blocking check or review needs attention label Nov 8, 2026
@github-actions

github-actions Bot commented Nov 8, 2026

Copy link
Copy Markdown
Contributor

Design Review (Opus 5.5) — ✅ PASS

Design-level review of 8f2216b6e420206d9a69ad0e983fcd95c2221288 — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS does not.

Design-Verdict: PASS

This backports fifteen fixes that already merged on main, cherry-picked unchanged and one commit each. Nothing blocks: no fix adds a new design decision, and nothing is taken away beyond what the description says.

Notes

[DESIGN-REVIEWED] 8f2216b

Description read

[DESCRIPTION-READ] 17e44cc39624f166b10253921b6e9774a033ef063fdfabdf557035534fb086cf

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

First Principles Review (Opus 5.5) — ✅ PASS

Premise-level review of 8f2216b6e420206d9a69ad0e983fcd95c2221288 — 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 backports fifteen fixes already merged on main, and each is tied to a reported defect and a regression test. The only thing that rides along is three PR screenshots.

Notes

What this change ships

Inventory (10 items) — 9 justified

Intent: give 0.12.0 users fifteen fixes that already landed on main, without them waiting for 0.12.x+1 — FIX. Capped at 10 of 16 items.

  1. On claude, manual, auto and fileMatch steering files follow their inclusion front matter instead of loading every turn (fix(context): honour steering inclusion front matter in the claude load #30113) — justified
  2. On claude, @server MCP grants and auto_deny_tools rules now apply (fix(acp): read claude's adapter-stamped mcp server as trusted identity #30105, fix(hooks): bind @server/tool auto_deny_tools to an mcp__ title #30079) — justified
  3. On claude, a 529 Overloaded error retries instead of failing the turn (fix(acp): retry a claude 529 overloaded failure as transient #30087) — justified
  4. On claude, incognito and temporary chats and released subagents delete their native transcripts (fix(acp): delete claude native transcripts for disposable sessions #30082) — justified
  5. On claude, the / menu lists claude's own commands and forwards them raw (fix(dashboard): list claude's commands in the / menu on acp_backend claude #30112) — justified
  6. Deleting a semantic memory entry now asks for confirmation first (fix(memory): ask before deleting a semantic memory entry #30022) — justified
  7. The PPTX "Install engine" banner stays up until the engine is on disk (fix(pptx-maker): keep the install banner up until the engine prompt is on disk #30185) — justified
  8. Toggling a custom MCP server no longer creates the shared mcp.json (fix(mcp): keep the toggle off the shared mcp.json for store-only servers #30102) — justified
  9. A bad byte or bad JSON in a cron history, job, alias or meta file no longer breaks the feature, and "Show in file manager" works under the Linux service again (fix(cron): read run history as bytes so a bad byte costs one line #30223, fix(cloud): hold off a second launch while a job file holds bad JSON #30045, fix(connections): read a non-UTF-8 tool-alias record as claiming nothing #30214, fix(artifacts): read hand-edited meta.json tags as strings #30221, fix(dashboard): restore the file manager under the installed Linux service #30083) — justified
  10. Three PR screenshots are committed to the release branch — rides along

[FIRST-PRINCIPLES-REVIEWED] 8f2216b

Description read

[DESCRIPTION-READ] 0d5c4d68feee999cd03753658bf38fc3422a9fbb152e040ddd5f91d021551d85

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

GPT 6.1 Review — ✅ no blocking findings

GPT 6.1 completed its review of 8f2216b6e420206d9a69ad0e983fcd95c2221288 and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] 8f2216b

False positive or not applicable? A repository writer can comment:
/ai-review override gpt 8f2216b6e420206d9a69ad0e983fcd95c2221288 false-positive: <reason, 20 to 500 characters>

@buluoray
buluoray merged commit 4061cb7 into release/0.12.0 Nov 8, 2026
20 of 37 checks passed
@buluoray
buluoray deleted the crew/slack/0.12.1-wide-3 branch November 8, 2026 03:40
@github-actions github-actions Bot removed the readiness: action required A blocking check or review needs attention label Nov 8, 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.

6 participants