Repository navigation
fix: backport 0.12.1 batch 53 (9 user-reported fixes) - #32791
Conversation
… tool_search (#30430) The kirocrew-team-lead spec mapped only the team-lead, goal-conductor and dashboard skills. It is a custom template, so its skill search scope is only what is mapped: widgets, living-ui and artifacts could not be found, and WIDGET_BLOCK_AGENTS left it without the per-surface widget pointer. Map the three skills the conductor maps and add the team lead to WIDGET_BLOCK_AGENTS. The spec mounted tool_search but did not auto-approve it, while its dispatch verbs are deferred under MCP Tool Search. Every load that reaches session_create prompted first in Normal permission mode. Grant tool_search beside the dispatch verbs in _mount_team_lead_dispatch, only when the inherited tools mount it; the governance ceiling still filters it in write_governed_spec. Only kirocrew-team-lead.json changes: three skill resources and one allowedTools entry. The spec-byte goldens for that file are re-pinned. Co-authored-by: Raymond Chen <chenmingwei23@users.noreply.github.com> (cherry picked from commit 4194602) Original author: Raymond Chen (#30430). test/test_agent_refactor_spec_bytes.py hand-resolved: the kirocrew-team-lead.json golden digests are re-pinned to the bytes this tree writes (release pins differ from main's in unrelated specs); the team-lead spec diff against release is only tool_search in allowedTools and the widgets, living-ui and artifacts skills.
) A crewmate's worker session reported back after every top-level turn end, including a turn that ended through nothing_to_do with no reply text. note_turn_end recorded a report whenever reply_text was not None, so an empty string still woke the crewmate's thread with a hidden "It ended with no reply text." turn. A worker patrolling with monitor_start ends most cycles that way, so the crewmate spent a full turn, with its whole context, on each quiet cycle, and those turns queued ahead of the person's own messages. The turn's outcome now carries quiet_end: the turn's _terminal_directive_applied, the applier's verdict that kirocrew-core's nothing_to_do ended the turn, with an empty _turn_full_reply_text. The verdict holds for a backend that names no tool on its frame. The text check reads the turn's whole reply, not the trailing segment, because a tool call flushes the text written before it out of that segment; it is the same quiet-end test the Slack mirror applies. note_turn_end records no report and wakes nobody on a quiet end with an empty reply, and a report already pending for that worker stays as it was. A worker that writes and then calls nothing_to_do still reports, an empty reply without nothing_to_do still reports, and the crewmate thread's own turn end still delivers what is pending. (cherry picked from commit 1b4b566) Original author: Wen Lin (#29847). docs/system-specs/modules/crew-mode.md hand-resolved: main's two additions applied to release's crewmate wording.
The app build, the enable/disable/uninstall lifecycle scripts and the office-slides soffice conversion each spawn their child in its own session and killed the tree only when their wait timed out. A cancel (the gateway shuts down or restarts, including during a first-boot default install) re-raised with the child still running: third-party app code, or LibreOffice parsing an untrusted deck, with no deadline left. Add the CancelledError arm papyrus uses: kill and reap the tree through the shielded platform_compat.kill_and_reap, then let the cancel propagate. The install script's own arm landed separately in #25535. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com> (cherry picked from commit 61d5dff)
…31642) Every dashboard tab replays the Subagents panel's history on connect and on each reconnect. That read ran on the event loop's default executor, so a few tabs could hold every worker and leave /api/health, chat switching and every other asyncio.to_thread caller waiting. The replay now runs on a two-worker pool of its own, in a copy of the request's context so its audited denials keep their SEL provenance. The shared discovery pool is not used because chat turns wait on it. (cherry picked from commit 9eaff06) Original author: Bolin Chen (#31642). docs/architecture/resource-protection.md hand-resolved: added only this PR's paragraph; the neighbouring discovery_executor paragraph belongs to a main change not on release.
UX Review (Opus 5.5) — ⏭️ skippedRevision |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks: these are byte-identical picks of fixes already on Notes
[DESIGN-REVIEWED] 55978be Description read[DESCRIPTION-READ] e530bb0b41376c7018bffaddbb69e3c9737722333f6a06dc8466bda320d0adb8 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 |
First Principles Review (Opus 5.5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Nine already-merged main fixes, each with a reported issue and a test. No item is new surface beyond its source fix. Notes
What this change shipsInventory (9 items) — 8 justifiedIntent: bring nine user-reported fixes from main to 0.12.x users in 0.12.1. FIX.
[FIRST-PRINCIPLES-REVIEWED] 55978be Description read[DESCRIPTION-READ] 5d038077f7bd56f530bc46d94ae7367eab522e0cfa143724df9db072d71e33b3 That sha256 names the PR description this verdict read. Recompute it with |
Opus 5.5 Review — ✅ no blocking findingsReviewed Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
GPT 6.1 Review — ✅ no blocking findingsGPT 6.1 completed its review of This comment is updated in place on each push. Review detailsNo findings. False positive or not applicable? A repository writer can comment: |
In a self-chat on an account with other devices, Messages can write extra is_from_me=0 copies of a sent message hours or days later. Each copy has a new guid and ROWID but the sent row's date. The echo ledger matches for 30 s after the send resolves and is consumed on its first match, so every late copy reached the agent as a new user message, and the answer to it got late copies too. Recognise a copy by when the agent sent it instead of when it arrives. Each possibly delivered send also goes into a bounded, never-consumed history of (body key, wall-clock send time). When no ledger record matches, a row with that body whose bridge created_at is within ECHO_TTL_S of the send is the echo. The ledger, the gate order in transport.receive, the TTL and the window size are unchanged, so the accepted false positive stays the exact text dated within 30 s of the send. Nothing old is dropped by age, which keeps the cursor's restart replay working. The body key the ledger and the history retain is a SHA-256 digest of the normalized handle and body, so each entry has one size however long the message is. The user guide and the messaging spec describe the date-keyed match. (cherry picked from commit e9e94fd)
…child (#32234) doc_parser regex-scanned every stream, font programs included, and returned their ( ... ) byte runs as text, while hex-string glyph ids under subset / CID fonts were never decoded. Route PDFs through pdf_extract's memory-bounded pdfplumber child (already used by file-grep and knowledge ingest) so only page text comes back and a failed parse yields "". Fixes #32141 (cherry picked from commit 4f63b4e) Original author: Zezhen Xu (#32234).
A server already in the agent spec kept its old command forever when ~/.kiro/settings/mcp.json switched it to another program: the rebuild merged onto its own previous output and never reconciled `command`, and the stored absolute path resolved as-is on every pass. The reconcile now moves `command`, `args` and `env` as one unit when the source names a different program. The replaced unit is the last resolution candidate, so a new command that does not resolve keeps the old launch with a warning instead of dropping the server. An entry with a provenance record, and a command the ~/.kiro/crew/mcp.json store declares, keep their program as before. Closes #29678 (cherry picked from commit 0693a32)
A dashboard Restart, an update apply or the Slack restart command cuts every running turn with close_all() before it execs or exits. The next boot showed those chats as interrupted, the same as after a crash, so the owner had to press Resume on each one. Each planned restart (_restart_gateway, the gateway's update restart and the Slack restart command) now writes the names of the slots that carry a turn-in-flight marker to resume-after-restart.json in the data home, just before close_all(). Once the next dashboard boot is ready, it reads and deletes that file, then resumes only those chats through the Resume-all run (start_resume_all with only=), so its concurrency bound and refusals (pending approval, queued message, running sub-agents) still apply. The file is deleted before the attempt, so a boot tries once and never loops. A restart with no running chat removes an earlier file, and a file older than 30 minutes resumes nothing, so a record from a restart that never ran is not replayed. A crash writes no file and stays manual. (cherry picked from commit cc5398d)
35a0dd5 to
55978be
Compare
|
Goal edited at the maintainer's request: ten fixes -> nine. #31408 is removed from this PR (head 35a0dd5 -> 55978be) because GPT 6.1's security-class finding on |
Problem / Motivation
Goal: Bring nine user-reported fixes that merged on
mainafterv0.12.0ontorelease/0.12.0, one commit each, in a PR of its own.Why it matters
Each fix closes a problem a user reported. Each faulty line was checked on
release/0.12.0@7ea28ed960before the pick; the right-hand column names it.connections/warm.py:995tool_searchasks for approval.agent_materialization/service_agents.py:679nothing_to_doturn still wakes its Bot with an empty report.dashboard/mate_thread_wakes.py:529apps/lifecycle_scripts.py:140dashboard/ws.py:1542imessage/client.py:418doc_parser.py:120agent_materialization/mcp_sources.py:263dashboard/handlers/updates_apply.py:305Not a goal
main; nothing is changed only here.CHANGELOG.mdchange./Applications). It was in this PR's first head and is removed: GPT 6.1's security-class finding oncomputer_use/launch_macos.py:354, upheld by adjudication, holds. The stock-parent exemption checks only the bundle's own directories, and an admin user (the macOS default) can rename entries in/Applications, so an agent could swap the bundle between the check and the launch and run its own code outside the sandbox.mainhas the same code.What changed (motivation → approach → change)
Each main squash commit was cherry-picked with
-xontorelease/0.12.0@7ea28ed960, in merge order onmain, one commit each.2df95d317d8a26e87a954194602e66ced92862ad1b4b566257bbdaf2b6ca61d5dff13be526c6b0539eaff06a395b55ab289fe9e94fddc4098f7c569d4f63b4e4957f5d50127b0693a327309d63273a32cc5398d76b55978bed2aHand-resolved hunks, docs and tests only:
test/test_agent_refactor_spec_bytes.py(8 golden digests)kirocrew-team-lead.jsonin each of the 8 scenarios to the bytes this tree writes. The team-lead spec's diff against release is exactlytool_searchinallowedToolsplus thewidgets,living-uiandartifactsskills; the other specs and the audit events are unchanged. The resulting pins equal main's.docs/system-specs/modules/crew-mode.md(2 hunks)docs/architecture/resource-protection.mddiscovery_executorparagraph belongs to a main change not on release.Every other file has added and removed lines identical to its main commit.
The three hand-resolved commits, #32234 (co-author trailer with an internal address, non-public author email) and #32518 (non-public author email) are authored with the public identity and name their original author in the message; the internal trailer line is removed. The other four keep their original author.
builder-security checked: #31408, the one pick that touched a sandbox boundary, is not in this PR. #24862 kills a cancelled child, #30065 drops a projection from a sandboxed spawn, #32518 adds a resume record on restart; no bind, listener, route, auth mechanism or storage grant changes.
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
release/*, so the related set was run here: every touched test file, every test file that imports a touched module, and the tests that read the touched docs, 189 files,-n 4 --dist loadgroup, run on the first head that still carried fix: launch root-owned apps under a stock /Applications #31408. 6504 passed, 12 skipped, 2 failed. Both are intest_apps_registry_identity_admission.py(test_install_script_rewriting_manifest_is_refused,test_provenance_signer_uses_post_script_manifest) and fail the same way on a cleanrelease/0.12.0@7ea28ed960on this host (the sandbox refuses the test's missingpolicy_cachedirectory).test_agent_refactor_spec_bytes.py: 8 of 8 pass with the re-pinned digests.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
N/A - no frontend file changes.
Related Issues
Backport to the 0.12 line for 0.12.1 of #30065, #30430, #29847, #24862, #31642, #31455, #32234, #32455 and #32518. Source issues: #19538, #30428, #29662, #24861, #24923, #26059, #32141, #29678.
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
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)Contribution License Agreement
Contributions are accepted under the project's Apache-2.0 license; see the Licensing section of CONTRIBUTING.md.