Repository navigation
fix: backport 0.12.1 fast batch 1 (21 fixes) - #32866
Conversation
44e3691 to
fe8bff3
Compare
Sensitive Change ReviewSensitive files changed:
How to clear this check:
Only the author's comment counts this way. A reviewer can instead |
|
Taken over from the previous worker to clear the conflict. Rebased all 22 picks onto |
GPT 6.1 Review — ✅ no blocking findings (all downgraded on adjudication)GPT 6.1 flagged blocking issues on This comment is updated in place on each push. Review detailsBLOCKING -- src/kiro_crew/security/rm_floor.py:235 -- Only scan a piped heredoc when its body reaches the shell
Adjudication (Opus 5.5) — is blocking on each finding proportionate?F1 is LOW harm. The failure is a false refusal, and it needs every one of these at once: a piped heredoc on False positive or not applicable? A repository writer can comment: |
First Principles Review (Opus 5.5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS This backports fixes already merged on Notes
What this change shipsInventory (10 items, 21 fixes) — 9 justifiedIntent: give 0.12.1 users 21 fixes already merged on
[FIRST-PRINCIPLES-REVIEWED] c362e6b Description read[DESCRIPTION-READ] a69593d80fa197ba00a6314ad36f2a36a6783666b9f366293c49ed11b81013f8 That sha256 names the PR description this verdict read. Recompute it with |
Security Scope Review -- ✅ PASSWhich legitimate operations does Verdict basis: adjudicated, zero confirmed regressions. Security scope review -- adjudicated candidates
No candidate flipped from allowed to refused. Scope-Verdict: PASS Reason: The tightening sits in the rm floor and the argv wrapper table, both inside the shell deny composite, so every boundary was submitted as a shell candidate and no confirmed regression was handed back. One change is outside the composite and nothing classifies it: the kiro-cli allowed-command grant parser now rejects non-ASCII escapes. It is named under UNADJUDICATED. A second untested path is the new fail-closed walk budget. TIGHTENING: The rm floor ( ADJUDICATED: 36 candidates, 0 confirmed newly-refused PLATFORM COVERAGE: posix (30 rows). These cover what the diff assumes about bash: quoted, unquoted and UNADJUDICATED:
[SCOPE-REVIEWED] c362e6b |
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: |
fe8bff3 to
c1f4ffd
Compare
|
Dropped #31016 from this batch (new head |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks: the four take-aways I traced (install no longer registers, the slot-push flush no longer raises, the Claude prompt keeps the Kiro name, the mate-conductor skill flag) have every in-tree reader updated. Notes
[DESIGN-REVIEWED] c362e6b Description read[DESCRIPTION-READ] 9783ad01056f74be703b9cc4d7b55122c6bda846b4f8fd4bf44d23363134f085 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 |
UX Review (Opus 5.5) — ✅ PASS (a BLOCK with no Trigger and no Outcome does not block)UX-level review of UX-Verdict: PASS cannot evaluate: missing screenshots of the App Library's offline state. The batch adds new dimmed controls and an offline notice to the Library page and its app-tile menu, and none of the seven supplied screenshots shows them. Evidence gaps
Notes
[UX-REVIEWED] c362e6b Description read[DESCRIPTION-READ] 0b6cd7a72b28ddc93368be2d6935f26de9acb549aada95460fdaf1c69b3a5284 That sha256 names everything this verdict read: the PR description, plus the 10 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 |
c1f4ffd to
a81f4f1
Compare
…30074) On the Claude Code backend, session start delivers the agent config's file:// steering resources as prompt text, because claude-agent-acp does not read an agent's resources itself. A compaction drops that text, and post-compaction re-injection never sent it back, so the session lost its global steering silently. kiro-cli keeps its natively loaded resources. Session start and post_compaction_parts now share one gate, inclusion.steering_resources_part (Claude Code by positive identity, default agent, project group, not a member chat). The first turn after a compaction re-sends the text as a REINJECTED block with its payload scrubbed of structural markers; minimal-context turns, custom agents, member chats and the kiro and KAS paths get nothing. Closes #30054 (cherry picked from commit 2a1c4b4) Original author: Bolin Chen
A dashboard /clear on a claude session went to claude-agent-acp as session/prompt text. The adapter sends no clear status frame and has no _kiro.dev/commands/execute, so Crew never saw a clear it may have run: the todo pill, the pending reset_conversation discard and the next turn's replay all stayed. Add claude to ACP_BACKENDS_LOCAL_CLEAR so the slash gate runs the dashboard's own reset, as on the other harnesses without an observable /clear. claude runs one adapter per session, so the discard drops that adapter's conversation either way. Closes #30058 (cherry picked from commit 691fb42) Original author: Bolin Chen
…led (#30628) A fresh registry, streamed or local-path install writes its record disabled, yet _app_may_run_after_install answered True for every fresh install, so the route registered the disabled app's agents, skills, MCP servers and crons and started its backend. GET /api/apps showed the app off while its agent specs sat in the agents directory until the next enable or disable. This is the reinstall path for the former built-ins. The gate now reads the live enabled flag for fresh installs and updates alike, so resources arrive at the guarded enable transition. Registry default installs, which enable right after installing, read the new _app_may_enable_after_install (record present, no consent pending) instead. The file-only `kirocrew app install` follows the same rule. Part of #26967 (cherry picked from commit 2c733d1) Original author: Bolin Chen
…sh (#30435) * perf(crew-mode): keep the Board's worker list read and push its refresh GET /api/chat/slots/{root}/crew-workers walked the session tree, read every closed worker's saved title and every crew member's ledger store on each call, and the Board polled it every 5 s (15 s for the sidebar row and breadcrumb). At 3000 descendants one read took about 650 ms. The handler now keeps, per root, the closed list (walked again only when the session tree's nodes object or the live keys change), each closed worker's title (dropped once the key is seen live) and each ledger store's facts (read again only when that store's change mark moves). The work-ledger store gives a store a fresh mark at the end of every conductor or item lock hold, which every write, rebuild and purge takes. Kept answers older than 300 s are read again, which bounds what a write from another process can leave stale. Live rows are read fresh on every call. The website re-reads a root's list when the slot store says its crew moved: a crew row's state, title or parent (the lineage push), or the root's crew_work (its ledger push). Bursts coalesce into one re-read; the poll drops to 60 s. * perf(crew-mode): push a sub-conductor's ledger report to the Board A grandchild's report lands on its sub-conductor's ledger, which no slot row carried, so the root's worker list waited for the 60 s fallback. The slot frame now follows each sub-conductor's board and puts its crew_work on its row, and the push key includes it. --------- Co-authored-by: Raymond Chen <chenmingwei23@users.noreply.github.com> (cherry picked from commit 1f44d3a) Original author: Raymond Chen
The cron store read acked_items as stored, while acked_item_ids beside it already loads a non-list as empty. A null from an imported archive or a hand edit, or a number, made the next ack of that job answer 500, and a string or object was split into characters or keys. Load a non-list as no items; a list loads as before. (cherry picked from commit f8be84a) Original author: Rohan Khanderia
…30974) skills.select sent one Choice of up to 100 skills plus "none". Laya's local server refuses a Choice of more than 100 options, so with 100 or more eligible skills every call was a provider error and selection fell back to trigger matching. Cap the menu at 99 skills so the Choice is at most 100 options. Strands Decider 2B is not affected: its shipped checkpoint reads options through a pointer head, which has no slot cap. (cherry picked from commit 9a9f5b6) Original author: Bolin Chen
A rename typed while the gateway was down was accepted and silently lost, and a late reply could overwrite a newer title. Split from a larger sweep. Co-authored-by: m <m@e> (cherry picked from commit 573a192) Original author: Robert Noack
Tapping an option button the agent offered in a chat-app thread could run the option's text as a command instead of answering the agent. A tap now always sends the option to the agent as the answer, as the other chat channels already do. Typed messages are unchanged. (cherry picked from commit 5cccf3c) Original author: Bolin Chen
On the KAS backend, one spelling of a toolsSettings.shell.allowedCommands entry was read as granting a command, where kiro-cli rejects the same entry and grants nothing. Crew now reads that spelling the way kiro-cli does: the entry grants nothing and the session-start notice names it, so Crew never answers a command without a prompt that kiro-cli would prompt for. The security spec describes the reading. Closes #28518 (cherry picked from commit 7137647) Original author: Bolin Chen
…te (#30730) A slot create that had already committed answered 500 when the slots broadcast after it failed on the coalescing window's leading edge, while inside the window the same failure was only logged and the create answered 200. The caller's answer was decided by a timer. suspend_slots_push now logs a failing exit flush instead of raising it, so every block's caller is answered as when the trailing timer carries the failure, and the flush cannot hide the body's own exception. A committed resume or import publishes its final push through such a block, so neither turns into a 500 after it landed. A direct push_slots_update() still raises, so callers that audit and retry a failed publication keep that signal. Closes #8745 (cherry picked from commit 31aa5b1) Original author: Bolin Chen
#31382) DELETE /api/terminal/sessions/{id} answered 403 "Terminal panel disabled" once dashboard.terminal.enabled was false, so a shell opened before the setting changed kept running and its owner could not stop it from the dashboard. The delete route drops only that feature check; sign-in, owner, session-id and still-opening checks stay, and creation, the WebSocket, completion and redaction stay gated. Builds on #21415 by ktreharrison. (cherry picked from commit e934e28) Original author: Bolin Chen
) In the macOS desktop app a mouse's raw Back and Forward side buttons reached the dashboard only as Chromium's own history pop. The draft guard catches that pop only through NavigationBackGuard's best-effort trap, which stays out of the stack while a Forward branch exists and is not re-armed after a reload until it recalibrates. In those states one press of the Back button left an open Bot editor and threw its unsaved edits away without asking, while Cmd+Left and the header arrow asked first. useNativeHistoryStep listens for mouseup with button 3 or 4 in the macOS desktop app, beside the swipe relay. When the dashboard has a route entry that way, it cancels the press so Chromium does not step, and takes the same guarded step the swipe relay takes, which asks first. A press with no route entry that way keeps Chromium's own step: after a gateway reconnect the window's history still holds the routes of the dashboard's earlier load, and only that step reaches them. It stays inert on a narrow viewport, like the chords and the swipe. Part of #8258 (cherry picked from commit 6a38a1e) Original author: Bolin Chen
The Library page's app management controls fired their gateway writes even while the dashboard was disconnected, so a click looked like it started and only failed later. Each one now uses the refusal the sidebar surfaces already use: dimmed, an offline reason in its own tooltip, and no request. Part of #10463 (cherry picked from commit 46b858c) Original author: Bolin Chen
a81f4f1 to
c362e6b
Compare
…ck (#33057) The capability store's weakref.finalize callback took the store's non-reentrant lock. The cyclic garbage collector can run that callback on a thread that already holds the lock (inside issue, or redeem's locked liveness check), and the thread then waits on itself forever, hanging the event loop or the subprocess executor. The finalizer now only queues the pickup's digest; the next issue, redeem or verify drops every queued pickup and its capability under the lock before it reads, so a retired pickup is never served. Refs #31016, #32866
…ck (#33057) The capability store's weakref.finalize callback took the store's non-reentrant lock. The cyclic garbage collector can run that callback on a thread that already holds the lock (inside issue, or redeem's locked liveness check), and the thread then waits on itself forever, hanging the event loop or the subprocess executor. The finalizer now only queues the pickup's digest; the next issue, redeem or verify drops every queued pickup and its capability under the lock before it reads, so a retired pickup is never served. Refs #31016, #32866 (cherry picked from commit 1d9a632) Original author: Ray Xu (#33057).
Problem / Motivation
Goal: Bring 21 fixes merged on
mainafterv0.12.0ontorelease/0.12.0for 0.12.1, one commit each, in a PR of its own.Why it matters
Each fix changes behaviour a 0.12.0 user can hit today: claude-harness hooks, steering and
/clear, app install and skill copies, cron ack lists, the Board's worker list, offline rename and library writes, heredoc commands the rm floor misread, and a terminal stop the disabled panel refused. They come from the value-2 rows of the 0.12.1 wide-bar list, second half by rank, freeze 2026-11-08 16:00Z.Not a goal
main; nothing is changed only here.CHANGELOG.mdchange.What changed (motivation → approach → change)
Each main squash commit was cherry-picked onto
release/0.12.0@aa7dfb4aae, in merge order onmain, one commit each, with the(cherry picked from commit …)trailer. The repository's leak guard requires the public identity, so every commit is authored as Ray Xu and names its original author on anOriginal author:line.eb966e8bf14c24370dc47970989c1f8708bbcb98556ad1710ebe257e5b284fbcb71f8f2ea62dcc092a1c4b4020f8f95781c7691fb422cf938674dee39b34938e5818cf34fcc32c733d14a6549f2bf62b1f44d3a942095f1e2d98f8be84a1e38b50d8097b9a9f5b6424054a596b1c573a1928c82e91cf0a0a5cccf3c421343fbb42db7137647c3b874f78ad0d31aa5b11ee0f518207334f5e21977c0d26e9e50040e8300affac162631d0b4943da199ce93577267e934e280fa1a427a63956a38a1eb88f015031a0646b858ce7944e369128cEvery production-code hunk applied cleanly except one comment. The hand-resolved hunks:
agent-host-contract.md,subagent.md,agent-spec-fields.md; fix(chat): answer /clear locally on the claude harness #30080agent-host-contract.md: main's added sentences merged word by word into release's newer text.test/fixtures/acp_frames/README.md: main'seffort-clear-live.jsonlsentence added after release'szero-tools-refused-live.jsonlone.docs/decisions/README.md: both index rows kept.docs/architecture/context-management.md: release's steering row kept (front matter stripped, withheld documents) and main's inclusion gate and compaction re-send added.dashboard/chat_turn/slash_commands.py: comment only. Main's claude wording plus release's kiro-cli-only lines; the code line is release's, unchanged.crew-mode.md; fix(cron): load a non-list stored ack list as no items #31031, fix(dashboard): a failing slots flush no longer 500s a committed create #30730, fix(terminal): let an owner stop a session while the panel is disabled #31382learn-cron-dashboard.md; fix(decisions): keep the skills menu within laya's 100-option limit #30974decisions.md; fix: treat a chat option tap as the answer, never a command #31008slack-gateway.md; fix: tighten an allowed-command entry check #30967agent-spec-fields.md; fix(sidebar): gate session rename while the gateway is offline #10920, fix(desktop): ask before raw mouse back or forward drops a draft #31539docs/feature-map/README.md: main's lines added beside release's. For fix(cron): load a non-list stored ack list as no items #31031 and fix(dashboard): a failing slots flush no longer 500s a committed create #30730 release's text has noInterval on create/slots_diff.pybullets, so only main's new lines go in.test/test_snapshot_invariants.py: main's docstring, which describes the new connection-level check.test/test_security_gate_liveness.py: line budget raised by this fix's growth as measured on release (+235) on top of release's current value, 34,701 → 34,936.apps/routes_registry.py: release's fix(apps): carry a registry update's crons into the running scheduler #30584 readswas_installedafter the install (to resync an update's crons), and fix(apps): register an installed app's resources only once it is enabled #30628 deletes that variable along with thefresh_installargument. Both routes keep release'swas_installedread and itsget_appimport, and take fix(apps): register an installed app's resources only once it is enabled #30628's one-argument_app_may_run_after_installcall, asmaindoes. Without this the registry install route raisedNameError(caught bytest_apps_install_registers_nothing_disabled.py).test/test_acp_bindings.py: the__init__signature hash is the one measured with both release's and this fix's parameters,6207f62346e1.Left out of this batch:
agent_ipc.py'sweakref.finalizecallback_retiretakes the store's non-reentrantthreading.Lock, so a GC that runs it on a thread already holding that lock (insideissueorredeem) deadlocks the event loop or executor.mainhas the same code.expires_at_epoch, from the pending-confirmation expiry release does not have.Backwards compatibility
Breaking: #30628 an installed but disabled app no longer has its agents, skills, MCP servers or crons registered until it is enabled. #30967 an
allowedCommandsentry in the refused spelling no longer grants its command on KAS or allows it underdenyByDefault. Writer sweep: each source PR's own sweep onmain(#30628 @0b18d86432); not re-run on release.Reader: src/kiro_crew/apps/default_installs.py:_install_transaction -- release -- default installs still enable and register (#30628's
test_default_app_installs.py)Reader: src/kiro_crew/cli_commands_app.py:_handle_app -- release --
app enableregisters (#30628'stest_app_cli.py)Reader: src/kiro_crew/skill_runtime/catalog.py:_iter_visible -- chat|subagent|crew page -- an adopted app's packaged copy is listed once (#30609's tests)
Reader: src/kiro_crew/decisions/points/skills_select.py:candidates_from_loader -- chat -- the menu is cut at
MAX_CANDIDATESafter ranking (#30974's tests)Reader: the KAS grant list and
denyByDefaultallowlist (src/kiro_crew/acp) -- chat|cron|subagent -- the refused entry allows nothing, like any unreadable entry (#30967's tests)Tests
Proof: not provable -- other: a backport of main commits whose proofs ran on their own PRs
release/*, so the related set ran here, one pytest process at a time with no xdist: every test file the batch changed, plustest/test_<module>*.pyfor each changed source module (1,139 files, rebased ontorelease/0.12.0@989d318f81). 13 failed, and each also fails on a cleanrelease/0.12.0@989d318f81on this host: copilot-backend cases intest_acp_agent_env.py,test_acp_copilot_backend.py,test_acp_harness_contract.py,test_acp_launch_goldens.py, plustest_acp_tool_gate.py::test_every_enforced_harness_reaches_the_spawn_preflight,test_gateway_secret_senders_prove_peer.py::…::test_list_goes_through_the_proof,test_harness_parity.py::test_is_kiro_cli_is_positive,test_auto_improvement_members.py(1) andtest_folder_steering_provider_dedup.py(2).tsc -p tsconfig.app.json --noEmitpasses; vitest on the changed test files plus the tests named after each changed source file, 272 files, all passed.release/0.12.0@dc384c6114(headc362e6b76b): the batch's changed test files, the new release commits' test files, the importers of source files both sides touch and everytest/test_apps*.py(182 files) reran sequentially. The only failures (test_acp_harness_contract.py::…[copilot],test_apps_registry_identity_admission.py×2) also fail on a clean release checkout here.tscpasses; vitest on thesrc/i18n,App*and changed test files (174 files) passes.c1f4ffd43b), the test files that import any module fix(acp): let kirocrew-core reach the gateway under an enforced harness #31016 touched (188 files) reran sequentially and all passed, as didtest_security_gate_liveness.py,test_dashboard_server_composition_contract.py,test_pid_reader_ratchet.pyand the docs-lint tests.check_file_line_capfails only ondashboard/chat_api/resume.py, which is already 2,008 lines on release.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 (#10920, #30435, #31539, #31842) 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 #30134, #30107, #29926, #30072, #30074, #30080, #30609, #30628, #30435, #31031, #30974, #10920, #31008, #30967, #30730, #30688, #30877, #30600, #31382, #31539 and #31842.
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.