Repository navigation
fix: backport 0.12.1 batch 54 (10 user-reported fixes) - #32801
Conversation
…30755) The Chromium download opened its URL without an SSL context, so the CAs it trusted depended on whether the startup prelude had patched ssl. On macOS, when that patch did not take effect, the download trusted only file-based CAs and failed verification behind a CA installed in the Keychain, and nothing said why. The download now passes an explicit context from _ssl_compat.system_trust_context(): a truststore context on macOS (an operator's SSL_CERT_FILE still wins), ssl.create_default_context() elsewhere. Both require a valid chain and a matching hostname. When the system trust store is not in use, _ssl_compat records why; a certificate failure in the download names that reason, and kirocrew doctor prints a "tls trust" row on macOS. Fixes #30450 ## Pattern harvest Not generalizable: the rule is specific to this downloader; other HTTPS clients already go through asset_downloader.make_ssl_context or the injected ssl module. (cherry picked from commit 7a09d86) Original author: Matthew Briones (#30755).
…0443) Managed MCP temp dirs sit at <data home>/run/mcp-tmp/<digest12>-<token8>. A server that binds a Unix socket under its temp dir (the MATLAB MCP server adds a 53-byte matlab-mcp-server-<n>/watchdog-<n>.sock) then exceeds sun_path (108 bytes on Linux, 104 on macOS, 108 for AF_UNIX on Windows), so a server that starts with the OS default temp dir fails under the managed one. Shorten the pieces that stay inside the sealed <data home>/run parent: the root name run/mcp-tmp -> run/t, the per-process digest prefix 12 -> 8 chars (the random token is kept whole, so collision safety is unchanged), and the probe child scratch subdir tmp -> t. The sweep now visits both run/t and the prior run/mcp-tmp so dirs left by an older gateway are still reclaimed. Containment is otherwise unchanged: one 0700 dir per process, the owner marker, the dead-owner + idle sweep, and a spec-declared TMP/TEMP/TMPDIR still wins. The root must stay under <data home>/run because the sandbox seals that parent read-only and carves the per-process dir writable only there; a root outside it would hand a sandboxed child an unwritable TMPDIR. This fits the default data home with bytes to spare but a longer home can still overflow; the full remedy is a short writable root outside the data home, which needs a sandbox-seal decision. Part of #30271 Co-authored-by: kirocrew-agent <agent@kirocrew.local> (cherry picked from commit 261ccd8) Original author: Raymond Chen (#30443).
The desktop shell spawns the gateway with windowsHide, but hands it the launch log as an inherited stdout/stderr descriptor. libuv adds CREATE_NO_WINDOW only when no stdio entry inherits a descriptor, so the gateway still got a console of its own, which can show and take the focus on launch. On Windows the shell now names the log in KIROCREW_DESKTOP_STDIO_LOG and gives the child null stdout and a piped stderr, which it copies into the log. Both entry points re-point descriptors 1 and 2 at that file early, which also closes the pipe, so tracebacks and child output still reach gateway-launch.log. Fixes #30570 (cherry picked from commit f1043f8) Original author: Zezhen Xu (#30962).
…26605) A wave member that calls spawn_run while the dispatch window is full gets its child accepted only as a task-store row, with no _queue entry and no _agents record. wave_has_live_nested_spawns read only memory, so the wave-close digest told the parent "This run is complete" while that child had not started. The completion consumer now awaits wave_has_live_nested_spawns_async for a completion that may close the wave. It adds a per-member taskq_overflow_async count, off the loop, to the in-memory checks, and answers True when the store cannot be read. The digest takes the result as _nested_live and keeps its own synchronous in-memory check. Closes #26599 (cherry picked from commit 352ea44) Original author: rohnkum (#26605).
…#20296) * fix(acp): disable codex background session MCP fleet via CODEX_CONFIG Background (kirocrew-lite) codex runtimes read the mcp_servers names from CODEX_HOME/config.toml off the event loop and inject a CODEX_CONFIG override with enabled:false per server, so high-churn background sessions stop starting an MCP fleet that session/close never reaps. An existing CODEX_CONFIG wins and the skip is logged. Project-level .codex/config.toml is a known residual. Fixes #13324 * fix(acp): keep codex mcp disable override valid after config edits codex re-reads config.toml on every thread/start while CODEX_CONFIG is fixed for the adapter's lifetime. A server removed or renamed after spawn left an override entry with only enabled=false, which codex rejects as "invalid transport", failing every background session/new. Each override entry now also carries an inert placeholder for the transport key the server had at spawn (command or url), so an orphaned entry still parses as a disabled server. Adds a test for the server-removed-after-spawn case and moves the tomllib import to module scope. --------- (cherry picked from commit 0a1d678) Original author: Zezhen Xu (#20296).
…answers again (#30907) When a control-plane server cannot read its session's tool policy, the resolver caches the failure for 60s and refuses the tool calls of every session that never resolved. The window could only end by expiring, so a gateway that came back seconds into an outage still had those sessions refused for the rest of the minute. Inside the window, one call per 10s (_NEGATIVE_CACHE_PROBE_INTERVAL) now dials the gateway as a half-open probe. Any answered read, a 200 or any 4xx, clears the window for every session in the process. A probe that fails, a 5xx included, re-arms the window and refuses as before. A call that returns before the read, such as one with no session key, does not spend the probe. The tool_policy.resolution_failed audit row also names the port it dialled and the resolution step that chose it (port=, port_source=). (cherry picked from commit 1ee9845) Original author: Pearcekieser (#30907).
…31718) A Reading title is display text: kiro-cli writes a file name with a line range or a label there (Reading listing workspace), not the path it opens. The gate turned that text into a filesystem.read item, the path matcher anchored it at the gateway's working directory, and an allow-mode read confinement refused a read whose real path was allowed. When the call's arguments name a read path, the gate now judges the read on that path and drops the title's item. A call whose arguments name no read path keeps the title's item as before. Closes #22275 (cherry picked from commit c7c409e)
Every gateway wrote its aggregate TasksMax/MemoryMax onto the one shared kirocrew-agents.slice at startup. A pod boots from an empty data home, so `kirocrew pod up` replaced the host gateway's configured ceiling with the defaults (131072 tasks down to 32768 on one host), and the live gateway's agents then failed fork with EAGAIN. A pod now writes its limits on its own per-instance child slice, which cgroup v2 still bounds by the parent, and sets a parent property only where systemd shows it unset, so a host that runs only pods keeps an aggregate cap. The sampler's self-heal reads the slice the ceiling was written on, and re-applies once per new login session, so a host gateway that outlived a user-manager restart takes the parent back from a pod that filled it first. Every slice write logs the previous pids.max and memory.max, at WARNING when it lowers either. A non-pod gateway writes the parent as before. Builds on #22636. (cherry picked from commit efb58fc)
Sensitive Change ReviewSensitive files changed:
How to clear this check:
Only the author's comment counts this way. A reviewer can instead |
UX Review (Opus 5.5) — ✅ PASSUX-level review of UX-Verdict: PASS Nothing to check. These are backend, Electron-launch and CLI fixes. The diff adds or changes no dashboard control, so no screenshot is required. The only new copy is the macOS [UX-REVIEWED] 6b75d11 Description read[DESCRIPTION-READ] 2c5da4cacdf9063139cf9a7315ffb098506f90f28294bbd8ca78ec342de02157 That sha256 names everything this verdict read: the PR description, plus the 3 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS No design concern holds: these are ten fixes already merged on Notes
[DESIGN-REVIEWED] 6b75d11 Description read[DESCRIPTION-READ] 85b5ad362042194bf2f7cdf1fce052e9e962c2ebf6cedcd88f5b92c4923fcaab 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 Ten main-merged fixes, each linked to a user-reported issue, carried over unchanged. Nothing new is decided here. What this change shipsInventory (10 items) — 10 justifiedIntent: bring ten user-reported fixes that already merged on main onto the 0.12 release line for 0.12.1. This is a FIX.
[FIRST-PRINCIPLES-REVIEWED] 6b75d11 Description read[DESCRIPTION-READ] 42218fd0edfd09d06e7f4d4fcde8e06df0350b5bbcb4183a0dc64dbe3c6bf47c That sha256 names the PR description this verdict read. Recompute it with |
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: |
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: |
Problem / Motivation
Goal: Bring ten 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@1e56b25e49before the pick; the right-hand column names it.browser_harness/chromium.py:320dashboard/chat_folders_routes.py:607sun_pathallows.mcp_gateway/backend_tmp.py:51website/electron/gateway-supervisor.js:839slack/gateway_runtime/subagent_report.py:588acp/harness/codex.py:297mcp_shared_policy.py:112platform/governance_parts/evaluator.py:338apps/builtins/meetings/backend/store.py:493TasksMaxfor every agent.sandbox_slice.py:779Not a goal
main; nothing is changed only here.CHANGELOG.mdchange.What changed (motivation → approach → change)
Each main squash commit was cherry-picked with
-xontorelease/0.12.0@1e56b25e49, in merge order onmain, one commit each.7a09d86156b419fff67709787ce84c8ef4db6d81261ccd8a57bad2760075f1043f8bd55ab7fd445d352ea448b830e18e30440a1d678d5a5005cefccc1ee984562a6eff6a9ee6c7c409e5953c71162beee28abd2785264b4084b9efb58fcc706b75d111daOne hunk is hand-resolved, tests only: in
test/link_screen_sites.py(#32190) only this PR's("platform_compat_files.py", "remove_with_retry")row is added; the neighbouringPinnedDirectory.is_linkrow belongs to a main change not on release. Every commit has added and removed lines identical to its main commit.Seven commits (#30755, #30443, #30962, #26605, #20296, #30907, #32190) carried a non-public author email, an internal co-author trailer, or a hand resolution, which the push guard refuses, so they are authored with the public identity and name their original author in the message; internal trailer lines are removed. The other three keep their original author.
builder-security checked: #30687 narrows which agent sessions may change sidebar folders and #31718 makes the governance file-read gate judge the argument path; both tighten an existing check and are the same changes already merged on
main. 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 readtest/link_screen_sites.pyor the touched docs, 853 files,-n 4 --dist loadgroup. 37119 passed, 345 skipped, 33 failed. All 33 fail the same way on a cleanrelease/0.12.0@1e56b25e49on this host (real-sandbox tests such astest_snapshot_read_only_real_view.py,test_pinned_order_real_sandbox.pyandtest_cron_run_scratch.py,test_design_tweak_devproc.py, and the release-onlytest_file_line_cap.pyred fromdashboard/chat_api/resume.pyat 2,008 lines).node --test electron/test/gateway-supervisor*.test.js: 162 passed, 0 failed.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 dashboard UI change; the one
website/change is the Electron gateway supervisor.Related Issues
Backport to the 0.12 line for 0.12.1 of #30755, #30687, #30443, #30962, #26605, #20296, #30907, #31718, #32190 and #32333. Source issues: #30450, #15618, #30570, #26599, #13324, #30906, #22275, #27832, #22635.
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.