Repository navigation
fix: 0.12.1 batch s3b4 (30 fixes from slice 4) - #33530
Conversation
…32554) A .pptx dropped onto the Knowledge page (or posted to the upload API) was answered 'processing' and then failed with no reason shown: .pptx has a reader but is kept out of FileReader.SUPPORTED because python-pptx is not a declared dependency, and the upload route never checked for that. The route now refuses an extension that has a dispatched reader but is outside SUPPORTED with a 400 naming the format, before anything is staged or a source row exists. Every other extension is handled as before. A drop closes the Add Source panel, which holds the only place a refused file's reason is shown, so a failed upload row now opens the panel again. Closes #30346 (cherry picked from commit 71bc66c)
…2933) One TaskRunner serves the whole gateway, but it kept the run's agent in a single self._agent field that plan, start_background, execute_plan and retry_from_task overwrote and every step read when it ran. Starting another run switched the agent of runs in flight, a run resumed after a restart (adopt_task_rows calls execute_plan with no agent) fell back to the default agent and lost its MCP servers, and the dashboard retry passed whichever agent started last. The agent is now Project.agent: set when the run is planned or started, persisted in runs.json, and read from the run by every step, review, decompose, replan and lesson session and by the taskq scope_ref. execute_plan applies its agent only to a planned run, the rule workspace_dir already follows, so Resume from the Projects page, which sends its selector value, cannot switch a run that has begun. retry_from_task re-targets only when given an agent, and the retry route passes none. status() reports each run's agent. (cherry picked from commit 07ccef8)
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) — ✅ 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 changed dashboard states. The blind read was not performed because the PR attaches no image and commits none. Evidence gaps
Notes
[UX-REVIEWED] 536167c Description read[DESCRIPTION-READ] 8ff8cba2295603a719b0ac5a100c01aa16b45eb9d3bc98f76fefbcd2c14c743a 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) — ✅ PASS (a BLOCK with no Trigger and no Outcome does not block)Design-level review of Design-Verdict: PASS On the 0.12 line, the cron-grant pick (#33456) silently drops a secret grant that an existing job already has stored, yet the description says "Removes nothing". BlockersTake-away with an unlisted broken reader: persisted cron secret grants
Fix: add this reader to Notes
[DESIGN-REVIEWED] 536167c Description read[DESCRIPTION-READ] 640d6cef6d3afbec46d3b2ba52eb52a04d088f6b7e0db6c3c07fb9b16fdb729d 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 This PR only carries 30 fixes that are already merged on Notes
What this change shipsInventory (30 items; 10 shown) — 10 justifiedIntent: give 0.12.x users the fixes that
(20 more rows, #11–#30 in the table, are each one already-merged [FIRST-PRINCIPLES-REVIEWED] 536167c Description read[DESCRIPTION-READ] ece8f244b50f9150ca999d344137c97bd1ef11e9f875ee0929646b8a8c943c89 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: |
…bound sweep cost (#33260) * test(sandbox): red tests for mount-source sweep pass budget and cost bounds A pin scan whose churn settles after four passes must prove coverage. A keyed pass must probe liveness once per pid within a root listing, examine at most a capped number of entries and resume where it stopped, rotate roots so one that spends the cap cannot starve the rest, and the periodic pass must be rate-limited. * fix(sandbox): let the mount-source sweep reclaim under churn and bound its cost The pin scan proves coverage only on a pass where no task of this uid exits between the listing and its read, and it allowed 3 passes. A host that starts short-lived processes all the time rarely got such a pass, so every dead-pid stand-in directory was held back and the pile grew to ~100k. Measured on such a host: 3 passes proved coverage on 4 of 10 scans, 8 on 10 of 10. A later pass reads only newly listed pids. The keyed pass now probes liveness once per pid within a root listing, uses one lstat per entry, examines at most 20,000 entries, resumes at the entry and the root where it stopped, and the periodic pass runs at most once every 300 s. (cherry picked from commit 6eb44d5)
DirectoryCheckpointStore.load documents that it never raises, but json.loads raises RecursionError on a payload nested past the interpreter limit and load only caught (OSError, ValueError), so one damaged savepoint file raised out of every member event-log read. Catch RecursionError in load and answer None like every other unreadable payload. save has the same gap around json.dumps and now answers False the same way. Adds regression tests for both sides. Co-authored-by: bayshanhai-dev <bayshanhai-dev@users.noreply.github.com> (cherry picked from commit 202f038)
) When the app info request failed, the migration page still rendered the migration card: the app has moved to a standalone package, your data has been preserved, and the not-in-registry Coming soon block. A failed load knows none of that, so the page told the user facts it had not read. Render only the error notice and Retry when no load has succeeded; a failed Refresh keeps the last good card, which is still true, beside the error. (cherry picked from commit a066c3a)
) AppPage replaced the previous page with a spinner and a 'Loading app' line that was not a status region, so a screen reader heard nothing until the app arrived. The line is now role=status with the spinner hidden from assistive tech, the same markup the app pop-out's loading line already uses. (cherry picked from commit fce27b7)
…33402) The Fork button on a remote artifact's page took a native `disabled` while the fork ran. A disabled button leaves the tab order, so pressing it from the keyboard dropped focus to the page, and a refused fork left focus there. The button now stays focusable with `aria-disabled` while it runs and ignores a second press, as the Skills refresh button does. Its look while busy is unchanged. (cherry picked from commit 3d8e027)
The Office, Panda Office, Underwater Lab and Wizard Tower scenes clip a session name to 8 characters for the desk nameplate and the whiteboard card with a plain UTF-16 slice. An emoji across that cut lost half of its surrogate pair, so "Deploy 🚀 now" showed "Deploy �". Each cut now uses sliceCodePoints, the helper the rest of the dashboard already truncates with, so an emoji is kept whole or left out. (cherry picked from commit e9a44c0)
Escape left keyboard focus on the page body, whether the bar was opened from the header button or with Ctrl+K in the message box. The shared focus trap reads the element to restore after mount, when the bar's own autofocused input already holds focus. The palette hook now records the focused control at the gesture and the host hands it to the overlay. A dismiss (Escape, an outside click, the gesture toggling the bar shut, a pane switch) focuses it again; a row that closes the bar leaves focus to the action it ran, as the legacy palette does, so a split-view session pick no longer risks handing focus back to the other pane's composer. An outside click dismisses on mousedown, and the browser's own handling of that press runs after the close and cleared focus again. Both the bar and the legacy palette now prevent that default. Builds on #29938 by @chenmingwei23. Closes #29716 Closes #33323 (cherry picked from commit 0196c1b)
…ld silence (#33392) * fix(autonudge): wake a gated PR watch on a blocker the judge would silence A gated pull-request loop screens each tick through a wake judge. A narrow custom brief can answer QUIET on a tick where a check just went red or the branch just started conflicting, and once a judge runs the typed fetcher's ACTIONABLE classification gets no say -- so an idle watcher slept through real blockers until the quiet-streak floor. The gate now wakes on a failing check, or a conflicting state, that the LAST DELIVERED reading did not carry, whatever the judge said. A blocker the owner was already shown (a PR under repair) stays quiet. The delivered basis lives on the monitor record beside the short-reading basis and is set only on a confirmed delivery; a re-aim clears it. The wake names the new blocker and withdraws the judge's suppression row so calibration is not mislabelled. * fix(autonudge): address Opus + GPT review on the new-blocker wake - Neutral wake-notice lead-in ('a blocker not present at the last delivery'); each new_blockers label already names its kind, so a conflict-only wake no longer claims a check failed. - The delivered-blocker basis is subject-dependent: _record_delivered_reading now records it whenever the monitor identity still matches, and applies the message/spec guard only to the question-dependent short-reading hold. Fixes a same-PR mid-flight reword leaving the basis stale (GPT 6.1 F1) and a same-PR re-aim wiping it (Opus). A true retarget replaces the monitor, so nothing is written to the old subject. - Document the delivered-blocker basis and the override in babysit-pr-watch.md, and correct the 'one deterministic mapping' sentence. - Tests: newly-failing/conflicting wake, same-set quiet, same-PR-re-aim keep, mid-flight reword records basis, true-retarget empty, fail-safe load. (cherry picked from commit 1d70a41)
…33239) The Shared MCP gateway card formatted its sizes and percents through fmtNumber but rendered the backend count, the stub count and the stub column as raw numbers, so a Bengali card mixed its own digits with Latin ones. Format the counts the same way; the pid stays raw as an identifier. (cherry picked from commit 153b1bb)
The aidlc task detail panel printed the raw backend status in its badge
and in the blocked banner ("in progress", "failed"), so every locale
read untranslated English and English itself showed internal values.
Read each status's word from the DAG legend's own table through one
exported helper, so the panel and the graph agree in every language
and there is a single status-to-word mapping.
(cherry picked from commit 2c66ab6)
A script cron's secret grant may not name a variable that changes how the child process runs. The deny set listed the POSIX names (HOME, SHELL, TMPDIR, PATH) but none of the Windows names that do the same jobs, so on Windows a grant could land a secret in USERPROFILE, TEMP, PATHEXT or COMSPEC and break every fire of the job. Add USERPROFILE, HOMEDRIVE, HOMEPATH, TEMP, TMP, PATHEXT, COMSPEC, SYSTEMROOT and WINDIR to the set on every host. The request-time validator and the delivery filter read the one set, so a new grant naming one is refused and a stored one is dropped at delivery. Closes #33290 (cherry picked from commit 5e0a4d3)
…33419) The Importing label appended dots from a 400 ms setInterval. That is JS-driven text the global prefers-reduced-motion CSS cut never reaches, so it kept ticking for users who asked for no motion. Read the live preference with useReducedMotion and skip the interval while it is set, as the Projects planning banner fix did. Co-authored-by: Zezhen Xu <zezhexu@amazon.com> (cherry picked from commit 00055e8)
#33441) With updates.unattended_apply: false, About told a managed-venv install "This install cannot apply updates itself" while its Install update button still worked. The switch's note read update_can_apply, which a managed venv always reports false because its Update is the arm and approve flow. The status frame now carries update_auto_manual_only, true for a notify that kept its route, and the switch reads it as well. unattended_apply_allowed() answers False when the governance context cannot be read; every other update pin still reads unpinned there. The legacy update loop logs the policy's notify at INFO, as the provider path does, and the governance spec says the desktop app's own updater does not read the key. (cherry picked from commit 9d54470)
…ns (#33427) A disabled button leaves the tab order, so disabling the remove (X) button while its removal runs dropped the focus of the keyboard user who pressed it to <body>. When the removal fails the card stays mounted but the user is no longer on the button. Make it aria-disabled/aria-busy instead and refuse a second press while busy. onDelete opens a modal confirm and only then starts the removal, which flips the busy flag before the next discrete event, so the busy flag alone is enough and a cancel that starts no removal leaves the button live. Visual disabled styling is unchanged. Co-authored-by: kirocrew-agent <agent@kirocrew.local> (cherry picked from commit 83f2ea4)
…29858) The column head on the pipeline board card counts a state's whole share, including the closed items the work fold let go. The rows under it only looked at the cards the fold still holds. So a column read "rejected · 5 of 300" above "no items", and "accepted · 291 of 300" above one row with no word about the other 290. _column_rows now takes the same share the head uses. When the share is larger than the cards held, it adds a "+N closed earlier, not listed" line, or prints only that line when the column holds no cards. Closes #29857 (cherry picked from commit eb7c3a3)
) A bulk approve re-reads each row's head before it submits the verdict. When that read failed (a provider timeout, a refused read), _refuse_if_head_moved returned the mapped 403/502 response, but the bulk loop treated every non-None answer as a moved head, so the row told the user to refresh and review the new commit. A refresh fixes neither failure. Only a 409 keeps that wording now; any other refusal keeps the provider's own error text, the text the per-PR review route already returns. Closes #28196 (cherry picked from commit 2e8bfac)
…ill_count (#31723) The dashboard export wrote contents.skill_count as the number of files under skills/, so a skill with a README and a script counted three times. The import preview shows this value as "Skills:", and the CLI backup manifest counts skill directories for the same key. The export now collects the distinct skill directory names it packages and writes their count, so two skills read as 2 however many files they hold. Closes #31709 (cherry picked from commit 1c20a71)
…#28151) "Copy link" on a skill's own SKILL.md gives a GitHub blob URL such as https://github.com/acme/widgets/blob/main/skills/reviewer/SKILL.md. parse_repo_spec kept the whole path, file included, as the skill directory, so the tree was requested as <commit>:skills/reviewer/SKILL.md, which is a blob, and the import answered "GitHub found no directory". A blob URL whose last segment is SKILL.md or AGENTS.md (the two files the bundle's root test looks for) now addresses the directory that holds it. Any other blob URL keeps today's behaviour. Closes #28148 (cherry picked from commit cd7a730)
…#28186) Crew Companion's import route calls import_bundle with nothing in front of it, and import_bundle checked the files a bundle CARRIES, never the files its manifest NAMES. The dashboard route adds that check, so the companion installed, with a 200, bundles the dashboard refuses: - a slot naming art the bundle does not carry (states.idle = missing.svg), which installed a pack that draws nothing in that slot; - two carried names that are the NFC and NFD spellings of one name, which differ under casefold but are one file on macOS, so the second write replaced the first. import_bundle now refuses a states/moods/random slot that names an art file it does not carry (the art side of _sound_reference_problem), and two carried names with the same pack_id_key (NFC, then casefold). save_pack compares carried names on pack_id_key instead of casefold, so the companion's save route refuses the NFC/NFD pair as well. Closes #28184 (cherry picked from commit bcaebc5)
…28250) advance() and skip_once() rolled a recurring reminder forward one interval at a time under a 100,000-step ceiling. A reminder staler than that, such as an every-minute reminder left for 70 days, stopped short of now: advance() left it due, so the one-second scheduler fired it again on the next tick, and skip_once() returned a past fireAt that fired at once. Both now compute the number of whole intervals with one division, so the result is always the first slot strictly after now and no ceiling is needed. Closes #28249 (cherry picked from commit 41ba859)
…hat edit (#31781) plan_to_chat_context wrote force_approval into the plan chat's Tasks JSON only when it was true. update_plan_tasks fills an absent force_approval from the old task at the same position. So inserting or deleting a step in that JSON moved an approval gate onto a different task, which then blocked even in auto-approve mode, while a deleted gate passed to the next task. plan_to_chat_context now always writes force_approval explicitly, so applying its own JSON never falls back to the position-based value. New tests insert and delete a step in the chat JSON and check that each gate stays with its task. Closes #31766 (cherry picked from commit a214d45)
…tion (#31753) When a workflow finishes, the gateway runs a turn whose prompt starts with "[Workflow `<name>` finished: <status>]" and stores it as a user row. The session summary's injected-prefix list had no entry for it. So extract_turns numbered it as a user turn, the user-turn counts included it, and the summary model read a gateway instruction as the person's request. The prefix list now includes "[workflow `", matched case-insensitively like the others. The row is flagged as automation, rendered as automation, and left out of every user-turn count. The spec's list of automation rows names it too. Closes #31739 (cherry picked from commit cf5d095)
) A Teams turn that called a tool shows one "🔧 <tool>…" message, and the answer's first chunk edits it in place. A reply made only of [OPTIONS:] chips or only of images has no text chunk, so nothing edited it and the "working" message stayed on screen above the finished reply. Edit it to a short done line in that case. (cherry picked from commit 0b741aa)
…ody runs (#28243) After admit hands back a handle, admitted_agent_fn and _execute_single_task wrote the running mark (and the TaskRunner told the workflow view the step started) before the try whose CancelledError arm cancels the handle. A run cancel, pause or timeout landing during those awaits propagated with the handle unsettled, so its lane slot was never released and the row stayed starting or running under this incarnation's lease until the next boot. The lane is shared by TaskRunner steps and workflow agent calls, so each lost slot lowered their combined concurrency. _taskq_begin_run had the same gap for the run's container row, whose handle is not in _run_handles yet for _taskq_end_run to settle. The running mark now has its own synchronous cancel arm in all three places, and the TaskRunner's start publication moved inside the body's try. Closes #28223 (cherry picked from commit 1d51232)
0445389 to
536167c
Compare
Problem / Motivation
Goal: Bring 31 user-facing fixes already merged on
mainontorelease/0.12.0for 0.12.1, one commit each, in a PR of their own.Why it matters
Each fix changes behaviour a 0.12.0 user can hit today: lost keyboard focus on busy buttons, untranslated labels, task runs that use the wrong agent, sandbox mount stand-ins that pile up, cron secret grants that accept Windows process env names, and the other cases in the table below.
Not a goal
main; nothing is changed only here.CHANGELOG.mdchange.What changed (motivation → approach → change)
Each main squash commit was cherry-picked with
git cherry-pick -xontorelease/0.12.0@7a89cc18d6, one commit each, inmainmerge order. Every pick applied with no conflict.skill_countis wrongSKILL.mdblob URL is not read as the skill's directoryforce_approvalonto another taskDropped from this batch:
0445389517: inGalleryPanel.tsx, a save that fails while the 450 ms lookup debounce is pending skips the new input's lookup, so a retried Use submits the older companion. It is taken out rather than fixed here.test/test_dashboard_chat_turn_wiring.py::TestRuntimeWiringTurn::test_first_slash_preserves_pending_replay_for_next_promptfails on top ofrelease/0.12.0and passes without it. It is taken out rather than fixed here.Excluded on purpose (they wait for 0.12.2):
queue_drain.pynames Spec Builder's decision-delivery meta keys.Backwards compatibility
Compatible: each fix keeps its main PR's compatibility notes; nothing here changes stored data or config beyond what those PRs already shipped on
main.Removes nothing: every commit is a fix already reviewed and merged on
main.Tests
Proof: not provable -- other: a backport of main commits whose proofs ran on their own PRs
Run on head
0445389517(before #33002 was taken out; dropping it changes onlyGalleryPanel.tsxand its test), one process at a time, no xdist:test/test_cron_secret_env.py(7 failed) andtest/test_taskrunner_retry_admission.py(9 errors). Both fail the same way on a cleanrelease/0.12.0checkout, so this batch did not cause them. After dropping fix: stop re-sending the session context after a streamed turn fails #32484, the chat-turn and session files it touched pass (485 passed).tsc -b --forcereports no errors. On head536167c75c(without fix: keep focus in the companion import dialog while it saves #33002):CrewCompanionGallery.coverage.test.tsx68 passed andtsc -b --forcereports no errors.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 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 the 30 PRs in the table above.
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.