Repository navigation
chore(release): 26.11.26 - #60
Merged
Merged
Conversation
A step binds only to the launch whose description names it, so a Claude Code crew launch that named none of its member's steps started with no step; the tech lead learned it only afterwards and had to stop and relaunch the member (twice in two days). The PreToolUse hook checks a main-session Agent or Task call for a devkit crew member with the matcher the binder uses after the launch (progress._bind_step: the member's steps, scored by progress._name_score on the description's task text). When the description names none and exactly one step waits for that member's launch, the hook puts the step's name in the description, records it in the launch record so the agent binds that step, writes crew.launch-named to the audit trail, and tells the tech lead in the call's context. With two or more steps waiting, or two or more named equally, the tech lead gets the exact description that takes each. The new description and the brief's required sections go out in one updatedInput. Codex launches are left as they are: Codex applies updatedInput only to shell commands, and its spawn's task_name selects the step. Verified: dev test run (23 files that send launches through the hook): 605 passed; dev test prove: the 6 new tests fail without the change; python3 and /usr/bin/python3 -m unittest tests.test_hooks_launch: 6 passed; dev docs check: 0 issues
GitHub keeps every earlier body of an edited comment in a public edit history, so record comments the workflow edited in place still showed the old shared voter digests, which link one person's votes across reports. Root cause: the registry workflow wrote each record with PATCH on the same comment (_write called edit_comment), and a comment's edit history goes only when the comment is deleted Each write posts the record as a new comment, then deletes the old one with its history; posting first means a stopped run leaves two records, never none. The workflow and devkit read the newest bot record, and the next write deletes the older one. The record's seen counter skips the bot's own comments, so a reposted record does not trigger another rewrite. The manual run also posts anew the record of every registry issue closed more than 90 days ago when its comment was ever edited or still uses the older digests, keeping its votes and counts; a second run writes nothing. GitHub requests that may succeed later are tried up to three times, except posts. Verified: python3 -m unittest of test_bug_registry_action, test_bugs_sync, test_bugs_render, test_ci_commands, test_workflow_ci, test_consistency, test_docs_check, test_security_corpus (281 passed); dev test prove (9 new or changed tests fail without the fix)
A step queued for a member's busy lane is announced to that member (it joins your lane once your step closes); when the tech lead then assigned it to another member, the queue was cleared and the change recorded, but the member was never told and kept expecting the step. dev progress assign now sends each running lane the step was queued for a devkit notice that names the step and its new owner (and the step that lane goes on with), through the run's messages that the lane's next hook and dev watch --wait deliver. A step queued for no lane sends no lane anything. Verified: dev test run (47 files that assign steps): 1177 passed; dev test prove: the new test fails without the change (0 != 1 notices); python3 and /usr/bin/python3 -m unittest tests.test_progress_reassign: OK
dev progress assign now sends each lane a queued step was waiting for a notice naming the step and its new owner, and the crew guide described only the queue Verified: dev docs check docs/crew.md: markdownlint and links clean, one Vale warning on line 3 that predates this change
Most lessons have no Fix and no Evidence (1,090 of 1,516 on this Mac), so nothing traces them to a real event; listing them in the review would add a thousand lines and bury what needs acting on, and the review did not show them at all. lessons.review returns the lessons that are not retired and have no Fix and no Evidence. dev lesson review shows them as one line with their count and the flag that lists them; dev lesson review --untraced lists them one line each, with how to retire one. The shell completions, the command reference, the lessons guide and the learning-loop skill name the flag. Verified: python3 -m unittest tests.test_lessons_cli: 42 OK (2 new, failing before the change: no count line, unknown flag); tests.test_consistency: 31 OK (D4 and D13 clean); /usr/bin/python3 ReviewUntracedTests: OK; dev docs check: 0 issues; bin/dev lesson review on this Mac: 665 lines, one count line for 1,090 lessons
The audit trail records each dev call's options by their class, and the test that every option is classified failed once dev lesson review gained --untraced. Root cause: the new --untraced option was added to the lesson review parser but not to audit.FLAG_OPTIONS, the list of option names that take no value Verified: dev test run tests/test_audit.py tests/test_lessons_cli.py tests/test_consistency.py: 198 passed (the classification test failed before: --untraced missing from FLAG_OPTIONS)
Vale flagged 'several' in the crew skill's sentence about the pre-tool hook naming a launch's step; the guide already says two or more. Verified: vale on plugin/skills/crew/SKILL.md: no finding on lines 154-158; dev docs check: 0 markdown issues, all links resolve
dev ci --run ran 'pytest > ~/.bashrc' as written, which overwrites the developer's shell settings with the test log, and skipped 'pytest | sudo tee /tmp/x.log' as runner setup, so its tests never ran while the evidence was recorded. Root cause: the step classifier dropped the redirect targets the shell reader returns (gitx._commands), so a redirect was never judged as a write, and it took any writing command run with sudo for the runner's own setup (the as_root branch of gitx._prep_one), whatever folder it wrote A shell redirect (>, >>, 2>, &>) is judged like tee: next to a check, one into a dotfile or dot folder of the home folder blocks the evidence; with no check beside it, a write into the home folder is the runner's setup; a log elsewhere in the home folder, in the checkout, or /dev/null still runs. A writing command run with sudo is the runner's setup only when every file it writes is in a system folder (/etc, /usr, /opt, /var outside its temporary folders, and the like): next to a check, any other root write blocks the evidence, so 'curl … | sudo tee /etc/apt/…' stays setup and 'pytest | sudo tee /tmp/x.log' blocks. The command reference and the local CI guide say so. Verified: python3 -m unittest tests.test_gitx tests.test_gitx_ci_shortcuts: 186 passed; test_ci_commands, test_cli_ci, test_workflow_ci, test_shellparse: 33 passed; test_security_corpus: 22 passed; dev test prove tests/test_gitx.py --runs 2: both new tests pass with the change and fail without it; dev docs check on both docs: 0 issues
On a 390 x 844 phone a project page showed its section tabs only after more than a screen of scrolling (tab bar at y 985): eight health reasons and four 2 x 2 chart tiles came first, so a link to a tab showed no line of it. At every width, a tab chosen from the stuck tab bar kept the old scroll offset, so the new panel started hidden above the window or the page jumped to its end. Under 481 px the reasons list shows the three most serious (the server's order) and a Show all N reasons button (aria-expanded, aria-controls) that shows the rest and reads Show fewer; it stays open while the page shows the project and closes when the person leaves it; print lists every reason. The four tiles become one row of figures, the number above its label, without charts. The Open tasks and Open PRs tiles carry their chart's words as hidden text at every width, so their accessible names keep the breakdown (2 doing, 1 blocked) where the chart is not drawn. On acme/widget the tabs move from y 985 to y 665 and the first Overview panel starts above the fold. At every width, a tab chosen from the stuck bar (by click or the arrow, Home and End keys) brings the bar back to its sticky top with the new panel right under it; the panel is then at least a window tall for the rest of the visit, so a short panel can bring the bar up too. With the bar not stuck the page does not move. From 481 px up nothing else changes. docs/dashboard.md, Project page, says how the page reads on a phone. Verified: dev test run on the 13 dashboard, page and dependent test files: 335 passed; after the last edit tests/test_dashboard.py, test_dashboard_now.py, test_dashboard_sessions.py: 158 passed; dev test prove: the 6 new browser tests pass 5 of 5 runs and fail without the change; dev visual captures at 390x844 and 1440x900 in light and dark; dev docs check docs/dashboard.md clean in my lines
A peer box that was not a JSON object stayed forever: delivery to its session failed, and in the canonical store the hourly sweep stopped at it, as it did at a newer devkit's box, so void claims and old escalations were never cleaned up. Every box and escalation the sweep removed also left its lock file behind. Root cause: the sweep had no rule for a box it cannot read: its canonical loop let _read's refusal escape and its legacy loop skipped a box _read returned as None; state.remove deleted only the record, and its lock file could not go safely because state._acquire never checked after its flock that the lock path still named the file it locked, so a waiter on a removed lock file became a second holder state._acquire checks after its flock that the path still names the file it locked and otherwise locks the file the path names now, within the same wait. state.remove(with_lock=True) deletes the lock file under its flock, only as the outermost hold, when no record the lock covers is left, and never in the canonical message store, whose acquisition refuses a replaced lock. The sweep uses it for the legacy boxes and escalations it removes, whose waiters never write a missing record. A box whose bytes are not a JSON object is renamed <name>.<time>.unreadable next to it, under its lock (state.set_aside), so the next delivery starts a new box and the bytes stay for a person to look at. A JSON object this copy cannot read, such as a newer copy's box, stays in place, and the legacy loop no longer rewrites a box with another messaging_schema. Verified: dev test: 277 files, 5722 passed, 2 skipped, 1 failed that passes on rerun (test_security_corpus read the dashboard while another branch commit changed it); dev test prove tests/test_peers.py tests/test_state.py: the 7 new tests fail without the change and pass 5 of 5 runs; the new tests pass 30 of 30 repeated runs
Comments in the audit, extension and lane routing modules named the requirement, finding and review labels of the plans that produced them, which mean nothing to a reader who never saw those plans. Each label is dropped or replaced by what the code does and why: section rules keep their titles, a label that stood before a comment's text goes, and a comment that was only a label is removed. The masker's comment keeps the id shapes it keeps in records (D<n>, M<n>, R<n>), which are data. Only comments and docstrings change: the syntax tree of each module, without docstrings, is identical before and after. Verified: dev test run tests/test_security_corpus.py tests/test_security.py tests/test_audit.py tests/test_audit_lane.py tests/test_ext.py tests/test_extension_auto.py tests/test_laneroute.py: 345 passed; test_hook_capture, test_mailbox_performance and test_consistency: 46 passed; ast.dump of each module without docstrings equal before and after
A test outside HookCase could still start a real detached dev sync that outlived it and kept running after the run. Root cause: the guard against background syncs lived in HookCase's patches of hooks._spawn_sync and work.request_sync, so a test module outside that harness reached the real detached Popen; the hermetic audit hook checked no sync job Verified: dev test prove: both new DetachedSync tests fail without the guard; dev test run --all: 5806 passed, 2 skipped, 0 tripped the guard
Comments in the progress, hooks, command-line and state modules cited the requirement, contract, finding, work item and lesson ids of the plans and reviews that produced them, which mean nothing to a reader who never saw those records. Each label is dropped or replaced by what the code does: section rules keep their titles, a parenthesis that held only labels goes, and a lesson or bug reference keeps the rule it stated. Real identifiers stay: the message and decision id examples (M3, D2@9448c4d0), D3 in the dev crew decide help, the C0 and C1 control-code sets, the step specs of dev progress add and the reference to the global instructions' status block section. Only comments and docstrings change: the syntax tree of each module, without docstrings, is identical before and after. Verified: dev test run on the 18 progress, hooks, cli and state test files with test_security_corpus, test_security, test_consistency and test_hook_capture: 468 passed; ast.dump of each module without docstrings equal before and after
the refusal text for DEVKIT_* variables in settings and two caller-identity messages named spec items the reader never saw; the messages already say the rule Verified: dev test run tests/test_config_change.py tests/test_caller_identity.py tests/test_security_corpus.py: 69 passed
The detached-sync guard refused a foreground shell line such as sh -c 'dev sync 2>&1 | cat', and its message showed the command cut at the redirect ('dev sync 2>').
Root cause: the shell line was split at every &, so the & of 2>&1 or &> ended the command and read as a background marker; the commands came from the same split
A shell line's commands and their background marker now come from one parse that keeps a redirect's & (right after > or <, or before a digit, - or >) inside its word; a detached form such as 'dev sync >/dev/null 2>&1 &' is still refused and shown whole.
Verified: tests.test_hermetic.DetachedSync against f44c6ff: 2 failures, 2 errors on the new cases; with the fix 27 of 27 pass; dev test prove: both tests fail without the guard; dev test run on 9 harness and sync-path modules: 249 passed
Comments and docstrings cited plan, spec and review references by their codes, which mean nothing to a reader who never saw those documents. Each reference is dropped where the sentence already says what the code does and why, or replaced by the words it stood for. Comments and docstrings only; no code changes. Verified: the syntax tree of each of the 29 files with docstrings removed is identical to HEAD; dev test run on 40 test files (security, security corpus, consistency and the changed modules' tests): 1204 passed, 1 skipped
dev ci --run still skipped 'pytest | sudo tee /var/log/pytest.log' and 'pytest | sudo dd of=/tmp/x.log' as runner setup, recording evidence for tests that never ran, and ran 'pytest >| ~/.bashrc' as written, which overwrites the developer's shell settings. Root cause: gitx._prep_one took a root write into a system folder for setup whatever fed it, _classify turned a pipeline of setup and a check into setup, dd was not among the writing commands, and shellparse.split_commands read >| as two words of the command instead of a redirect A writing command run as root into a system folder is setup only when nothing but a download or key unpacking (curl, wget, gpg) feeds it, so 'curl … | gpg --dearmor | sudo tee /etc/apt/keyrings/…' stays setup; next to any other check it blocks the evidence like a root write elsewhere. dd joins the writing commands with its of= target. >| and N>| are truncating redirects. Verified: dev test run test_gitx, test_gitx_ci_shortcuts, test_shellparse, test_hooks_launch, test_security_corpus: 239 passed; 27 more files that use the shell reader or the classifier: 819 passed; dev test prove: the 2 new tests fail without the change
Comments and docstrings in gitx.py and shellparse.py named plan labels (spec and finding numbers), and a test comment named a lesson id, which mean nothing to a reader who never saw that plan. The labels are gone, and each sentence says the behavior instead; the test comment says what the description it uses guards. Verified: each file's syntax tree with docstrings removed is the same before and after (ast dump hash 89f770591bcc78d9 for gitx.py, 22ef98e7333481c0 for shellparse.py); dev test run of the five named files: 239 passed
dev ci --run now blocks a check whose output goes to sudo tee or sudo dd into any folder, and keeps only a download written as root into a system folder as setup; the two rows still described the earlier rule Verified: dev docs check docs/commands.md docs/git-and-github.md: 0 issues
the id points into the private handoff repository; the comments already say why bash's reading is used Verified: syntax tree without docstrings unchanged against HEAD; dev test run tests/test_shellparse.py tests/test_consistency.py: 55 passed
Release the registry repost, the peer box, dev ci --run write and test guard fixes, with the crew launch, queued step notice, lesson review and dashboard phone changes. The CHANGELOG gains the 26.11.26 entry, and VERSION, desktop/manifest.json and both plugin manifests read 26.11.26. Verified: dev docs check CHANGELOG.md (exit 0, no finding in the new entry); json.load of the three manifests (ok)
CI tests Python 3.9 and 3.13; the detached-sync guard's probe used Popen(process_group=0), which only Python 3.11 and later accept, so PR #60's 3.9 job failed Root cause: subprocess.Popen gained process_group in Python 3.11; on 3.9 the probe raised TypeError, ended as an error instead of the guard's failure, and the outer test's failures=5 count did not hold Verified: /usr/bin/python3 (3.9.6) -m unittest tests.test_hermetic: 27 OK; Python 3.13.18 -m unittest tests.test_hermetic: 27 OK
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What and why
This releases devkit 26.11.26. It fixes four bugs: bug registry records kept old voter digests in GitHub's public
edit history, unreadable peer boxes blocked the hourly cleanup,
dev ci --runmistook some writes for setup, and atest could leave a real
dev syncrunning. It also ships the crew, lessons and dashboard improvements listed below.Changes
no longer keeps the old shared voter digests. A manual workflow run reposts, once, the records of issues closed
more than 90 days ago that were ever edited or still use the older digests.
<name>.<time>.unreadable, a box a newer devkit wroteis kept, and the sweep removes the lock files of the boxes and escalations it removes, safely.
dev ci --run: shell redirects (including>|) count as writes, and a check written as root (sudo tee,sudo dd of=) blocks the evidence instead of being skipped as runner setup.dev syncis refused from any test module, and the&of2>&1is no longer read as abackground marker.
dev lesson reviewcounts untraced lessons in one line;--untracedlists them.chosen from the stuck tab bar starts the new panel right under it.
desktop/manifest.jsonand both plugin manifests read 26.11.26.How to verify
dev ci --run --resume --step-timeout 60with Python 3.13.18 → passed: 5,810 unit tests OK (2 skipped) in 35m42s; gitleaks, semgrep, osv-scanner, trivy and zizmor cleandev ci audit: 0 critical, 0 high, 0 medium findings;dev docs check CHANGELOG.mdexits 0.Risk and rollback
Low to medium. The registry change rewrites record comments on the public registry; a stopped run leaves two record
comments, never none, and devkit reads the newest. The
dev ci --runchange can block evidence that a run recordedbefore; the refusal names the step. Roll back by releasing a new version that reverts the change; published tags
are never moved.
Links