Session Handoff [default]: #3671 Thread-Aware Held Wait Released in 2.0.1286, Carried Backlog Next

Next steps, in priority order

Re-read each issue and pull request rather than trusting these summaries.

Carried hub backlog (default). Re-rank against the open backlog first, since it moves every round. Recent unlabeled pr_review.py issues are cheap, well-scoped picks: Name a Merged or Closed Pull Request's State in pr_review.py status #3677 (name a merged or closed pull request's state in status), pr_review.py Leaves a Copilot Overview's Outstanding Findings List Unplaced #3672, and Skip a Null GraphQL Node in pr_review.py reviewer_nodes Instead of Crashing wait #3392. The older carried list is Read Action Files or Name the Cause When UV_NO_CACHE Removes the Resolved PyYAML Interpreter #2358, Let an Unattended Handoff Run the Strict-Review Pass for a Worker That Cannot Dispatch #2342, Add safety hooks for opencode and codex #781, then Validate the Interpreter the Agent-Safety Installer Writes Into settings.json #1172, A Parent Pinned to UTF-8 Spawning a Python Child Leaves the Child's Encoder on the Locale #1577, skills_install.py runs a bare claude argv, which may miss a Windows .cmd shim #1754, Python devcontainer snippet: mount path join and mktemp -t portability #1763. Done means one issue fixed through a feature pull request into develop and a mergeable promotion.
Check why Copilot reviewed both feature pull requests into develop this round. Exit pr_review.py wait When Unresolved Threads Land on an Attested Head #3680 and Walk a Cut Thread Page Once per Poll Read #3682 each drew a Copilot round (rounds=1) although wait requested none on an attested head. If an automatic-review setting causes it, the fleet's "hosted review runs once per change, at the promotion" cost model does not hold on this repository. Done means the cause is named, and an issue is filed if it is a defect or a setting to change.
Downstream, each in its own repository's session: unchanged from round 66. ptr727/Blog's lib/fleet/**.sh root .editorconfig section goes at its next resync. ptr727/HomeAutomation-Config#1435 and Read the Release Env File Through the Hub's Carried Readers Blog#553 each start with a hub pull request declaring host-secrets in registry/repos.json, and the round-60 set stands.

External blockers

A feature merge into develop needs the maintainer's hand. The auto mode classifier denied the agent's gh pr merge --squash --match-head-commit again this round. Both feature merges, Exit pr_review.py wait When Unresolved Threads Land on an Attested Head #3680 and Walk a Cut Thread Page Once per Poll Read #3682, ran as ! lines the maintainer typed. The promotion merge and the release dispatch ran from the agent seat. Ask for the ! line with the head SHA read live, rather than retrying.
The nested hub clone hub-resync, write grants downstream, a native Windows host, and cleanup --sweep --apply are unchanged from round 60. This round's sweep would remove two other sessions' detached hub worktrees, and --apply stays the maintainer's call.

Internal dependencies

scripts/pr_review.py is at 2115 lines against its 2200 ceiling, and scripts/pr_review_checks.py is at 376 of 450. A change that grows pr_review.py much further needs a split along module boundaries first, never a raised ceiling.
Step 3's hub pull requests each edit registry/repos.json, so they go one at a time or in one pull request.

State

Re-derive with git worktree list, git log origin/main..origin/develop, gh pr list, and gh release list.

main and develop are level, at promotion Promote Develop: Exit pr_review.py wait When Unresolved Threads Land on an Attested Head #3681's merge commit. 2.0.1286 is the latest release.
Open pull requests: none.
Worktrees: both task worktrees went through pr_review.py cleanup. The session's own hub clone on main remains under the worktree directory. The primary checkout is clean, on develop at origin/develop. The session's opening status had it on main.
Other repositories touched: none.

The parked decision queue

Zero open decision issues in this repository at close, read live. This round put one question, whether to merge #3681 and dispatch the release. The answer was "Merge and release".

What the last round did

Implemented Exit pr_review.py wait When Unresolved Threads Land on an Attested Head #3671 via Exit pr_review.py wait When Unresolved Threads Land on an Attested Head #3680. On an attested held head, wait's poll now reads each thread's state and author. An open thread ends the poll with exit 46 and status=THREADS_UNRESOLVED, which names the authors and outranks 30, 44, and 45. A merged or closed pull request still exits 0. A thread read cut at 100 is walked to its end through unresolved_threads().
Walk a Cut Thread Page Once per Poll Read #3682 caches that walk on the payload as _open, answering Copilot's one finding on promotion Promote Develop: Exit pr_review.py wait When Unresolved Threads Land on an Attested Head #3681.
Promotion Promote Develop: Exit pr_review.py wait When Unresolved Threads Land on an Attested Head #3681 merged into main, closing Exit pr_review.py wait When Unresolved Threads Land on an Attested Head #3671. Release 2.0.1286 published.
Local strict review ran six passes at the frontier tier across three pushes, and every finding was given an outcome on the pull requests.
Filed nothing.

What not to repeat

Failing closed on a cut thread page. d7413e6 made a cut page with none open in it exit 46. That loops forever on a pull request whose threads are all resolved and number more than 100, since nothing is left to dispose of. Walk the page instead.
Declining a cost finding that has a three-line fix. The doubled walk was declined on Exit pr_review.py wait When Unresolved Threads Land on an Attested Head #3680 and then came back from Copilot on the promotion, which cost a second feature cycle. Where the fix is that small, take it in the same pull request.
Typing a --match-head-commit SHA by hand. One merge call this round carried a SHA written out rather than read. The classifier happened to deny that call. Capture the SHA from gh pr view --json headRefOid into the command.

New learnings

None beyond the entries above, each specific to this round's attempt.

Parked decision queue answered at session close: #2353 took the deletion-only exemption to local-strict-review's two-round cap, which applies to attended and unattended seats alike. Its decision label is removed, and it stays open as the work to write the rule into local-strict-review's budget bullet. The queue now holds zero issues.

Round 49, closed 2026-10-04: Session Handoff [default]: Four Follow-Ups Landed and Released as 2.0.826, Downstream Adoption Next

Next steps, in priority order

Re-read each issue and pull request rather than trusting these summaries.

Downstream adoption of the two releases. 2.0.822 and 2.0.826 carry the merge-bot closed event, test-matrix, the live comments label read, and the uvx YAML reader. Each step needs a fresh write grant for its repository.
- Add closed to the Merge-Bot Stub Types of Adopters Setting delete-branch #2336: five adopters add closed to their merge-bot stub types.
- homeassistant-purpleair adopts test-matrix (Add a Dependency-Version Matrix Axis to validate-task.yml for the Gates Stage #2334) for Decide How to Adopt the Hub's Reusable Release Chain homeassistant-purpleair#365.
- Every carrier trails the hub's skills. Re-carry them with the next resync rather than one repository at a time.
Grant pull-requests: read So the Live comments Label Read Reaches Private Repositories #2354: grant pull-requests: read so the live label read reaches private repositories. Today a private repository logs the fallback line and keeps A comments Label Applied by gh pr create Misses the Event That Starts the First Run #1703's behavior. The grant must land in each downstream stub before that repository's pin reaches the change, so roll it out with the resync in step 1.
Let an Unattended Handoff Run the Strict-Review Pass for a Worker That Cannot Dispatch #2342: the digest-handback line in unattended-handoff's return table, per the maintainer's answer recorded on the issue.
Small follow-ups filed this round: Bound the In-Process Composite-Actions Read Against an Alias-Expanded Step Name or Shell #2357 (the in-process composite-actions read can hang on an alias-expanded step name or shell) and Read Action Files or Name the Cause When UV_NO_CACHE Removes the Resolved PyYAML Interpreter #2358 (the UV_NO_CACHE host finding names a vanished path, plus an untested probe-start failure).
Close the Remaining CommonMark Gaps in prose_lint's Sentence-Length Block Parser #2341, low value. Keep a hard scope cap.

External blockers

Write grants: every downstream step needs a fresh per-repository grant.
Let an Unattended Worker File a Wording-Only Finding Past the Two-Round Cap Instead of Parking #2353 is a decision issue (the wording-only exception to the two-round cap), filed by the HomeAutomation-Config session. This round added its data as a comment.
A native Windows host is still needed for the bootstrap.ps1 check (Session Handoff [default]: WORKFLOW.md Corrections and the pyproject.toml Decision Released as 2.0.672 #1804). The uvx reader's -I -X utf8 behavior on Windows was reasoned about, not run.

Internal dependencies

Grant pull-requests: read So the Live comments Label Read Reaches Private Repositories #2354 and the step-1 resync go together. A stub that does not grant pull-requests: read fails at startup once its pin reaches a lint job that declares it.
Bound the In-Process Composite-Actions Read Against an Alias-Expanded Step Name or Shell #2357 and Read Action Files or Name the Cause When UV_NO_CACHE Removes the Resolved PyYAML Interpreter #2358 both touch repo_gate.py's composite-actions read. Take them in one lane, or in sequence.

State

Hub:
- main is at f75f9b52, through promotion Promote develop to main: Live comments Label Read, uvx YAML Reader, and Review Digest Tolerance #2360.
- Releases today: 2.0.822 (from Promote develop to main: Composite Action Linting, Ref-Scoped Publisher Concurrency, and Review Digest Fixes #2349) and 2.0.826 (from Promote develop to main: Live comments Label Read, uvx YAML Reader, and Review Digest Tolerance #2360).
- develop is at cf053cb5.
- There are no open hub pull requests, and the primary checkout is on develop, current and clean.
- This round's worktrees and its local and remote branches are all removed. Re-derive with git worktree list.
- This machine's Skills snapshot reads current at f75f9b52.
Unattended lanes untouched: auto-2176, auto-2167, auto-2157, auto-2080, python-contract.
Other repositories written: none.

The parked decision queue

One issue carries the decision label: #2353. It was filed today by another session and presented when this link was written.

What the last round did

Released 2.0.822 from Promote develop to main: Composite Action Linting, Ref-Scoped Publisher Concurrency, and Review Digest Fixes #2349 (the previous round's promotion).
Merged into develop, then promoted as Promote develop to main: Live comments Label Read, uvx YAML Reader, and Review Digest Tolerance #2360: Assert the Fallback Fetch URL in hub-fetch-run.py's Tests #2352 (Assert the Fallback Fetch URL in hub-fetch-run.py's Tests #2339), Tolerate a Per-Field workflowRun Error in the Check Query #2355 (Tolerate a Per-Field workflowRun Error in pr_review.py's Check Query #2338), Read the Comments Label Live Rather Than off the Event Payload #2356 (A comments Label Applied by gh pr create Misses the Event That Starts the First Run #1703), Provision the Composite-Actions Check's YAML Parser Through uvx #2359 (Provision the Composite-Actions Check's YAML Parser Deterministically #2348).
Filed: Grant pull-requests: read So the Live comments Label Read Reaches Private Repositories #2354, Bound the In-Process Composite-Actions Read Against an Alias-Expanded Step Name or Shell #2357, Read Action Files or Name the Cause When UV_NO_CACHE Removes the Resolved PyYAML Interpreter #2358. Data was added to Let an Unattended Worker File a Wording-Only Finding Past the Two-Round Cap Instead of Parking #2353, and a probe-start test gap was added to Read Action Files or Name the Cause When UV_NO_CACHE Removes the Resolved PyYAML Interpreter #2358.
Maintainer decisions:
- Read the Comments Label Live Rather Than off the Event Payload #2356 stays at contents: read, so the live read is public-only and private repositories follow in Grant pull-requests: read So the Live comments Label Read Reaches Private Repositories #2354.
- Three lanes that hit the two-round cap each got one bounded fix, and a final pass whose findings were filed rather than fixed.
A dispatcher-run strict pass on every push caught two security regressions before any push:
- A cwd fix that ran the child in the shared /tmp under hub-fetch-run.py.
- A sys.path strip that Python 3.13+'s pre-run linecache import bypasses.
- Provision the Composite-Actions Check's YAML Parser Through uvx #2359 settled on python -I -X utf8 for both children.

What not to repeat

Defending a python -c child by setting its working directory. The gate can run from a temp file, so the "trusted" directory was /tmp. Isolated mode (-I -X utf8) is the fix, and the tests in tests/test_repo_gate.py plant linecache.py for that reason.
A JSON round trip of a whole YAML tree. It expands shared aliases and diverges on key types. Project only the fields the check reads.
A per-file uvx call. A uvx failure then repeats per file, and a timeout orphans uv's child. Resolve the interpreter once and run the reader as a direct child.
Trusting a worker's full SHA. One READY report carried a fabricated full SHA beside a correct abbreviation. Read the head live before merging.

New learnings

Cap data and the case for one bounded edit in every finding class are on Let an Unattended Worker File a Wording-Only Finding Past the Two-Round Cap Instead of Parking #2353.
The process-level lessons are in this link's "What not to repeat", and the code carries its tests. Nothing here needed governance prose.

Round 50, closed 2026-10-23: Session Handoff [default]: Four Tracks Folded into One Standing Handoff, Downstream Adoption Next

This link folds the default, windows, host-instructions, and python-contract tracks into one standing default track, under the maintainer's decision on #3391 item 5. Every claim below is a pointer to re-read, not an answer.

Next steps, in priority order

Re-read each issue and pull request rather than trusting these summaries.

Downstream adoption of the 2.0.822 and 2.0.826 releases (default). Each step needs a fresh write grant for its repository.
- Add closed to the Merge-Bot Stub Types of Adopters Setting delete-branch #2336: five adopters add closed to their merge-bot stub types.
- homeassistant-purpleair adopts test-matrix (Add a Dependency-Version Matrix Axis to validate-task.yml for the Gates Stage #2334) for Decide How to Adopt the Hub's Reusable Release Chain homeassistant-purpleair#365.
- Re-carry the hub's skills with the next resync rather than one repository at a time.
Grant pull-requests: read So the Live comments Label Read Reaches Private Repositories #2354: grant pull-requests: read so the live label read reaches private repositories (default). Roll it out with the resync in step 1, since a stub without the grant fails at startup once its pin reaches a lint job that declares it.
install-tools.ps1: dotnet status doesn't see pre-release SDKs Visual Studio installs outside winget #694 closes through promotion PR Promote develop to main: Read Copilot's Emoji-Prefixed Review Effort Line #2580 (windows). Name a Newer dotnet SDK From Outside the Managed Package in the Windows Report #2578 merged to develop with Closes on promotion: #694. Promote develop to main: Read Copilot's Emoji-Prefixed Review Effort Line #2580 belongs to the auto-2487 lane (Session Handoff [auto-2487]: Fix pr_review.py per #2487 #2559). Done means install-tools.ps1: dotnet status doesn't see pre-release SDKs Visual Studio installs outside winget #694 is closed after Promote develop to main: Read Copilot's Emoji-Prefixed Review Effort Line #2580 merges. If it merged without the keyword, close install-tools.ps1: dotnet status doesn't see pre-release SDKs Visual Studio installs outside winget #694 by hand and cite Name a Newer dotnet SDK From Outside the Managed Package in the Windows Report #2578.
PlexCleaner adopts the python gate (python-contract). It runs in PlexCleaner's own session. Pin the validator to 2.0.737 or later and pass python-directories: RegressionTests on both callers. RegressionTests/ needs a tests/ suite and a mypy config first, with mypy's strict flags if mypy is its only CI checker. Done means CI is green and the (python.directories.declared) driftNote is deleted.
Fix the Stale Docstring on test_a_created_file_that_cannot_be_removed_is_named_as_incomplete #2499: fix the one stale test docstring (host-instructions). The docstring of test_a_created_file_that_cannot_be_removed_is_named_as_incomplete should describe the error naming the removal failure too.
Let an Unattended Handoff Run the Strict-Review Pass for a Worker That Cannot Dispatch #2342: the digest-handback line in unattended-handoff's return table (default), per the maintainer's answer recorded on the issue.
Refuse a Commented or Multi-Line python-directories Value in the Audit #2013: refuse a commented or multi-line python-directories value in the audit (python-contract). It is small and in the function Refuse an Escaped or Folded python-directories Value Wherever It Starts #2010 touched.
Bound the In-Process Composite-Actions Read Against an Alias-Expanded Step Name or Shell #2357 and Read Action Files or Name the Cause When UV_NO_CACHE Removes the Resolved PyYAML Interpreter #2358 (default). The in-process composite-actions read can hang on an alias-expanded step name or shell, and the UV_NO_CACHE host finding names a vanished path with an untested probe-start failure. Both touch repo_gate.py's composite-actions read, so take them in one lane.
Name the Mypy Suppression Mechanisms Beside Pyright's in python-codestyle #2004: name the mypy suppression mechanisms beside the pyright ones in python-codestyle (python-contract). Blocked on the auto-1947 lane, see Internal dependencies.
The Windows lane, oldest first (windows): Give the Windows python Tool a Family So an Off-Line Host Reads Correctly #1171, Validate the Interpreter the Agent-Safety Installer Writes Into settings.json #1172, checkout_provenance Passes an Absolute Path as a Git Pathspec, Unverified on Windows, Where a Failure Reports a Tracked Clean Copy as unknown #1419, A Parent Pinned to UTF-8 Spawning a Python Child Leaves the Child's Encoder on the Locale #1577, Requirement 8 reaches Linux and WSL only, with no measured answer for macOS or Windows #1625, skills_install.py runs a bare claude argv, which may miss a Windows .cmd shim #1754, Python devcontainer snippet: mount path join and mktemp -t portability #1763. Roll out global ~/.claude/CLAUDE.md safety rules to all machines #365 rewrites the Windows host's hooks, settings.json and CLAUDE.md, so ask before running it.
Judge before starting (host-instructions): Report reply's --all-identical Batch Count When a GraphQL Call Exits the Process #2421, codify Codify: Let the Real Operation Be the Check, and Close a Defect Class Through One Primitive #2431, and Add safety hooks for opencode and codex #781, each against GOVERNANCE.md "Trust Boundaries and Hardening Effort", since Report reply's --all-identical Batch Count When a GraphQL Call Exits the Process #2421 is an error-path report case.
Three Test Harnesses Build a Stand-In PATH With an Empty Fallback When PATH Is Unset #1862 then configure.sh Resolves a Registry Entry by Name Rather Than by owner/repo Identity #1894 (python-contract). Three Test Harnesses Build a Stand-In PATH With an Empty Fallback When PATH Is Unset #1862 is three test harnesses building a child PATH with an empty fallback. configure.sh Resolves a Registry Entry by Name Rather Than by owner/repo Identity #1894 is configure.sh matching a registry entry by name rather than by owner/repo. Neither is Python-contract work.
Close the Remaining CommonMark Gaps in prose_lint's Sentence-Length Block Parser #2341 (default), low value. Keep a hard scope cap.

External blockers

Write grants: every downstream step needs a fresh per-repository grant.
A native Windows host is needed for the bootstrap.ps1 check (Session Handoff [default]: WORKFLOW.md Corrections and the pyproject.toml Decision Released as 2.0.672 #1804) and for every Windows-lane fix. A fix is not proven until the Windows host or the hosted windows-latest job has run it. The uvx reader's -I -X utf8 behavior on Windows was reasoned about, not run.
Promote develop to main: Read Copilot's Emoji-Prefixed Review Effort Line #2580 is not this lane's to drive, edit, or merge unless the maintainer hands it over.
Recognize Copilot's Collapsible "N Open Findings" Overview Block in pr_review.py #2480: Copilot's collapsible "N open findings" overview still blocks pr_review.py on some rounds as UNRECOGNIZED_REVIEWER_OUTPUT. Each PR it hits needs its own maintainer answer to merge over it.
Each PlexCleaner adoption runs in another repository's session.

Internal dependencies

Grant pull-requests: read So the Live comments Label Read Reaches Private Repositories #2354 and the step 1 resync go together, as step 2 states.
Bound the In-Process Composite-Actions Read Against an Alias-Expanded Step Name or Shell #2357 and Read Action Files or Name the Cause When UV_NO_CACHE Removes the Resolved PyYAML Interpreter #2358 share repo_gate.py's composite-actions read.
Name the Mypy Suppression Mechanisms Beside Pyright's in python-codestyle #2004 edits .agents/skills/python-codestyle/SKILL.md, and the auto-1947 lane (handoff Session Handoff [auto-1947]: State the Python Build Profile by Its Trait, Not One of Its Forms #1981) has an unmerged branch rewriting that file's profile and local-loop sections. Land Name the Mypy Suppression Mechanisms Beside Pyright's in python-codestyle #2004 after that lane merges, or coordinate the rebase.
Use tests/host_capability.py and bash_or_skip before writing any new Windows-sensitive test. Pass a path into a PowerShell test body as environment or argv, never quoted into the script. To test an install-tools.ps1 branch running a native command, put a stub on PATH alone, as TestWindowsUntrackedDotnetNote in tests/test_install_tools.py does.

State

Re-derive all of this at resume with git worktree list, git log, and gh pr list.

Folded tracks and their last links: default Session Handoff [default]: #3671 Thread-Aware Held Wait Released in 2.0.1286, Carried Backlog Next #2361, windows Session Handoff [windows]: #2578 Landed for #694, Closing Through Promotion #2580, #1171 Next #2581, host-instructions Session Handoff [host-instructions]: Hardening Rule and Merge Guard Released as 2.0.867 and 2.0.868, #2499 Docstring Next #2514, python-contract Session Handoff [python-contract]: Escaped and Folded python-directories Values Refused, Released as 2.0.738 #2014. Their closed chains stay readable with chain.
Releases the folded links record: 2.0.738, 2.0.822, 2.0.826, 2.0.867, 2.0.868, plus whatever the windows lane landed. Open PR recorded: Promote develop to main: Read Copilot's Emoji-Prefixed Review Effort Line #2580 (develop to main, auto-2487 lane).
Unattended lanes untouched: auto-2176, auto-2167, auto-2157, auto-2080, auto-1947 (unmerged feature/auto-1947), and the other auto-* lanes with unmerged remote branches.
Windows open lane issues: install-tools.ps1: dotnet status doesn't see pre-release SDKs Visual Studio installs outside winget #694 (awaiting promotion), Roll out global ~/.claude/CLAUDE.md safety rules to all machines #365 (both hosts), Give the Windows python Tool a Family So an Off-Line Host Reads Correctly #1171, Validate the Interpreter the Agent-Safety Installer Writes Into settings.json #1172, checkout_provenance Passes an Absolute Path as a Git Pathspec, Unverified on Windows, Where a Failure Reports a Tracked Clean Copy as unknown #1419, A Parent Pinned to UTF-8 Spawning a Python Child Leaves the Child's Encoder on the Locale #1577, Requirement 8 reaches Linux and WSL only, with no measured answer for macOS or Windows #1625, skills_install.py runs a bare claude argv, which may miss a Windows .cmd shim #1754, Python devcontainer snippet: mount path join and mktemp -t portability #1763.
Other repositories written by the folded rounds: none.

The parked decision queue

The folded links recorded #2353 as the one decision issue (the wording-only exception to the two-round cap, filed by the HomeAutomation-Config session). Re-read the live decision label count and present the questions, since this consolidation asked none.

What the last round did

Consolidation: four tracks folded into this one standing track. One line per folded track follows.
- default Session Handoff [default]: #3671 Thread-Aware Held Wait Released in 2.0.1286, Carried Backlog Next #2361: released 2.0.822 and 2.0.826, merged Assert the Fallback Fetch URL in hub-fetch-run.py's Tests #2352, Tolerate a Per-Field workflowRun Error in the Check Query #2355, Read the Comments Label Live Rather Than off the Event Payload #2356, Provision the Composite-Actions Check's YAML Parser Through uvx #2359 through Promote develop to main: Live comments Label Read, uvx YAML Reader, and Review Digest Tolerance #2360, filed Grant pull-requests: read So the Live comments Label Read Reaches Private Repositories #2354, Bound the In-Process Composite-Actions Read Against an Alias-Expanded Step Name or Shell #2357, Read Action Files or Name the Cause When UV_NO_CACHE Removes the Resolved PyYAML Interpreter #2358.
- windows Session Handoff [windows]: #2578 Landed for #694, Closing Through Promotion #2580, #1171 Next #2581: landed Name a Newer dotnet SDK From Outside the Managed Package in the Windows Report #2578 for install-tools.ps1: dotnet status doesn't see pre-release SDKs Visual Studio installs outside winget #694 (Windows dotnet row reads dotnet --list-sdks), closed Both Menus Judge a Cached Hub Fresh Against main Whatever --ref Names #2541 by hand citing Judge a Cached Hub Fresh Against the Session's Own Ref in Both Menus #2571.
- host-instructions Session Handoff [host-instructions]: Hardening Rule and Merge Guard Released as 2.0.867 and 2.0.868, #2499 Docstring Next #2514: released 2.0.867 (hardening-effort rule, Encode Trust Boundaries So Hardening Effort Follows the Real Threat, Not the Maintainer's Own Inputs #2501) and 2.0.868 (--match-head-commit merge guard, Merge a Reported Promotion With --match-head-commit Set to the Reported Head, Not a Fresh Read #2505). Verify on Windows That fstat and lstat Agree on File Identity in write_regular_file's Cleanup #2494 and Stamp CLAUDE.md's Digest When It Is Written, Not Only When the Whole Install Completes #2498 closed as not planned, Fix the Stale Docstring on test_a_created_file_that_cannot_be_removed_is_named_as_incomplete #2499 narrowed.
- python-contract Session Handoff [python-contract]: Escaped and Folded python-directories Values Refused, Released as 2.0.738 #2014: released 2.0.738 (Refuse an Escaped or Folded python-directories Value Wherever It Starts #2010 refuses escaped or folded python-directories values), found Count a setup.cfg [mypy] Section as a Declared Directory's Type Checker #1915 already fixed by Settle the Python Gate's Behavior-Changing Edge Cases From #1936 #1942, filed Refuse a Commented or Multi-Line python-directories Value in the Audit #2013.

What not to repeat

Defending a python -c child by setting its working directory. The gate can run from a temp file, so the "trusted" directory was /tmp. Use isolated mode (-I -X utf8), and the tests in tests/test_repo_gate.py plant linecache.py for that reason.
A JSON round trip of a whole YAML tree. It expands shared aliases and diverges on key types. Project only the fields the check reads.
A per-file uvx call. A failure repeats per file and a timeout orphans uv's child. Resolve the interpreter once and run the reader as a direct child.
Trusting a worker's full SHA. One READY report carried a fabricated full SHA beside a correct abbreviation. Read the head live before merging.
A git push, or git commit ... >/dev/null, anywhere in a compound command run from the primary checkout. gh-write-guard refuses the whole command, so none of its steps run. Run the remote delete from a worktree and git worktree remove from the base clone, as separate commands.
Redirecting PowerShell streams as 2>&1 1>$null to read stderr alone. It merges stderr into stdout and discards both. Use Start-Process with -RedirectStandardOutput and -RedirectStandardError to two files.
Asserting a version selection with each tie listed in one order only. List the release-and-preview case in both orders.
Illustrating "cannot occur" with concurrency on a fleet host. Concurrent agent sessions are the fleet's normal deployment.
Keying a merge guard to whether a report exists. Key it to the head the go-ahead covers. An unattended worker also holds drive-pr's report and an attended re-ask makes it stale.
Budgeting a new rule section as converging in one pass. It took four local passes (9, 3, 3, 0 findings) plus one Copilot finding, all substantive.
Starting an open issue without checking whether it already landed. Run git log -S '<the fix's key string>' before opening a worktree, as Count a setup.cfg [mypy] Section as a Declared Directory's Type Checker #1915 was already fixed by Settle the Python Gate's Behavior-Changing Edge Cases From #1936 #1942.
Matching an input's shape on the key's own line. Read the value's text through the same indentation walk that finds the key.
A regression case that passes by accident. Mutation-test each new case against the old code and read the finding it produces, not only the count.

New learnings

pr_review.py reads "fleet-review" in a Copilot file-table summary as a coverage-marker mention, so a PR touching fleet-code-review reports COVERAGE_IS_UNSTATED despite a full table. The failure is loud and costs one prompt, so under the hardening-effort rule it was declined and not filed.
Cap data and the case for one bounded edit per finding class are on Let an Unattended Worker File a Wording-Only Finding Past the Two-Round Cap Instead of Parking #2353.

Round 51, closed 2026-10-24: Session Handoff [default]: The #3391 Review-Cost Program Shipped as 2.0.1148 and 2.0.1158, Shared Markdown Reader Next

Next steps, in priority order

Re-read each issue and pull request rather than trusting these summaries. This round ran the #3391 review-cost program end to end, so the earlier steps below are re-ranked beneath its follow-ups.

Move the shared Markdown region reading onto markdown-it-py for all four tools at once (default). spec/carried_text.py keeps the one hand-written fence and section reader because carry.py, build_dist.py, and canonical_review.py import it under plain python3. Done means each of the four carries PEP 723 metadata, every caller runs them as uv run --no-project --script, carried_text.py is deleted, and the full audit sweep's findings are unchanged before and after. Also folds in validate.py's own Markdown readers and the references() versus _LINK_DEF split the audit lane recorded on Move Audit Tests to tests/, Parse YAML and Markdown, Add Check Registry #3416.
Brief rule for lanes: a pull request body draft lives under a file named for its branch (default). Two lanes this round wrote pr-body.md in one scratchpad and one gh pr edit published the wrong body onto another lane's pull request, twice. One sentence in backlog-burndown and drive-pr where they tell a lane to draft the body. Done means the sentence is in both skills and the mirrors are regenerated.
Let an Earlier Partial Outrank an Unknown Head Reading in pr_review.py #3412, Read a Run-Block Heredoc Body Without Letting Its Apostrophes Hide Later Comments #3413, Skip Inline Code Spans When Numbering an Inline Markdown Comment #3414 (default). Three introduced findings left after edit budgets this round: a head round reading unknown beside an earlier partial, an apostrophe in a run-block heredoc hiding later comments, an inline comment numbered on an earlier code span's line. Each is small and has a constructed case on its issue.
Stale wording the lanes found and did not fix (default). .agents/skills/python-codestyle/references/profiles.md still calls the lint-only profile "stdlib-only" after the uv mechanism, and spec/project-types.json and README.md carry the same phrase for downstream lint-only subtrees. One pull request.
Weekly backlog report mailed from the automation host. HomeAutomation-Config Tracker: .agents/skills/shell-codestyle/SKILL.md #1353 carries the timer and the notify_send_report wiring. The hub's scripts/backlog_metrics.py is the renderer and its invocation is on that issue. Runs in that repository's session.
Compile-time C++ lint on ESPHome generated code. ESPHome-Config Remove TelemetryImporter from the catalog #256, wanted work, runs in that repository's session.
Downstream adoption carried from round 50 (default): Add closed to the Merge-Bot Stub Types of Adopters Setting delete-branch #2336 (five adopters add closed to their merge-bot stub types), Grant pull-requests: read So the Live comments Label Read Reaches Private Repositories #2354 (pull-requests: read for the live label read on private repositories), and homeassistant-purpleair test-matrix for Decide How to Adopt the Hub's Reusable Release Chain homeassistant-purpleair#365. Roll them into the next resync rather than one repository at a time.
PlexCleaner adopts the python gate. Runs in PlexCleaner's own session. Pin the validator to the current release and pass python-directories: RegressionTests on both callers, with a tests/ suite and a mypy config first. Done means CI is green and the (python.directories.declared) driftNote is deleted.
Bound the In-Process Composite-Actions Read Against an Alias-Expanded Step Name or Shell #2357 and Read Action Files or Name the Cause When UV_NO_CACHE Removes the Resolved PyYAML Interpreter #2358 (default). Both touch repo_gate.py's composite-actions read, so one lane.
Fix the Stale Docstring on test_a_created_file_that_cannot_be_removed_is_named_as_incomplete #2499, Let an Unattended Handoff Run the Strict-Review Pass for a Worker That Cannot Dispatch #2342, Refuse a Commented or Multi-Line python-directories Value in the Audit #2013, Name the Mypy Suppression Mechanisms Beside Pyright's in python-codestyle #2004 (default). Each small and independent. Name the Mypy Suppression Mechanisms Beside Pyright's in python-codestyle #2004 edits python-codestyle, so check the auto-1947 branch state first.
The Windows lane, oldest first: Give the Windows python Tool a Family So an Off-Line Host Reads Correctly #1171, Validate the Interpreter the Agent-Safety Installer Writes Into settings.json #1172, checkout_provenance Passes an Absolute Path as a Git Pathspec, Unverified on Windows, Where a Failure Reports a Tracked Clean Copy as unknown #1419, A Parent Pinned to UTF-8 Spawning a Python Child Leaves the Child's Encoder on the Locale #1577, Requirement 8 reaches Linux and WSL only, with no measured answer for macOS or Windows #1625, skills_install.py runs a bare claude argv, which may miss a Windows .cmd shim #1754, Python devcontainer snippet: mount path join and mktemp -t portability #1763. Roll out global ~/.claude/CLAUDE.md safety rules to all machines #365 rewrites the Windows host's hooks and settings, so ask before running it. A fix is not proven until a Windows host or the hosted windows-latest job has run it.
Judge before starting: Report reply's --all-identical Batch Count When a GraphQL Call Exits the Process #2421, Codify: Let the Real Operation Be the Check, and Close a Defect Class Through One Primitive #2431, Add safety hooks for opencode and codex #781 against GOVERNANCE.md "Trust Boundaries and Hardening Effort". Close the Remaining CommonMark Gaps in prose_lint's Sentence-Length Block Parser #2341, low value, hard scope cap. Three Test Harnesses Build a Stand-In PATH With an Empty Fallback When PATH Is Unset #1862 then configure.sh Resolves a Registry Entry by Name Rather Than by owner/repo Identity #1894.

External blockers

The agent-safety kit on every host reads STALE and only its own installer refreshes it. That installer rewrites hooks and settings, so the maintainer runs python3 host-setup/agent-safety/claude/install.py from a checkout at main. Until then each host runs the 21,000-line guard and lacks the Bash(uv run --no-project --script scripts/pr_review.py:*) allow rule, so auto mode may prompt on cleanup and reply.
The worktree sweep lists prunable entries from other sessions' scratch directories and one idle detached hub worktree. pr_review.py cleanup --sweep --apply reaches other sessions' finished work, so it waits for the maintainer's go-ahead.
Write grants: every downstream step needs a fresh per-repository grant.
A native Windows host for the Windows lane and the bootstrap.ps1 check (Session Handoff [default]: WORKFLOW.md Corrections and the pyproject.toml Decision Released as 2.0.672 #1804).

Internal dependencies

Step 1 before any further Markdown reading change in carry.py, canonical_review.py, or validate.py, since each would otherwise add a reader step 1 deletes.
Step 2 before the next multi-lane dispatch, since the collision recurs under the same brief.
Steps 7 and 8 ride the next resync, which also carries the new hub_fetch_run.py to any repository that later adopts the hook snippet.
Name the Mypy Suppression Mechanisms Beside Pyright's in python-codestyle #2004 waits on the auto-1947 branch, as in round 50.

State

Re-derive all of this at resume with git worktree list, git log, gh pr list, and gh release list.

Releases this round: 2.0.1148 (promotion Promote Write Guard Threat Model and Review Cost Changes to Main #3405: guard threat model State the Write Guard's Threat Model and Document Two Allowed Shapes #3399, pre-existing findings in the body List Pre-Existing Review Findings in the PR Body Instead of Filing Them #3401, standing handoff issue Keep One Standing Handoff Issue per Track #3402, backlog metrics Add Weekly Backlog and Review-Cost Report Script #3400, Copilot at the promotion only Run Copilot Review Once per Change at the Promotion #3403, aligned rule text Align Review Rule Text with Promotion-Only Copilot Review #3404, fix Fix Promotion Review Findings on Local Pass and Backlog Report #3406) and 2.0.1158 (promotion Promote the Write Guard Rewrite, uv Scripts, and Parser Gates to Main #3417: guard rewrite Rewrite the Write Guard as a Small Lexer With Plain-Shape Rules #3407, CODESTYLE rules and uv run --no-project --script Add Test and Parser Rules and Run Hub Scripts Through uv #3409, pr_review split Split pr_review.py into a Gate, a Copilot Adapter, and a Cleanup Script #3410, prose gate on Pygments and markdown-it-py Restructure the Prose Gate into Extract, Normalize, and Evaluate Layers #3411, acknowledgment hold Hold the Merge Gate on Unplaced Review Text Until Acknowledged #3415, audit refactor Move Audit Tests to tests/, Parse YAML and Markdown, Add Check Registry #3416, fixes Fix Promotion Review Findings on Escalation, Test Deps, and uvx Assert #3418 Read an Oversized Prose Coverage Count as Unknown in the Copilot Adapter #3419 Pin the Prose Gate and Its Data to One Commit via git ls-remote #3420 Read an Oversized Finding Count as Unknown in the Copilot Adapter #3421).
Primary checkout: clean on develop at origin/develop, level with main after Promote the Write Guard Rewrite, uv Scripts, and Parser Gates to Main #3417.
Skills snapshot on this machine: current at the promoted commit. Kit: STALE, see blockers.
Rulesets: Copilot's automatic review is removed from the develop ruleset on the hub and 21 other cataloged repositories through configure.sh apply. main keeps it.
Other repositories written this round: Financial-Modeling (35 issues closed, standing handoff Migrate MediaTools to the plan-job publisher so its registry entry records reality #1478, live 401k track Tracker: WORKFLOW.md > 5. Test Methodology #1370 left), HomeAutomation-Config (6 closed, standing handoff Tracker: catalog/snippets/workflows/merge-bot-pull-request.yml #1358, issue Tracker: .agents/skills/shell-codestyle/SKILL.md #1353 filed), ESPHome-Config (Remove TelemetryImporter from the catalog #256 filed, standing handoff Adopt the job-suffix naming for the PR-gate aggregator #237), ESPHome-NonRoot (pull request Add agent safety rules to shared AGENTS.md (repo-scope + no writes-as-probes) #364 merged into its develop: the new hub_fetch_run.py carried byte-identical, uv added to its devcontainer, its OPERATIONS.md corrected, not yet promoted there), and twenty small repositories (25 issues closed, handoffs folded to one default each).
Handoff tracks folded this round: hub engineering Session Handoff [engineering]: Callout Tier, Subprocess Rule, and Guard Allow Gaps Released as 2.0.1141, Review-Cost Research Next #3398 (closed by the promotion merge before the fold, its one open step was Move the Remaining Prose-Matched Subprocess Failures Onto the Status-Code Rule #3378), windows Session Handoff [windows]: #2578 Landed for #694, Closing Through Promotion #2580, #1171 Next #2581, host-instructions Session Handoff [host-instructions]: Hardening Rule and Merge Guard Released as 2.0.867 and 2.0.868, #2499 Docstring Next #2514, python-contract Session Handoff [python-contract]: Escaped and Folded python-directories Values Refused, Released as 2.0.738 #2014. Fleet-wide 32 tracks folded.
Lane briefs for the program live outside the repository under the maintainer's plans directory, named guard-rewrite, pr-review-split, prose-gate-refactor, audit-refactor, handoff-consolidation.

The parked decision queue

Zero open decision issues in this repository at close, read live. The two filed this round (#3382, #3385) closed through #3399 at promotion #3405.

What the last round did

Posted the Review Anthropic's Code Review Practices Against the Fleet Review Workflow for Cost and Backlog #3391 comparison, the ranked proposal, and the landed record.
Closed 240 hub issues (31 guard out of scope, 49 frozen prose, 151 under the fleet audit, 9 unworked Why the Backlog Grows: Prose Review Has No Fixed Point, and a Plan to Keep the Rigor Without It #1311 items, Why the Backlog Grows: Prose Review Has No Fixed Point, and a Plan to Keep the Rigor Without It #1311, P0: Triage the Open Backlog Into a Label Set, One Tracker Per Canonical Unit, and One Decision List #1315, 4 guard shapes the rewrite covers) and 66 downstream issues, each with the decision comment. Hub open non-handoff: 387 to 150.
Merged 17 hub pull requests into develop and 2 promotions to main, driven by 9 Opus lanes, each filing zero issues under the new outcome 4.
Filed: HomeAutomation-Config Tracker: .agents/skills/shell-codestyle/SKILL.md #1353, ESPHome-Config Remove TelemetryImporter from the catalog #256, hub Run the Prose Gate Through uv in hub_fetch_run.py and the Local Shim #3408 (closed by Restructure the Prose Gate into Extract, Normalize, and Evaluate Layers #3411 at promotion), Let an Earlier Partial Outrank an Unknown Head Reading in pr_review.py #3412, Read a Run-Block Heredoc Body Without Letting Its Apostrophes Hide Later Comments #3413, Skip Inline Code Spans When Numbering an Inline Markdown Comment #3414 by lanes.
Consolidated every repository's handoffs to one standing default issue.

What not to repeat

One shared pr-body.md across concurrent lanes. Two incidents, each publishing a foreign draft onto a pull request for seconds, restored from edit history. Step 2 above.
Defining a file and running a guarded write in one Bash call. gh-write-guard denied the whole command, the file never existed, and 217 closes ran with an empty comment, which cost two retry loops. Write the file in its own call.
A tight loop of gh issue comment writes. GitHub's secondary limit fails about half with "was submitted too quickly". Space them five seconds apart and retry from the loop's own FAILED lines.
Briefing uv run --script without --no-project. Inside a checkout it syncs the project and writes uv.lock and .venv. Lane G measured it.
Assuming a script is only run, never imported. The audit brief's parser swap would have broken three importers under python3. Grep for importers before giving a script a dependency block.
A security-reducing brief without the maintainer's words in the transcript. The classifier refused the guard-rewrite dispatch as "Security Weaken" until he approved it in a message.
Landing a develop pull request while a promotion is in review. It moves the promotion's content and costs a Copilot round. Lanes held at mergeable until Promote Write Guard Threat Model and Review Cost Changes to Main #3405 merged, which worked.
The guard's own over-denials while rewriting it. The old live guard denied heredocs holding example commands, the signing probe's git -C "$d", and a cd <worktree> && git push compound. Lanes used the Write tool and ran commands singly.
An Opus lane on a 20-commit refactor runs five to six hours (the audit lane, 334 tool uses before its first hand-back) and can hit a weekly or monthly spend limit mid-review. Resume it with a message rather than redispatching.

New learnings

The write guard, not prose, was the hub's issue pump in October: 135 issues and 41 pull requests in one month from a hand-written bash reader growing toward the full grammar. The fix was a threat model stated in the README and a 2,500-line ceiling test, recorded in host-setup/agent-safety/README.md and tests/test_guard_size.py.
Two fleet rules now in CODESTYLE.md: tests live in tests/ and a deployed script keeps only a smoke --selftest, and a program reading a structured format uses a maintained parser. Mechanism: PEP 723 metadata and uv run --no-project --script.
Blocking until answered rather than until parsed: pr-review-conduct Merge Gate item 4 holds on unplaced Copilot text until an owner acknowledges it with comment --fingerprint, and the bot's vocabulary is a JSON edit.
Only ESPHome-NonRoot carried hub_fetch_run.py, so the prose gate's dependency change broke one repository's local hook, not twenty. Check which repositories carry a snippet before sizing a rollout.
Orchestration lessons for this host are in the maintainer's memory file on parallel lanes, since they are environment nuance rather than fleet rules.

Round 52, closed 2026-10-24: Session Handoff [default]: Five Fixes Released as 2.0.1163, Prose Gate and pr_review Follow-Ups Next

Next steps, in priority order

Re-read each issue and pull request rather than trusting these summaries.

Prose gate follow-ups, one lane (default). Skip the Run-Block Heredoc Split for a Non-Shell Step #3430 (skip the run-block heredoc split for a non-shell step), Read a Heredoc Whose Delimiter Opens on a Digit in the Prose Gate #3435 (a delimiter opening on a digit, such as <<123), Read a Second Heredoc on One Opener Line Without Letting Its Apostrophes Hide Later Comments #3428 (a second heredoc on one opener line), Number Duplicate Inline Comments Correctly Past a Stray Backtick #3429 (duplicate inline comments past a stray backtick). All four edit .github/actions/prose-gate/prose_lint.py, so they share one lane. The gate is 1839 lines against a ceiling the maintainer set at 1840 on Lex Heredoc Bodies Apart and Skip Code Spans When Numbering Inline Comments #3427, so the lane compacts as it goes or asks before moving the ceiling again. Done means each issue's constructed case is a test that fails before and passes after.
pr_review.py and pr_cleanup.py follow-ups (default). Re-Read a Deleted Ref Before Reporting DELETE_NOT_CONFIRMED in pr_cleanup #3436 (cleanup re-reads a deleted ref before reporting DELETE_NOT_CONFIRMED) is in pr_cleanup.py. Fix the COVERAGE_IS_PARTIAL Status Text on the Carry Paths in pr_review.py #3424 (the COVERAGE_IS_PARTIAL text on the carry paths) is in pr_review.py, which sits at 2999 of its 3000-line ceiling, so its fix must shrink the file or ask first.
Read validate.py's Markdown and the README Link Definitions Through markdown-it-py #3426, validate.py onto markdown-it-py (default). The remainder Read Carried Markdown Through markdown-it-py and Delete carried_text #3431 deferred, including the references() versus _LINK_DEF split. repo-config/configure.sh runs spec/resolve_description.py under a probed Python 3.7+ and that imports validate, so a dependency block there breaks it unless that caller moves too. Needs its own before-and-after comparison of the carried-link check.
Refuse an Include Marker Indented Past Three Spaces Rather Than Skipping It #3432 (default). build_dist.py silently skips an include marker indented past three spaces. Small, pre-existing.
Stale "stdlib-only" wording (default). .agents/skills/python-codestyle/references/profiles.md, spec/project-types.json, and README.md still call the lint-only profile stdlib-only after the uv mechanism. One pull request.
Downstream, each in its own repository's session: HomeAutomation-Config Tracker: .agents/skills/shell-codestyle/SKILL.md #1353 (weekly backlog report), ESPHome-Config Remove TelemetryImporter from the catalog #256 (C++ lint on generated code), PlexCleaner adopting the python gate, and the round-50 adoption set Add closed to the Merge-Bot Stub Types of Adopters Setting delete-branch #2336, Grant pull-requests: read So the Live comments Label Read Reaches Private Repositories #2354, and homeassistant-purpleair Roll out global ~/.claude/CLAUDE.md safety rules to all machines #365 rolled into the next resync. ESPHome-NonRoot Add agent safety rules to shared AGENTS.md (repo-scope + no writes-as-probes) #364 is merged into its develop and not yet promoted there.
Carried hub backlog (default): Bound the In-Process Composite-Actions Read Against an Alias-Expanded Step Name or Shell #2357 with Read Action Files or Name the Cause When UV_NO_CACHE Removes the Resolved PyYAML Interpreter #2358 (one lane, repo_gate.py), then Fix the Stale Docstring on test_a_created_file_that_cannot_be_removed_is_named_as_incomplete #2499, Let an Unattended Handoff Run the Strict-Review Pass for a Worker That Cannot Dispatch #2342, Refuse a Commented or Multi-Line python-directories Value in the Audit #2013, Name the Mypy Suppression Mechanisms Beside Pyright's in python-codestyle #2004 (check the auto-1947 branch state first). Judge before starting: Report reply's --all-identical Batch Count When a GraphQL Call Exits the Process #2421, Codify: Let the Real Operation Be the Check, and Close a Defect Class Through One Primitive #2431, Add safety hooks for opencode and codex #781 against GOVERNANCE.md "Trust Boundaries and Hardening Effort", Close the Remaining CommonMark Gaps in prose_lint's Sentence-Length Block Parser #2341 with a hard scope cap, Three Test Harnesses Build a Stand-In PATH With an Empty Fallback When PATH Is Unset #1862 then configure.sh Resolves a Registry Entry by Name Rather Than by owner/repo Identity #1894.
The Windows lane, oldest first: Give the Windows python Tool a Family So an Off-Line Host Reads Correctly #1171, Validate the Interpreter the Agent-Safety Installer Writes Into settings.json #1172, checkout_provenance Passes an Absolute Path as a Git Pathspec, Unverified on Windows, Where a Failure Reports a Tracked Clean Copy as unknown #1419, A Parent Pinned to UTF-8 Spawning a Python Child Leaves the Child's Encoder on the Locale #1577, Requirement 8 reaches Linux and WSL only, with no measured answer for macOS or Windows #1625, skills_install.py runs a bare claude argv, which may miss a Windows .cmd shim #1754, Python devcontainer snippet: mount path join and mktemp -t portability #1763. Roll out global ~/.claude/CLAUDE.md safety rules to all machines #365 rewrites the Windows host's hooks and settings, so ask before running it.

External blockers

The worktree sweep (pr_review.py cleanup --sweep) lists prunable entries from other sessions' scratch directories and two removable detached worktrees. --apply waits for the maintainer's go-ahead, since it reaches other sessions' finished work. One listed worktree was idle under an hour at round close and may be a live peer session.
The auto-mode classifier refuses a standard-tier lane's develop merge as "Merge Without Review" even with the gate green. Both Sonnet lanes stopped at READY, and the main seat merged after a maintainer prompt. Budget one prompt per such lane.
Write grants for every downstream step, and a native Windows host for the Windows lane and Session Handoff [default]: WORKFLOW.md Corrections and the pyproject.toml Decision Released as 2.0.672 #1804.

Internal dependencies

Step 1 lands as one lane, since all four issues edit one file and the ceiling is shared.
Step 3 waits on nothing, but any further Markdown-reading change to validate.py waits on it.
Fix the COVERAGE_IS_PARTIAL Status Text on the Carry Paths in pr_review.py #3424 needs line budget in pr_review.py, so it pairs well with any change that shrinks that file.
Name the Mypy Suppression Mechanisms Beside Pyright's in python-codestyle #2004 waits on the auto-1947 branch, as in rounds 50 and 51.

State

Re-derive all of this at resume with git worktree list, git log, gh pr list, and gh release list.

Release this round: 2.0.1163, promotion Promote the markdown-it Shared Reader, Heredoc Lexing, and Exit-42 Fixes to Main #3434 at merge commit 56ea2d4, carrying Name a Pull Request Body Draft After Its Repository and Branch #3422, Let an Earlier Partial Outrank an Unknown Head Reading in pr_review #3423, Lex Heredoc Bodies Apart and Skip Code Spans When Numbering Inline Comments #3427, Read Carried Markdown Through markdown-it-py and Delete carried_text #3431, Say When an Earlier Partial Round Exits 42 in pr_review Docs #3433. It closed Let an Earlier Partial Outrank an Unknown Head Reading in pr_review.py #3412, Read a Run-Block Heredoc Body Without Letting Its Apostrophes Hide Later Comments #3413, Skip Inline Code Spans When Numbering an Inline Markdown Comment #3414, Update pr-review-conduct to Say an Unbelievable Head Reading Beside a Partial Exits 42 #3425.
Primary checkout: clean on develop at origin/develop 4c77db2. No open pull requests at close.
This host: the skills snapshot is current at 56ea2d4, and the agent-safety kit reads CURRENT.
Worktrees: every worktree this round created is removed. The remote feature/auto-* branches and feature/wsl-2299 predate this round.
Other repositories: none written this round.

The parked decision queue

Zero open decision issues in this repository at close, read live. The decisions this round asked were answered in session and recorded on their pull requests: merge #3423 and #3427 from the main seat, accept the prose gate's ceiling at 1840, defer #3429 and #3430, and merge and release #3434.

What the last round did

Installed the agent-safety kit on this host from a main checkout, which added the uv run --no-project --script scripts/pr_review.py allow rule, and refreshed the skills snapshot from main.
Merged five pull requests into develop: Name a Pull Request Body Draft After Its Repository and Branch #3422 (body draft per repository and branch, step 2 of round 51), Let an Earlier Partial Outrank an Unknown Head Reading in pr_review #3423 (Let an Earlier Partial Outrank an Unknown Head Reading in pr_review.py #3412), Lex Heredoc Bodies Apart and Skip Code Spans When Numbering Inline Comments #3427 (Read a Run-Block Heredoc Body Without Letting Its Apostrophes Hide Later Comments #3413 and Skip Inline Code Spans When Numbering an Inline Markdown Comment #3414), Read Carried Markdown Through markdown-it-py and Delete carried_text #3431 (step 1 of round 51, the shared markdown-it reader), Say When an Earlier Partial Round Exits 42 in pr_review Docs #3433 (Update pr-review-conduct to Say an Unbelievable Head Reading Beside a Partial Exits 42 #3425). Promoted as Promote the markdown-it Shared Reader, Heredoc Lexing, and Exit-42 Fixes to Main #3434 and released as 2.0.1163.
Filed: Fix the COVERAGE_IS_PARTIAL Status Text on the Carry Paths in pr_review.py #3424, Update pr-review-conduct to Say an Unbelievable Head Reading Beside a Partial Exits 42 #3425, Read validate.py's Markdown and the README Link Definitions Through markdown-it-py #3426, Read a Second Heredoc on One Opener Line Without Letting Its Apostrophes Hide Later Comments #3428, Number Duplicate Inline Comments Correctly Past a Stray Backtick #3429, Skip the Run-Block Heredoc Split for a Non-Shell Step #3430, Refuse an Include Marker Indented Past Three Spaces Rather Than Skipping It #3432 by lanes, Read a Heredoc Whose Delimiter Opens on a Digit in the Prose Gate #3435 from the promotion review, Re-Read a Deleted Ref Before Reporting DELETE_NOT_CONFIRMED in pr_cleanup #3436 from cleanup.

What not to repeat

A lane recording a local-review receipt over commits its pass never read. The prose-gate lane read the digest after two later commits and recorded it under the earlier pass. A fresh Opus pass over the real head then found a regression and a better design. The lane brief now says to read the digest after the last commit, review that, then record. Keep that line in every lane brief.
A standard lane on a lexer change. The Sonnet lane wrote a regex-and-parity heredoc guesser where the file already had a shell-state lexer, which cost a full rework. Brief a parser change with the existing lexer named, or run it at frontier.
Rewording a condition without a state table. Say When an Earlier Partial Round Exits 42 in pr_review Docs #3433 took four review rounds because each wording fixed one head state and broke another. Brief the reviewer, or enumerate in the draft, every input state the sentence must hold for, before the first push.
Piping a gate into tail -2. It hid five of seven sentence-length hits. Read a gate's output whole.

New learnings

pr_cleanup.py's ref read-back can lag a landed delete. Recorded as Re-Read a Deleted Ref Before Reporting DELETE_NOT_CONFIRMED in pr_cleanup #3436, and rerunning cleanup completes the local half.
The stock Pygments heredoc rule takes \w+ while the gate's patched rule takes a letter first. Recorded on Read a Heredoc Whose Delimiter Opens on a Digit in the Prose Gate #3435.

Round 53, closed 2026-10-25: Session Handoff [default]: Prose Gate Heredoc and Backtick Fixes Released as 2.0.1166, Size Guard Redesign Next

Next steps, in priority order

Re-read each issue and pull request rather than trusting these summaries.

Replace the Prose Gate's Line Ceiling with a Check for Vocabulary in the Script #3441, replace the prose gate's line ceiling (default). The maintainer chose this round to replace test_the_gate_stays_under_its_size_ceiling's line count with a check for vocabulary or exemption data in prose_lint.py, which is what its docstring says the cap guards. Until it lands, the gate sits at exactly its 1862 ceiling with no headroom, so any prose_lint.py change compacts to fit and never asks to raise it. Done means the line-count assertion is gone and the new check fails on a constructed script carrying a vocabulary list.
pr_review.py and pr_cleanup.py follow-ups (default). Re-Read a Deleted Ref Before Reporting DELETE_NOT_CONFIRMED in pr_cleanup #3436 (cleanup re-reads a deleted ref before reporting DELETE_NOT_CONFIRMED) is in pr_cleanup.py. Fix the COVERAGE_IS_PARTIAL Status Text on the Carry Paths in pr_review.py #3424 (the COVERAGE_IS_PARTIAL text on the carry paths) is in pr_review.py, which sits at 2999 of its 3000-line ceiling, so its fix shrinks the file. Do not ask to raise that ceiling either.
Read validate.py's Markdown and the README Link Definitions Through markdown-it-py #3426, validate.py onto markdown-it-py (default). The remainder Read Carried Markdown Through markdown-it-py and Delete carried_text #3431 deferred, including the references() versus _LINK_DEF split. repo-config/configure.sh runs spec/resolve_description.py under a probed Python 3.7+, which imports validate, so a dependency block there breaks it unless that caller moves too.
Refuse an Include Marker Indented Past Three Spaces Rather Than Skipping It #3432 (default). build_dist.py silently skips an include marker indented past three spaces. Small, pre-existing.
Stale "stdlib-only" wording (default). .agents/skills/python-codestyle/references/profiles.md, spec/project-types.json, and README.md still call the lint-only profile stdlib-only after the uv mechanism. One pull request.
Downstream, each in its own repository's session: HomeAutomation-Config Tracker: .agents/skills/shell-codestyle/SKILL.md #1353, ESPHome-Config Remove TelemetryImporter from the catalog #256, PlexCleaner adopting the python gate, and the round-50 adoption set Add closed to the Merge-Bot Stub Types of Adopters Setting delete-branch #2336, Grant pull-requests: read So the Live comments Label Read Reaches Private Repositories #2354, and homeassistant-purpleair Roll out global ~/.claude/CLAUDE.md safety rules to all machines #365 rolled into the next resync. ESPHome-NonRoot Add agent safety rules to shared AGENTS.md (repo-scope + no writes-as-probes) #364 is merged into its develop and not yet promoted there.
Carried hub backlog (default): Bound the In-Process Composite-Actions Read Against an Alias-Expanded Step Name or Shell #2357 with Read Action Files or Name the Cause When UV_NO_CACHE Removes the Resolved PyYAML Interpreter #2358 (one lane, repo_gate.py), then Fix the Stale Docstring on test_a_created_file_that_cannot_be_removed_is_named_as_incomplete #2499, Let an Unattended Handoff Run the Strict-Review Pass for a Worker That Cannot Dispatch #2342, Refuse a Commented or Multi-Line python-directories Value in the Audit #2013, Name the Mypy Suppression Mechanisms Beside Pyright's in python-codestyle #2004 (check the auto-1947 branch state first). Judge before starting: Report reply's --all-identical Batch Count When a GraphQL Call Exits the Process #2421, Codify: Let the Real Operation Be the Check, and Close a Defect Class Through One Primitive #2431, Add safety hooks for opencode and codex #781 against GOVERNANCE.md "Trust Boundaries and Hardening Effort", Close the Remaining CommonMark Gaps in prose_lint's Sentence-Length Block Parser #2341 with a hard scope cap, Three Test Harnesses Build a Stand-In PATH With an Empty Fallback When PATH Is Unset #1862 then configure.sh Resolves a Registry Entry by Name Rather Than by owner/repo Identity #1894.
The Windows lane, oldest first: Give the Windows python Tool a Family So an Off-Line Host Reads Correctly #1171, Validate the Interpreter the Agent-Safety Installer Writes Into settings.json #1172, checkout_provenance Passes an Absolute Path as a Git Pathspec, Unverified on Windows, Where a Failure Reports a Tracked Clean Copy as unknown #1419, A Parent Pinned to UTF-8 Spawning a Python Child Leaves the Child's Encoder on the Locale #1577, Requirement 8 reaches Linux and WSL only, with no measured answer for macOS or Windows #1625, skills_install.py runs a bare claude argv, which may miss a Windows .cmd shim #1754, Python devcontainer snippet: mount path join and mktemp -t portability #1763. Roll out global ~/.claude/CLAUDE.md safety rules to all machines #365 rewrites the Windows host's hooks and settings, so ask before running it.

External blockers

A nested hub clone, hub-resync, under the fleet worktrees directory was left in place. It is clean with nothing unpushed. The harness's rm safety check refused its removal because the shell's working directory sat inside it at the time, and that refusal is the maintainer's to clear, so remove it by hand.
Write grants for every downstream step, and a native Windows host for the Windows lane and Session Handoff [default]: WORKFLOW.md Corrections and the pyproject.toml Decision Released as 2.0.672 #1804.

Internal dependencies

Step 1 lands before any other prose_lint.py change, since every one of them meets the zero-headroom ceiling until it does.
Step 3 waits on nothing, but any further Markdown-reading change to validate.py waits on it.
Fix the COVERAGE_IS_PARTIAL Status Text on the Carry Paths in pr_review.py #3424 needs line budget in pr_review.py, so it pairs well with a change that shrinks that file.
Name the Mypy Suppression Mechanisms Beside Pyright's in python-codestyle #2004 waits on the auto-1947 branch, as in rounds 50 to 52.

State

Re-derive all of this at resume with git worktree list, git log, gh pr list, and gh release list.

Release this round: 2.0.1166, promotion Promote the Prose Gate's Digit, Second, and Non-Shell Heredoc and Backtick Fixes to Main #3438 at merge commit 63cf1c9, carrying Read Digit-Labelled and Second Heredocs, Skip Non-Shell Run Splits, and Pair Backticks as CommonMark Does #3437, Lex Each of Two Heredoc Bodies Alone and Keep Autolinks Out of the Code-Span Mask #3439, Say Which HEREDOC Branch Drops a Digit Label #3440. It closed Read a Second Heredoc on One Opener Line Without Letting Its Apostrophes Hide Later Comments #3428, Number Duplicate Inline Comments Correctly Past a Stray Backtick #3429, Skip the Run-Block Heredoc Split for a Non-Shell Step #3430, Read a Heredoc Whose Delimiter Opens on a Digit in the Prose Gate #3435.
Primary checkout: clean on develop at origin/develop 687ccd9. No open pull requests at close.
This host: the skills snapshot is current at 63cf1c9, and the agent-safety kit reads CURRENT.
Worktrees: every worktree this round created is removed. The cleanup sweep ran --apply on the hub, Financial-Modeling, HomeAutomation-Config, LanguageTags, and loopj-home-assistant-vantage, and 52 idle, clean, fully pushed standalone hub clones were deleted from the fleet worktrees directory. A sweep keeps anything idle under 24 hours or holding unmerged commits, so the downstream repositories' remaining worktrees are their own sessions' work.
Other repositories: remote feature branches of merged pull requests were deleted by the sweep in Financial-Modeling and HomeAutomation-Config. Nothing else was written.

The parked decision queue

Zero open decision issues in this repository at close, read live. The decisions this round asked were answered in session: raise the prose gate's ceiling to 1862 for #3437, then no further raise, a compaction holding it there for #3439, then replace the line ceiling with a vocabulary check (#3441), and merge and release #3438.

What the last round did

Swept worktrees fleet-wide, per State above.
Merged three pull requests into develop: Read Digit-Labelled and Second Heredocs, Skip Non-Shell Run Splits, and Pair Backticks as CommonMark Does #3437 (Skip the Run-Block Heredoc Split for a Non-Shell Step #3430, Read a Heredoc Whose Delimiter Opens on a Digit in the Prose Gate #3435, Read a Second Heredoc on One Opener Line Without Letting Its Apostrophes Hide Later Comments #3428, Number Duplicate Inline Comments Correctly Past a Stray Backtick #3429, step 1 of round 52), Lex Each of Two Heredoc Bodies Alone and Keep Autolinks Out of the Code-Span Mask #3439 (the two findings the promotion review raised, two heredoc bodies lexed together and autolink backticks), Say Which HEREDOC Branch Drops a Digit Label #3440 (a comment naming which HEREDOC branch drops a digit label). Promoted as Promote the Prose Gate's Digit, Second, and Non-Shell Heredoc and Backtick Fixes to Main #3438 and released as 2.0.1166.
Filed: Replace the Prose Gate's Line Ceiling with a Check for Vocabulary in the Script #3441.

What not to repeat

Asking the maintainer to raise a size ceiling a few lines at a time. This round asked twice inside one lane and the maintainer pushed back: a ceiling moved per fix is a moving target. Compact to fit, and where compaction cannot reach, say so once with the design question rather than another number.
A digit-label heredoc rule taken from the issue as written. Read a Heredoc Whose Delimiter Opens on a Digit in the Prose Gate #3435 suggested \w[\w-]*, which would have read every (( x << 2 )), $[1<<2], and a[1<<2]=v shift as a heredoc. Probe the real shell and the lexer on the shapes before taking an issue's proposed fix.
A second-opener rule that scans the whole opener line. The first version let a <<EOF inside a quote or a comment queue a body that ran to a later heredoc's EOF. The frontier local pass caught it. Keep that pass on every regex change.
Assuming Copilot's reading of a nested regex conditional. It twice read |(?(3).*|(?!)) as the else branch of (?(7)...). A direct re.match with the groups printed disproves that in one command, and outcome 5's comment fix (Say Which HEREDOC Branch Drops a Digit Label #3440) is the response to a recurrence.
gh run list --jq with --arg. gh's --jq takes no jq arguments, and the loop silently matched nothing. Filter with jq itself, or use the run URL gh workflow run prints.

New learnings

A CommonMark-aligned mask is best built from markdown-it-py's own grammar (HTML_TAG_RE, AUTOLINK_RE's classes) rather than a hand-rolled copy, and the patterns hold no capturing group today, which the code-span \1 depends on. Recorded in Read Digit-Labelled and Second Heredocs, Skip Non-Shell Run Splits, and Pair Backticks as CommonMark Does #3437 and Lex Each of Two Heredoc Bodies Alone and Keep Autolinks Out of the Code-Span Mask #3439.
Re-matching a lexer token's own text with the regex that produced it is exact here only because HEREDOC uses no ^, $, or lookbehind before its start. Recorded in Lex Each of Two Heredoc Bodies Alone and Keep Autolinks Out of the Code-Span Mask #3439.

Round 54, closed 2026-10-25: Session Handoff [default]: pr_review.py Split Into Leaf Readers and the Held-Wait Run Hold on Develop, Promotion Next

Next steps, in priority order

Re-read each issue and pull request rather than trusting these summaries.

Promote develop to main (default). develop carries 11 unreleased merges, among them Split pr_review.py's Check and Round Readers Into Leaf Modules #3462 (the pr_review.py split) and Hold a Held Wait While a Head's Workflow Run Has Not Completed #3470 (the held-wait run hold). Drive a develop -> main promotion pull request with drive-pr to mergeable. Its Closes lines carry each feature pull request's "Closes on promotion" issue, pr_review.py wait Reports Checks Complete Before a Fresh Head's Jobs Register #3451, Correct the Lint Snippets' Claim That the Hook's EOL Gate Is Diff-Scoped #3369 and Move the Remaining Prose-Matched Subprocess Failures Onto the Status-Code Rule #3378 among them. Merging and dispatching the release stay merge-and-release, each on an explicit go-ahead. Done means the promotion is mergeable, or merged and released if the maintainer says so.
Replace the Prose Gate's Line Ceiling with a Check for Vocabulary in the Script #3441, replace the prose gate's line ceiling (default). Unchanged from round 53. The maintainer chose to replace test_the_gate_stays_under_its_size_ceiling's line count with a check for vocabulary or exemption data in prose_lint.py. Until it lands, the gate has no headroom, so a prose_lint.py change compacts to fit and never asks to raise it.
pr_review.py and pr_cleanup.py follow-ups, one at a time (default). The split left room, so these no longer shrink the file to fit. Take them in this order:
- Fix the COVERAGE_IS_PARTIAL Status Text on the Carry Paths in pr_review.py #3424, the COVERAGE_IS_PARTIAL text, which round 53 held for budget.
- Re-Read a Deleted Ref Before Reporting DELETE_NOT_CONFIRMED in pr_cleanup #3436, in pr_cleanup.py.
- Stop pr_review.py cleanup From Loading a Second Copy of the Gate #3461, cleanup loading a second copy of the gate. This is a real module-state hazard.
- A Held Wait Exits 0 Under a DIRTY or BEHIND Merge Whose Required Checks Never Ran #3471, exit 0 where the required checks never ran: a DIRTY or BEHIND merge, and an unreadable rollup per its comment.
- Render pr_review.py's Digest From a Pure Leaf and Build Its Header From a Field List #3457 to Give the pr_review Readers One Accessor Each for Commit Oid, Page Info, and Head Reviews #3460, the design survey's follow-ups: the digest renderer leaf and header field list, wait out of main with a flag table, the gate's repeated shapes as helpers, and the payload accessors.
Read validate.py's Markdown and the README Link Definitions Through markdown-it-py #3426, validate.py onto markdown-it-py (default). Unchanged. repo-config/configure.sh runs spec/resolve_description.py under a probed Python 3.7+, which imports validate, so a dependency block breaks it unless that caller moves too.
Refuse an Include Marker Indented Past Three Spaces Rather Than Skipping It #3432 (default). build_dist.py silently skips an include marker indented past three spaces. Small, and pre-existing.
Stale "stdlib-only" wording (default). .agents/skills/python-codestyle/references/profiles.md, spec/project-types.json, and README.md, unless a round since fixed it. Re-check first.
Downstream, each in its own repository's session:
- HomeAutomation-Config Tracker: .agents/skills/shell-codestyle/SKILL.md #1353 and ESPHome-Config Remove TelemetryImporter from the catalog #256.
- PlexCleaner adopting the python gate.
- The round-50 adoption set (Add closed to the Merge-Bot Stub Types of Adopters Setting delete-branch #2336, Grant pull-requests: read So the Live comments Label Read Reaches Private Repositories #2354, and homeassistant-purpleair Roll out global ~/.claude/CLAUDE.md safety rules to all machines #365), rolled into the next resync.
- ESPHome-NonRoot Add agent safety rules to shared AGENTS.md (repo-scope + no writes-as-probes) #364, merged into its develop and not yet promoted there.
Carried hub backlog (default): Read Action Files or Name the Cause When UV_NO_CACHE Removes the Resolved PyYAML Interpreter #2358 (repo_gate.py) and Let an Unattended Handoff Run the Strict-Review Pass for a Worker That Cannot Dispatch #2342. Judge Add safety hooks for opencode and codex #781 against GOVERNANCE.md "Trust Boundaries and Hardening Effort" before starting it. Every other item round 53 listed here has closed.
The Windows lane, oldest first: Validate the Interpreter the Agent-Safety Installer Writes Into settings.json #1172, A Parent Pinned to UTF-8 Spawning a Python Child Leaves the Child's Encoder on the Locale #1577, skills_install.py runs a bare claude argv, which may miss a Windows .cmd shim #1754, Python devcontainer snippet: mount path join and mktemp -t portability #1763. Roll out global ~/.claude/CLAUDE.md safety rules to all machines #365 rewrites the Windows host's hooks and settings, so ask before running it.

External blockers

The nested hub clone hub-resync is still under the fleet worktrees directory. It was clean with nothing unpushed at round 53. Its removal is the maintainer's to do by hand.
Write grants are needed for every downstream step, and a native Windows host for the Windows lane.

Internal dependencies

Step 3's issues all edit pr_review.py or its readers, so they go one at a time. Render pr_review.py's Digest From a Pure Leaf and Build Its Header From a Field List #3457 to Give the pr_review Readers One Accessor Each for Commit Oid, Page Info, and Head Reviews #3460 also reshape the code Fix the COVERAGE_IS_PARTIAL Status Text on the Carry Paths in pr_review.py #3424 and A Held Wait Exits 0 Under a DIRTY or BEHIND Merge Whose Required Checks Never Ran #3471 touch, so take those two first.
scripts/pr_review_checks.py sits at 425 of its 450-line ceiling. A Held Wait Exits 0 Under a DIRTY or BEHIND Merge Whose Required Checks Never Ran #3471 or a new check shape may reach it. Its test docstring says a ceiling reached is a prompt to redesign along the module's boundaries, never to raise it.
Step 2 lands before any other prose_lint.py change, since each meets the zero-headroom ceiling until it does.
Step 4 waits on nothing, but any further Markdown-reading change to validate.py waits on it.

State

Re-derive all of this at resume with git worktree list, git log origin/main..origin/develop, gh pr list, and gh release list.

Develop is at 35bedae, 11 merges ahead of main. The latest release is still 2.0.1166.
Open pull requests: none at close.
Primary checkout: clean on develop at origin/develop.
Worktrees: every worktree and branch this round created is removed, through pr_review.py cleanup for both merged pull requests. The detached hub checkouts that git worktree list still shows belong to other sessions' scratchpads.
Sizes against their ceilings in tests/test_pr_review_size.py:
- pr_review.py: 2118 of 2200
- pr_review_checks.py: 425 of 450
- pr_review_rounds.py: 625 of 700
A parallel unattended-handoff run in this repository merged its own lanes into develop while this round ran, and has ended. Its auto-* tracks are closed, so tracks lists only default.

The parked decision queue

Zero open decision issues in this repository at close, read live. This round put three decisions to the maintainer, all answered in session:

Split pr_review.py rather than raise its ceiling or trim it.
Merge Split pr_review.py's Check and Round Readers Into Leaf Modules #3462.
Merge Hold a Held Wait While a Head's Workflow Run Has Not Completed #3470.

What the last round did

Merged Split pr_review.py's Check and Round Readers Into Leaf Modules #3462. The pure CI rollup readers moved into scripts/pr_review_checks.py, and the pure review-round readers into scripts/pr_review_rounds.py. Both are leaf modules that make no call and never import the gate. The digest's check lines are now a table-driven check_lines. Tests guard each module's imports, including indirect ones through the Copilot adapter.
Merged Hold a Held Wait While a Head's Workflow Run Has Not Completed #3470, fixing pr_review.py wait Reports Checks Complete Before a Fresh Head's Jobs Register #3451 on promotion. On a covered held head, wait holds while any workflow run on the head has not completed, whatever the merge reads. A check run waiting on its environment reads as a new GATED shape. Measured live, it held 781 seconds until every run finished.
Filed:
- Render pr_review.py's Digest From a Pure Leaf and Build Its Header From a Field List #3457 to Stop pr_review.py cleanup From Loading a Second Copy of the Gate #3461, from a frontier design survey of the split.
- A Held Wait Exits 0 Under a DIRTY or BEHIND Merge Whose Required Checks Never Ran #3471, a pre-existing exit 0 where the required checks never ran.
Coordinated with the unattended orchestrator over its exclusions while pr_review.py was moving.

What not to repeat

Adding exceptions to the run hold. Releasing a run waiting on its environment, or one with a starved job, or gating that on sibling activity: each reopened pr_review.py wait Reports Checks Complete Before a Fresh Head's Jobs Register #3451's early exit 0 through another state, and cost three review rounds. Every open run holds now, and the digest line names why. Don't reintroduce a release without a live repro of the state it claims to clear.
Rewording one help sentence a pass at a time. Six passes narrowed "every run has completed" against a smaller condition each time. When help restates a code condition, name the condition the code reads (covered) on the first try.
Trusting wait's exit 0 from a checkout without the pr_review.py wait Reports Checks Complete Before a Fresh Head's Jobs Register #3451 fix. It returned while two checks were still running. Until the promotion releases the fix, confirm the checks directly.
Typing a SHA into gh pr merge --match-head-commit from memory. The guard refused a mistyped one. Capture it from gh pr view --json headRefOid into a variable.
git worktree remove against the primary checkout. The write guard denies it. Run it from a linked worktree of the same repository.
gh api --jq with --arg. Like gh run list, it takes no jq arguments. Filter with jq itself.

New learnings

A size ceiling is a tripwire for redesign, not a target. The tests/test_pr_review_size.py docstring says so, and Split pr_review.py's Check and Round Readers Into Leaf Modules #3462's body carries the design reasoning: cut along purity, keep every name a test patches in the gate, and turn data into tables.
A job-less suite left queued forever is real on this repository's heads, but only from other apps (azure-pipelines, claude, coderabbitai). Every Actions suite measured carried jobs. That is recorded in Hold a Held Wait While a Head's Workflow Run Has Not Completed #3470's Decisions section, and a test pins the other-app case.

Round 55, closed 2026-10-25: Session Handoff [default]: Promotion #3472 Released as 2.0.1178, Prose Gate Ceiling and pr_review Follow-Ups Next

Next steps, in priority order

Re-read each issue and pull request rather than trusting these summaries.

Replace the Prose Gate's Line Ceiling with a Check for Vocabulary in the Script #3441, replace the prose gate's line ceiling (default). Unchanged since round 53. The maintainer chose to replace test_the_gate_stays_under_its_size_ceiling's line count with a check for vocabulary or exemption data in prose_lint.py. Until it lands, the gate has no headroom, so a prose_lint.py change compacts to fit and never asks to raise it.
pr_review.py and pr_cleanup.py follow-ups, one at a time (default). Take them in this order:
- Fix the COVERAGE_IS_PARTIAL Status Text on the Carry Paths in pr_review.py #3424, the COVERAGE_IS_PARTIAL text.
- Re-Read a Deleted Ref Before Reporting DELETE_NOT_CONFIRMED in pr_cleanup #3436, in pr_cleanup.py.
- Stop pr_review.py cleanup From Loading a Second Copy of the Gate #3461, cleanup loading a second copy of the gate. This is a real module-state hazard.
- A Held Wait Exits 0 Under a DIRTY or BEHIND Merge Whose Required Checks Never Ran #3471, exit 0 on a held head whose required checks never ran. Say What a Held Wait's Exit 0 Does Not Prove #3473 already made the help name every such path, per its latest comment, so what is left is the exit decision itself. Done means those paths exit a code naming the state, or the maintainer keeps exit 0 and the issue closes on the help.
- Render pr_review.py's Digest From a Pure Leaf and Build Its Header From a Field List #3457 to Give the pr_review Readers One Accessor Each for Commit Oid, Page Info, and Head Reviews #3460, the design survey's follow-ups: the digest renderer leaf and header field list, wait out of main with a flag table, the gate's repeated shapes as helpers, and the payload accessors.
Read validate.py's Markdown and the README Link Definitions Through markdown-it-py #3426, validate.py onto markdown-it-py (default). Unchanged. repo-config/configure.sh runs spec/resolve_description.py under a probed Python 3.7+, which imports validate, so a dependency block breaks it unless that caller moves too.
Refuse an Include Marker Indented Past Three Spaces Rather Than Skipping It #3432 (default). build_dist.py silently skips an include marker indented past three spaces. Small, and pre-existing.
Stale "stdlib-only" wording (default). .agents/skills/python-codestyle/references/profiles.md, spec/project-types.json, and README.md, unless a round since fixed it. Re-check first.
Downstream, each in its own repository's session:
- HomeAutomation-Config Tracker: .agents/skills/shell-codestyle/SKILL.md #1353 and ESPHome-Config Remove TelemetryImporter from the catalog #256.
- PlexCleaner adopting the python gate.
- The round-50 adoption set (Add closed to the Merge-Bot Stub Types of Adopters Setting delete-branch #2336, Grant pull-requests: read So the Live comments Label Read Reaches Private Repositories #2354, and homeassistant-purpleair Roll out global ~/.claude/CLAUDE.md safety rules to all machines #365), rolled into the next resync.
- ESPHome-NonRoot Add agent safety rules to shared AGENTS.md (repo-scope + no writes-as-probes) #364, merged into its develop and not yet promoted there.
Carried hub backlog (default): Read Action Files or Name the Cause When UV_NO_CACHE Removes the Resolved PyYAML Interpreter #2358 (repo_gate.py) and Let an Unattended Handoff Run the Strict-Review Pass for a Worker That Cannot Dispatch #2342. Judge Add safety hooks for opencode and codex #781 against GOVERNANCE.md "Trust Boundaries and Hardening Effort" before starting it.
The Windows lane, oldest first: Validate the Interpreter the Agent-Safety Installer Writes Into settings.json #1172, A Parent Pinned to UTF-8 Spawning a Python Child Leaves the Child's Encoder on the Locale #1577, skills_install.py runs a bare claude argv, which may miss a Windows .cmd shim #1754, Python devcontainer snippet: mount path join and mktemp -t portability #1763. Roll out global ~/.claude/CLAUDE.md safety rules to all machines #365 rewrites the Windows host's hooks and settings, so ask before running it.

External blockers

The nested hub clone hub-resync is still under the fleet worktrees directory per round 53. Its removal is the maintainer's to do by hand.
Write grants are needed for every downstream step, and a native Windows host for the Windows lane.

Internal dependencies

Step 2's issues all edit pr_review.py or its readers, so they go one at a time. Render pr_review.py's Digest From a Pure Leaf and Build Its Header From a Field List #3457 to Give the pr_review Readers One Accessor Each for Commit Oid, Page Info, and Head Reviews #3460 also reshape the code Fix the COVERAGE_IS_PARTIAL Status Text on the Carry Paths in pr_review.py #3424 and A Held Wait Exits 0 Under a DIRTY or BEHIND Merge Whose Required Checks Never Ran #3471 touch, so take those two first.
A Held Wait Exits 0 Under a DIRTY or BEHIND Merge Whose Required Checks Never Ran #3471 changes the help Say What a Held Wait's Exit 0 Does Not Prove #3473 just wrote. Whatever exit it picks, rewrite the exit-0 sentences in the wait help and in scripts/README.md's exit table and held-wait paragraph together, since they state the same list.
scripts/pr_review_checks.py is near its 450-line ceiling. A Held Wait Exits 0 Under a DIRTY or BEHIND Merge Whose Required Checks Never Ran #3471 or a new check shape may reach it. Its test docstring says a ceiling reached is a prompt to redesign along the module's boundaries, never to raise it.
Step 1 lands before any other prose_lint.py change, since each meets the zero-headroom ceiling until it does.
Step 3 waits on nothing, but any further Markdown-reading change to validate.py waits on it.

State

Re-derive all of this at resume with git worktree list, git log origin/main..origin/develop, gh pr list, and gh release list.

Released 2.0.1178 from main at a245dd1, the merge of promotion Promote the pr_review.py Leaf-Module Split, the Held-Wait Run Hold, and Nine Fixes to Main #3472. develop is at 412ce42 and nothing on it is unreleased.
Open pull requests: none at close.
Primary checkout: clean on develop at origin/develop.
Worktrees: this round's feature worktree went through pr_review.py cleanup, and its Skills refresh worktree was removed. cleanup --sweep reported problems=0. The detached hub checkouts it keeps belong to other sessions.
This machine: Skills snapshot current at a245dd1, host-safety kit CURRENT.
Sizes against their ceilings in tests/test_pr_review_size.py, by wc -l at close:
- pr_review.py: 2122 of 2200
- pr_review_checks.py: 418 of 450
- pr_review_rounds.py: 625 of 700

The parked decision queue

Zero open decision issues in this repository at close, read live. This round put one decision to the maintainer, answered in session: merge and release #3472.

What the last round did

Merged Say What a Held Wait's Exit 0 Does Not Prove #3473 into develop. The wait help and scripts/README.md no longer claim a held head's exit 0 proves its required checks ran or that every run completed. They name the CHECKS UNREADABLE path, the rules-read and CHECKS TRUNCATED conditions on the unregistered-required-check hold, the merges that release short of the holds, and the completed-run read's reach. It answered Promote the pr_review.py Leaf-Module Split, the Held-Wait Run Hold, and Nine Fixes to Main #3472's one finding and two Copilot findings on itself.
Merged promotion Promote the pr_review.py Leaf-Module Split, the Held-Wait Run Hold, and Nine Fixes to Main #3472 and released 2.0.1178. It carried the pr_review.py leaf-module split (Split pr_review.py's Check and Round Readers Into Leaf Modules #3462), the held-wait run hold (Hold a Held Wait While a Head's Workflow Run Has Not Completed #3470), and ten other changes, and closed Keep Untracked Files Out of the Pre-Commit Snippet's Prose Gate #3370, Exempt the BuildKit check= Parser Directive in the Prose Gate #3372, Compare Intent Staleness Against the Hub's Own Copy and Name the Moved Canonical #3371, Correct the Lint Snippets' Claim That the Hook's EOL Gate Is Diff-Scoped #3369, Move the Remaining Prose-Matched Subprocess Failures Onto the Status-Code Rule #3378, Remove the Semicolon in WORKFLOW.md's Filesystem Output-Seam Bullet #3454, Nothing Reports an Operational Caller That Omits the workflow-model Input #3332, skills_install --report follow-ups: stamp-less live report, bootstrap remedy text, --intended guard test #1759, and pr_review.py wait Reports Checks Complete Before a Fresh Head's Jobs Register #3451.
Commented on A Held Wait Exits 0 Under a DIRTY or BEHIND Merge Whose Required Checks Never Ran #3471 with what Say What a Held Wait's Exit 0 Does Not Prove #3473 settled and what remains.

What not to repeat

Writing a guarantee into help from the one branch a change touched. Say What a Held Wait's Exit 0 Does Not Prove #3473's exit-0 sentences took three local passes and a Copilot round, each finding one more path: DRAFT, then the rules-read and truncation guards, then BLOCKED short of them, then the query windows. Before writing what an exit proves, walk every return path of the deciding function, held_checks_open here, and every window in the query feeding it, and write the claim from that list in one pass.
Trusting wait's exit 0 on a held head. It returned with the Windows job still running on both Say What a Held Wait's Exit 0 Does Not Prove #3473 and Promote the pr_review.py Leaf-Module Split, the Held-Wait Run Hold, and Nine Fixes to Main #3472. 2.0.1178 ships the run hold, so recheck this once a fresh head uses it, and still read gh pr checks before merging until then.
Re-answering exit 50 on a promotion. It repeats on the head after the comment answer by design, since nothing reads one. Answer once, then watch the checks with the bounded gh pr checks --watch it prints.

New learnings

A Copilot "needs a closer look" round that counts no finding and only asks for human approval is answered with pr_review.py comment naming the maintainer's merge go-ahead as the gate it describes. Recorded on Promote the pr_review.py Leaf-Module Split, the Held-Wait Run Hold, and Nine Fixes to Main #3472.

Round 67, closed 2026-10-29: Session Handoff [default]: #3432 Include Indent Refusal Ready in Promotion #3631, Merge and Release Next

Next steps, in priority order

Re-read each issue and pull request rather than trusting these summaries.

Merge promotion Promote Develop: Refuse an Include Marker Indented Past the Three-Space Bound #3631 and dispatch the release (default). It carries Refuse an Include Marker Indented Past the Three-Space Bound #3630 and closes Refuse an Include Marker Indented Past Three Spaces Rather Than Skipping It #3432. It read CLEAN at head 15d4c306, 13 of 13 checks, and Copilot's table covers all 7 files with no finding. Merging and releasing is merge-and-release, on the maintainer's go-ahead. Re-read the head before merging.
A Rollup of Only Unreadable Nodes Is Not Read as Empty by the Held Wait #3488, then Read validate.py's Markdown and the README Link Definitions Through markdown-it-py #3426 (default). A Rollup of Only Unreadable Nodes Is Not Read as Empty by the Held Wait #3488 is the rollup of only unreadable nodes in scripts/pr_review_checks.py. Read validate.py's Markdown and the README Link Definitions Through markdown-it-py #3426 moves spec/validate.py and the _LINK_DEF readers onto markdown-it-py, and needs a before-and-after audit comparison.
Downstream, each in its own repository's session: unchanged from round 66. ptr727/Blog's lib/fleet/**.sh root .editorconfig section goes at its next resync. ptr727/HomeAutomation-Config#1435 and Read the Release Env File Through the Hub's Carried Readers Blog#553 each start with a hub pull request declaring host-secrets in registry/repos.json, and the round-60 set stands.
Carried hub backlog and the Windows lane (default): Read Action Files or Name the Cause When UV_NO_CACHE Removes the Resolved PyYAML Interpreter #2358, Let an Unattended Handoff Run the Strict-Review Pass for a Worker That Cannot Dispatch #2342, Add safety hooks for opencode and codex #781, then Validate the Interpreter the Agent-Safety Installer Writes Into settings.json #1172, A Parent Pinned to UTF-8 Spawning a Python Child Leaves the Child's Encoder on the Locale #1577, skills_install.py runs a bare claude argv, which may miss a Windows .cmd shim #1754, Python devcontainer snippet: mount path join and mktemp -t portability #1763.

External blockers

Step 1 waits on the maintainer for the merge and release go-ahead.
A feature pull request's squash merge into develop is the maintainer's to run at the prompt, since this host's auto-mode classifier denies the agent's. Ask up front with the exact command in the question.
The nested hub clone hub-resync, write grants downstream, a native Windows host, and cleanup --sweep --apply are unchanged from round 60.

Internal dependencies

A Rollup of Only Unreadable Nodes Is Not Read as Empty by the Held Wait #3488 touches scripts/pr_review_checks.py, at 444 lines against its 450 ceiling, so take it on its own. A ceiling reached is answered by redesign along module boundaries, never by raising it.
Step 3's hub pull requests each edit registry/repos.json, so they go one at a time or in one pull request.

State

Re-derive with git worktree list, git log origin/main..origin/develop, gh pr list, and gh release list.

Develop is at 15d4c306, one commit ahead of main, carried by promotion Promote Develop: Refuse an Include Marker Indented Past the Three-Space Bound #3631. 2.0.1265 is the latest release.
Open pull requests: only Promote Develop: Refuse an Include Marker Indented Past the Three-Space Bound #3631.
Primary checkout: clean on develop at origin/develop.
Worktrees: this round's went through pr_review.py cleanup. The detached hub checkouts that remain belong to other sessions.
Other repositories touched: none.

The parked decision queue

Zero open decision issues in this repository at close, read live. This round puts one question at close, whether to merge #3631 and dispatch the release. A comment on this issue records the answer.

What the last round did

Implemented Refuse an Include Marker Indented Past Three Spaces Rather Than Skipping It #3432 via Refuse an Include Marker Indented Past the Three-Space Bound #3630. _marker in scripts/build_dist.py measures indentation once in columns, a tab reaching the next four-column stop. Outside a code block it refuses a marker past three columns on the exit-2 path, and names a deep near miss as a near miss first. _unindented is folded in. The refusal lists in scripts/README.md and skill-lifecycle gain the case.
Tests pin the issue's list-item case at four and six spaces and two tab forms, a deep marker in a source, the region-end probe, a deep near miss, and a list item's own fenced and indented samples staying content. The full suite ran 3087 tests, OK (skipped=7).
The local strict review ran two rounds at the frontier tier. Round one raised a stale comment and docstring, a doubled indentation measure, the near-miss message order, and two test gaps, all fixed. Round two found nothing. Copilot's one nit on Refuse an Include Marker Indented Past the Three-Space Bound #3630, a possessive in a method name, was disproved in a comment.
Dropped step 2's stale "stdlib-only" wording item. spec/validate.py's "stdlib only" is accurate, since Read validate.py's Markdown and the README Link Definitions Through markdown-it-py #3426 keeps its import lazy for configure.sh, and scripts/README.md names markdown-it-py as the exception.

What not to repeat

Editing a # comment line to match new behavior. The prose gate's comment-added rule counts a modified comment line as added. Delete a stale line or move the explanation into a docstring, which the rule does not read.
Two indentation measures for one bound. A space count plus a separate tab-aware width disagreed on a tab-led line. Measure columns once with expandtabs(4).
Asserting where a refusal fires without tracing the call order. A source file's own region scan runs before its body is read, so the body-scan call site never sees a deep marker first.

New learnings

None beyond what the entries above hold. Each is specific to this round's attempt rather than a governance lesson.

Round 68 close: the maintainer answered the one question this round put, "Merge and release". Promotion #3633 merged into main as d6c2e13, closing #3488. Release 2.0.1267 published from run 40982979436, concluding success. This machine's Skills snapshot reads current at d6c2e13, and the host-safety kit reads CURRENT. Step 1 of this round's next steps is done, so the next session starts at step 2, #3426.

Parked decision queue at close: zero open decision issues, read live. The cleanup sweep would remove nothing, since every detached hub checkout it lists is under its 24-hour idle bound.

Round 68, closed 2026-10-30: Session Handoff [default]: #3488 Unread-Only Rollup Hold Ready in Promotion #3633, Merge and Release Next

Next steps, in priority order

Re-read each issue and pull request rather than trusting these summaries.

Merge promotion Promote Develop: Hold a Rollup of Only Unreadable Nodes as an Empty One #3633 and dispatch the release (default). It carries Hold a Rollup of Only Unreadable Nodes as an Empty One #3632 and closes A Rollup of Only Unreadable Nodes Is Not Read as Empty by the Held Wait #3488. It read CLEAN at head e15559fb, 13 of 13 checks, and Copilot's table covers all 4 files with no finding. Merging and releasing is merge-and-release, on the maintainer's go-ahead. Re-read the head before merging.
Read validate.py's Markdown and the README Link Definitions Through markdown-it-py #3426 (default). It moves spec/validate.py and the _LINK_DEF readers onto markdown-it-py, and needs a before-and-after audit comparison.
Downstream, each in its own repository's session: unchanged from round 66. ptr727/Blog's lib/fleet/**.sh root .editorconfig section goes at its next resync. ptr727/HomeAutomation-Config#1435 and Read the Release Env File Through the Hub's Carried Readers Blog#553 each start with a hub pull request declaring host-secrets in registry/repos.json, and the round-60 set stands.
Carried hub backlog and the Windows lane (default): Read Action Files or Name the Cause When UV_NO_CACHE Removes the Resolved PyYAML Interpreter #2358, Let an Unattended Handoff Run the Strict-Review Pass for a Worker That Cannot Dispatch #2342, Add safety hooks for opencode and codex #781, then Validate the Interpreter the Agent-Safety Installer Writes Into settings.json #1172, A Parent Pinned to UTF-8 Spawning a Python Child Leaves the Child's Encoder on the Locale #1577, skills_install.py runs a bare claude argv, which may miss a Windows .cmd shim #1754, Python devcontainer snippet: mount path join and mktemp -t portability #1763.

External blockers

Step 1 waits on the maintainer for the merge and release go-ahead.
A feature pull request's squash merge into develop is the maintainer's to run at the prompt, since this host's auto-mode classifier denies the agent's. Ask up front with the exact command in the question.
The nested hub clone hub-resync, write grants downstream, a native Windows host, and cleanup --sweep --apply are unchanged from round 60.

Internal dependencies

scripts/pr_review_checks.py is at 447 lines against its 450 ceiling. The next change that grows it needs a redesign along module boundaries first, never a raised ceiling.
Step 3's hub pull requests each edit registry/repos.json, so they go one at a time or in one pull request.

State

Re-derive with git worktree list, git log origin/main..origin/develop, gh pr list, and gh release list.

Develop is at e15559fb, one commit ahead of main, carried by promotion Promote Develop: Hold a Rollup of Only Unreadable Nodes as an Empty One #3633. 2.0.1266 is the latest release.
Open pull requests: only Promote Develop: Hold a Rollup of Only Unreadable Nodes as an Empty One #3633.
Primary checkout: clean on develop at origin/develop.
Worktrees: this round's went through pr_review.py cleanup. The detached hub checkouts that remain belong to other sessions.
Other repositories touched: none.

The parked decision queue

Zero open decision issues in this repository at close, read live. This round puts one question at close, whether to merge #3633 and dispatch the release. A comment on this issue records the answer.

What the last round did

Implemented A Rollup of Only Unreadable Nodes Is Not Read as Empty by the Held Wait #3488 via Hold a Rollup of Only Unreadable Nodes as an Empty One #3632. held_checks_open and checks_never_ran count readable checks through checks_tally, so a rollup of only unreadable types holds a BLOCKED, BEHIND, or DRAFT merge and exits 45 under DIRTY. The DIRTY release names the unread types. checks_unread returns each type once, sorted, so that message and the digest's CHECKS PARTIALLY UNREAD line share one form. The wait help, the CHECKS_PENDING status line, and scripts/README.md agree.
Tests hold an unread-only rollup under each merge, and pin a mixed rollup and two distinct types in reverse order. The full suite ran 3091 tests, OK (skipped=7).
The local strict review ran five recorded passes at the frontier tier. Rounds one and two raised stale help and status text, a repeated type name, sentences over the bound, and a duplicated dedupe, all fixed. Round three's one finding spent the edit budget, went in the body, and was fixed on the next push. Copilot raised it too on the older commit, plus a missing sort-order test, fixed in 31a86e7.

What not to repeat

Running a targeted test module while the full suite runs in the same tree. It recompiles scripts/__pycache__, and test_cli_usage's side-effect check fails on the changed .pyc files. Run the full suite alone.
Rewording help text at the edges of a sentence to fit the length bound. Dropping "Either" from a list of cases made the first case read as true of every exit. Restate the cases as "One case is ... Another is ... The last is ...".
Naming the reader as "it" in help or status text. It binds to the nearest noun, such as the head or the local pass. These modules say "this".

New learnings

None beyond what the entries above hold. Each is specific to this round's attempt rather than a governance lesson.

Round 69 close: the maintainer answered the one question this round put, "Merge and release". Promotion #3636 merged into main as 5525f3f, closing #3426. Release 2.0.1269 published from run 40999961956, concluding success. This machine's Skills snapshot reads current at 5525f3f, and the host-safety kit reads CURRENT. Step 1 of this round's next steps is done, so the next session starts at step 2, #3634.

One answer on #3636 landed imperfect: the comment answering the review-body finding with no thread names the review commit 04bd8ae and the fixing commit correctly, but links the pull request rather than the review round, since a sed substitution failed before it posted. It was left as posted.

Parked decision queue at close: zero open decision issues, read live. The cleanup sweep would remove nothing, since every detached hub checkout it lists is under its 24-hour idle bound.

Round 69, closed 2026-10-30: Session Handoff [default]: #3426 markdown-it-py Link Reading Ready in Promotion #3636, Merge and Release Next

Next steps, in priority order

Re-read each issue and pull request rather than trusting these summaries.

Merge promotion Promote Develop: Read validate.py's Markdown and the README Link Definitions Through markdown-it-py #3636 and dispatch the release (default). It carries Read validate.py's Markdown and the README Link Definitions Through markdown-it-py #3635 and Read Backslash Escapes in a Definition's Label and Angle Destination #3637 and closes Read validate.py's Markdown and the README Link Definitions Through markdown-it-py #3426. It read CLEAN at head 19548f49, 13 of 13 checks. Copilot's second round on that head recommends approval with no finding, and both findings from its first round are answered. Merging and releasing is merge-and-release, on the maintainer's go-ahead. Re-read the head before merging.
Exempt a PEP 723 Script Metadata Block From comment-added #3634 (default). Exempt a PEP 723 script metadata block from comment-added. This round met it adding the header to spec/validate.py, and took the comments label instead.
Downstream, each in its own repository's session: unchanged from round 66. ptr727/Blog's lib/fleet/**.sh root .editorconfig section goes at its next resync. ptr727/HomeAutomation-Config#1435 and Read the Release Env File Through the Hub's Carried Readers Blog#553 each start with a hub pull request declaring host-secrets in registry/repos.json, and the round-60 set stands.
Carried hub backlog and the Windows lane (default): Read Action Files or Name the Cause When UV_NO_CACHE Removes the Resolved PyYAML Interpreter #2358, Let an Unattended Handoff Run the Strict-Review Pass for a Worker That Cannot Dispatch #2342, Add safety hooks for opencode and codex #781, then Validate the Interpreter the Agent-Safety Installer Writes Into settings.json #1172, A Parent Pinned to UTF-8 Spawning a Python Child Leaves the Child's Encoder on the Locale #1577, skills_install.py runs a bare claude argv, which may miss a Windows .cmd shim #1754, Python devcontainer snippet: mount path join and mktemp -t portability #1763.

External blockers

Step 1 waits on the maintainer for the merge and release go-ahead.
A feature pull request's squash merge into develop is the maintainer's to run at the prompt, since this host's auto-mode classifier denies the agent's. Ask with the exact command in the question once the pull request is green.
The nested hub clone hub-resync, write grants downstream, a native Windows host, and cleanup --sweep --apply are unchanged from round 60.

Internal dependencies

scripts/pr_review_checks.py is at 447 lines against its 450 ceiling. The next change that grows it needs a redesign along module boundaries first, never a raised ceiling.
Step 3's hub pull requests each edit registry/repos.json, so they go one at a time or in one pull request.

State

Re-derive with git worktree list, git log origin/main..origin/develop, gh pr list, and gh release list.

Develop is at 19548f49, two commits ahead of main, carried by promotion Promote Develop: Read validate.py's Markdown and the README Link Definitions Through markdown-it-py #3636. 2.0.1267 is the latest release.
Open pull requests: only Promote Develop: Read validate.py's Markdown and the README Link Definitions Through markdown-it-py #3636.
Primary checkout: clean on develop at origin/develop.
Worktrees: both of this round's went through pr_review.py cleanup. The detached hub checkouts that remain belong to other sessions.
Other repositories touched: none.

The parked decision queue

Zero open decision issues in this repository at close, read live. This round puts one question at close, whether to merge #3636 and dispatch the release. A comment on this issue records the answer.

What the last round did

Implemented Read validate.py's Markdown and the README Link Definitions Through markdown-it-py #3426 via Read validate.py's Markdown and the README Link Definitions Through markdown-it-py #3635. spec/validate.py reads link targets, the verbatim-section region, and section-declaration headings through spec/fidelity_text.py, and imports it lazily so spec/resolve_description.py still runs under plain python3. The validator runs under uv run --no-project --script in CI and OPERATIONS.md. The README link-naming, shield, and 3rd Party Tools checks read definitions from the parser, first wins per folded label. validate.py's output and the full audit sweep matched before and after, apart from live repository movement.
Read Backslash Escapes in a Definition's Label and Angle Destination #3637 fixed the two escape findings Copilot raised on promotion Promote Develop: Read validate.py's Markdown and the README Link Definitions Through markdown-it-py #3636 against _DESTINATION.
Filed Exempt a PEP 723 Script Metadata Block From comment-added #3634, the PEP 723 comment-added gap.
Local strict review ran seven recorded passes on Read validate.py's Markdown and the README Link Definitions Through markdown-it-py #3635 and two on Read Backslash Escapes in a Definition's Label and Angle Destination #3637, at the frontier tier. Every introduced finding was fixed. Three consolidation findings were declined with evidence in Read validate.py's Markdown and the README Link Definitions Through markdown-it-py #3635's body.

What not to repeat

Putting a URL into sed's s#...#...# replacement. A review URL carries #, which ends the expression early. The chain then stopped before the thread reply, while a separate ; command posted its comment with the placeholder link. Build such text with the Write tool, or use a delimiter the value cannot hold.
Glueing a reference definition under a paragraph line in a test fixture. CommonMark does not read it as a definition, so a fixture that leaned on the old line scan broke the moment the parser read it.
Branching from origin/develop and then pushing with a bare git push. The branch tracks origin/develop until the first push names its own upstream, so always push with -u origin <branch>.

New learnings

None beyond what the entries above hold. Each is specific to this round's attempt rather than a governance lesson.

Round 70 close: the maintainer answered the one question this round put, "Merge and release". Promotion #3675 merged into main as ee7839b, closing #3634. Release 2.0.1283 published from run 41123104774, concluding success. This machine's Skills snapshot reads current at ee7839b, and the host-safety kit reads CURRENT. Step 1 of this round's next steps is done, so the next session starts at step 2, #3673.

Parked decision queue at close: zero open decision issues, read live. The cleanup sweep would remove nothing.

Round 70, closed 2026-10-30: Session Handoff [default]: #3634 PEP 723 Comment-Rule Exemption Ready in Promotion #3675, Merge and Release Next

Next steps, in priority order

Re-read each issue and pull request rather than trusting these summaries.

Merge promotion Promote Develop: Exempt the PEP 723 Script Metadata Block From the Comment Rules #3675 and dispatch the release (default). It carries Exempt a PEP 723 Script Metadata Block From the Comment Rules #3674 and closes Exempt a PEP 723 Script Metadata Block From comment-added #3634. It read CLEAN at head e8ea7b19, 13 of 13 checks. Copilot's round on that head recommends approval with no finding. Merging and releasing is merge-and-release, on the maintainer's go-ahead. Re-read the head before merging.
Read a uv run Shebang as Python in the Prose Gate #3673 (default). Read a uv run shebang as Python in the prose gate, so an extensionless PEP 723 script gets the Python reading, the metadata block exemption among it. Filed this round as pre-existing from Exempt a PEP 723 Script Metadata Block From the Comment Rules #3674's local review.
Downstream, each in its own repository's session: unchanged from round 66. ptr727/Blog's lib/fleet/**.sh root .editorconfig section goes at its next resync. ptr727/HomeAutomation-Config#1435 and Read the Release Env File Through the Hub's Carried Readers Blog#553 each start with a hub pull request declaring host-secrets in registry/repos.json, and the round-60 set stands.
Carried hub backlog and the Windows lane (default): Read Action Files or Name the Cause When UV_NO_CACHE Removes the Resolved PyYAML Interpreter #2358, Let an Unattended Handoff Run the Strict-Review Pass for a Worker That Cannot Dispatch #2342, Add safety hooks for opencode and codex #781, then Validate the Interpreter the Agent-Safety Installer Writes Into settings.json #1172, A Parent Pinned to UTF-8 Spawning a Python Child Leaves the Child's Encoder on the Locale #1577, skills_install.py runs a bare claude argv, which may miss a Windows .cmd shim #1754, Python devcontainer snippet: mount path join and mktemp -t portability #1763.

External blockers

Step 1 waits on the maintainer for the merge and release go-ahead.
The nested hub clone hub-resync, write grants downstream, a native Windows host, and cleanup --sweep --apply are unchanged from round 60.

Internal dependencies

scripts/pr_review_checks.py is at 447 lines against its 450 ceiling. The next change that grows it needs a redesign along module boundaries first, never a raised ceiling.
Step 3's hub pull requests each edit registry/repos.json, so they go one at a time or in one pull request.

State

Re-derive with git worktree list, git log origin/main..origin/develop, gh pr list, and gh release list.

Develop is at e8ea7b19, one commit ahead of main, carried by promotion Promote Develop: Exempt the PEP 723 Script Metadata Block From the Comment Rules #3675. 2.0.1282 is the latest release.
Open pull requests: only Promote Develop: Exempt the PEP 723 Script Metadata Block From the Comment Rules #3675.
Worktrees: this round's task worktree went through pr_review.py cleanup. The session's own hub clone on main remains under the worktree directory, and the detached hub checkouts beside it belong to other sessions.
Other repositories touched: none.

The parked decision queue

Zero open decision issues in this repository at close, read live. This round puts one question at close, whether to merge #3675 and dispatch the release. A comment on this issue records the answer.

What the last round did

Implemented Exempt a PEP 723 Script Metadata Block From comment-added #3634 via Exempt a PEP 723 Script Metadata Block From the Comment Rules #3674. The Python language data in prose-gate.json declares a PEP 723 block's form in metadata_block, opener # /// script, content line, and closer. check_file drops each closed block's lines from every comment unit before any rule runs, so comment-added, comment-wrap, comment-case, sentence-length, spelling, dupword, and issue-ref all skip it. A block declaring one key no longer reports comment-wrap with no remedy.
Filed Read a uv run Shebang as Python in the Prose Gate #3673, the extensionless uv run shebang read as bash, labeled pre-existing.
Local strict review ran four passes at the frontier tier, three of them recorded, and Copilot raised two findings on Exempt a PEP 723 Script Metadata Block From the Comment Rules #3674. Both were fixed in 03e0ba6, the second as a class fix moving the skip from two rules to the dispatch.
The feature merge into develop ran from the agent seat this round, gh pr merge --squash --match-head-commit, where earlier rounds had the maintainer run it.

What not to repeat

Fixing a comment-rule exemption one rule at a time. Round one put the skip in comment-added, round two added comment-wrap, and Copilot then found sentence-length still reading the block. Dropping the lines from the unit at check_file covered every comment rule at once. Sweep every rule a unit kind reaches before choosing where the skip goes.
A scripted exact-string replace against a file ruff format has since rewritten. The replace missed, and the edits before it in the same script had already applied. Re-read the region after a format, or use the Edit tool.
pr_review.py reply against a thread Copilot resolved itself after a push. It refuses with NO_MATCH, since it reads unresolved threads only. Answer such a thread by its id captured from a live GraphQL read.

New learnings

None beyond what the entries above hold. Each is specific to this round's attempt rather than a governance lesson.

Round 71 close: the maintainer answered the one question this round put, "Merge and release". Promotion #3678 merged into main as b19bf38, closing #3673. Release 2.0.1284 published from run 41138904960, concluding success. This machine's Skills snapshot reads current at b19bf38, and the host-safety kit reads CURRENT. Step 1 of this round's next steps is done, so the next session starts at step 2, the carried hub backlog.

The feature merge of #3676 into develop was run by the maintainer as a ! line after the auto mode classifier denied it from the agent seat. The promotion merge and the release dispatch both ran from the agent seat.

Parked decision queue at close: zero open decision issues, read live. The cleanup sweep reports only, listing detached hub checkouts other sessions left, and --apply stays the maintainer's call.

Round 71, closed 2026-10-30: Session Handoff [default]: #3673 uv run Shebang Lexer Ready in Promotion #3678, Merge and Release Next

Next steps, in priority order

Re-read each issue and pull request rather than trusting these summaries.

Merge promotion Promote Develop: Read a uv run Shebang as Python in the Prose Gate #3678 and dispatch the release (default). It carries Read a uv run Shebang as Python in the Prose Gate #3676 and closes Read a uv run Shebang as Python in the Prose Gate #3673. It read CLEAN at head 90813a0e, 13 of 13 checks. Copilot's round on that head recommends approval with no finding. Merging and releasing is merge-and-release, on the maintainer's go-ahead. Re-read the head before merging.
Carried hub backlog and the Windows lane (default): Read Action Files or Name the Cause When UV_NO_CACHE Removes the Resolved PyYAML Interpreter #2358, Let an Unattended Handoff Run the Strict-Review Pass for a Worker That Cannot Dispatch #2342, Add safety hooks for opencode and codex #781, then Validate the Interpreter the Agent-Safety Installer Writes Into settings.json #1172, A Parent Pinned to UTF-8 Spawning a Python Child Leaves the Child's Encoder on the Locale #1577, skills_install.py runs a bare claude argv, which may miss a Windows .cmd shim #1754, Python devcontainer snippet: mount path join and mktemp -t portability #1763. Re-rank against the open backlog first, since earlier rounds' picks came from issues filed after this list.
Downstream, each in its own repository's session: unchanged from round 66. ptr727/Blog's lib/fleet/**.sh root .editorconfig section goes at its next resync. ptr727/HomeAutomation-Config#1435 and Read the Release Env File Through the Hub's Carried Readers Blog#553 each start with a hub pull request declaring host-secrets in registry/repos.json, and the round-60 set stands.

External blockers

Step 1 waits on the maintainer for the merge and release go-ahead.
A feature merge into develop may need the maintainer's hand. This round the auto mode classifier denied the agent's gh pr merge --squash --match-head-commit on Read a uv run Shebang as Python in the Prose Gate #3676, and the maintainer ran it as a ! line. Round 70 ran the same command from the agent seat. Ask for the ! line when the merge is denied rather than retrying it.
The nested hub clone hub-resync, write grants downstream, a native Windows host, and cleanup --sweep --apply are unchanged from round 60.

Internal dependencies

scripts/pr_review_checks.py is at 447 lines against its 450 ceiling. The next change that grows it needs a redesign along module boundaries first, never a raised ceiling.
Step 3's hub pull requests each edit registry/repos.json, so they go one at a time or in one pull request.

State

Re-derive with git worktree list, git log origin/main..origin/develop, gh pr list, and gh release list.

Develop is at 90813a0e, one commit ahead of main, carried by promotion Promote Develop: Read a uv run Shebang as Python in the Prose Gate #3678. 2.0.1283 is the latest release.
Open pull requests: only Promote Develop: Read a uv run Shebang as Python in the Prose Gate #3678.
Worktrees: this round's task worktree went through pr_review.py cleanup. The session's own hub clone on main remains under the worktree directory. The primary checkout is clean.
Other repositories touched: none.

The parked decision queue

Zero open decision issues in this repository at close, read live. This round puts one question at close, whether to merge #3678 and dispatch the release. A comment on this issue records the answer.

What the last round did

Implemented Read a uv run Shebang as Python in the Prose Gate #3673 via Read a uv run Shebang as Python in the Prose Gate #3676. prose-gate.json gains files.runner, { "uv": "python" }, and lexer_name maps the shebang's program through it before asking Pygments. An extensionless uv run --script file now gets the Python reading, the PEP 723 block exemption among it. Tests cover four uv shebang forms and the extensionless PEP 723 case, each failing with the lookup removed.
Local strict review ran two passes at the frontier tier, both recorded. The first raised one introduced finding, that test_every_file_mapping_names_a_declared_language did not cover the new map, so a mistyped runner value would silently fall back to bash. It was fixed in 5ac8f52. Copilot raised nothing on Read a uv run Shebang as Python in the Prose Gate #3676 or Promote Develop: Read a uv run Shebang as Python in the Prose Gate #3678.
Filed nothing.

What not to repeat

Reading pr_review.py wait exit 0 on a Copilot-covered head as checks settled. It returned in one second with two checks still running, as scripts/README.md states for that exit. Wait on the checks separately, with a bounded gh pr checks --watch.
Adding a data map without adding it to the mapping-language test. Every map under files that names a lexer belongs in named there, or a typo degrades to plain text with no error.

New learnings

None beyond the entries above, each specific to this round's attempt.

Round 72 close: the maintainer answered the one question this round put, "Merge and release". Promotion #3681 merged into main as 5540e38, closing #3671. Release 2.0.1286 published from run 41163336141, concluding success. This machine's Skills snapshot reads current, and the host-safety kit reads CURRENT.

Both feature merges, #3680 and #3682, were run by the maintainer as ! lines after the auto mode classifier denied them from the agent seat. The promotion merge and the release dispatch both ran from the agent seat.

Parked decision queue at close: zero open decision issues, read live. The cleanup sweep reports only, and --apply stays the maintainer's call.

Session Handoff [default]: #3671 Thread-Aware Held Wait Released in 2.0.1286, Carried Backlog Next #2361

Description

Next steps, in priority order

External blockers

Internal dependencies

State

The parked decision queue

What the last round did

What not to repeat

New learnings

Activity

ptr727 commented on Oct 4, 2026

ptr727 commented on Oct 23, 2026

Round 49, closed 2026-10-04: Session Handoff [default]: Four Follow-Ups Landed and Released as 2.0.826, Downstream Adoption Next

Next steps, in priority order

External blockers

Internal dependencies

State

The parked decision queue

What the last round did

What not to repeat

New learnings

ptr727 commented on Oct 24, 2026

Round 50, closed 2026-10-23: Session Handoff [default]: Four Tracks Folded into One Standing Handoff, Downstream Adoption Next

Next steps, in priority order

External blockers

Internal dependencies

State

The parked decision queue

What the last round did

What not to repeat

New learnings

ptr727 commented on Oct 24, 2026

Round 51, closed 2026-10-24: Session Handoff [default]: The #3391 Review-Cost Program Shipped as 2.0.1148 and 2.0.1158, Shared Markdown Reader Next

Next steps, in priority order

External blockers

Internal dependencies

State

The parked decision queue

What the last round did

What not to repeat

New learnings

ptr727 commented on Oct 25, 2026

Round 52, closed 2026-10-24: Session Handoff [default]: Five Fixes Released as 2.0.1163, Prose Gate and pr_review Follow-Ups Next

Next steps, in priority order

External blockers

Internal dependencies

State

The parked decision queue

What the last round did

What not to repeat

New learnings

ptr727 commented on Oct 25, 2026

Round 53, closed 2026-10-25: Session Handoff [default]: Prose Gate Heredoc and Backtick Fixes Released as 2.0.1166, Size Guard Redesign Next

Next steps, in priority order

External blockers

Internal dependencies

State

The parked decision queue

What the last round did

What not to repeat

New learnings

ptr727 commented on Oct 25, 2026

Round 54, closed 2026-10-25: Session Handoff [default]: pr_review.py Split Into Leaf Readers and the Held-Wait Run Hold on Develop, Promotion Next

Next steps, in priority order

External blockers

Internal dependencies

State

The parked decision queue

What the last round did

What not to repeat

New learnings

ptr727 commented on Oct 25, 2026

Round 55, closed 2026-10-25: Session Handoff [default]: Promotion #3472 Released as 2.0.1178, Prose Gate Ceiling and pr_review Follow-Ups Next

Next steps, in priority order

External blockers

Internal dependencies

State

The parked decision queue

What the last round did