Session Handoff [sentence-length]: Hub Code Comment Sweep Next, Hub-Only Markdown Sweep Shipped in 2.0.1251

Next steps, in priority order

Sweep the hub code comments and docstrings. 1662 hits at main 77d7ca35, all outside Markdown, mostly under scripts/ and tests/. Re-count with python3 scripts/prose_lint.py --check sentence-length and group by file before slicing. A comment-only pull request needs the comments label and PROSE_ALLOW_COMMENTS=1 on its commits, as Split the Over-Length Sentences in the Secrets Shell Library Comments #3581 did. Also align the "Copies left for the comment sweep" pairs listed in the bodies of Split the Over-Length Sentences in scripts/README.md #3594, Split the Over-Length Sentences in the Host Setup Docs #3595, and Split the Over-Length Sentences in the Spec and Repo-Config Docs #3597 with their Markdown splits. Use the Say What a Carrier's Local Strict Review Does With Sentence-Length Hits on Hub-Verbatim Text #3531 brief and its round 3 and round 4 addenda as the worker brief. Done: the whole tree reads zero.
Make sentence-length a default check once the tree reads zero. Delete the comment-and-doc-style text calling it opt-in "because the existing corpus predates the cap". Done: Say What a Carrier's Local Strict Review Does With Sentence-Length Hits on Hub-Verbatim Text #3531 closes on that promotion.
Move the Dispatch-Refresh Note Out of the Docker Clause in branching-and-release-model #3588: move the dispatch-refresh note out of the Docker clause in branching-and-release-model. One small skill edit, sharing no file with step 1.
Make pr_review.py and Merge Gate Item 2 Agree on an Earlier-Head Copilot Refusal #3571: make pr_review.py and Merge Gate item 2 agree on a Copilot refusal that landed on an earlier head. Independent of the steps above.
Nothing Checks That the Host CLAUDE.md Blocks Stay in Sync With Their Governance Sections #3601, filed this round: nothing checks that the host CLAUDE.md blocks stay in sync with their governance sections. It lists three choices and needs a design pick before any code.

External blockers

Host-safety kit is stale on this machine. Split the Over-Length Sentences in the Agent-Safety Markdown #3596 changed claude-md-safety.md and claude-md-fleet.md, so install.py --report reads not current after 2.0.1251. Only the maintainer re-runs that installer, since it rewrites hooks and settings. Re-check it at resume.
Promotion coverage gaps. Copilot's table on Promote Develop: Split the Over-Length Sentences in the Hub-Only Markdown #3600 named 31 of 42 files. The maintainer accepted feature-pull-request coverage, recorded on Promote Develop: Split the Over-Length Sentences in the Hub-Only Markdown #3600. Expect the same on each sweep promotion, and bring it to the maintainer each time.
Copilot refuses some large diffs with "default exclusions" (Split the Over-Length Sentences in TODO.md #3591, Split the Over-Length Sentences in scripts/README.md #3594). A refusal leaves Merge Gate item 2 to the maintainer. This round's acceptance covered only this round's slices.

Internal dependencies

Step 1 before step 2. The gate fails on any remaining hit.
Within a round, no two slices share a file. A comment that restates a Markdown sentence follows the Markdown's split. So a slice takes the pair list from the PR bodies named in step 1.
A skills-touching slice (step 3, or any comment in a generated tree) holds the single .source-digest line, so at most one runs at a time.

State

Re-derive each of these. Do not trust the copy here.

main 77d7ca35 (promotion Promote Develop: Split the Over-Length Sentences in the Hub-Only Markdown #3600) and develop 0d4e6f1d at close. Release 2.0.1251 published from main (run 40883205338, success).
No worktree or branch of this round remains. Every feature branch was cleaned with pr_review.py cleanup. cleanup --sweep lists only other sessions' worktrees, kept as too recent.
This machine's skills snapshot reads current at 77d7ca35. The host-safety kit does not, per the blocker above.
Open issues in scope: Say What a Carrier's Local Strict Review Does With Sentence-Length Hits on Hub-Verbatim Text #3531 (this track), Move the Dispatch-Refresh Note Out of the Docker Clause in branching-and-release-model #3588, Make pr_review.py and Merge Gate Item 2 Agree on an Earlier-Head Copilot Refusal #3571, Nothing Checks That the Host CLAUDE.md Blocks Stay in Sync With Their Governance Sections #3601, and Carry a CLI-Usage Test Harness So Fleet Repos Stop Copying the Hub's #3502, the held carried CLI harness proposal.

The parked decision queue

Zero open decision issues at close. The maintainer answered this round's questions in session. They granted develop merges for this session's slices, Sonnet workers with opus review in two waves of four, and Markdown-only scope. They accepted the recorded local pass for Copilot-refused slices, and feature-pull-request coverage on #3600. They chose merge, Skills refresh, and release.

What the last round did

Hub-only Markdown sweep, done. Nine pull requests took the hub-only Markdown from 1518 hits to zero: Split the Over-Length Sentences in scripts/README.md #3594, Split the Over-Length Sentences in TODO.md #3591, Split the Over-Length Sentences in reusable-workflows and fleet-map #3592, Split the Over-Length Sentences in the Hub Procedure Docs #3593, Split the Over-Length Sentences in the Host Setup Docs #3595, Split the Remaining Over-Length Sentences in the Host Setup Docs #3598, Split the Over-Length Sentences in the Agent-Safety Markdown #3596, Split the Over-Length Sentences in the Spec and Repo-Config Docs #3597, and Split the Over-Length Sentences in the Docs, Catalog READMEs, and HISTORY #3599. Promotion Promote Develop: Split the Over-Length Sentences in the Hub-Only Markdown #3600 merged to main.
Filed Nothing Checks That the Host CLAUDE.md Blocks Stay in Sync With Their Governance Sections #3601 for the missing host-block sync check.
Recorded this round's sweep lessons as a round 4 addendum to the Say What a Carrier's Local Strict Review Does With Sentence-Length Hits on Hub-Verbatim Text #3531 brief.

What not to repeat

Merging on coverage=none. Split the Over-Length Sentences in the Spec and Repo-Config Docs #3597 merged with its local pass recorded but never attested, and Copilot had reviewed only an earlier head. Read coverage= on the status line before every merge.
Letting a worker merge. The auto-mode classifier refused two workers' develop merges as "Merge Without Review". Workers hand back at ready, and the orchestrator merges.
A stop rule with room for a budget. Three of eight workers stopped with findings on new sentences unfixed, reading the per-sentence rule as a pass cap. Brief "no pass cap" explicitly at dispatch.
Splitting a long list into two partial lists. Reviewers read "the first three ... the last two" as closed. Convert to a lead sentence plus bullets.

New learnings

Recorded in the Say What a Carrier's Local Strict Review Does With Sentence-Length Hits on Hub-Verbatim Text #3531 round 4 brief addendum: no pass cap, orchestrator merges, attest before ready, and Copilot refusals on large Markdown. It also covers lists for list sentences, the copies rule, and settling a contested condition against the code.

Added to this track, to be ranked into the next round's next steps: #3577, the 23 sentence-length hits in lib/sh/secrets.sh at b72f50f3. The file is carried byte-locked into lib/fleet/sh, so every carrier's first carry pull request reports these hits and cannot fix them. That ranks it with the skill-source sweep (step 2) rather than with the hub-only text (step 3). The slice is that one file, and it shares no file with the split audits in flight. The Blog session filed it while carrying lib/fleet/sh (Blog #561).

Round 1, closed 2026-10-28: Session Handoff [sentence-length]: Split Audit of the Swept GOVERNANCE, AGENTS and WORKFLOW Text Next

Next steps, in priority order

Split-audit the sweeps already on main: Bring GOVERNANCE.md Sentences to 25 Words or Fewer #3551 (GOVERNANCE.md), Bring AGENTS.md and copilot-instructions.md Sentences to 25 Words or Fewer #3559 (AGENTS.md, .github/copilot-instructions.md), Bring WORKFLOW.md Sentences to 25 Words or Fewer #3561 (WORKFLOW.md). Compare every split each made with its original (git show <squash>) and repair any that lost meaning. Done: one PR per source file group, each split listed in its body, with a Copilot round whose table names every changed file. Each audit of a later sweep found 12 to 16 broken splits, so expect some here.
Sweep the remaining skill sources under .agents/skills/: 484 hits across 29 files. The largest are local-strict-review, comment-and-doc-style (with its references/), skill-lifecycle, repo-worktree, branching-and-release-model/references/release-publish-mechanics.md, and merge-and-release. Done: zero hits per file. Each brief requires split-by-split comparison before review, per the Say What a Carrier's Local Strict Review Does With Sentence-Length Hits on Hub-Verbatim Text #3531 brief comment.
Sweep the hub-only text: 1685 hits in code comments and docstrings, and 1518 in other Markdown (scripts/README.md, TODO.md, docs/, reports/). Lower priority, since no carrier copies it, but the ruling covers the whole hub.
Make sentence-length a default check once the tree reads zero. Delete the comment-and-doc-style text calling it opt-in "because the existing corpus predates the cap". Done: Say What a Carrier's Local Strict Review Does With Sentence-Length Hits on Hub-Verbatim Text #3531 closes on that promotion.
Make pr_review.py and Merge Gate Item 2 Agree on an Earlier-Head Copilot Refusal #3571: make pr_review.py and Merge Gate item 2 agree on a Copilot refusal that landed on an earlier head. Independent of the steps above.

External blockers

Copilot quota. On 2026-10-28 Copilot returned error refusals read as possible quota on Name the Undeclared Root Default and Bring CODESTYLE.md and Codestyle Skill Sentences to 25 Words or Fewer #3567, Bring PR Review Family Sentences to 25 Words or Fewer and Name the Sent and Spent Readings #3568, Drop the Registry Python Profile in Favor of Per-Directory Shape #3569, and Bring the Handoff Skills' Sentences to 25 Words or Fewer #3570, then reviewed normally from Promote Develop: Registry Python Profile, Codestyle and PR Review Sentence Length, Handoff Skills, and Blog host-secrets #3572 on. A refusal leaves Merge Gate item 2 to the maintainer. Each merge over one needs that maintainer's go-ahead for that pull request, recorded on it. Meanwhile, work and attest local passes, and stop at the merge.
Promotion coverage gaps. Copilot's promotion tables leave out the largest sweep diffs. Only the maintainer can accept other coverage, and that acceptance is per pull request.

Internal dependencies

Step 1 before step 4. The gate locks in whatever the text says, so meaning drift is fixed first.
Step 2 before step 4, and step 3 before step 4. The gate fails on any remaining hit.
Within a round, no two slices share a file, generated copies included. Each GOVERNANCE.md, AGENTS.md, CODESTYLE.md, or WORKFLOW.md section included by a skill regenerates that skill. Read the include markers (grep -rhoE 'include: [A-Z]+\.md > [^>]*' .agents/skills) before grouping.

State

Re-derive each of these. Do not trust the copy here.

main was 1abb46f2 and develop b72f50f3 at close. No promotion is open, no release was dispatched.
No worktree or branch of this run remains. pr_review.py cleanup --sweep reported four detached worktrees older than 24 hours from other sessions: ProjectTemplate-lint-homepage, hub-w1493, and two Financial-Modeling scratch hubs. They were not applied, since removal needs the maintainer.
Open issues in scope: Say What a Carrier's Local Strict Review Does With Sentence-Length Hits on Hub-Verbatim Text #3531 (this track), Make pr_review.py and Merge Gate Item 2 Agree on an Earlier-Head Copilot Refusal #3571, and Carry a CLI-Usage Test Harness So Fleet Repos Stop Copying the Hub's #3502, the held carried CLI harness proposal.
The run touched ptr727/Blog only by message. Its session waits on nothing from the hub now that Declare host-secrets for Blog and Name It the Overlay's First Adopter #3534 is on main.

The parked decision queue

Zero open decision issues at close. #3502 is a held design proposal, not a parked decision.

What the last round did

Driven by the backlog-burndown skill over four rounds and four promotions. Promotion #3544 went to main 65adbf17, #3553 to 54f477a6, #3562 to a5bfc9c2, and #3572 to 1abb46f2.

Resync-filed issues closed: Make the Model Tiers Bullets Agree on What a Dispatch Naming No Model Runs At #3529, Bring the Durable Knowledge Handoff Bullets Onto the Standing-Issue Lifecycle #3504, Triage Financial-Modeling's tests/test_audit.py as a Path Collision, and Consider Keeping Hub tests/ Out of the Hub-Only Detector #3528, Tell a Transient Copilot Fetch Error From the Quota Body Before Advising No Re-Request #3510, State That a Python Profile Belongs to a Declared Directory, and When a Repo May Declare Several #3492.
Follow-ups the run filed and closed: Make the Model Tiers Tool-Fallback Sentence and the Spec Note Agree With the No-Model Rule #3537, Read a Transient Copilot Error Inside the Digest's Refusal Block #3541, Name the Carried Trees Exclusion Wherever the Hub-Only Set Is Defined #3543, Name the Build Trait as a Deliverable in python-codestyle SKILL.md #3548, Stop Calling the Toolchain Bundle a Profile in python-codestyle profiles.md #3550, Bring the Spec's Python Profile Text Onto the Per-Directory Rule #3552, State the Root Gating Condition in the python-codestyle Skill #3554, Say What the Registry's Python Profile Means Once the Profile Is Per Directory #3556, Name the Sent and Spent Exceptions in pr_review.py's refusal= Usage Text #3560, Name the Undeclared Root Default in CODESTYLE.md Python #3563.
Pull requests: Make the Model Tiers Bullets Agree on What a Dispatch Naming No Model Runs At #3538 through Restore Reasons, Antecedents, and Condition Scope Lost in the CODESTYLE and PR Review Family Sentence Splits #3574, plus Blog's Declare host-secrets for Blog and Name It the Overlay's First Adopter #3534, taken over at the maintainer's direction.
Maintainer decisions recorded on their issues: Triage Financial-Modeling's tests/test_audit.py as a Path Collision, and Consider Keeping Hub tests/ Out of the Hub-Only Detector #3528 closes the class with repoOwnedTrees. Tell a Transient Copilot Fetch Error From the Quota Body Before Advising No Re-Request #3510 reads the run log. State That a Python Profile Belongs to a Declared Directory, and When a Repo May Declare Several #3492 gives one build directory per deliverable, with lint-only Python beside its component. Say What the Registry's Python Profile Means Once the Profile Is Per Directory #3556 drops the registry python profile. Say What a Carrier's Local Strict Review Does With Sentence-Length Hits on Hub-Verbatim Text #3531 says fix the text.
Filed and left open: Make pr_review.py and Merge Gate Item 2 Agree on an Earlier-Head Copilot Refusal #3571.

What not to repeat

Merging a sweep on local passes alone. Bring the Handoff Skills' Sentences to 25 Words or Fewer #3570, Name the Undeclared Root Default and Bring CODESTYLE.md and Codestyle Skill Sentences to 25 Words or Fewer #3567, and Bring PR Review Family Sentences to 25 Words or Fewer and Name the Sent and Spent Readings #3568 did, under maintainer authorization while Copilot refused. Each later audit found 28 broken splits between them. Cost: two extra promotion-fix PRs and a held promotion.
Posting text before its check passes. Three times this run, a PR body, an issue, and a comment went out with a sentence over 25 words. The command counted the hits instead of gating on the exit status. Gate the post as prose_lint.py --check sentence-length F && gh ....
Hiding a write's output. The write-guard refused a gh pr edit ... >/dev/null. Read every write's response.
Trusting a truncated read. A cut -c1-500 hid the end of WORKFLOW.md:142, and the orchestrator wrongly told a worker that D1.2 did not state a condition it does state.

New learnings

The sweep brief and both run lessons, with the drift classes and promotion table gaps: the Say What a Carrier's Local Strict Review Does With Sentence-Length Hits on Hub-Verbatim Text #3531 comment of 2026-10-28.
An earlier-head Copilot refusal reads coverage=local in the tool but unmet in the rule: Make pr_review.py and Merge Gate Item 2 Agree on an Earlier-Head Copilot Refusal #3571.

Closing answers for round 2, from the maintainer on 2026-10-28:

Worktree sweep: approved. cleanup --sweep --apply removed ProjectTemplate-lint-homepage and hub-w1493. It kept the two dirty Financial-Modeling scratch hubs.
Host-safety kit: the maintainer is running the installer from the primary checkout. Re-check it with install.py --report at resume rather than assuming it reads CURRENT.

Parked decision queue: zero open decision issues.

Round 2, closed 2026-10-28: Session Handoff [sentence-length]: Skill-Source Sweep and lib/sh/secrets.sh Next, Split Audits Shipped in 2.0.1234

Next steps, in priority order

Sweep the remaining skill sources under .agents/skills/: 484 hits across 29 files at develop b82a4a0d. The largest are local-strict-review (54), comment-and-doc-style (48, plus 24 in references/line-endings.md), skill-lifecycle (36), repo-worktree (35), branching-and-release-model/references/release-publish-mechanics.md (34), and merge-and-release (32). Two known leftovers sit there: 12 hits in agent-conduct outside its include regions, and one 28-word sentence in the workflow-ci-contract references/d-guarantees.md preamble. Done: zero hits per file. Each worker compares every split with its original before review, per the Say What a Carrier's Local Strict Review Does With Sentence-Length Hits on Hub-Verbatim Text #3531 brief comment.
Split the Over-Length Comment Sentences in lib/sh/secrets.sh #3577: the 23 hits in lib/sh/secrets.sh. The file is carried byte-locked into lib/fleet/sh, so every carrier's first carry PR reports them and cannot fix them. One-file slice, sharing no file with step 1. Done: prose_lint.py --check sentence-length lib exits 0.
Sweep the hub-only text: about 1685 hits in code comments and docstrings, and 1518 in other Markdown (scripts/README.md, TODO.md, docs/, reports/). Lower priority, since no carrier copies it, but the ruling covers the whole hub. Re-count before slicing.
Make sentence-length a default check once the tree reads zero. Delete the comment-and-doc-style text calling it opt-in "because the existing corpus predates the cap". Done: Say What a Carrier's Local Strict Review Does With Sentence-Length Hits on Hub-Verbatim Text #3531 closes on that promotion.
Make pr_review.py and Merge Gate Item 2 Agree on an Earlier-Head Copilot Refusal #3571: make pr_review.py and Merge Gate item 2 agree on a Copilot refusal that landed on an earlier head. Independent of the steps above.

External blockers

Copilot quota. Copilot refused on 2026-10-28 for Name the Undeclared Root Default and Bring CODESTYLE.md and Codestyle Skill Sentences to 25 Words or Fewer #3567 to Bring the Handoff Skills' Sentences to 25 Words or Fewer #3570, then reviewed normally from Promote Develop: Registry Python Profile, Codestyle and PR Review Sentence Length, Handoff Skills, and Blog host-secrets #3572 through Promote Develop: Restore Meaning Lost in the GOVERNANCE, AGENTS, and WORKFLOW Sentence Splits #3580. A refusal leaves Merge Gate item 2 to the maintainer, per pull request. Meanwhile, work and attest local passes, and stop at the merge.
Promotion coverage gaps. Earlier promotion tables left out the largest sweep diffs. Promote Develop: Restore Meaning Lost in the GOVERNANCE, AGENTS, and WORKFLOW Sentence Splits #3580's table named all 22 files, so expect it but do not assume it. Only the maintainer accepts other coverage, per pull request.

Internal dependencies

Steps 1, 2 and 3 before step 4. The gate fails on any remaining hit.
Within a round, no two slices share a file, generated copies included. Read the include markers (grep -rhoE 'include: [A-Z]+\.md > [^>]*' .agents/skills) before grouping. A skill-source slice edits text outside the include regions, so its only generated files are that skill's own copies and digest.

State

Re-derive each of these. Do not trust the copy here.

main 4aae620c (promotion Promote Develop: Restore Meaning Lost in the GOVERNANCE, AGENTS, and WORKFLOW Sentence Splits #3580) and develop b82a4a0d at close. Release 2.0.1234 published from main (run 40774865902, success).
No worktree or branch of this round remains. pr_review.py cleanup --sweep still lists two detached worktrees from other sessions it would remove (ProjectTemplate-lint-homepage, hub-w1493), and keeps two dirty Financial-Modeling scratch hubs. Removal needs the maintainer's go-ahead.
This machine's skills snapshot is current at 4aae620c. The host-safety kit reports STALE (installed from hub f60f57d). Its installer rewrites hooks and settings, so only the maintainer runs it.
Open issues in scope: Say What a Carrier's Local Strict Review Does With Sentence-Length Hits on Hub-Verbatim Text #3531 (this track), Split the Over-Length Comment Sentences in lib/sh/secrets.sh #3577, Make pr_review.py and Merge Gate Item 2 Agree on an Earlier-Head Copilot Refusal #3571, and Carry a CLI-Usage Test Harness So Fleet Repos Stop Copying the Hub's #3502, the held carried CLI harness proposal.
ptr727/Blog filed Carry an Indent Setting With lib/sh So Tab-Indented Carriers Pass shfmt #3576 and Split the Over-Length Comment Sentences in lib/sh/secrets.sh #3577 while carrying lib/fleet/sh (Blog Ask for a blocked decision instead of reporting it #561). Carry an Indent Setting With lib/sh So Tab-Indented Carriers Pass shfmt #3576, a .editorconfig for lib/sh, went to the default track (Session Handoff [default]: #3671 Thread-Aware Held Wait Released in 2.0.1286, Carried Backlog Next #2361) as a comment.

The parked decision queue

Zero open decision issues at close.

What the last round did

Split audit of the sweeps on main. Restore Meaning Lost in the GOVERNANCE and AGENTS Sentence Splits #3579 repaired 15 broken splits from Bring GOVERNANCE.md Sentences to 25 Words or Fewer #3551 and Bring AGENTS.md and copilot-instructions.md Sentences to 25 Words or Fewer #3559 in GOVERNANCE.md, AGENTS.md, and .github/copilot-instructions.md. Seven had a reason left on one half of a split. Restore Antecedents, Reasons, and Scope Lost in the WORKFLOW.md Sentence Splits #3578 repaired 10 of the 107 splits Bring WORKFLOW.md Sentences to 25 Words or Fewer #3561 made in WORKFLOW.md. Six had lost an antecedent.
Both had a Copilot round covering every changed file, and both merged into develop. Promotion Promote Develop: Restore Meaning Lost in the GOVERNANCE, AGENTS, and WORKFLOW Sentence Splits #3580 merged to main with all 22 files in Copilot's table.
Recorded Split the Over-Length Comment Sentences in lib/sh/secrets.sh #3577 on this track and Carry an Indent Setting With lib/sh So Tab-Indented Carriers Pass shfmt #3576 on default, at the Blog session's request.

What not to repeat

Trusting a sweep's local passes for meaning. All five audits so far found 10 to 16 broken splits per sweep. A sweep worker compares each split with its original before review. An audit is still owed for any sweep merged without that.
Running hub scripts with bare python3 on this host. build_dist.py and pr_review.py import markdown_it, which the system Python lacks. Use uv run --no-project --script scripts/<name>.py.
Posting text before its check passes. Gate the post as prose_lint.py --check sentence-length F && gh .... The promotion worker's first body draft failed the gate, and the gate stopped it.

New learnings

Nothing new beyond the Say What a Carrier's Local Strict Review Does With Sentence-Length Hits on Hub-Verbatim Text #3531 brief comment of 2026-10-28, which still holds. The audit counts above confirm its split-drift lesson.

Closing answers for round 3, from the maintainer on 2026-10-28:

Develop merges: authorized for this session's sentence-length slices only. Each slice was merged green and resolved, and the promotion stopped for the maintainer.
Worker tier: Sonnet workers with opus local strict review.
Promotion Promote Develop: Split the Over-Length Sentences in the Skill Sources and the Secrets Shell Library #3589 coverage: feature-pull-request coverage accepted for the 35 files Copilot's table omitted, recorded on Promote Develop: Split the Over-Length Sentences in the Skill Sources and the Secrets Shell Library #3589.
Promotion scope: merge, Skills refresh, and release. Release 2.0.1242 published.

Parked decision queue: zero open decision issues.

Round 3, closed 2026-10-29: Session Handoff [sentence-length]: Hub-Only Markdown and Comment Sweep Next, Skill-Source Sweep Shipped in 2.0.1242

Next steps, in priority order

Sweep the hub-only Markdown: 1518 hits at main 13c2b8b1. The largest are scripts/README.md (303), TODO.md (178), docs/reusable-workflows.md (126), STANDUP.md (83), docs/fleet-map.md (78), AUDIT.md (73), docs/host-setup.md (70), and README.md (61). Re-count before slicing. Done: zero hits per file. Use the Say What a Carrier's Local Strict Review Does With Sentence-Length Hits on Hub-Verbatim Text #3531 brief and its round 3 addendum as the worker brief.
Sweep the hub code comments and docstrings: 1662 hits, mostly under scripts/ and tests/. A comment-only pull request needs the comments label and PROSE_ALLOW_COMMENTS=1 on its commits, as Split the Over-Length Sentences in the Secrets Shell Library Comments #3581 did. Done: the whole tree reads zero.
Make sentence-length a default check once the tree reads zero. Delete the comment-and-doc-style text calling it opt-in "because the existing corpus predates the cap". Done: Say What a Carrier's Local Strict Review Does With Sentence-Length Hits on Hub-Verbatim Text #3531 closes on that promotion.
Move the Dispatch-Refresh Note Out of the Docker Clause in branching-and-release-model #3588: move the dispatch-refresh note out of the Docker clause in branching-and-release-model. One small skill edit, sharing no file with steps 1 and 2.
Make pr_review.py and Merge Gate Item 2 Agree on an Earlier-Head Copilot Refusal #3571: make pr_review.py and Merge Gate item 2 agree on a Copilot refusal that landed on an earlier head. Independent of the steps above.

External blockers

Promotion coverage gaps. Copilot's table on Promote Develop: Split the Over-Length Sentences in the Skill Sources and the Secrets Shell Library #3589 named 71 of 106 files, and its round on the final head carried no table. The maintainer accepted feature-pull-request coverage, recorded on Promote Develop: Split the Over-Length Sentences in the Skill Sources and the Secrets Shell Library #3589. Expect the same on each sweep promotion, and bring it to the maintainer each time.
Copilot quota. No refusal this round. A refusal leaves Merge Gate item 2 to the maintainer, per pull request.

Internal dependencies

Steps 1 and 2 before step 3. The gate fails on any remaining hit.
Within a round, no two slices share a file. A hub-only doc may be an include source, so read the include markers (grep -rhE '^<!--\s*include:' .agents/skills) before grouping. Keep a doc that feeds a skill in the same slice as that skill's regenerated copies.

State

Re-derive each of these. Do not trust the copy here.

main 13c2b8b1 (promotion Promote Develop: Split the Over-Length Sentences in the Skill Sources and the Secrets Shell Library #3589) and develop 7364e141 at close. Release 2.0.1242 published from main (run 40804010827, success).
No worktree or branch of this round remains. All eight feature branches were cleaned with pr_review.py cleanup. The cleanup --sweep report lists only other sessions' worktrees, kept as dirty or too recent.
This machine's skills snapshot is current at 13c2b8b1, and the host-safety kit reads CURRENT.
Open issues in scope: Say What a Carrier's Local Strict Review Does With Sentence-Length Hits on Hub-Verbatim Text #3531 (this track), Move the Dispatch-Refresh Note Out of the Docker Clause in branching-and-release-model #3588, Make pr_review.py and Merge Gate Item 2 Agree on an Earlier-Head Copilot Refusal #3571, and Carry a CLI-Usage Test Harness So Fleet Repos Stop Copying the Hub's #3502, the held carried CLI harness proposal. Split the Over-Length Comment Sentences in lib/sh/secrets.sh #3577 closed on Promote Develop: Split the Over-Length Sentences in the Skill Sources and the Secrets Shell Library #3589.

The parked decision queue

Zero open decision issues at close. The maintainer answered this round's two questions in session. They accepted feature-pull-request coverage on #3589, and chose merge, Skills refresh, and release.

What the last round did

Skill-source sweep, done. Seven slices took the skill sources and lib/sh/secrets.sh from 507 hits to zero: Split the Over-Length Sentences in the Local Review, Agent Conduct, and Check This Repo Skills #3584, Split the Over-Length Sentences in the Comment and Doc Style Skill #3586, Split the Over-Length Sentences in the Repo Procedure Skills #3587, Split the Over-Length Sentences in the Git and Worktree Skills #3582, Split the Over-Length Sentences in the Release Skills #3585, Split the Over-Length Sentences in the Codestyle and Guard Skills #3583, and Split the Over-Length Sentences in the Secrets Shell Library Comments #3581. Promotion fix Restore a Relative Pronoun and Split a Long Reference Sentence in agent-conduct #3590 answered two Copilot findings on Promote Develop: Split the Over-Length Sentences in the Skill Sources and the Secrets Shell Library #3589, and Promote Develop: Split the Over-Length Sentences in the Skill Sources and the Secrets Shell Library #3589 merged to main.
Filed Move the Dispatch-Refresh Note Out of the Docker Clause in branching-and-release-model #3588, a pre-existing misplaced parenthetical the release-skills worker found and left alone.
Recorded this round's sweep lessons as an addendum to the Say What a Carrier's Local Strict Review Does With Sentence-Length Hits on Hub-Verbatim Text #3531 brief.

What not to repeat

A stop rule that does not say what it counts. "More than two fix rounds" was read as two review passes over the branch. One worker stopped unpushed, and another pushed with known drifts. State it per sentence.
Letting workers settle a contested sentence. On slice A, opus reviewers flipped one description sentence for four rounds. An orchestrator ruling against the original, marked final, ended it in one pass.
Calling a pull request clean at 0 threads. The body of a Copilot round can carry [P2] items with no thread, and on Split the Over-Length Sentences in the Codestyle and Guard Skills #3583 both were real drifts.

New learnings

Recorded in the Say What a Carrier's Local Strict Review Does With Sentence-Length Hits on Hub-Verbatim Text #3531 round 3 brief addendum: the per-sentence stop rule, orchestrator rulings, body-only Copilot findings, and the gate not counting quoted strings.

Closing answers for round 4, 2026-10-29:

Host-safety kit: the maintainer re-ran install.py after release 2.0.1251. install.py --report reads CURRENT at a tree identical to main 77d7ca35, so the round 4 "Host-safety kit is stale" blocker is resolved.

Parked decision queue: zero open decision issues.

Outcome: track complete (round 5, closed 2026-10-29)

The sentence-length track's goal is done. The whole hub tree reads 0 sentence-length hits, down from 6964 when the track started. The rule is now a default prose-gate check fleet-wide, and #3531 is closed. Shipped in release 2.0.1262, from main f24e9f17 (promotion #3619, publish run 40935484439, success).

What round 5 did

Hub code comment sweep. Eight slices took code comments and docstrings from 1662 hits in 101 files to zero: Split the Over-Length Comment Sentences in the Local and Canonical Review Engines #3607, Split the Over-Length Comment Sentences in build_dist, carry, handoff, and skills_install #3608, Split the Over-Length Comment Sentences in pr_review and host_gate #3606, Split the Over-Length Comment Sentences in the Prose and Repo Gates #3609, Split the Over-Length Comment Sentences in the Spec Scripts and Workflows #3611, Split the Over-Length Comment Sentences in the Host Install Scripts #3612, Split the Over-Length Comment Sentences in the Host Menu, Bootstrap, and Setup Scripts #3613, and Split the Over-Length Comment Sentences in the Agent-Safety Kit, Repo Config, and Snippets #3615.
Make sentence-length a Default Prose Gate Check #3618 made sentence-length default in prose-gate.json, and dropped the opt-in text from comment-and-doc-style, local-strict-review, GOVERNANCE.md, OPERATIONS.md, docs/fleet-map.md, and scripts/README.md. It closed Say What a Carrier's Local Strict Review Does With Sentence-Length Hits on Hub-Verbatim Text #3531.
Move the Dispatch-Refresh Note Out of the Docker Clause in branching-and-release-model #3603 closed Move the Dispatch-Refresh Note Out of the Docker Clause in branching-and-release-model #3588, the dispatch-refresh parenthetical in branching-and-release-model.
Describe the sentence-length Default in Present Tense in the Scripts README #3621 applied a promotion finding: present tense in scripts/README.md.
Coverage. The maintainer accepted feature-PR coverage for the 27 files Copilot's table omitted on Promote Develop: Split the Over-Length Comment Sentences and Make sentence-length a Default Check #3619, recorded on Promote Develop: Split the Over-Length Comment Sentences and Make sentence-length a Default Check #3619.
Filed or annotated:
- ruff@latest in the Lint-Only Validator Turns Carriers Red on a Ruff Release With No Change of Their Own #3616 (decision): ruff@latest in the lint-only validator turns carriers red on a ruff release.
- Align the Comment Copies the Sentence Sweep Split Two Ways, and Fix Three Stale Counts #3622: align the comment copies F1 and F2 split differently, and fix three stale counts.
- Drop the Stale RegressionTests Clause From PlexCleaner's driftNote Once PlexCleaner #1064 Reaches main #3623: drop PlexCleaner's stale RegressionTests driftNote clause once PlexCleaner Tighten Dependabot Exclude-Patterns Rationale Comment #1064 reaches its main.
- Commented qodo_open Double-Counts and Never Clears the Open-Source Qodo App's Threaded Findings #2252, the Qodo dismiss-by-number evidence from PlexCleaner pr-review-conduct: check budget/quota status before reporting a reviewer bot as broken #1060, plus the todo to verify pr_review.py against a live Qodo PR.
- Commented Adopting the Gates omits three preconditions of validate-task, and each fails silently or confusingly #1516: the lint-only mypy step needs files.

Where the remaining work lives

Each is an ordinary backlog issue now rather than a step on this track:

Align the Comment Copies the Sentence Sweep Split Two Ways, and Fix Three Stale Counts #3622 (follow-ups from this sweep)
Make pr_review.py and Merge Gate Item 2 Agree on an Earlier-Head Copilot Refusal #3571 (pr_review.py and Merge Gate item 2 on an earlier-head Copilot refusal)
Nothing Checks That the Host CLAUDE.md Blocks Stay in Sync With Their Governance Sections #3601 (host CLAUDE.md block sync check, needs a design pick)
qodo_open Double-Counts and Never Clears the Open-Source Qodo App's Threaded Findings #2252 (Qodo qodo_open, verify on PlexCleaner)
ruff@latest in the Lint-Only Validator Turns Carriers Red on a Ruff Release With No Change of Their Own #3616 (decision)
Carry a Newly Declared Verbatim Section in carry.py Instead of Refusing the Resync #3614, Decide How the Workspace-Edit Rule Reads an Unknown Window Session State #3617, Triage Five CodeRabbit Findings on Carried Skills From a Downstream Resync #3620 (from the PlexCleaner resync)
Drop the Stale RegressionTests Clause From PlexCleaner's driftNote Once PlexCleaner #1064 Reaches main #3623 (blocked on PlexCleaner's promotion)

State at close

main is f24e9f17 and develop is 487deaa6. The primary checkout is on develop, current, and clean.
Every branch and worktree this round created is removed. cleanup --sweep lists only other sessions' worktrees, kept as too recent.
The skills snapshot reads current at f24e9f17.
The host-safety kit reads STALE, since Split the Over-Length Comment Sentences in the Agent-Safety Kit, Repo Config, and Snippets #3615 changed its comments. Only the maintainer reruns host-setup/agent-safety/claude/install.py.

What not to repeat

Recording a local pass no reviewer ran. One worker recorded digests for two unreviewed commits "itself". A real whole-branch pass afterwards found about 15 genuine split drifts. The brief now says a recorded pass must be one a reviewer actually ran.
Handing back before ready. Two workers called the handback while their own reviewer or CI wait was still running, which ends the run. Brief "do not hand back until ready or blocked".
Force-pushing after a rebase. One worker tried it, and the permission system denied it. A squash merge needs no rebase, so merge develop in instead.
Assuming a develop merge leaves a promotion alone. A promotion's head is develop, so any develop merge moves it and triggers a new Copilot round.
A shared scratch helper. Two workers overwrote each other's apply.py. Namespace helpers per branch.
Splitting a shared .ps1/.sh comment in two parallel slices. F1 and F2 split the same copies differently (Align the Comment Copies the Sentence Sweep Split Two Ways, and Fix Three Stale Counts #3622). Put both halves of a copy pair in one slice.

New learnings

Orchestrator rulings against the original text, marked final, ended every repeated-flag loop in one pass. That covered six sentences across Split the Over-Length Comment Sentences in the Prose and Repo Gates #3609 and Split the Over-Length Comment Sentences in the Spec Scripts and Workflows #3611. Recorded here, as the round 3 addendum on Say What a Carrier's Local Strict Review Does With Sentence-Length Hits on Hub-Verbatim Text #3531 already states it.
A Copilot "needs a closer look" headline on Split the Over-Length Comment Sentences in the Local and Canonical Review Engines #3607 named a real pre-existing defect ("Two kinds" describing three). It was worth fixing, even though the PR was comment-only.
Plain uv run pytest fails to spawn on this host. uv run --no-project --with-requirements tests/requirements.txt python tests/<file>.py works. This is recorded in this machine's memory.

The parked decision queue

There is one open decision issue: #3616, ruff pinning, filed this round. It is put to the maintainer in this session, and the answer is commented here.

Closing answer, 2026-10-29: the maintainer decided #3616 as a Dependabot-managed ruff pin with no fleet canary, recorded on #3616, and its decision label is removed.

Parked decision queue after this round: zero open decision issues.

Session Handoff [sentence-length]: Hub Code Comment Sweep Next, Hub-Only Markdown Sweep Shipped in 2.0.1251 #3575

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 28, 2026

ptr727 commented on Oct 28, 2026

Round 1, closed 2026-10-28: Session Handoff [sentence-length]: Split Audit of the Swept GOVERNANCE, AGENTS and WORKFLOW Text 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 28, 2026

ptr727 commented on Oct 29, 2026

Round 2, closed 2026-10-28: Session Handoff [sentence-length]: Skill-Source Sweep and lib/sh/secrets.sh Next, Split Audits Shipped in 2.0.1234

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 29, 2026

ptr727 commented on Oct 29, 2026

Round 3, closed 2026-10-29: Session Handoff [sentence-length]: Hub-Only Markdown and Comment Sweep Next, Skill-Source Sweep Shipped in 2.0.1242

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 29, 2026

ptr727 commented on Oct 29, 2026

Outcome: track complete (round 5, closed 2026-10-29)

What round 5 did

Where the remaining work lives

State at close

What not to repeat

New learnings

The parked decision queue

ptr727 commented on Oct 29, 2026

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions