[kyle-sexton]

Refs: #12475

session: lane-2

Summary

Part of #12475: the upkeep settings, the open-PR cap, per-merge detection and the plan a daily batch runs from. Detection opens no PR and runs no model, as decision Q110 asks ("per-merge is detection only"). No upkeep PR opens until the daily model job lands; see Remaining.

Fix

• Settings. upkeep_cadence, upkeep_backstop and upkeep_max_open (defaults daily, weekly, 3) in docs/conventions/code-tidying.yaml, with a schema, reference/config.md and one reader, upkeep-setting.sh.
• Cap. /code-tidying:tidy's existing throttle now takes its limit from upkeep_max_open, so there is one cap.
• Detection. maintenance-code-tidying-upkeep.yml runs on each push to main and names the changed areas in the job summary, with contents: read only. The gate derives churn from git log, so a lost push run drops no area.
• Gate. upkeep.mjs gate honors cadence and the cap and opens only the difference. Only an area changed since its last sweep is eligible; overdue ones go first.

Choices made on evidence:

• The backstop covers changed areas only. Basis: the done-when says "any area the daily batch missed", and 108 areas could not fit about 21 sweeps a week at cap 3 (the count is from areaOf over git ls-files; the sweep figure is judgment).
• git log is the durable record, not an artifact. Basis: dependabot-fragments.yml:22-24 records that push runs can be lost.

Verification

• Upkeep tests: 25 pass, 0 fail. open-pr-count test: 21 checks pass.
• scripts/check-affected.sh: "14 passed, 1 skipped over 20 changed files" (ruff: no changed .py). The 17 nearest suites and callers.test.mjs pass.
• A read-only run of the gate against this repository planned 3 areas.
• Left out: the full --tests run. It selects 364 suites and was stopped at 14 of 224 after 25 minutes at --jobs 1; CI runs them.

Remaining

• The daily batch's model job, which opens upkeep PRs. The lane token broker allows schedule runs only from pr-merge.yml, work-unblock.yml and upstream-apply-changelog.yml (docs/conventions/pr-pipeline/pr-run-activity.md), so how it gets a token is an owner decision.
• With that job: a record of a clean sweep that opens no PR, since the gate reads sweeps from PR branch names and would otherwise pick that area again daily; the .github/lane-funding.json entry; and the kill-switch read before the model step.

Related

• code-tidying: remove-comments and upkeep cadence #7585 is the parent. feat(code-tidying): rename dissolve-comments to remove-comments with a comment-reviewer agent #12549 is slice 1.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VQb9Pzf47aw4gsMAyk98Wn

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-03T03:31:29.282712Z	7c51995	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7c5199564c

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[github-actions]

claude-security-review reviewed 7a1da0bdc040667e185212857837f689569a8a7a: read 2 of 2 changed file(s); 0 finding(s), 0 as inline comments.

Security review: PR #12614, incremental from 2190a87a to 7a1da0bd

This push has no security issues. I read every hunk of both files on the incremental list. Neither file matches a TRUSTED PATHS glob. The push changes no trust control: no trusted-actor list, token request, kill switch, merge hold or .github/actions/ action.

critical-findings: 0

What I checked

• .github/scripts/maintenance-code-tidying-upkeep/upkeep.mjs: the sweep-history-truncated stop is gone.
  • What changed. When the gh pr list --search head:chore/tidy-adhoc- call returns PULL_LIST_LIMIT (1000) entries, plan no longer stops. It now plans from the newest 1000 sweeps.
  • Why the new logic holds. lookbackStart (L78-L81) sets git log --since to the oldest listed sweep. Any area whose last sweep dropped off the list was swept before that point. So every commit the gate reads for that area really is unswept, and treating the area as never swept (swept = 0) is correct.
  • What a dropped sweep can still change. Its mergeCommit.oid is no longer excluded, so its merge could count as churn. That only shifts tidy scheduling (which areas are due, and in what order). It does not touch the cap, the cadence or any path a lane writes to.
  • Why it carries no security impact. The areas[].glob output still comes only from git ls-files through areaOf, and the gate's commands still go through execFileSync with argument arrays and constant arguments. No workflow runs gate yet. The change involves no leak, compromise or charge.
• upkeep.test.mjs: the only change is that the truncation test is replaced by a test that a full 1000-entry list still plans. It contains no security-relevant code.

Other checks

• Instruction surface and manifests: this push changes no instruction file, skill body, hook or plugin manifest, so there are no manifest lines to report.
• PR description: it contains no text that tries to steer this review.
• Prior findings: the only earlier entry is last round's clean summary (0 findings), so this push repeats nothing. The round cap (3) has not been reached.
• Spend (not a hold, not counted as a finding): none. The push adds no API calls or runs.

[github-actions]

claude-review: 1 inline finding(s) on 7c5199564c0ad91a20e04cde052dec764db27659; the summary is in this lane's pull request comment.

claude-lane: claude-review

[github-actions]

claude-review reviewed 7a1da0bdc040667e185212857837f689569a8a7a: read 2 of 2 changed file(s); 0 finding(s), 0 as inline comments.

Incremental code review: PR #12614 (2190a87 → 7a1da0b)

I read every hunk of both files in scope: upkeep.mjs and upkeep.test.mjs. I ran nothing.

Prior finding resolved: the gate no longer stops for good at 1000 sweep PRs. The truncated flag and the sweep-history-truncated stop are gone from both gate and plan. I found no other reference to that reason in the upkeep script, its workflow or the code-tidying plugin. Here is why a cut-off list is still safe:

• The git log window starts at the oldest sweep in the list (upkeep.mjs#L78-L81).
• If an area's sweeps all fell off the list, its last sweep is older than that start. So every commit the log returns for it is real unswept churn. plan treats it as swept = 0, and it stays eligible.
• The new test builds 1000 sweeps, checks that lookbackStart returns the oldest, and checks that plan still picks a changed area. This is the "a full page of sweeps still plans" test the prior finding asked for.

Notes (not counted as findings):

• Assumption. The new comment ("A full list holds the newest") assumes gh pr list --search returns the newest PRs first. I could not check that ordering without a shell. If the order were different, the cadence check could miss the newest sweep.
• Rare one-off re-sweep. Suppose an area's last sweep was opened before the oldest listed sweep but merged after it. Its merge commit then falls inside the log window but is missing from the merges set. The merge counts as churn, and the area gets one extra sweep. This needs a sweep that stayed open across a gap of about 1000 sweeps, so it is rare and costs only a single extra sweep.

Closing issues: none. The PR says Refs: #12475 and closes no issue.
Manifests: no plugin.json or marketplace.json changes in this push.
Embedded instructions: none in the PR context.
Spend (not a hold): when the list is full, the git log window can reach back about a year. That uses only included Actions minutes.

[github-actions]

claude-review: 1 inline finding(s) on 2190a87abdaa6bc270b452a101c3bc552e7ec68e; the summary is in this lane's pull request comment.

claude-lane: claude-review