Affected Repositories

• PyAutoBrain (primary)

Branch Survey

Suggested branch: feature/repo-settings-org-enumeration

Implementation Steps

1. .github/workflows/repo_settings.yml

• Add ORG: ${{ github.repository_owner }} to the step env: alongside GH_TOKEN, rather than interpolating ${{ }} inline in the run: block.
• Replace the single targets= line (currently line 84) with two lists:

  in_org=$(gh api --paginate "orgs/$ORG/repos?per_page=100" \
             --jq '.[] | select(.archived == false) | .full_name')
  outside=$(python3 bin/branch_sweep_targets.py --outside-owner "$ORG" .mind/repos.yaml)

• Retarget the refuse-to-continue guard: an empty in_org is an error, never a clean run.
• Failure-policy split in the loop: in-org PATCH refusal or unreadable setting keeps failed=1; out-of-org refusal emits ::warning:: only. Today every refusal sets failed=1, which would make the weekly schedule permanently red once personal-account repos are included.
• Add a Scope/Owner column to the $GITHUB_STEP_SUMMARY table so the split is visible in the run.
• Rewrite the header comment: the "27 repos" count and the "a repo added to the body map after today is corrected on the next run" sentence both become false. Replace with the real reason enumeration is live — a new repo is not in the body map at the moment it is created.

2. bin/branch_sweep_targets.py

• Add --outside-owner <owner> → body-map slugs whose owner part != <owner>. Deliberately no category filter: the org enumeration has no category boundary, so its complement must not have one; filtering would yield nothing, since the only two foreign repos are categories admin and assistant.
• Remove --include-self-sweeping and the include_self_sweeping parameter. Sole consumer verified to be repo_settings.yml:84. An unused widening flag on a boundary whose other modes delete refs is a trap.
• Do not touch targets() or SWEEPABLE_CATEGORIES — they still serve branch_sweep_all.yml:98 and branch_archive.yml:110, where the action is deleting refs and the narrow boundary is correct.
• Update the module docstring: drop the --include-self-sweeping paragraph, add the --outside-owner rationale.

3. tests/test_branch_sweep_targets.py

• Delete test_include_self_sweeping_keeps_the_organs_that_sweep_themselves with the flag.
• Repoint test_cli_flag_matches_the_function at --outside-owner.
• Add: foreign-owner slugs yielded regardless of category; an all-in-org body map yields empty; CLI/function parity for the new flag.
• Keep the fail-closed tests as they are (unknown flag → exit 2, missing body map → exit 1, unknown/missing category excluded).
• Fixtures keep invented repo names — this file is .py under a framework organ and the tenant firewall scans it.

4. Stale prose

• skills/repo_cleanup/SKILL.md:133 — "body-map-wide" becomes org-wide.
• skills/repo_cleanup/reference.md:164 — one sentence noting the settings sweep no longer uses this module's category boundary.
• skills/prm/reference.md:311 and skills/GITHUB_ACCESS.md:81 say only that the setting is kept on per repo — still true, unchanged.

Key Files

• .github/workflows/repo_settings.yml — the sweep; target derivation and failure policy
• bin/branch_sweep_targets.py — body-map derivation; gains --outside-owner, loses --include-self-sweeping
• tests/test_branch_sweep_targets.py — the boundary's test surface
• skills/repo_cleanup/{SKILL.md,reference.md} — prose that goes stale

Trade-offs / risks

• Reverses part of PR#290, merged hours earlier. That PR added a test asserting the category boundary must not widen. This task does not widen it — it bypasses it for one consumer that enumerates the org instead — but the --include-self-sweeping flag it introduced is retired. The PR body must say so and why.
• Org enumeration sweeps repos nobody classified (.github, PyAutoLogo, both -template repos, the typo repo autolens_workspace_developer-). Flipping a reversible boolean there is harmless, but it is a real widening.
• Main risk: PAT_PYAUTOLABS may lack admin on some org repo a personal token can reach. Unknown until the workflow runs — it never has.

Testing

• python3 bin/branch_sweep_targets.py --outside-owner PyAutoLabs ../PyAutoMind/repos.yaml → exactly Jammy2211/admin_jammy and Jammy2211/euclid_assistant.
• python3 bin/branch_sweep_targets.py ../PyAutoMind/repos.yaml | wc -l → still 25 (measured before the change; the plain derivation must not move).
• pytest tests/ -q → baseline measured at 542 passed (4m25s), minus the one removed test, plus the new ones.
• bash bin/check_skill_line_counts.sh; YAML parse of the workflow; repos_sync.py --check --only "tenant firewall (organ code)".
• Post-merge: dispatch the workflow in audit mode first (it has never run) — expect ~44 in-org + 2 out-of-org rows, all "already on", green. Only then dispatch apply.

The repo-settings sweep should enumerate the org, not the body map

Type: feature
Target: pyautobrain
Repos:

• PyAutoBrain
  Difficulty: small
  Autonomy: supervised
  Priority: normal
  Status: formalised
  Filed: 2026-08-26

The original request (verbatim)

I want to make it so that when a branch merges via PR, on all repos, its auto
deleted. Can we do that here or do I need to manually go to each repo on
Github.com and do it?

and, once told PyAutoBrain#290 had already shipped most of it:

make it so new repos do inherit this, we should of also just made it so prm
doesnt try to delete anymore, I think the mobile chat pushed but chec

What is already done — do not redo it

PyAutoBrain#290 (merged 5ac5586, 2026-08-26, from a mobile session) shipped:

• .github/workflows/repo_settings.yml — a weekly + dispatchable sweep that
  PATCHes delete_branch_on_merge=true, targets derived from the body map.
• bin/branch_sweep_targets.py --include-self-sweeping — the widened organ
  boundary that sweep uses, plus tests/test_branch_sweep_targets.py.
• /prm deletes no remote branch on any surface (skills/prm/prm.md:43,
  step 5.6). The second half of the request above is therefore already
  satisfied — do not reopen it.

Separately, all 46 live repos (44 PyAutoLabs/* + Jammy2211/admin_jammy +
Jammy2211/euclid_assistant) were PATCHed to true by hand in the filing
session and verified by re-reading the API. This task is about durability for
repos the sweep cannot currently see — not about today's state.

The hole

The sweep derives its targets from PyAutoMind/repos.yaml. Two consequences:

1. A new repo is invisible until someone registers it. Org repos are created
   by hand on github.com — there is no gh repo create path in the organism for
   them (checked: the four assistants' start-new-project.md create a visiting
   scientist's own private project repo under their account; the clone
   conductor has no create call at all, clone/DESIGN.md:148 still lists repo
   creation as an open v0 question; /spawn force-syncs template repos that
   already exist). So birth-time is not a hook we have, and registration is the
   only thing standing between a new repo and the sweep.
2. Nine live org repos are absent from the body map and so are never swept:
   .github, PyAutoLogo, PyAutoMemory-template, PyAutoMind-template,
   PyAutoProject, autolens_jax_joss, autoproject_workspace,
   autoproject_workspace_test, and the typo repo
   autolens_workspace_developer-.

A third, smaller gap: the category boundary excludes assistant (5),
pipeline (1), project (3) and admin (1), so 10 of the 37 registered
repos are skipped too — 27 of 37 swept. Those exclusions are correct for the
branch sweep, whose action is deleting refs, and PR#290 added a test
asserting they must not widen. They do not transfer to flipping a reversible
boolean on repos that demonstrably merge PRs (autolens_assistant#115,
autogalaxy_assistant#19, autocti_assistant#26 all merged this week).

Decision (human, at filing time)

Enumerate the org live. Targets become every non-archived repo under
${{ github.repository_owner }} read from the API, union any body-map repo
whose owner is not the org. This closes all three gaps at once and needs no
category boundary at all. The owner comes from the GitHub context, so no
instance fact enters organ code and the tenant firewall stays satisfied.

Do not touch the assistants' start-new-project.md. That repo belongs to
the visiting scientist, under their own account, and the skill presents it as a
backup/collaboration surface that may not use PRs. Imposing our merge policy on
it is our convention leaking into someone else's workspace.

Notes for the implementer

• --include-self-sweeping has exactly one consumer, repo_settings.yml:84.
  Org enumeration removes that consumer, so the flag becomes dead code on a
  safety-relevant boundary — retire it with its two tests rather than leaving it
  loaded. This reverses part of a PR merged hours earlier; say so in the PR body.
• The plain derivation still has two consumers (branch_sweep_all.yml:98,
  branch_archive.yml:110) and its narrow category boundary must not move.
• Out-of-org repos are the user's own (Jammy2211/*); PAT_PYAUTOLABS admin
  scope there is unverified. A PATCH refusal on those must warn, not fail —
  the job currently sets failed=1 on any refusal and would go red every week.
• Stale claims to update once the source changes: the workflow's own header
  comment ("27 repos", body-map framing), skills/repo_cleanup/SKILL.md:133
  ("body-map-wide"), and the derivation paragraph at
  skills/repo_cleanup/reference.md:164 if it describes the settings consumer.
• The workflow has never actually run. Dispatching it after merge is the smoke
  test: with every repo already true, apply should report all-already-on.