ci: add a PR test workflow (and the sizing skill wrapper that blocks it)

[Jammy2211]

Overview

PyAutoBrain has no PR test workflow. .github/workflows/ holds only
docs.yml (path-filtered to docs/** + .readthedocs.yaml, so it never fires
for agents/, bin/ or tests/) and nightly-release.yml (a cron release
scheduler). A Brain PR's head sha returns total_count: 0 check runs — #193
merged that way earlier today. So the ~193 tests in tests/ execute in no CI
anywhere, and a Brain PR's only gate is whatever the authoring session
happens to run locally.

Turning CI on requires fixing the suite first: test_skill_install.py:: test_every_public_agent_has_a_skill_wrapper is already red on main (the
sizing faculty has no skills/sizing/SKILL.md) — which is exactly what no CI
looks like. Both parts land in one PR, because merging a workflow that is known
to fail would be shipping a red gate.

Plan

• Add the missing sizing skill wrapper, mirroring the sibling faculty
  wrappers — sizing is listed under "Faculties (read-only … also runnable
  directly)" in pyauto-brain help and ships its own sizing.sh, so the
  wrapper is the missing piece; delisting it from help would be the wrong fix.
• Add a tests.yml modelled on PyAutoHeart/.github/workflows/heart-tests.yml,
  which is the house pattern for an organ's own unit suite.
• Check out PyAutoBrain and PyAutoMind side by side — measured, not assumed:
  a lone-repo checkout cannot even collect the suite.
• Keep the workflow to pytest only (no pyauto-brain agent runs, no network),
  the same discipline heart-tests.yml states for itself.

Detailed implementation plan

Affected Repositories

• PyAutoBrain (primary, only)

Branch Survey

Repository	Current Branch	Dirty?
./PyAutoBrain	main	clean

No worktree claim on PyAutoBrain.

Suggested branch: feature/pyautobrain-pr-test-ci
Worktree root: ~/Code/PyAutoLabs-wt/pyautobrain-pr-test-ci/

Measured findings (2026-08-04, before writing anything)

Cloned PyAutoBrain alone into a scratch dir and ran the suite with a clean
env — it does not fail, it does not even collect:

ERROR collecting tests/test_policy_seams.py
ERROR collecting tests/test_sizing_paths.py
E   FileNotFoundError: .../ci-probe/PyAutoMind/repos.yaml
agents/faculties/sizing/_sizing.py:105: LIBRARY_REPOS, WORKSPACE_REPOS, ORGANISM_REPOS = _target_sets()

_sizing.py reads the body map at import time and is deliberately strict
(BODY_MAP_PATH = BRAIN_HOME.parent / "PyAutoMind" / "repos.yaml"). Adding
PyAutoMind alongside:

192 passed, 1 failed in 29.02s
FAILED tests/test_skill_install.py::test_every_public_agent_has_a_skill_wrapper — assert ['sizing'] == []

So: two repos are required, no others; the suite needs only pytest + PyYAML;
it takes ~30s; and the single failure is the sizing wrapper. Both repos are
public, so the default GITHUB_TOKEN can check out the sibling.

Implementation Steps

1. skills/sizing/SKILL.md — frontmatter (name, description) plus the
   3-line body the other faculty wrappers use: read
   ../../agents/faculties/sizing/AGENTS.md, then run
   bin/pyauto-brain sizing. Must stay under the 200-line primary-file budget
   (bin/check_skill_line_counts.sh) and its markdown links must resolve
   (test_local_skill_links_resolve).
2. skills/sizing/agents/openai.yaml — the Codex descriptor every sibling
   faculty carries (display_name / short_description / default_prompt).
3. .github/workflows/tests.yml:
   • on: push to main + pull_request; concurrency group keyed on the ref
     with cancel-in-progress only off main (a cancelled main run reads as
     red CI in Heart's FAILURE_CONCLUSIONS).
   • matrix python 3.12, 3.13 (matches heart-tests.yml).
   • actions/checkout@v4 twice, with path: PyAutoBrain and
     repository: PyAutoLabs/PyAutoMind + path: PyAutoMind, so
     BRAIN_HOME.parent/PyAutoMind resolves.
   • pip install pytest PyYAML; run pytest tests/ -q with
     working-directory: PyAutoBrain.

Verification

• Re-run the isolated two-repo probe after the wrapper lands: expect
  193 passed, 0 failed.
• bash bin/check_skill_line_counts.sh — the new SKILL.md within budget.
• bash bin/install.sh into a tmp HOME still passes
  test_installer_keeps_commands_and_installs_both_skill_homes.
• The PR itself is the real proof: it must show a non-zero, green check
  count — the first Brain PR ever to have one.

Key Files

• PyAutoBrain/.github/workflows/tests.yml — new.
• PyAutoBrain/skills/sizing/SKILL.md, skills/sizing/agents/openai.yaml — new.
• PyAutoHeart/.github/workflows/heart-tests.yml — read-only, the model.
• PyAutoBrain/tests/test_skill_install.py — the test this makes green
  (unchanged).

Sizing note

Brain scored this large (9) / split-into-phases. Not taken, same as #192: the
score is prose-driven, repos_affected is one repo, and the change is a
workflow file plus a 3-line skill wrapper.

Original Prompt

Click to expand starting prompt

Filed from the #193 ship (PyAutoMind/draft/maintenance/pyautobrain/no_pr_test_ci.md).
Follow-up items 1 and 2 surfaced there: PyAutoBrain has no PR test CI, and the
sizing SKILL.md failure blocks it from ever being green.

[Jammy2211]

PR created

PR: #195 (branch feature/pyautobrain-pr-test-ci, 3 files)

Both parts in one PR, since merging the workflow alone would ship a knowingly-red gate:

1. .github/workflows/tests.yml — modelled on PyAutoHeart/heart-tests.yml; pytest only, python 3.12/3.13, two-repo checkout (PyAutoBrain + PyAutoMind) because a lone checkout cannot even collect the suite.
2. skills/sizing/SKILL.md + agents/openai.yaml — the wrapper sizing never got when docs: 7th organ + slim AGENTS.md + wire sizing faculty (#84) #141 wired it into pyauto-brain help, which is what kept test_skill_install red on main.

Verified in the exact two-repo layout the workflow builds (scratch copy of the branch, fresh PyAutoMind clone, clean env + isolated HOME): 193 passed. Same layout on main: 192 passed / 1 failed — so the wrapper is what closes it. check_skill_line_counts.sh → all 45 skills within budget.

The PR itself is the acceptance test: it should be the first PyAutoBrain PR with a non-zero, green check count. If no checks appear, the workflow is wrong and it must not merge.

[Jammy2211]

Follow-up shipped in #196 (a2264fe): the nightly's exit-code/conclusion contract and the stop-summary None/count bug — both surfaced by the triage in PyAutoMind/draft/triage/nightly_release_blocked_eight_nights.md.

0/2/3 now leave the nightly run green (a blocked night is the gate working — with a named Blocked at a gate step, a ::warning:: and a job summary), and only a driver error turns it red. overnight_status.sh reads that step so a blocked night shows as ⏸, not a plain green — otherwise the fix would have swapped a false alarm for a silent one.

The page detail now reads 1 failed: autolens database/start_here.py; verify_install FAILED instead of …, None verify_install, verified against the real stage_report.json from the 2026-08-04 failure.

#196 was also the first PR gated by the CI this issue added — both matrix legs green.

Still open: the releases themselves are blocked (Class A rotating script failures, Class B gate-vs-release disagreement). Items 1-2 made the signal legible; they unblocked nothing.