[Jammy2211]

Closes #482.

Why

Claude Code ≥ 2.1.277 loads AGENTS.md natively (built-in plugin cc-plugin-agents-md, default mode claude-md-or-agents-md). Measured on 2.1.292 (2026-10-07): in that mode any CLAUDE.md in the cwd or an ancestor makes the engine ignore every AGENTS.md, so the 46 content-free CLAUDE.md → @AGENTS.md pointer files (and the unversioned workspace-root CLAUDE.md) must all be retired together. A file CLAUDE.md imports is never loaded twice, so this is hygiene, not a token saving. Codex, Cursor and other agents read AGENTS.md directly and never read CLAUDE.md, so nothing changes for them.

What

• scripts/repos_sync.py: the "CLAUDE.md → AGENTS.md pointers" check leg (name byte-identical, so PyAutoHeart's manifest_drift parser needs no change) now reports a repo that has AGENTS.md and a CLAUDE.md. write_claude_md_pointers → remove_claude_md_pointers: deletes only content-free pointers; a CLAUDE.md with real content is reported (KEPT) and never deleted.
• .github/workflows/session_hook_propagate.yml: carries the removal on workflow_dispatch only (never on push), honours dry_run, stages only git rm --cached CLAUDE.md.
• .github/workflows/firewall_gate.yml: TEMPORARY skip of the pointer leg (organ mains keep their pointers until the wave runs) plus a replacement step asserting PyAutoMind carries no CLAUDE.md and running the retirement tests. Remove with the wave's close-out.
• scripts/spawn.py / docs/pyautobrain/spawn_spec.md: templates no longer generate wiki/example/CLAUDE.md; new policy/hpc_ral.md is DROPped from the public template (names a user, hosts and SSH aliases).
• policy/hpc_ral.md: the RAL/HPC rules moved verbatim out of the workspace-root AGENTS.md, which shrinks to a three-line pointer stanza (root files are unversioned; applied by hand after merge).
• PyAutoMind's own CLAUDE.md removed here (the propagation job never targets Mind).
• Tests: 11 new in test_repos_sync_structure_lint.py; updates in test_session_hook_sync.py, test_repos_sync_standards_block.py, test_spawn_template_contract.py, test_spawn_privacy.py. Mind suite 684 passed; PyAutoHeart/tests/test_manifest_drift.py 8 passed.

Follow-ups (after merge)

1. Dispatch session_hook_propagate with dry_run: true, review the diff stat, then dispatch for real → removes the 44 reachable pointers.
2. euclid_assistant (session_hook: false, personal namespace): delete its CLAUDE.md by hand.
3. Re-spawn the Mind/Memory templates (spawn_drift flags wiki/example/CLAUDE.md until then).
4. Remove the TEMPORARY firewall skip, its replacement step and the matching TEMPORARY tests.
5. Nested pointers: 17 tracked CLAUDE.md pointers sit below repo roots (assistant scripts/ and wiki/*, PyAutoMemory wiki/*, workspace_test scripts/). Measured: a pointer below the cwd suppresses nothing and still loads lazily, so they are harmless for root-started sessions, but a session started inside such a folder would lose its ancestors' AGENTS.md. Extend the removal leg to nested pointers (or hand-clean) in a follow-up.
6. Root workspace: apply the HPC stanza to AGENTS.md, delete root CLAUDE.md, re-run bin/install.sh (Brain#492 supplies the corrected model-delegation block).

API Changes

None for the published libraries (PyAutoNerves/Fit/Array/Galaxy/Lens are untouched). Internal PyAutoMind tooling only: scripts/repos_sync.py renames write_claude_md_pointers → remove_claude_md_pointers and inverts the "CLAUDE.md → AGENTS.md pointers" check leg (leg name unchanged). See full details below.

Test Plan

• PyAutoMind python -m pytest tests/ — 684 passed (clean env; worktree branch head 9f39fb2)
• PyAutoHeart tests/test_manifest_drift.py — 8 passed (leg name byte-identical, parser unchanged)
• Note: under the worktree activate.sh env, tests/test_repos_sync_root_default.py::test_a_root_that_does_not_hold_this_checkout_is_refused fails — pre-existing env leak (PYAUTO_ROOT), reproduced identically on base c567af45; not touched by this PR
• CI: firewall_gate (pointer leg TEMPORARILY skipped; replacement step asserts Mind has no CLAUDE.md)
• After merge: session_hook_propagate dry_run: true dispatch, review diff stat, then real dispatch

Full API Changes (for automation & release notes)

Removed

• scripts/repos_sync.write_claude_md_pointers — replaced by remove_claude_md_pointers
• PyAutoMind/CLAUDE.md

Added

• scripts/repos_sync.remove_claude_md_pointers — deletes content-free CLAUDE.md pointers; reports (KEPT) and never deletes a CLAUDE.md with real content
• policy/hpc_ral.md — RAL HPC rules moved verbatim from the workspace-root AGENTS.md; DROPped from the spawned public template

Changed Behaviour

• repos_sync.py --check leg "CLAUDE.md → AGENTS.md pointers" now flags a repo that has AGENTS.md and a CLAUDE.md (previously flagged a missing pointer)
• session_hook_propagate.yml carries the CLAUDE.md removal on workflow_dispatch only, honouring dry_run
• spawn.py no longer generates wiki/example/CLAUDE.md

Migration

• Downstream: the 44 propagation-reachable repos lose their pointer via the human-dispatched wave; euclid_assistant by hand; workspace root by hand (see Follow-ups)

Downstream impact

• PyAutoHeart manifest_drift: no change needed (leg name unchanged; 8 tests pass). Until the wave runs, Heart readiness reports manifest drift: CLAUDE.md → AGENTS.md pointers — 45 mismatch(es) (YELLOW) — expected.
• Workspace scripts: none affected (no library symbol changed) → option (iii); no smoke test applies to Mind tooling.

🤖 Generated with Claude Code

https://claude.ai/code/session_01QmiavJnNzmKb6EJmgYATfx