[Jammy2211]

Closes #492. Sibling of PyAutoMind#482 (CLAUDE.md pointer retirement).

Why

Claude Code ≥ 2.1.277 loads AGENTS.md natively, so every CLAUDE.md is being retired (PyAutoMind#482), including the unversioned workspace-root CLAUDE.md whose "Claude-specific notes" need a versioned, agent-agnostic home. That home is policy/workspace_model_delegation.md, which bin/install.sh already renders into the root AGENTS.md as the pyauto:model-delegation block. Measured on 2.1.292: a fresh delegated subagent does receive the full AGENTS.md hierarchy plus the skill-description list (~18-22k tokens per delegation), so the per-delegation cost is what to trim; nothing needs restating in subagent prompts.

What

• policy/workspace_model_delegation.md: absorbs the root notes, corrected — the subagent template / "Progress heartbeat" that the old notes pointed at left WORKFLOW.md on 2026-09-18 (a7d8fce), so the block points at MODEL_DELEGATION.md's bounded-worker contract and heartbeat; the Sonnet floor is ship_* step 4; slash commands are the installed skills. Body 1069 B; renders and --check-workspace-policy OK on a scratch copy of the root file.
• skills/WORKFLOW.md / skills/MODEL_DELEGATION.md: "Model delegation" and "No OpenAI API fallback" collapsed into MODEL_DELEGATION.md with one-line pointers; the independent-review rule stays in WORKFLOW.md (pinned by test_chat_orchestration_contract.py) with a pointer from MODEL_DELEGATION.md.
• Stale CLAUDE.md references dropped or marked legacy in ORGANISM.md, BUILD_CAPABILITIES.md, HEART_CAPABILITIES.md, repo_cleanup/SKILL.md, docs/agent_harness_smoke.md, bin/install.sh, tests/test_memory_surfaces.py. _clone.py / regroup_workspace.py keep the pattern as tolerated-legacy until fix: honor absent witnesses and unresolved triage in sizing #482's wave removes the assistants' scripts/CLAUDE.md (their clone-boundary.yml CI would otherwise fail on unclassified files); _memory.py reads only wiki/AGENTS.md.
• Skill descriptions > 300 chars trimmed to ≤ 250 (cortex, sampler_pipeline, ci_speedup, community, intake, batch, board, issue_cleanup, hygiene, eyes, feedback); trigger phrases kept, no renames, YAML valid. Descriptions are listed in every session and every delegated subagent, and SKILL.md installs into both the Claude and Codex skill roots.
• Tests: 1291 passed in a clean shell (one test_grouped_organ_consumers failure appears only when activate.sh is sourced — pre-existing env leak). repos_sync.py --check: all Brain-owned generated blocks OK.

Follow-ups

• After fix: honor absent witnesses and unresolved triage in sizing #482's wave: drop the legacy CLAUDE.md patterns in _clone.py and regroup_workspace.py.
• Root workspace: re-run bin/install.sh so the root AGENTS.md carries the new block; delete root CLAUDE.md (with fix: honor absent witnesses and unresolved triage in sizing #482).

API Changes

None — internal changes only. Docs, policy text and skill-description trims; no public Python API added, removed or changed (_clone.py, _memory.py, regroup_workspace.py edits are internal file-pattern lists).

Test Plan

• python3 -m pytest -q -n auto in PyAutoBrain (clean shell, no activate.sh): 1291 passed
• repos_sync.py --check from the bundle root: all generated blocks OK (only the bundle-root autofit_inference / autofit_profiling checkout-manifest mismatch, unrelated to this branch)
• After merge: re-run bin/install.sh in the root workspace and confirm the pyauto:model-delegation block renders

🤖 Generated with Claude Code

https://claude.ai/code/session_01QmiavJnNzmKb6EJmgYATfx