Remove remaining provider assumptions from agent metadata and docs

Type: maintenance
Target: organs
Repos:

• @PyAutoBrain
• @PyAutoMind
• @PyAutoMemory
• @autofit_assistant
• @autogalaxy_assistant
• @autolens_assistant
• @autocti_assistant
• @autolens_workspace_test
  Difficulty: medium
  Autonomy: supervised
  Priority: high
  Status: draft
  Filed: 2026-09-17
  Parent: draft/maintenance/organs/agent_harness_agnostic_setup.md
  Blocked-by: codex_skill_discovery_parity

Request

Can you check that all Agent setup (skills, md files, etc) is agnostic to whether its claude or codex, been using claude for a while but gonna codex for the forseeable so worth an agnostic sweep

Scope

Move the three remaining canonical nested CLAUDE.md bodies to AGENTS.md
with thin Claude adapters; accept both claude/** and codex/** ledger branch
names; make resume metadata provider-aware; remove mandatory false
Claude/Anthropic commit attribution; and correct stale Codex hook/support
claims. Preserve historical records verbatim.

Acceptance

• Shared canonical guidance lives in AGENTS.md; CLAUDE.md only adapts it.
• Ledger-only session branches from either supported provider reach the same
  guarded merge path.
• New resume and attribution metadata never guesses a provider identity.
• Current setup/support documentation is backed by a bounded Codex smoke
  record and distinguishes tested behavior from intended support.

Resume request — 2026-09-19

Find and finish the work to make the agentic AI ecosystem agent agnostic, we should be on phase 4

Phase 3 is still PyAutoBrain#386 (approved 2026-09-17, resumed 2026-09-19).
This phase remains gated on that work; the human approved this phase in the Codex session on 2026-09-19.
Work follows phase 3, with shared-repo PRs stacked on its branch until merged.

Proposed implementation plan

Branch: feature/provider-agnostic-metadata across the listed repositories.
Classification: workspace/infrastructure. Use task worktrees inside the workspace.

1. Move PyAutoMemory/wiki/CLAUDE.md,
   autolens_workspace_test/scripts/CLAUDE.md, and
   autolens_workspace_test/scripts/misc/database/scrape/CLAUDE.md to sibling
   AGENTS.md bodies. Replace the old files with thin imports; fix current
   links and references without rewriting historical records.
2. Extend PyAutoMind/.github/workflows/mind_ledger_merge.yml to trigger for
   codex/** as well as claude/**. Preserve the existing default-deny path
   classifier, permissions, conflict resolver, and review boundary. Extend
   tests/test_ledger_merge.py with both branch namespaces and a workflow
   trigger assertion; verify code/hook/workflow edits remain ineligible.
3. Replace assumed Claude resume commands in Brain start-dev/start-workspace
   registry templates and Mind REFERENCE with explicit harness/session fields.
   Record a resume command only when the active harness actually supplies it;
   unknown IDs stay unknown. Audit current consumers for assumptions before
   changing the schema. Preserve historical session records.
4. Update the four assistant AGENTS commit-attribution rules to identify the
   actual harness/model when known, never fabricate an Anthropic identity.
   Update active setup/support prose in Brain, Mind, Memory and assistants to
   distinguish shared instructions, generated adapters, and harness-specific
   mechanisms. Verify current product claims against local installed evidence
   and official documentation where needed.
5. Record a bounded Codex smoke: discover representative generated skills,
   resolve their canonical instructions, exercise a harmless workflow/read-only
   command, and test the applicable hook fixtures. State explicitly which
   checks exercised a real harness and which only tested adapter code. Do not
   claim full scientific-workflow or every-harness support from this smoke.
6. Run targeted Mind ledger/hook tests, Brain discovery/template tests, adapter
   drift checks, and instruction-link checks; obtain an independent review and
   Heart verdict, then prepare one PR per affected repo. Merge is a separate
   human action. Update the parent task only against observed completion.

Branch survey — 2026-09-19

All eight phase-4 main checkouts are on main. Brain and the four assistants
are now claimed by phase 3, so phase 4 must follow it. Memory and
autolens_workspace_test are clean. Mind has an unrelated local deletion of
draft/maintenance/pyautobrain/workspace_resolver_fanout.md; preserve it.
The two assistant personal scripts noted in phase 3 also remain out of scope.