WORKBOOK v3.1 — Real-Loop Hardening & Ship (H1–H4)

[CTlanston]

What

Adds WORKBOOK_v3.1.md — the next-cycle plan for Codex to execute. It is a hardening + ship workbook, not a new-feature one.

Why

An independent, evidence-grounded audit of the P0–P7 commit (a7d400f) found the code is substantially real (6/8 phases solid) — but two verified gaps and one doc-overclaim remain:

1. P4 — the isolated Gemini hard gate has only ever been mock-tested. The block is real and fails-closed in code, but every committed real-smoke shows Gemini as GEMINI_NOT_CONFIGURED (FAIL) or "not provided/run". A real Gemini has never judged a real diff (no real PASS, no real FAIL).
2. P7 — the real E2E ran on a /tmp sandbox repo, not a registered repo, and had no real Gemini verdict. planner=claude / coder=codex (real tokens) / worktree diff were real; the validator leg was not.
3. README overclaims "external Gemini validator … all passed", and its lower 2/3 still describes a deleted Python dual-kernel + Agent-Mesh/Sentinel as "production grade", contradicting docs/PARKED.md.

The plan (H1–H4, each = one commit + one PR, actually pushed)

• H1 — wire AEDEV_GEMINI_API_KEY into the daemon (validator only; coder/planner stay key-stripped) and make a real Gemini produce one FAIL and one PASS on real diffs, with verdict artifacts under evidence/.
• H2 — run the full conversational loop on a registered repo (hermus-agent), not /tmp, with a real Gemini verdict.
• H3 — fix the README overclaim and reconcile the stale Python/Agent-Mesh body with docs/PARKED.md.
• H4 — small fixes (stale DAG comment, memory into planner, dedup memory noise) + confirm PR-flow closure.

Process rules this enforces (fixing last cycle)

• G1 every phase is its own commit + PR, verified on origin (last cycle squashed 8 phases into one [P7] commit and never pushed).
• G2 no "validator passed" claim without a real Gemini verdict artifact in evidence/.
• G4 PR flow, not direct-to-main (last cycle pushed a7d400f straight to main).

See WORKBOOK_v3.1.md §3 for the full per-phase L1 acceptance criteria.

🤖 Generated with Claude Code