M1 canonical specs (opencode reviewer research, task-plan template,
Review Loop subroutine, Phase 1-10 prose, secret-scan regex list)
are embedded in the M2 claude-code SKILL.md as the canonical
reference; later variants (M3 codex, M4 cursor, M5 opencode) will
fork from this file.
Reviewer: codex / gpt-5.4. Approved after 3 rounds:
- Round 1: 2 P1 + 3 P2 → REVISE
- Round 2: 1 P2 → REVISE
- Round 3: 0 P0/P1/P2, 1 P3 (non-blocking) → APPROVED
Key design properties:
- Plan-review payload strips Runtime State and Review History to
prevent reviewer session-ID leakage across rounds.
- Secret-scan step 1a redacts matched text to
[REDACTED:<pattern-label>:<match-length>-chars] before any user
surfacing; never echoes raw match content.
- Brainstorming required for any behavior-changing task; auto-skip
limited to pure-documentation and pure-comment-whitespace-rename.
- Phase 3 reviewer config defaults to codex / gpt-5.4 with
MAX_ROUNDS=10 when user opts for defaults; explicit interactive
default also gpt-5.4 for internal consistency.
- Template Metadata records Branch Name and Worktree Path.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Stefano Fiorini
437b2024cd