claude-codex-orchestration

Use when coordinating Claude Code + Codex as dual agents with Gemini as frontend/UI specialist consultant (via gemini-cli MCP), delegating parallel implementation of bounded modules to Codex, splitting work by boundary-clarity not domain, controlling token consumption, cross-harness setup (Cursor/Codex/OpenCode), pre-task thinking (/co-think), strategic plan review (/co-plan-review), engineering principles enforcement (Hyrum, Beyoncé, Chesterton, trunk-based, shift-left, feature-flags, deprecation, maintainability harness, testing quality), UI style standard (shadcn/radix-nova), and knowledge compounding. Self-contained; optional plugins and Gemini MCP auto-detected.

Pre-task thinking for claude-codex-orchestration (grill / office-hours style). Source-first: inspect repo/docs before asking. Product/Design, Technical/Approach, and Domain/Docs lanes. Ask only user-owned or genuinely ambiguous questions, one at a time, with recommendation + confidence. High-verification-risk fuzzy work becomes a Run Contract before dispatch. Always ends with Premise Challenge. Optional Codex cold-read second opinion. Delegates to references/thinking-decision.md.

Knowledge compounding for claude-codex-orchestration. Capture a solved non-trivial problem to docs/solutions/[category]/[slug]-[date].md with YAML frontmatter. Two modes: Full (4 parallel subagents — Context Analyzer / Solution Extractor / Related Docs Finder / Session Historian; overlap detection; Phase 2.5 selective refresh of older docs; discoverability check) or Lightweight (single pass). Delegates to claude-codex-orchestration references/knowledge-compounding.md.

Autonomous background refinement loop for claude-codex-orchestration. Runs /co-eval + /co-review, if candidate ≥ 6 runs /co-promote and commits, then schedules next cycle via ScheduleWakeup (270s, stays within cache TTL). Never self-stops — continues until user manually interrupts. Darwin ratchet: next session scoring the promoted dimension must improve or rule is reverted. Strategy escalation for persistent weak points (5/8/12/15+ sessions).

CEO-mode plan review for claude-codex-orchestration. Run after an Execution Plan is drafted. User chooses one of four modes (EXPAND/SELECTIVE/HOLD/REDUCE) dictating posture — dream bigger vs baseline+cherry-pick vs make-bulletproof vs minimum-viable. Seven Prime Directives apply in all modes: zero silent failures / named errors / shadow paths / interaction edge cases / observability-is-scope / deferred-is-written-down / security-is-scope. Critical rule: every scope change is explicit user opt-in. Delegates to references/thinking-decision.md.

Cross-session history search for claude-codex-orchestration. Before starting complex work, dispatch a general-purpose subagent to search ~/.claude/projects/ and ~/.codex/sessions/ for prior investigations of the same problem. Returns prior approaches tried, dead ends and why they failed, key decisions + rationale, related context. Avoids repeating failed paths from past sessions. Delegates to claude-codex-orchestration references/knowledge-compounding.md.

Layer 1 session self-eval for claude-codex-orchestration. Two-axis rating (Ambition × Execution) with locked 3×3 composite matrix (score 1–5), mandatory Devil's Advocate (argue lower AND higher before committing), anti-inflation check (last 5 entries, 4+ identical flags). Persists to .eval-scores.jsonl. Run at session end or when long-session prompt triggers (≥10 dispatches OR ≥6h since last eval). Delegates to claude-codex-orchestration references/self-correction.md Layer 1.

Layer 3 promote-to-SKILL for claude-codex-orchestration. Take a candidate identified by /co-review (Durability + Impact + Scope ≥ 6), distill from descriptive to prescriptive, write the rule into SKILL.md or appropriate reference file, remove source .error-log.jsonl entries, trigger Layer 0 auto-score and README sync. Ratchet rule enforces next same-dimension session must improve or revert. Delegates to claude-codex-orchestration references/self-correction.md Layer 3.

Layer 3 promote-candidate review for claude-codex-orchestration. Every 5 sessions (or when .eval-scores.jsonl has 5+ new entries), scan the weak-point + root-cause distribution in recent logs, identify recurring patterns (≥ 2× occurrences), score each candidate on Durability + Impact + Scope (0–3 each), flag those ≥ 6 for /co-promote. Delegates to claude-codex-orchestration references/self-correction.md Layer 3.

Layer 0 skill-modification score for claude-codex-orchestration. Auto-fire after any commit touching SKILL.md / references/*.md / README.md / CLAUDE.md.template / AGENTS.md template. Two-step process — Step 1 README coherence check (grep README for changed skill-file references, verify descriptions still accurate, update in same commit if stale), Step 2 8-dimension weighted score (Design 20% / Docs 15% / Self-Containment 15% / Cross-Harness 10% / Executability 15% / Validation 10% / Engineering 10% / Usability 5%). Single-dim drop ≥3 requires justification. Append to git-tracked .skill-scores.jsonl. Delegates to references/self-correction.md Layer 0.