claude-collaboration-looplisted

# Claude Collaboration Loop Use this skill for complex, high-stakes, or ambiguous tasks where Codex and Claude should cover each other's blind spots. Workflow: 1. Codex reads repo state and writes or updates `task_plan.md`, `findings.md`, and `progress.md` when file-backed planning applies. 2. Codex runs: ```bash node "${CODEX_PLUGIN_ROOT}/scripts/claude-companion.mjs" plan "$ARGUMENTS" ``` 3. Codex reconciles Claude's plan against local evidence: - adopt concrete missing tests, - adopt safer task ordering when justified, - reject unsupported speculation, - record the reconciliation in `findings.md`. 4. Codex implements the reconciled plan. 5. Codex runs: ```bash node "${CODEX_PLUGIN_ROOT}/scripts/claude-companion.mjs" adversarial-review "$ARGUMENTS" ``` 6. Codex reports: - implemented files, - verification commands, - Claude findings adopted, - Claude findings rejected, - residual risk. ## Natural-Language Claude Routing <!-- routing:collaboration-loop routing:codex-reconciles-claude --> - Do not ask the user to write `--quality`, `--model`, or `--effort` unless troubleshooting the plugin itself. - Keep Codex responsible for reconciling Claude output before edits. - Use this workflow when the user wants Codex and Claude to plan, implement, review, and reconcile together. - Keep Claude findings advisory until Codex verifies them against local evidence. - Do not substitute strong local Claude routing with `claude ultrareview`; ultrareview re