Coordinates the full improvement pipeline: Generator → Discriminator → Evaluator → Executor → Gate.
When to Use
- Run a full improvement cycle on one or more skills
- Coordinate the 5-stage pipeline end-to-end (with optional evaluator)
- Retry failed improvements with trace-aware feedback (Ralph Wiggum loop)
When NOT to Use
- Only want to check skill quality scores → use
improvement-learner - Only want to manually score candidates → use
improvement-discriminator - Only want to modify one file → use
improvement-executor - Only want to check benchmark data → use
benchmark-store
Pipeline
propose → discriminate → evaluate* → execute → gate
↻ Ralph Wiggum: fail → inject trace → retry (max 3)
* evaluate is optional — skipped if no task_suite.yaml exists
CLI
python3 scripts/orchestrate.py \
--target /path/to/skill \
--state-root /path/to/state \
--max-retries 3 \
--auto
Output Artifacts
| Request | Deliverable |
|---|---|
| Full pipeline | JSON with all stage outputs, final scores, execution trace |
| Retry cycle | Updated candidates with injected failure traces |
Related Skills
- improvement-generator: Produces candidate proposals (stage 1)
- improvement-discriminator: Multi-reviewer panel scoring (stage 2)
- improvement-evaluator: Task suite execution validation (stage 3, optional)
- improvement-executor: Applies changes with backup/rollback (stage 4)
- improvement-gate: 6-layer quality gate (stage 5)
- benchmark-store: Frozen benchmarks and Pareto front data
References
- Architecture — System design and data flow
- Guardrails — Safety rules and protected targets
- End-to-End Demo — Complete walkthrough