// Pi-native evaluate→improve→persist harness. Provides a systematic loop for generating candidates, judging them against rubrics, and accumulating knowledge as playbooks. Use for iterative output improvement, not for one-off edits (use continuity) or single PR review (use zereight-review).
Run iterative agent evaluation and improvement loops (judge, improve, scenarios, playbooks) via pi-autocontext or autoctx CLI. Use when improving repeatable agent workflows (review rubrics, skill quality, scenario-based feedback), checking run status, or accumulating knowledge—not for one-off code edits or repo pattern drift (use continuity) or single PR review (use zereight-review).
Audit agent skills — token cost, duplicates, outdated plugin versions, unused skills, and overly long descriptions. Use when trimming skill prompt budget, finding duplicate or unused skills, auditing plugin versions, or deciding which skills to remove.
Diagnose why a skill is missing from Pi /skill or Cursor lists. Use when a skill exists on disk but does not appear, Pi shows only project or extension skills, or after adding a repo skill under .agents/skills/.
Comprehensive code review skill for practical PR feedback. Use for feature, bugfix, and refactor reviews. Prioritizes correctness, edge cases, logic invariants, fallback-chain safety, async state transitions, architecture analysis, OWASP security, and clear actionable feedback.
Skill map and inventory workflow. Use for scanning installed skills, skill overview, 스킬 목록, skill map.
Inspects Cursor agent transcript history, analyzes repeated assistant mistakes, and turns durable learnings into corrective skills or context updates. Use when the user asks to review previous Cursor chats, learn from a failed interaction, run /session-lessons, or stop the model repeating a mistake.
| name | harness |
| description | Pi-native evaluate→improve→persist harness. Provides a systematic loop for generating candidates, judging them against rubrics, and accumulating knowledge as playbooks. Use for iterative output improvement, not for one-off edits (use continuity) or single PR review (use zereight-review). |
| argument-hint | [solve|knowledge|status] |
Builds on autoctx (judge/improve core) + Pi tools to provide a coherent harness.
| Layer | Source | Role |
|---|---|---|
| Extension | ~/.pi/agent/extensions/harness/index.ts | Tools + TUI + lifecycle |
| Judge engine | autoctx (via pi-autocontext dep) | LLM-based evaluation |
| Skill | .agents/skills/harness/SKILL.md | Usage guide + routing |
| Tool | Description |
|---|---|
harness_solve | Full loop: generate → judge → improve → persist playbook |
harness_knowledge | Read/write/list playbooks (knowledge/<scenario>/playbook.md) |
harness_status | Recent runs, scores, and scenarios overview |
# Solve a task with rubric-driven improvement
goal: "Improve zereight-review checklist for RN PRs"
rubric: "Actionable findings, Effect anti-patterns, security when relevant"
→ Best output + playbook saved to knowledge/<scenario>/
# Check previous runs
→ Recent runs with scores and scenarios
# Browse accumulated knowledge
action: list → all scenarios with playbooks
action: read, scenario: zereight-review → full playbook
| Need | Use |
|---|---|
| Systematic output improvement | harness_solve |
| Review accumulated lessons | harness_knowledge read |
| Check past run history | harness_status |
| One-off judge/improve | autocontext_judge / autocontext_improve (pi-autocontext) |
harness_solve consumes Pi provider quota per generation round.gens: 1 to calibrate rubric, then increase.harness_knowledge and harness_status are zero-cost.