The ultimate AI coding harness on top of pi.dev.

ultimate-pi adds a governed coding workflow to Pi: bootstrap the repo, plan with evidence, execute only against an approved PlanPacket, then run an independent review gate before merge.

Documentation Paths

Use DOCS_BY_AUDIENCE.md as the routing source of truth. This README stays focused on the package itself, not on repeating the full doc map.

Quick start

Requirements: Node 18+, npm 9+, git, and Pi.

1. Install the package in your project:

pi install npm:ultimate-pi
/reload

2. Bootstrap the harness once per project:

/harness-setup

3. Run the strict end-to-end pipeline:

/harness-auto "implement feature X safely"

/harness-auto runs plan → execute → review → optional steer loop. It may prepare commit/PR work when gates pass, but it never auto-merges.

Core workflow

Recommended: one command

/harness-auto "your task" [--quick] [--risk low|med|high]

Use this for most feature, fix, and refactor work. The parent orchestrator handles the phase handoffs and keeps active run context in .pi/harness/active-run.json plus run artifacts under .pi/harness/runs/.

Manual: phase by phase

/harness-plan "your task" [--risk low|med|high] [--quick]
/harness-run
/harness-review [--quick]

Manual mode is useful when you want to inspect or approve each handoff. On the happy path you do not pass --plan or a run id; the harness restores the active PlanPacket and run context.

Repair loop

If /harness-review returns implementation_gap, run:

/harness-steer
/harness-review

/harness-steer uses artifacts/repair-brief.yaml and respawns the executor in repair mode without widening the approved plan scope.

Command reference

| Command | Purpose | |---|---| | /harness-setup [--skip-graphify] [--skip-tools] [--non-interactive] [--force] | Idempotent project bootstrap: Graphify, harness-web/Scrapling, CLI tools, settings, contracts, Sentrux, ls-lint naming, harness lens, and verification. | | /harness-auto "<task>" [--quick] [--risk low\|med\|high] | Strict full pipeline: plan, execute, review, steer when appropriate. | | /harness-plan "<task>" [--risk low\|med\|high] [--quick] | PM-grade planning: clarifies the task with you first, then reconnaissance, decomposition, hypothesis, external research, ExecutionPlan, DAG validation, Review Gate debate, approve_plan, create_plan. | | /harness-run | Executes the approved active PlanPacket by spawning harness/running/executor; no inline implementation. | | /harness-review [--run <id>] [--quick] [--readonly] [--trace <ref>] | Post-run verification gate: deterministic checks, benchmark evaluator, policy verdict, adversary, optional tie-breaker. | | /harness-steer [--attempt N] | Post-review repair pass for implementation_gap; executor reads repair-brief.yaml, then you re-run /harness-review. | | /harness-abort [reason] | Safely aborts the active run, clears plan readiness, and re-locks mutation until a fresh plan is approved. |

| /harness-clear | Deletes all .pi/harness/runs/<run_id>/ directories, including the active run, after mandatory confirmation; non-affirmative/outage confirmation paths are no-op. | | /harness-trace [--run <id>] [--phase plan\|execute\|evaluate\|adversary\|merge] | Summarizes run traces and artifact handoffs for replay/forensics. | | /harness-incident --trigger <reason> [--run <id>] [--severity low\|med\|high\|critical] | Records incident, rollback, and override trail for harness failures. | | /harness-sentrux-steward [--run <id>] | Ad-hoc architectural intent review for Sentrux manifest/rule alignment. | | /harness-ls-lint-sync | Regenerate .ls-lint.yml from .pi/harness/ls-lint/naming.manifest.json. | | /harness-ls-lint-steward [--run <id>] | Ad-hoc naming-intent review for ls-lint manifest/rule alignment. | | /graphify [directory] | Bootstraps or updates the Graphify knowledge graph. | | /wiki-autoresearch [topic] | Runs autonomous web research and builds a Graphify-backed research wiki. | | /wiki-save | Saves the current conversation or insight as a structured wiki note. | | /release [patch\|minor\|major] [--dry-run] | Maintainer release helper. |

Harness phases and agents

• Planning uses agents under .pi/agents/harness/planning/ plus parent-led Graphify → sg → ccc reconnaissance. Legacy tool-tied planning/scout-* agents have been removed; planning context is captured in artifacts/planning-context.yaml.
• Running uses .pi/agents/harness/running/executor.md via agent id harness/running/executor.
• Reviewing uses .pi/agents/harness/reviewing/ via harness/reviewing/evaluator, harness/reviewing/adversary, and harness/reviewing/tie-breaker.
• Support agents such as harness/incident-recorder, harness/sentrux-steward, harness/ls-lint-steward, and harness/trace-librarian remain under .pi/agents/harness/.

Subagents run isolated from the parent session. They persist canonical YAML through submit_* tools; the parent gates with harness_artifact_ready and writes only orchestrator-owned merge artifacts.

Artifacts and layout

| Path | Description | |---|---| | .pi/harness/active-run.json | Active run pointer for happy-path commands. | | .pi/harness/runs/<run_id>/plan-packet.yaml | Approved execution baseline. | | .pi/harness/runs/<run_id>/research-brief.yaml | Planning evidence and research merge. | | .pi/harness/runs/<run_id>/artifacts/ | Planning context, decomposition, research, benchmark, verdict, adversary, repair, Sentrux, and ls-lint signal artifacts. | | .pi/harness/runs/<run_id>/handoff/executor-summary.yaml | Executor handoff written by submit_executor_handoff. | | .pi/harness/incidents/ | Incident records and rollback/override trail. | | .pi/harness/docs/adrs/ | Harness architectural decisions. | | .pi/harness/specs/ | Artifact contracts and schemas seeded into projects. |

Safety defaults

• Graph before grep: planning consults graphify-out/GRAPH_REPORT.md and Graphify queries before raw file reads.
• Plan before mutate: mutating tools are blocked until /harness-plan approves and creates a plan.
• No inline execution: /harness-run delegates to harness/running/executor only.
• No inline review: /harness-review delegates verdicts to isolated reviewing agents.
• No auto-merge: final merge remains a human/operator decision.
• Sentrux is the architecture signal: structural baselines and gates inform review; executor does not optimize metrics as a goal.
• ls-lint is the naming signal: manifest-driven .ls-lint.yml enforces filesystem conventions; steward evolves rules via chair-approved proposals (ADR 0052).
• pi-lens is edit-time diagnostics: LSP/lint/format/ast feedback complements Sentrux and does not replace architecture gating.

Troubleshooting

| Problem | Try | |---|---| | Setup fails | Confirm node --version is 18+, npm --version is 9+, then rerun /harness-setup. | | No approved plan | Run /harness-plan "<task>", then /harness-run. | | Need to inspect handoff | Run /harness-trace or inspect .pi/harness/runs/<run_id>/. | | Need to restart safely | Run /harness-abort [reason], then create a fresh plan. |

| Need to prune old run history safely | Run /harness-clear; all run directories, including the active run, are eligible and confirmation failure/cancel deletes nothing. | | Review says implementation_gap | Run /harness-steer, then /harness-review. | | Review says plan_gap | Revise with /harness-plan "<updated task>". | | Sentrux missing | Install/configure Sentrux or keep it skipped; harness verification still reports the status. | | ls-lint failures | Run node .pi/scripts/harness-ls-lint-cli.mjs, fix paths or propose manifest changes via /harness-ls-lint-steward. |

Optional integrations can be configured by copying .env.example to .env; /harness-setup appends missing keys without overwriting existing values.

Contributing

Local development, harness internals, and quality gates: CONTRIBUTING.md, .pi/scripts/README.md, and .pi/harness/docs/adrs/.