ultimate-pi
Ultimate AI coding harness for pi.dev — extensible skills, Obsidian wiki knowledge layer, compressed context, deterministic output
Package details
Install ultimate-pi from npm and Pi will load the resources declared by the package manifest.
$ pi install npm:ultimate-pi- Package
ultimate-pi- Version
0.19.0- Published
- May 24, 2026
- Downloads
- 5,551/mo · 3,477/wk
- Author
- aryaniyaps
- License
- MIT
- Types
- extension, skill, prompt
- Size
- 18 MB
- Dependencies
- 15 dependencies · 1 peer
Pi manifest JSON
{
"extensions": [
"./.pi/extensions"
],
"skills": [
"./.agents/skills",
"./.pi/skills"
],
"prompts": [
"./.pi/prompts"
]
}Security note
Pi packages can execute code and influence agent behavior. Review the source before installing third-party packages.
README

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.
Quick start
Requirements: Node 18+, npm 9+, git, and Pi.
- Install the package in your project:
pi install npm:ultimate-pi
/reload
- Bootstrap the harness once per project:
/harness-setup
- 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, 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: 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-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. |
/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→cccreconnaissance. Legacy tool-tiedplanning/scout-*agents have been removed; planning context is captured inartifacts/planning-context.yaml. - Running uses
.pi/agents/harness/running/executor.mdvia agent idharness/running/executor. - Reviewing uses
.pi/agents/harness/reviewing/viaharness/reviewing/evaluator,harness/reviewing/adversary, andharness/reviewing/tie-breaker. - Support agents such as
harness/incident-recorder,harness/sentrux-steward, andharness/trace-librarianremain 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, and Sentrux 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.mdand Graphify queries before raw file reads. - Plan before mutate: mutating tools are blocked until
/harness-planapproves and creates a plan. - No inline execution:
/harness-rundelegates toharness/running/executoronly. - No inline review:
/harness-reviewdelegates 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.
- 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. |
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. |
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/.