pi-goal

Unattended goal orchestration for pi. Give it a checkable completion condition with /goal and walk away — it decomposes the goal into reviewable tasks, works them in dependency-ordered waves, merges the results, runs a cleanup sweep, and reviews the integrated whole against the original goal until it reaches a clear terminal state: achieved, unmet, halted, or budget_limited.

Each task is modeled as an OpenSpec change, so it carries a spec with acceptance criteria. Reviewers gate against the real source of truth (running tests, reading files, checking specs) rather than judging the transcript.

Status: the orchestration engine (src/) is implemented and unit-tested. The pi extension that drives it unattended is wired up; a few command subcommands (runs, resume, continue) are still being connected.

Install

pi install npm:@peeraponw/pi-goal

The extension loads automatically in every pi session. To try it without installing, use a one-off run instead:

pi -e npm:@peeraponw/pi-goal

Requires pi (Node 22+).

Quick start

Set a goal that carries checkable acceptance criteria:

/goal Add a file `hello.txt` at the repo root whose entire contents are the
single line `pong`. Acceptance: `[ -f hello.txt ] && [ "$(cat hello.txt)" = "pong" ]`
exits with status 0.

pi decomposes the goal, spawns worker and reviewer subagents per task in isolated git worktrees, merges each wave, and reports the terminal verdict. Progress is tailed under goal-runs/<run-id>/.

Command	Action
/goal <condition>	Set the goal and start working. Replaces any active goal.
/goal	Status — condition, current wave, task progress, elapsed time, spend.
/goal clear	Clear the active goal (aliases: stop, off, reset, none, cancel). Records the run as halted.
/goal resume	Restart from the last persisted checkpoint. (in progress)
/goal continue	Resume after human intervention. (in progress)
/goal runs	Browse past runs and open a report. (in progress)

Per-run overrides go anywhere in the invocation:

/goal --worker-model anthropic/claude-sonnet-4-5 --max-cost 2 \
      migrate src/api/users.ts to v2; npm test -- users exits 0

How it works

The system is split into three layers that keep decisions testable and transport-free:

• Thin pi extension (.pi/extensions/goal.ts) owns the /goal command, session-scoped state, the live status indicator, and resume. It hands each run to the engine.
• Pure engine (src/) makes every decision: config resolution, wave computation, the task state machine, budget enforcement, blast-radius excursion detection, merge-strategy selection, the debt ledger, and run reports. It has no dependency on pi and is fully unit-tested.
• pi-subagents performs each agent stage (planner, worker, reviewer, merger, goal-reviewer, refactor, arbiter).

The workflow per goal:

1. Decompose — one planner pass over the entire goal produces tasks (scope, checkable acceptance criteria, predicted blast radius, dependency edges). Waves are the topological levels of the dependency graph.
2. Execute a wave — tasks within a wave run in parallel, each in its own system-owned git worktree. Per task: worker → read-only reviewer → bounded rework, then blocked if still unresolved. Deferred findings accumulate in a run-level debt ledger.
3. Merge the wave — clean or disjoint blast radii fast-forward; generated files (lockfiles, build artifacts) are regenerated; other overlaps go to a merger pass, then a planner-as-arbiter, then human escalation.
4. Refactor sweep — one worker consumes the debt ledger (cleanup only).
5. Goal review — a whole-system reviewer gates the integrated result against the original goal. On unmet, recovery escalates: targeted re-plan → full re-plan → human.

A single task failing is isolated — it never aborts the goal; it is reported at the end. Budget is checked at every wave, rework, and amendment boundary.

Configuration

Sources, in ascending precedence (highest wins):

1. Global — ~/.pi/agent/goal.config.yaml
2. Project — .pi/goal.config.yaml
3. CLI flags — on the /goal invocation

Models and scalar limits merge per-field, so a project file that sets only the worker model keeps the global planner. See docs/goal.md for the full schema, defaults, CLI flags, and resource-limit semantics (per-role memory / CPU-time / task-count ceilings enforced at agent-launch boundaries).

Development

Prerequisites: Node.js 22+ and npm.

npm install                 # install dev dependencies
npm test                    # run the unit suite (252 tests)
npm run typecheck           # tsc --noEmit
npm run build:extension     # bundle the extension → dist/goal.extension.js

The unit suite is self-contained and never calls a model. The engine is pi-free, so it is exercised with fake launchers in src/**/*.test.ts.

End-to-end smoke test

test/smoke.test.ts spawns real child-pi subagents for every role against a throwaway git repo, so it spends real model calls. It is skipped by default:

GOAL_SMOKE=1 npx vitest run test/smoke.test.ts

Bundling

build-extension.mjs bundles .pi/extensions/goal.ts and the entire src/ engine into a single self-contained dist/goal.extension.js (the yaml dependency is inlined; the only pi import is a type that is erased). That bundle is what pi install npm:@peeraponw/pi-goal loads.

Project layout

.pi/extensions/goal.ts   # thin pi extension: /goal command, state, status
src/                     # pure, pi-free orchestration engine (unit-tested)
  orchestrator.ts        #   end-to-end run: plan → waves → refactor → review
  planning/              #   planner contract, wave computation, amendments
  execution/             #   worker→reviewer loop, concurrency, debt ledger
  merge/                 #   adaptive merge, regeneration, arbiter, escalation
  refactor/              #   debt-ledger cleanup sweep
  review/                #   goal-level reviewer + recovery tiers
  config/                #   file/CLI config resolution and merging
  budget/                #   budget enforcement
  state/                 #   task state machine, checkpoints, terminal state
  trace/                 #   event stream, run report, status formatting
  runtime/               #   child-pi subagent launcher (the pi adapter)
  vcs/                   #   system-owned git worktree lifecycle
docs/goal.md             # full configuration reference and workflow detail
openspec/                # design docs, specs, and task breakdowns
test/                    # end-to-end / live integration tests (opt-in)

Documentation

• docs/goal.md — full configuration reference, resource limits, CLI flags, and the workflow in detail.
• openspec/changes/ — the design docs, specs, and validation findings that shaped the architecture.
• pi documentation — the coding agent this extends.

License

MIT