pi-dev-loop

Autonomous development loop engine for pi — verify-driven iteration with subagent isolation and error tracking.

Architecture

Three-layer isolation:

┌─ Orchestrator (main session) ───────────────────┐
│  Tracks error registry, dispatches work,         │
│  never writes code directly.                     │
├─ Impl Subagents (forked) ───────────────────────┤
│  Follow TDD: failing test → implement → verify   │
├─ Review Subagents (fresh, read-only) ───────────┤
│  Independently evaluate code with zero context   │
└──────────────────────────────────────────────────┘

Quick Start

Start a loop from conversation (the agent picks it up):

loop_start({ goal: "Fix type errors in src/user.ts" })

Or via commands:

/loop goal "Implement registration module" --verify "bun run typecheck" --verify "bun run test"

Subcommands

Command	Description
/loop goal <desc> [options]	Start a verification-driven loop
/loop stop	Stop active loop
/loop pause	Pause loop (preserve state)
/loop resume	Resume paused loop
/loop status	Show loop state + error registry
/loop history	Show iteration history

Options for /loop goal:

• --verify <cmd> — add verification command (repeatable)
• --max-iterations <N> — max iterations before auto-pause (default 20)
• --from-config [path] — load .pidev.yml

Config File (.pidev.yml)

loop:
  mode: goal
  maxIterations: 20
  maxConsecutiveZeroProgress: 3

verify:
  - command: "bun run typecheck"
    runsOn: impl
  - command: "bun run test -- --related={files}"
    runsOn: impl
    timeout: 120000

guardrails:
  gitAutoSnapshot: true
  rollbackOnRegression: true
  maxFileChangesPerSubagent: 20

Verification-Driven Stopping

Unlike pure LLM-judged loops, pi-dev-loop requires verification commands to pass before advancing. Errors are tracked across iterations with SHA-256 fingerprinting, detecting:

• Regressed errors — previously fixed errors that reappeared
• New errors — introduced by the current iteration's changes
• Persistent errors — unresolved across multiple iterations

Project Layout

pi-dev-loop/
├── src/                     # Core modules
│   ├── state.ts             # DevLoopState, ErrorRecord types + serialization
│   ├── error-registry.ts    # Fingerprinting, mergeRegistry, parseOutput
│   ├── verify-config.ts     # --verify arg parsing, mergeConfigs
│   ├── load-config.ts       # .pidev.yml parsing
│   ├── git.ts               # Snapshot, rollback, pruning
│   ├── subagent-task.ts      # Context packing for impl/review subagents
│   ├── session-prompt.ts    # Iteration prompt builder
│   └── auto-detect.ts       # Auto-detect verify commands from project
├── extensions/pi-dev-loop/  # pi extension entry point
│   └── index.ts             # Commands, tools, events, decision engine
├── skills/pi-dev-loop/      # Agent behavior guide
├── prompts/                 # Prompt templates
├── tests/                   # Bun test suite (86 tests)
└── docs/
    ├── specs/               # Design documents
    └── plans/               # Implementation plans

Development

bun test              # Run all tests (86 tests, 196 expect calls)
bun run typecheck     # TypeScript check

This project follows design-confirm-implement → TDD → planned execution → review → verify. See docs/specs/ for design docs and docs/plans/ for implementation plans.

License

MIT © 2026 Sason Wong