# AGENTS.md

Project conventions for agent-assisted development.

## Project Overview

This repository provides Pi coding agent skills and extensions. The primary component is **goal-driven-loop** — a skill + extension pair that keeps a session focused on a user-defined goal with explicit acceptance criteria.

## Repository Structure

```
extensions/goal-driven-loop/    # TypeScript extension (Pi Extension API)
.agents/skills/goal-driven-loop/ # Skill instructions (Markdown)
.pi/skills/ → symlink            # Pi skill discovery
.pi/extensions/ → symlink        # Pi extension discovery
docs/superpowers/                # Design specs and implementation plans
openspec/                        # OpenSpec change archive
```

Agent directories (`.claude/`, `.agents/`, `.pi/`) are **gitignored**. Only the symlinks and source files at top-level directories are committed.

## Tech Stack

- **Skills**: Markdown (Agent Skills standard), no build step
- **Extensions**: TypeScript, loaded via jiti (no compilation needed), depends on `@earendil-works/pi-coding-agent` Extension API
- **Testing**: Manual behavioral tests via `pi -p` / `pi -e`; unit tests run with `npx jiti` against individual modules
- **Workflow**: OpenSpec → comet lifecycle for change tracking

## Development Commands

```bash
# Verify skill loads
pi -p "/skill:goal-driven-loop"

# Verify extension loads
pi -e extensions/goal-driven-loop/index.ts -p "test"

# Run unit tests for trigger-detector
npx jiti /tmp/test-trigger-detector.mjs

# Run unit tests for state-injector
npx jiti /tmp/test-state-injector.mjs

# End-to-end test (create a temp project first)
mkdir -p /tmp/test && cd /tmp/test && git init
pi -e /path/to/pi-setup/extensions/goal-driven-loop/index.ts \
   -p "当前目标是实现 reverseString(s) 函数"
```

## Key Design Decisions

1. **Skill + Extension dual layer**: Skill = instructions (agent reads). Extension = mechanics (event handlers). Do not mix concerns.
2. **File-based state** (`.agents/goal/GOAL.md`): Survives context compaction. Deleted on goal completion.
3. **Extension never blocks user**: All confirms use timed auto-defaults (10s timeout). Fallbacks are safe (continue existing goal).
4. **Anti-triggers are mechanical**: Code blocks, quote blocks, reported speech, trivial keywords — all checked programmatically, not by agent judgment.
5. **Per-iteration retry**: Exponential backoff counter resets each iteration. 10 retries / ~37min coverage.
6. **Comet delegation**: Coarse-grained goals auto-delegate to `comet-open`. Fallback to multi-stage mode if comet unavailable.

## When Modifying This Project

### Adding a new skill

1. Create directory at `.agents/skills/<name>/SKILL.md`
2. Follow the Agent Skills standard: YAML frontmatter with `name` and `description` (start with "Use when...")
3. Keep SKILL.md ≤ 500 words; move detailed reference to `references/` subdirectory
4. Create symlink: `ln -s ../../.agents/skills/<name> .pi/skills/<name>`
5. Test with `pi -p "/skill:<name>"`

### Adding a new extension

1. Create directory at `extensions/<name>/`
2. Write `index.ts` that exports a default function receiving `ExtensionAPI`
3. Add `package.json` with `{ "pi": { "extensions": ["./index.ts"] } }`
4. Create symlink: `ln -s ../../extensions/<name> .pi/extensions/<name>`
5. Test with `pi -e extensions/<name>/index.ts -p "test"`

### Modifying goal-driven-loop

- **Skill layer** (`.agents/skills/goal-driven-loop/`): Changes to instructions, acceptance criteria rules, error handling reference
- **Extension layer** (`extensions/goal-driven-loop/`):
  - `trigger-detector.ts` — trigger phrase patterns and anti-trigger logic
  - `state-injector.ts` — GOAL.md parsing and injection text generation
  - `ui-prompts.ts` — timed confirms and notification helpers
  - `index.ts` — event wiring (input, before_agent_start)

After any change, re-run unit tests and E2E verification.

## Commit Conventions

- `feat(skill):` / `feat(ext):` — new features
- `fix(skill):` / `fix(ext):` — bug fixes
- `docs:` — documentation changes
- `chore:` — maintenance (gitignore, archive, etc.)

## Known Limitations (v1)

1. Mid-loop granularity upgrade needs skill-side signal (`status: needs-upgrade` in GOAL.md) — extension cannot auto-detect goal complexity increase
2. Agent "estimates < 5 min" anti-trigger is instruction-level only, not mechanical
3. Stagnation dual-signal precise detection needs skill-side passing count history in GOAL.md — extension relies on skill's own stagnation detection