gsd-2-agent-framework

GSD 2 — Autonomous Spec-Driven Agent Framework

Skill by ara.so — Daily 2026 Skills collection

GSD 2 is a standalone CLI that turns a structured spec into running software autonomously. It controls the agent harness directly — managing fresh context windows per task, git worktree isolation, crash recovery, cost tracking, and stuck detection — rather than relying on LLM self-loops. One command, walk away, come back to a built project with clean git history.

---

Installation

npm install -g gsd-pi

Requires Node.js 18+. Works with Claude (Anthropic) as the underlying model via the Pi SDK.

---

Core Concepts

Work Hierarchy

Milestone  →  a shippable version (4–10 slices)
  Slice    →  one demoable vertical capability (1–7 tasks)
    Task   →  one context-window-sized unit of work

Iron rule: A task must fit in one context window. If it can't, split it into two tasks.

Directory Layout

project/
├── .gsd/
│   ├── STATE.md          # current auto-mode position
│   ├── DECISIONS.md      # architecture decisions register
│   ├── LOCK              # crash recovery lock file
│   ├── milestones/
│   │   └── M1/
│   │       ├── slices/
│   │       │   └── S1/
│   │       │       ├── PLAN.md        # task breakdown with must-haves
│   │       │       ├── RESEARCH.md    # codebase/doc scouting output
│   │       │       ├── SUMMARY.md     # completion summary
│   │       │       └── tasks/
│   │       │           └── T1/
│   │       │               ├── PLAN.md
│   │       │               └── SUMMARY.md
│   └── costs/
│       └── ledger.json   # per-unit token/cost tracking
├── ROADMAP.md            # milestone/slice structure
└── PROJECT.md            # project description and goals

---

Commands

/gsd auto — Primary Autonomous Mode

Run the full automation loop. Reads .gsd/STATE.md, dispatches each unit in a fresh session, handles recovery, and advances through the entire milestone without intervention.

/gsd auto
# or with options:
/gsd auto --budget 5.00        # pause if cost exceeds $5
/gsd auto --milestone M1       # run only milestone 1
/gsd auto --dry-run            # show dispatch plan without executing

/gsd init — Initialize a Project

Scaffold the .gsd/ directory from a ROADMAP.md and optional PROJECT.md.

/gsd init

Creates initial STATE.md, registers milestones and slices from your roadmap, sets up the cost ledger.

/gsd status — Dashboard

Shows current position, per-slice costs, token usage, and what's queued next.

/gsd status

Output example:

Milestone 1: Auth System  [3/5 slices complete]
  ✓ S1: User model + migrations
  ✓ S2: Password auth endpoints
  ✓ S3: JWT session management
  → S4: OAuth integration  [PLANNING]
    S5: Role-based access control

Cost: $1.84 / $5.00 budget
Tokens: 142k input, 38k output

/gsd run — Single Unit Dispatch

Execute one specific unit manually instead of running the full loop.

/gsd run --slice M1/S4            # run research + plan + execute for a slice
/gsd run --task M1/S4/T2          # run a single task
/gsd run --phase research M1/S4   # run just the research phase
/gsd run --phase plan M1/S4       # run just the planning phase

/gsd migrate — Migrate from v1

Import old .planning/ directories from the original Get Shit Done.

/gsd migrate                        # migrate current directory
/gsd migrate ~/projects/old-project # migrate specific path

/gsd costs — Cost Report

Detailed cost breakdown with projections.

/gsd costs
/gsd costs --by-phase
/gsd costs --by-slice
/gsd costs --export costs.csv

---

Project Setup

1. Write ROADMAP.md

# My Project Roadmap

## Milestone 1: Core API

### S1: Database schema and migrations
Set up Postgres schema for users, posts, and comments.

### S2: REST endpoints
CRUD endpoints for all resources with validation.

### S3: Authentication
JWT-based auth with refresh tokens.

## Milestone 2: Frontend

### S1: React app scaffold
...

2. Write PROJECT.md

# My Project

A REST API for a blogging platform built with Express + TypeScript + Postgres.

## Tech Stack
- Node.js 20, TypeScript 5
- Express 4
- PostgreSQL 15 via pg + kysely
- Jest for tests

## Conventions
- All endpoints return `{ data, error }` envelope
- Database migrations in `db/migrations/`
- Feature modules in `src/features/<name>/`

3. Initialize

/gsd init

4. Run

/gsd auto

---

The Auto-Mode State Machine

Research → Plan → Execute (per task) → Complete → Reassess → Next Slice

Each phase runs in a fresh session with context pre-inlined into the dispatch prompt:

---

Must-Haves: Mechanically Verifiable Outcomes

Every task plan includes must-haves — explicit, checkable criteria the LLM uses to confirm completion. Write them as shell commands or file existence checks:

## Must-Haves

- [ ] `npm test -- --testPathPattern=auth` passes with 0 failures
- [ ] File `src/features/auth/jwt.ts` exists and exports `signToken`, `verifyToken`
- [ ] `curl -X POST http://localhost:3000/auth/login` returns 200 with `{ data: { token } }`
- [ ] No TypeScript errors: `npx tsc --noEmit` exits 0

The execute phase ends only when the LLM can check off every must-have.

---

Git Strategy

GSD manages git automatically in auto mode:

main
 └── milestone/M1          ← worktree branch created at start
      ├── commit: [M1/S1/T1] implement user model
      ├── commit: [M1/S1/T2] add migrations
      ├── commit: [M1/S1] slice complete
      ├── commit: [M1/S2/T1] POST /users endpoint
      └── ...
 
 After milestone complete:
main ← squash merge of milestone/M1 as "[M1] Auth system"

Each task commits with a structured message. Each slice commits a summary commit. The milestone squash-merges to main as one clean entry.

---

Crash Recovery

GSD writes a lock file at .gsd/LOCK when a unit starts and removes it on clean completion. If the process dies:

# Next run detects the lock and auto-recovers:
/gsd auto

# Output:
# ⚠ Lock file found: M1/S3/T2 was interrupted
# Synthesizing recovery briefing from session artifacts...
# Resuming with full context

The recovery briefing is synthesized from every tool call that reached disk — file writes, shell output, partial completions — so the resumed session has context continuity.

---

Cost Controls

Set a budget ceiling to pause auto mode before overspending:

/gsd auto --budget 10.00

The cost ledger at .gsd/costs/ledger.json:

{
  "units": [
    {
      "id": "M1/S1/research",
      "model": "claude-opus-4",
      "inputTokens": 12400,
      "outputTokens": 3200,
      "costUsd": 0.21,
      "completedAt": "2025-01-15T10:23:44Z"
    }
  ],
  "totalCostUsd": 1.84,
  "budgetUsd": 10.00
}

---

Decisions Register

.gsd/DECISIONS.md is auto-injected into every task dispatch. Record architectural decisions here and the LLM will respect them across all future sessions:

# Decisions Register

## D1: Use kysely not prisma
**Date:** 2025-01-14
**Reason:** Better TypeScript inference, no code generation step needed.
**Impact:** All DB queries use kysely QueryBuilder syntax.

## D2: JWT in httpOnly cookie, not Authorization header
**Date:** 2025-01-14  
**Reason:** Better XSS protection for the web client.
**Impact:** Auth middleware reads `req.cookies.token`.

---

Stuck Detection

If the same unit dispatches twice without producing its expected artifact, GSD:

1. Retries once with a deep diagnostic prompt that includes what was expected vs. what exists on disk
2. If the second attempt fails, stops auto mode and reports:

✗ Stuck on M1/S3/T1 after 2 attempts
Expected: src/features/auth/jwt.ts (not found)
Last session: .gsd/sessions/M1-S3-T1-attempt2.log
Run `/gsd run --task M1/S3/T1` to retry manually

---

Skills Integration

GSD supports auto-detecting and installing relevant skills during the research phase. Create SKILLS.md in your project:

# Project Skills

- name: postgres-kysely
- name: express-typescript  
- name: jest-testing

Skills are injected into the research and plan dispatch prompts, giving the LLM curated knowledge about your exact stack without burning context on irrelevant docs.

---

Timeout Supervision

Three timeout tiers prevent runaway sessions:

Configure in .gsd/config.json:

{
  "timeouts": {
    "softMinutes": 8,
    "idleMinutes": 3,
    "hardMinutes": 15
  },
  "defaultModel": "claude-opus-4",
  "researchModel": "claude-sonnet-4"
}

---

TypeScript Integration (Pi SDK)

GSD is built on the Pi SDK. You can extend it programmatically:

import { GSDProject, AutoRunner } from 'gsd-pi';

const project = await GSDProject.load('/path/to/project');

// Check current state
const state = await project.getState();
console.log(state.currentMilestone, state.currentSlice);

// Run a single slice programmatically
const runner = new AutoRunner(project, {
  budget: 5.00,
  onUnitComplete: (unit, cost) => {
    console.log(`Completed ${unit.id}, cost: $${cost.toFixed(3)}`);
  },
  onStuck: (unit, attempts) => {
    console.error(`Stuck on ${unit.id} after ${attempts} attempts`);
    process.exit(1);
  }
});

await runner.runSlice('M1/S4');

---

Custom Dispatch Hooks

Inject custom context into any dispatch prompt:

// .gsd/hooks.ts
import type { DispatchHook } from 'gsd-pi';

export const beforeTaskDispatch: DispatchHook = async (ctx) => {
  // Append custom context to every task dispatch
  return {
    ...ctx,
    extraContext: `
## Live API Docs
${await fetchInternalAPIDocs()}
    `
  };
};

Register in .gsd/config.json:

{
  "hooks": "./hooks.ts"
}

---

Roadmap Reassessment

After each slice completes, GSD runs a reassessment pass that may:

• Re-order upcoming slices based on discovered dependencies
• Split a slice that turned out larger than expected
• Mark a slice as no longer needed
• Add a new slice for discovered work

The LLM edits ROADMAP.md in place. You can review diffs with:

git diff ROADMAP.md

To disable reassessment:

{
  "reassessment": false
}

---

Troubleshooting

Auto mode stops immediately with "no pending slices"

All slices in ROADMAP.md are marked [x]. Reset a slice: remove [x] from its entry and delete .gsd/milestones/M1/slices/S3/SUMMARY.md.

LLM keeps failing must-haves

Check .gsd/sessions/ for the last session log. Common causes: must-have references wrong file path, or test command needs environment variable. Adjust must-haves in the task's PLAN.md and re-run with /gsd run --task M1/S3/T2.

Cost ceiling hit unexpectedly

The research phase on large codebases can be expensive. Set researchModel to a cheaper model in config, or reduce codebase index depth.

Lock file left after clean exit

rm .gsd/LOCK
/gsd auto

Git worktree conflicts

git worktree list          # see active worktrees
git worktree remove .gsd/worktrees/M1 --force
/gsd auto                  # recreates cleanly

Session file too large for recovery

If .gsd/sessions/ grows large, GSD compresses sessions older than 24h automatically. Manual cleanup:

/gsd cleanup --sessions --older-than 7d

---

Links

• GitHub: gsd-build/GSD-2
• npm: gsd-pi
• Pi SDK
• Original GSD v1