llms.txt · git:20260415.506989f · 2026-04-15 · sha256 4ae491812290406e

llms.txt git:20260415.506989fA

Immutable. This exact content never changes and is served at /api/v1/blob/4ae491812290406e.

# forge

> Forge is a Claude Code plugin that adds structured planning, parallel
> execution in isolated git worktrees, deep validation with API contract
> checks, intelligent retry via a debugger agent, and cross-session memory
> for learned patterns. It turns a one-line objective into a validated
> dependency graph of modules, each executed, verified, and reviewed before
> being merged back.

## Install

```bash
claude plugin marketplace add TT-Wang/forge
claude plugin install forge@tt-wang-plugins
```

## Usage

```
/forge <objective>          # full workflow: plan → execute → validate → learn
/forge-status               # show plan progress, iteration state, memory
/forge-validate <module>    # run validation for a specific module
```

## Architecture

- **Orchestrator skill** (`skills/forge/`) — top-level workflow
- **Agents** (`agents/`) — planner, worker, reviewer, debugger
- **MCP server** (`forge-mcp-server/`) — 7 tools: validate, validate_plan,
  memory_recall, memory_save, iteration_state, forge_logs, session_state
- **`.forge/` directory** — per-project state: plans, iterations, logs,
  session snapshots, project + global memory

## MCP tools

| Tool | Purpose |
|------|---------|
| `validate` | Run verify commands + contract checks against a worktree |
| `validate_plan` | DAG cycle detection, file-overlap warnings, schema checks |
| `memory_recall` | Search project + global memory |
| `memory_save` | Persist learned patterns (dedup by category+pattern) |
| `iteration_state` | Track per-run retry counts, scores, stagnation |
| `forge_logs` | Query structured JSONL logs (run/module/phase/severity) |
| `session_state` | Save/load orchestrator state for resumability |

## Defaults worth knowing

- Node.js ≥ 20, ESM-only, zero build step (`node index.mjs`)
- `validate`'s `cwd` arg redirects checks to a worker's worktree
- Iteration state is scoped per `runId` so module IDs don't collide across plans
- stdio-only MCP transport (no network surface)
- State writes are atomic (`tmp + rename`)

## Links

- Repo: https://github.com/TT-Wang/forge
- Issues: https://github.com/TT-Wang/forge/issues
- Changelog: https://github.com/TT-Wang/forge/blob/main/CHANGELOG.md
- Security policy: https://github.com/TT-Wang/forge/blob/main/SECURITY.md
- Contributing: https://github.com/TT-Wang/forge/blob/main/CONTRIBUTING.md
- License: MIT

## Companion project

- **memem** — https://github.com/TT-Wang/memem — persistent cross-session
  memory for Claude Code. Recommended pairing: forge plans + executes,
  memem remembers. Forge's `memory_save` patterns are readable by memem's
  recall index, so learnings compound across runs.