GitHub as a knowledge graph for AI agents. Autonomous dev pipeline for Claude Code - investigate, build, review, merge. Issue in, PR out.
# Add to your Claude Code skills
git clone https://github.com/RapierCraftStudios/ForgeDockForgeDock is an open-source ai agents skill for AI coding assistants such as Claude Code, Codex CLI, and ChatGPT, built by RapierCraftStudios. GitHub as a knowledge graph for AI agents. Autonomous dev pipeline for Claude Code - investigate, build, review, merge. Issue in, PR out. It has 90 GitHub stars.
ForgeDock's catalog security scan is still queued. You can run an instant dependency and prompt-injection check now with the "Scan for vulnerabilities" button above.
Clone the repository with "git clone https://github.com/RapierCraftStudios/ForgeDock" and add it to your Claude Code skills directory (see the Installation section above).
ForgeDock is primarily written in JavaScript. It is open-source under RapierCraftStudios on GitHub, so you can review or fork the full source.
Yes. SkillsLLM lists many other AI Agents skills you can browse and compare side by side. Open the AI Agents category from the badge at the top of this page, or use the Related Skills and comparison links further down to weigh ForgeDock against similar tools.
No comments yet. Be the first to share your thoughts!
Unlocks once the catalog security scan passes (runs nightly).
The deep catalog scan for this skill is still queued. Run an instant dependency check now instead.
AI coding agents forget everything between sessions. They re-investigate the same bugs, miss context from past PRs, and make mistakes that were already caught and fixed last week. There's no institutional memory.
ForgeDock fixes this by using GitHub itself as the memory layer. Every pipeline stage writes structured FORGE: annotations to issues and PRs. Every downstream agent reads them. When a new session starts — even after Claude's context resets — the agent queries GitHub and picks up exactly where the last one left off.
ForgeDock is not another AI coding agent. It's a set of prompt-engineered command specs (.md files) that run inside Claude Code. No new runtime, no separate process, no vendor lock-in beyond what you already use.
| ForgeDock | Plain Claude Code | Cursor / Windsurf | Devin / Sweep | |
|---|---|---|---|---|
| Memory across sessions | Structured annotations on GitHub | CLAUDE.md + manual notes | Per-project context | Proprietary cloud state |
| Autonomous pipeline | Full lifecycle: investigate → merge | Manual, step by step | Autocomplete + chat | Autonomous but opaque |
| Review quality | 9 domain-specialist agents | You review everything | Basic suggestions | Varies |
| Infrastructure needed | None — just npx forgedock |
None | IDE-specific | Cloud service |
| Codebase visibility | Everything stays on GitHub | Local | Local + cloud sync | Cloud-only |
| Capability | How it works |
|---|---|
| Full-lifecycle automation | /work-on #42 — investigates the issue, architects a fix, builds it, runs quality gates, opens a PR, and reviews it. You click merge. |
| Persistent agent memory | Structured FORGE: annotations on GitHub issues/PRs survive context resets and session boundaries. Agents never start blind. |
| 9 specialist review agents | Security, billing, database, concurrency, auth, frontend, API, performance, infrastructure — every PR gets domain-expert review. |
| Cross-issue knowledge graph | Agent fixing issue #43 reads the investigation from #42 and applies the known pattern — no re-investigation. |
| Self-improving pipeline | Review agents learn from past findings — recurring patterns automatically become new quality gate checks. |
| Parallel orchestration | /orchestrate decomposes milestones into waves and runs /work-on on each in parallel. |
Cost note: ForgeDock itself is free and open-source. It orchestrates Claude Code sessions, so you pay your normal Anthropic API usage. A typical
/work-onrun on a straightforward bug uses roughly the same tokens as a 15–20 minute manual Claude Code session.
Here's what a real run looks like on issue #619 — a performance bug where command specs were burning ~200K tokens in context:
FORGE:INVESTIGATOR → CONFIRMED. All 27 command spec files (848KB) load into
context at session start via symlinks. ~200K tokens wasted.
FORGE:CONTRACT → Replace symlink-based install with stub-file pattern.
Installer parses frontmatter, writes minimal stubs.
FORGE:CONTEXT → Issue #577: install() had overly broad catch{} — fixed to
check err.code === 'ENOENT'. Issue #587: Windows writes
regular files, not symlinks — keep both paths working.
FORGE:ARCHITECT → 3 new functions in bin/forgedock.mjs: parseFrontmatter(),
generateStubContent(), updated install() flow.
FORGE:BUILDER → Branch feat/stub-install-pattern-619, 1 file changed.
FORGE:REVIEW → Auto-merged to staging.
FORGE:TRAJECTORY → Full audit trail recorded.
The context phase surfaced two historical bugs (#577, #587) in the same module — preventing the builder from repeating known mistakes. View the full issue →
Requirements: Node.js 18+, Claude Code, GitHub CLI (gh), and yq (YAML parser used by pipeline commands to read forge.yaml)
# Install pipeline commands
npx forgedock
# Generate config for your repo
npx forgedock init
This symlinks 25+ pipeline commands into ~/.claude/commands/ and generates a forge.yaml config in your project root. That's it — open Claude Code and run /work-on #42.
Claude Code Plugin Marketplace (v2.1.143+):
/plugin marketplace add RapierCraftStudios/ForgeDock
/plugin install forgedock@forgedock
CLI commands:
npx forgedock update # Pull latest commands
npx forgedock uninstall # Remove all ForgeDock commands from ~/.claude/commands/
npx forgedock help # Show all available commands
AI-powered setup (inside Claude Code):
/forgedock-init # Guided config walkthrough — scans your repo, queries GitHub, auto-fills forge.yaml
Issue → Investigate → Architect → Build → Quality Gate → Review → Merge
↓ ↓ ↓ ↓ ↓
writes to reads from reads from reads from writes to
GitHub GitHub GitHub GitHub GitHub
Each stage writes a structured annotation (<!-- FORGE:INVESTIGATOR -->, <!-- FORGE:CONTRACT -->, etc.) to the GitHub issue or PR. Each downstream stage reads what came before. The gh CLI is the query interface.
| Stage | What it does |
|---|---|
| Investigate | Traces root cause via git blame, related issues/PRs. Writes verdict, affected files, severity. |
| Context | Surfaces historical bugs and known pitfalls from the same module. Institutional memory. |
| Architect | Produces ordered implementation plan with exact file/function/line targets. |
| Build | Writes code, creates branch, makes commits. Follows the architect's plan. |
| Quality Gate | 14+ domain-specific checks (security, auth, DB, concurrency, etc.) |
| Review | 9 specialist agents review the PR diff with confidence-rated findings. |
| Close | Records full audit trail as FORGE:TRAJECTORY. |
Labels track workflow state (workflow:investigating, workflow:building, workflow:in-review, workflow:merged). The pipeline resumes from whatever state GitHub says it's in — restart-safe by design.
| Command | What it does |
|---|---|
/work-on #N |
Full issue lifecycle: investigate → build → review → merge |
/issue |
Create a pipeline-ready GitHub issue |
/orchestrate |
Parallel execution across a milestone's issues |
/review-pr |
PR review with 9 domain-specialist agents |
/quality-gate |
Pre-commit checks across 14+ domains |
/milestone |
Plan and ship milestones |
/deploy-info |
Staging vs main diff with risk assessment |
/review-pr-staging |
Staging-to-main review gate |
/rollback |
Automated revert PR for production incidents |
/incident-response |
P0 coordination: hotfix, timeline, postmortem |
/autopilot |
Autonomous improvement: recon → triage → fix |
/pipeline-health |
Self-analysis and prompt tuning |
/security-audit |
4-phase security posture audit |
/qa-sweep |
Full platform QA via browser automation |
/analytics |
Pull metrics from GSC, Clarity, Umami, Stripe |
/cleanup |
Sweep stale issues, branches, worktrees |
npx forgedock uninstall
Removes all ForgeDock command symlinks from ~/.claude/commands/. Your forge.yaml config and any FORGE: annotations on GitHub issues/PRs are left untouched.