Autonomous orchestrator for Claude Code. Write a one-line feature description; pipelinekit ships a merged PR. Charter → Analyze → Plan → Implement → Review → Merge, all driven by multi-agent dispatch with no human in the loop unless you opt in.
⚠️ v0.0.1 — pre-release. Actively developed; API not yet stable. Pin to a specific commit SHA when integrating into your own workflows.
Stuck on install or a first run? See documentation/getting-started.html#troubleshooting for the most common failure modes (/pipeline not found after install, cloud-provider CLI auth halts, review cycles not converging, a skill not firing).
The complete documentation lives in documentation/. Every page is a single self-contained HTML file — no CDN, no remote assets, no build step. After cloning, open any page directly via file:// URL:
xdg-open documentation/index.html # Linux
open documentation/index.html # macOS
start documentation/index.html # Windows / WSLOr download just the documentation/ folder as a zip and open index.html — same result, no clone needed.
Start with the master guide: documentation/getting-started.html — install + 4 worked tutorials (hello-world, TDD, Vercel deploy, meta-walkthrough).
- 45 skills + 22 agents + 32 hooks wired together so
/pipelinecan take a feature from charter to merged PR without intervention. - 5 cloud deployment providers (Azure, Vercel, Railway, Render, DigitalOcean) on a shared
deployment-engineer.mdbase. Your charter selects one; /pipeline routes deployment automatically. - Sandbox-isolated execution (default-on) — every subprocess in a podman/docker/worktree sandbox with auto-detected provider chain. Engine-when-present: prefer
podman, elsedocker, else graceful fallback to worktree-only with a one-line stderr log. Opt out per run viaSANDBOX_PROVIDER=worktree-onlyorPIPELINE_NO_SANDBOX=1. - TDD doctrine baked in —
dev-class features auto-route throughtdd-test-writer→tdd-implementerfor red/green pairing, enforced bytdd-red-phase-gate.shhook. - Multi-agent /pipeline-review — 2–6 specialized review agents (code-reviewer, security-auditor, test-engineer, performance-tuner, spec-tracer, symbol-verifier) with charter-aware finding classification + Path A/B/M/N routing.
- Research loop — Karpathy-style autoresearch via
/research: hypothesize → mutate → benchmark → keep-or-reset, with optional auto-publish to a branch via/ppr --research. - GitHub-native —
--issuesmode pulls features from open issues; PR bodies auto-includeCloses #N; squash-merge viagh pr merge --squash --auto.
git clone https://github.com/Bruce188/pipelinekit.git
cd pipelinekit
./scripts/install.shinstall.sh is idempotent. Re-run to update.
# Oracle Cloud Free Tier ARM A1 (aarch64, 24 GB RAM):
curl -fsSL https://raw.githubusercontent.com/Bruce188/pipelinekit/main/scripts/cloud/oracle-bootstrap.sh | bash
# Hetzner CX22 (x86_64, 4 GB RAM — swap auto-created):
curl -fsSL https://raw.githubusercontent.com/Bruce188/pipelinekit/main/scripts/cloud/hetzner-bootstrap.sh | bashBoth scripts are idempotent. Secrets read from env, never baked in. Full walkthrough: documentation/cloud-setup.html.
Open the repo in Codespaces — the devcontainer post-create runs install.sh automatically. Then claude in the terminal to start.
"Clone https://github.com/Bruce188/pipelinekit.git into ~/work,
cd into it, and run ./scripts/install.sh non-interactively."
Full install variants (local, cloud cold-start, Codespaces / devcontainer, non-interactive): documentation/installation.html.
After install, drop a feature file in a target repo:
mkdir -p docs
cat > docs/features.md <<'EOF'
## fix/readme-typo
**Description:** Fix the typo in README.md where "occured" should be "occurred".
**Constraints:** Single-file edit. No new tests. Non-dev feature class.
### Run Log
EOF
claudeIn the Claude session:
/pipeline docs/features.md --no-charter
That's it. The orchestrator handles the rest: analyze → plan → branch → implement → review → push → PR → squash-merge. Watch state transitions in real time. Total runtime: ~3–5 minutes for a typo fix; ~15–30 minutes for a real feature.
Full tutorial set: documentation/getting-started.html.
/pipeline routes per feature based on <type>/<name> prefix in docs/features.md:
| Prefix | Mode | TDD |
|---|---|---|
feat, fix, refactor, perf, test |
dev | yes — tdd-test-writer → tdd-implementer |
docs, chore, style, build, ci, content, ops, research |
non-dev | skipped |
Override with **Type:** dev|non-dev per-feature.
| Layer | Contents |
|---|---|
| Rules | claude/CLAUDE.md.template, claude/rules/workflow.md, claude/rules/agents-worktrees.md |
| Skills (45) | pipeline, analyze, create-plan, implement-plan, pipeline-review, ppr, research, tdd, incident, landing-report, learn, azure-ops, vercel-ops, railway-ops, render-ops, digitalocean-ops, expo, ios, docs-writer, document-release, claude-md-enhancer, write-a-skill, zoom-out, simplify, caveman-mode, ... |
| Agents (22) | code-reviewer, security-auditor, symbol-verifier, spec-tracer, tdd-test-writer, tdd-implementer, mobile-dev, deployment-engineer (base), azure-deployment-engineer, vercel-deployment-engineer, railway-deployment-engineer, render-deployment-engineer, digitalocean-deployment-engineer, incident-responder, claude-md-guardian, ... |
| Hooks (32) | validate-commit-msg, strip-ai-attribution, block-push-main, block-stage-sensitive, block-dangerous-commands, tdd-order-check, tdd-red-phase-gate, claude-md-guard, env-scrub, notify-emit, context-warning, context-budget-advisor, verify-worktree-commit, ... |
| MCP | context7, serena, sequential-thinking, agentmemory, codegraph, graphify, optional local-rag + claude-context |
| LSP | pyright, typescript, csharp, gopls, rust-analyzer |
| Model overlays | Per-model tuning for opus-4-7, sonnet-4-6, haiku-4-5 (consumed by phase skills) |
| Host adapters | claude + codex (concrete), cursor/gemini (stubs) — interface scaffold for multi-host dispatch |
| Sandbox | Pluggable SandboxProvider.sh (auto default: podman → docker → worktree-only) with sandbox_wrap shared helper |
Five providers, all driven from Charter Topic 10:
| Agent | Skill | Provider | Guide |
|---|---|---|---|
@azure-deployment-engineer |
azure-ops |
Azure App Service / Container Apps / Function Apps / AKS | docs |
@vercel-deployment-engineer |
vercel-ops |
Vercel (Next.js / SvelteKit / Astro / Remix) | docs |
@railway-deployment-engineer |
railway-ops |
Railway (zero-Dockerfile container PaaS) | docs |
@render-deployment-engineer |
render-ops |
Render (managed container PaaS) | docs |
@digitalocean-deployment-engineer |
digitalocean-ops |
DigitalOcean App Platform | docs |
claude/agents/deployment-engineer.md is the shared base agent (dispatch with provider: <name>) describing the shared contract (auth posture, secret hygiene, health-check polling, no-direct-REST). Each named agent extends it for its provider.
Publish research keep-rows from a TSV to a dedicated branch. Dry-run by default; pass --no-dry-run to publish:
/ppr --research # dry-run (default)
/ppr --research --no-dry-run --research-tag my-experimentReference: documentation/ppr-research-flag.html.
pipelinekit/
├── .devcontainer/ # Codespaces / VS Code devcontainer
├── claude/ # Overlay installed to ~/.claude/
│ ├── CLAUDE.md.template
│ ├── rules/
│ ├── skills/ # 45 native skills
│ │ ├── pipeline/ # The autonomous orchestrator
│ │ ├── docs-writer/ # template.html + render.py (rich-template HTML)
│ │ ├── research/ # Karpathy autoresearch loop + tsv-viewer.sh
│ │ ├── {railway,render,digitalocean}-ops/ # Cloud provider ops skills
│ │ └── ...
│ ├── agents/ # 22 specialized agents (incl. deployment-engineer base)
│ ├── hooks/ # 32 production hooks (29 harness + 3 git)
│ ├── lib/
│ │ ├── pipeline/ # ac_lint, charter_revalidate, decomposition_check, ...
│ │ └── sandbox/ # SandboxProvider.sh + providers/ + sandbox_wrap.sh
│ ├── model-overlays/ # opus-4-7.md, sonnet-4-6.md, haiku-4-5.md, claude.md
│ ├── host-adapters/ # claude.sh + codex.sh (concrete), cursor.sh / gemini.sh (stubs)
│ ├── memory/ # MEMORY.md scaffold (empty by design; populated per project)
│ └── config/ # never-stage.txt, etc.
├── scripts/
│ ├── install.sh # Idempotent installer
│ ├── sandbox/ # Container base image (Containerfile + build.sh)
│ └── cloud/ # Oracle + Hetzner bootstrap scripts
├── documentation/ # User-facing docs — all self-contained HTML
│ ├── index.html # Landing page
│ ├── getting-started.html # Master guide (start here)
│ ├── changelog.html # v0.0.1 dev log
│ ├── installation.html, pipeline.html, cloud-setup.html
│ ├── deployment-{azure,vercel,railway,render,digitalocean}.html
│ └── SKILL-AUTHORING-STANDARD.html, SKILL_PIPELINE.html, NOTICE.html # Vendored reference standards + attribution
├── .mcp.json.template # MCP server config (copy to project root on install)
├── LICENSE # MIT
└── README.md
Note: docs/ is intentionally NOT in the repo tree above — it's gitignored and exclusively for AI-workflow artifacts (analysis-vN.md, plan-vN.md, prompts-vN.md, review-vN.md, charter.md, progress.md, pipeline-state.md, features-vN.md, etc.). User-facing docs live in documentation/ and are committed as HTML.
Default verbosity in pipelinekit sessions: caveman wenyan-ultra. Drops articles, filler, hedging. Code, commits, and security-critical text remain normal English.
/caveman lite | full | ultra
stop caveman # revert to normal
Memory ships as an empty scaffold. After install, the agentmemory MCP (registered by install.sh) is the canonical memory store; legacy flat files under ~/.claude/projects/<project-slug>/memory/ stay read-only as a rollback window (dual-write via PIPELINE_KEEP_FLAT_FILE_MEMORY=1). See claude/memory/MEMORY.md for the flat-file schema reference (recency-weighted confidence decay, four memory types: user / feedback / project / reference).
- Master guide — install + 4 worked tutorials (start here)
- Documentation index — every reader-facing page
- Changelog v0.0.1 — every PR organized by theme
- Pipeline workflow — /pipeline phase-by-phase reference
- SKILL authoring standard — vendored standard (10-pattern skill DNA)
- SKILL pipeline lifecycle — Intent → ... → Verify
Pipelinekit stands on the shoulders of several upstream projects, plus inspiration from individual researchers and the Claude platform itself.
Inspirational-but-not-vendored:
- Andrej Karpathy — the autoresearch-loop pattern (hypothesize → mutate one file → benchmark → keep-or-reset → append row → repeat) is the design behind
/research. Thesymbol-verifieranti-hallucination agent (Agent 6 of/pipeline-reviewon medium+ diffs) is inspired by the same first-principles verification ethos. - Anthropic — the entire runtime: Claude Code, the Agent SDK, multi-agent dispatch, Skills, Hooks, MCP,
PushNotification, and the underlying Claude models that every pipelinekit phase calls. Pipelinekit is a workflow overlay on Claude Code — without Anthropic's platform, there's no pipelinekit.
Vendored upstreams (full table with pinned SHAs + licenses + scope + re-vendor procedures at documentation/credits.html):
- wshobson/agents — 2 specialist agents: deployment-engineer, incident-responder (MIT, © 2024 Seth Hobson)
- alirezarezvani/claude-skills — skill-authoring DNA + lifecycle reference + personas (MIT, © 2025 Alireza Rezvani)
- alirezarezvani/ClaudeForge — claude-md-guardian + claude-md-enhancer (MIT, © 2025 Alireza Rezvani)
- mattpocock/skills — TDD doctrine pack + write-a-skill + zoom-out (MIT, © Matt Pocock)
Per-directory NOTICE.md files carry the full upstream license text and re-vendor procedures.
MIT. See LICENSE. Pipelinekit's own original code (orchestrator, hooks, docs-writer, install scripts, documentation) is MIT-licensed; vendored upstream files retain their original licenses linked above.