An AI-agent skill for the review → plan → execute → verify → release cycle. AutoDev is a self-contained workflow you drop into your own agent harness — not a CLI app you drive by hand. Once installed, your agent gains a structured pipeline it runs with its own native tools.
Most "vibe coding" stops at a first working draft. AutoDev takes an existing concept and cycles it to done — review → code → test (locally and on CI) → repeat, until you decide the project is good enough to release. Each loop tightens the code instead of shipping the lucky first pass.
- Loop until release-ready. Review → execute → verify re-runs every iteration, so defects found late get fixed, not deferred.
- Local + CI, not just local. Fixes are verified by your test suite and GitHub Actions.
- Reproducible, file-based trail. Every review, plan, and CI report lands
in
dev-notes/as plain markdown — traceable, diffable, git-friendly. - Multi-harness by design. The skill is just
SKILL.md+references/; any agent harness can load it. Bundled Rust binaries are optional accelerators, not a requirement. - Developed on itself. Every AutoDev release was produced and hardened by AutoDev's own review cycles — see docs/development-by-cycles.md.
One command — no checkout needed:
curl -fsSL https://raw.githubusercontent.com/ni9aii/AutoDev/main/install.sh | bash -s -- --remoteThe installer auto-detects your harness (Hermes, Claude Code), downloads a
release tarball, and installs a self-contained copy of the skill with its
references. Re-run the same command (or install.sh --update from a checkout)
to upgrade; --uninstall removes it; a version stamp in the installed skill
dir tells you which release you have.
From a checkout instead:
git clone https://github.com/ni9aii/AutoDev && cd AutoDev
./install.sh # auto-detects harness
./install.sh --harness hermes | claude-code # force one
./install.sh --list # supported harnesses + paths
./install.sh --check # verify install without changesAfter install, load the skill in your agent and run a phase, e.g.
/autodev /path/to/project review.
| Harness | Install path | Invoke with |
|---|---|---|
| Hermes | ~/.hermes/skills/autonomous-ai-agents/autodev |
/skill autodev |
| Claude Code | ~/.claude/skills/autodev |
/autodev |
run-pipeline /path/to/project review --project myproject # 4 reviewers → reports
run-pipeline /path/to/project plan --project myproject # aggregate → fix plan
run-pipeline /path/to/project full --project myproject # all phases incl. verifyEach run leaves a timestamped trail under $DEV_NOTES_ROOT/<project>/
(reviews, plans as .md + machine-readable .json, CI reports). Your agent
executes the plan; the pipeline verifies locally and on CI.
Full documentation lives in docs/:
| Doc | Purpose |
|---|---|
| docs/how-it-works.md | Architecture: skill layer vs Rust accelerators, execution model |
| docs/installation.md | Installer details: flags, manual install, update/uninstall |
| docs/dev-notes-layout.md | Artifact tree, file formats, plan sidecar contract |
| docs/configuration.md | Env vars (GITHUB_TOKEN, DEV_NOTES_ROOT, …) |
| docs/jev-classification.md | Optional --jev AI classification for review findings (endpoints, env vars, guarantees) |
| docs/project-structure.md | Repository layout (generated, drift-checked) |
| docs/development-by-cycles.md | How AutoDev develops itself, cycle by cycle |
| references/ | Deep-dive notes: JSON output contract, troubleshooting, patterns |
A fully worked sample: four review reports → a generated fix plan
(examples/sample-project/plans/) and a
machine-readable --json summary.
See CONTRIBUTING.md. Adding support for a new agent harness is a single overlay file plus installer wiring — see docs/new-harness.md.
MIT — see LICENSE.