warmstart

Stop your AI agent from relearning your repo every session.

Every coding agent — Claude Code, Cursor, Codex, Copilot — pays the same tax on every repo, every session: it burns its first several turns guessing the test command, the build incantation, the package manager, the entry point, and stepping on gotchas the last session already figured out and threw away.

warmstart ends that. One command discovers your repo's real dev-loop, runs the commands to prove they work, and writes a committed AGENTS.md plus a live MCP server — so any agent reads the verified truth in one shot instead of fumbling.

npx warmstart init

Watch an agent waste 8 turns finding the test command → watch it read AGENTS.md and nail it in 1.

Why it's different from a hand-written AGENTS.md

A hand-written AGENTS.md rots the moment a command changes, so agents don't trust it and re-verify by trial anyway. warmstart keeps itself honest: every command carries a precise, scoped stamp of what was actually proven.

The stamp states what was proven, where, and when — never a bare "verified."

Related MCP server: instagit

How it works

1. Scan — mines your package.json scripts, Makefile targets, and pyproject.toml for candidate commands.

2. Consent — previews the exact command plan and asks before running anything.

3. Verify — runs the safe candidates and records honest, scoped results.

4. Emit — writes a delimited block in AGENTS.md (never clobbering your prose) and registers a warmstart serve MCP server for Claude Code.

Plays nice with your existing AGENTS.md

Already have a hand-written AGENTS.md? warmstart never overwrites it and never errors — it appends a single delimited block and leaves everything else untouched. Your file ends up with two zones that coexist:

# My Project
Hand-written notes you care about.   ← yours forever, never touched

## Gotchas
- Never run migrations on Friday.

<!-- warmstart:start -->
  verified dev-loop commands         ← warmstart owns / refreshes only this
<!-- warmstart:end -->

• First touch of a hand-written file → backs it up to AGENTS.md.bak, then appends the block.

• Re-runs are idempotent → warmstart replaces only the content between its markers (re-verifying the commands), so sections never stack up as duplicates.

• Your prose, architecture notes, and conventions stay yours. warmstart only manages the "how to actually run it" block.

Curated human knowledge above, machine-verified commands below — kept honest and current without stepping on your work.

Safety & privacy

• Runs fully locally. No telemetry. No network except your repo's own commands.

• Never auto-runs destructive commands. deploy, publish, rm -rf, git push, drop, and similar are listed but never executed.

• No shell execution. Commands run via cross-spawn with tokenized args — no shell-injection surface, and Windows .cmd shims work correctly.

• Never commits secrets. Command output is redacted (tokens, credentials, home paths) and raw output is never embedded in AGENTS.md.

• Never clobbers your files. An existing hand-written AGENTS.md is backed up and preserved; only a delimited block is managed.

Configuration (.warmstartrc.json or .warmstartrc.yaml)

{
  "timeoutMs": 120000,
  "flakyRuns": 1,
  "network": false,
  "disable": ["make deploy"],
  "env": {}
}

Commands

• warmstart init — scan, verify (with consent), emit AGENTS.md + register MCP. --yes for non-interactive/CI.

• warmstart serve — run the MCP server exposing the manifest.

• warmstart doctor — check toolchain availability.

Supported sources

Now: package.json (+ pnpm/yarn/npm/bun detection), Makefile, pyproject.toml. Planned: justfile, Taskfile.yml, tox.ini, CI workflows (as hint sources), passive observation, staleness re-verification, CI-stamped manifests.

License

MIT