agentscar

Your coding agent makes the same mistake twice. agentscar makes it the last time.

Auto-memory remembers. It doesn't learn. agentscar runs a blameless postmortem on every agent incident — root-causes it, then writes the guardrail (a hook, a rule, or a skill) at the point where the failure happens, not in a file the agent has to remember.

agentscar init   # sets up .agentscar/ — log, guardrail templates
agentscar new    # guided postmortem → routed guardrail + log entry
agentscar log    # your agent's scar tissue, greppable

Bash + markdown. Zero dependencies. Adapters for Claude Code and AGENTS.md; the rest is plain markdown. I haven't tested it against other CLIs.

Why

Every practitioner running coding agents daily knows the loop: the agent hits a rejected git push, "fixes" it with --force, and wipes commits it never saw; you add a rule, and three weeks later it happens again — because the rule was prose, buried deep in context, losing the attention war against file dumps and test output.

Notes don't change behavior. Enforcement does. A guard that runs before the push beats any instruction about pushing: the guard can exit 1, the instruction can only ask.

agentscar is the discipline SRE teams use for outages, applied to agents:

1. What happened — one honest paragraph.

2. Root cause — 3–5 whys. ("Agent is careless" is never a root cause.)

3. Guardrail — routed to the strongest enforcement layer that fits:

   hook (deterministic block) → rule (instruction) → skill (procedure) → test (regression)

4. Log entry — dated, typed, linked to the guardrail it produced.

If a lesson matters, it wants the strongest layer that fits — and that's a hook more often than you think.

vs. what you already have

| | CLAUDE.md / rules | auto-memory | hooks alone | agentscar | |---|---|---|---|---| | Captures lessons | manually | automatically | no | guided, per incident | | Changes behavior | if attention holds | no — storage | yes, narrowly | yes — routed to strongest layer | | Root-cause discipline | no | no | no | yes (blameless postmortem) | | History you can grep | no | partially | no | agentscar log |

agentscar doesn't replace any of these — it's the loop that decides what goes where, and why.

Quickstart

npm i -g agentscar   # or: pipx install agentscar (or: uv tool install agentscar)
# or from source:
git clone https://github.com/agentheavy/agentscar && cd agentscar && ./install.sh   # or: copy agentscar to PATH
cd your-project
agentscar init
# next time your agent does something you never want repeated:
agentscar new

Windows: agentscar is a bash script — run it from Git Bash (ships with Git for Windows, which Claude Code already requires) and keep bash on PATH. The npm and pipx installs add shims, but the shims still call bash — if you hit bash: command not found, switch to Git Bash.

agentscar new walks you through the postmortem and drops:

• an entry in .agentscar/log.md
• a guardrail skeleton in .agentscar/rules/, with the intended enforcement layer (rule, hook, or test) noted inside

Everything agentscar writes is plain markdown; the minimal frontmatter it carries is OKF v0.2-compatible, so catalog tooling can index a .agentscar/ bundle as-is.

Claude Code users: agentscar init --claude (or plain init in a repo that already has .claude/) also installs a skill so the agent runs the postmortem flow itself after an incident and proposes the guardrail as a diff — you approve, it lands.

Other agents (Codex, Cursor, …): agentscar init --agentsmd adds an agentscar section to AGENTS.md (created if missing; plain init also updates an existing one), pointing anything that reads the file at the postmortem flow.

Skill without installing the CLI: npx skills add agentheavy/agentscar (via skills.sh) drops the postmortem skill straight into Claude Code, Cursor, Codex and other agents — no agentscar install needed. Run agentscar init once when the skill needs a log to write incidents to.

Six failure types

Six types cover everything I have hit so far: wrong-assumption · destructive-action · verification-skip · instruction-drift · spec-drift · context-loss.

Each type has a guardrail shape that holds it better than the others do. Full breakdown: docs/failure-types.md.

Included templates

• hook-push-guard — fetch + compare before any push; blocks on divergence.
• hook-confirm-destructive — force-push / rm -rf / hard reset require explicit human approval.
• rule-skeleton — constraint / why (incident link) / how-to-apply / last-reviewed.

A CI bot's commits outran my local branch, and remembering to check wasn't enough — that's where push-guard came from. I run a version of it against my own agent. The other two I don't run; they're starting points.

FAQ

Can't I just put these rules in CLAUDE.md / AGENTS.md? You can — but prose competes for attention, and attention degrades with context length. agentscar's point is routing: most lessons people write as rules wanted to be hooks.

Does it phone home / need an API key? No. It's bash and markdown files in your repo. Nothing leaves your machine.

Why "agentscar"? Scar tissue is permanent memory of damage — grown so the same wound doesn't open twice.

Roadmap? agentscar lint (stale-rule detection: last-reviewed > 90 days → warn), contradiction checks, shared team logs. One thing at a time.

License

MIT.

agentscar is one discipline of a larger verification loop — the rest lives at agentheavy.dev, field notes at agentheavy.substack.com.