Machine Forensics Loop — AI skill for Claude Code

machine-forensics-loop

**Your machine will break again in a way it has already broken before.**

This is what stops your agent from starting over when it does.

An agent using this skill reads what is already known about *this* machine before it touches anything, verifies with machine evidence instead of narration, and writes down what it found — so the next session, or the next agent, starts where the last one stopped.

npx skills add timmyjoe917/machine-forensics-loop

That is the whole install. No account, no service, nothing to configure.

**The manual is not the notes.** This repository is the procedure, and it contains no facts about any machine. The knowledge base is built on your machine, by your agent, and stays there.

---

What it looks like

A knowledge base entry after one real investigation. This one is invented to show the shape — yours will name your own tooling:

### The sync daemon refuses to start, twice a week, with "another instance is running"

- 症状 / Symptom — exits within a second, no window; 6 times in 3 weeks, always after an unclean stop
- 归因 / Attribution
  - 【已验证】 a lock file survives the crash — 证据: its mtime is minutes *before* the failing start
  - 【推断】 the daemon checks the lock before it checks whether the owning process is alive
  - 【假设】 an unclean stop is the only trigger — 验证方法: kill -9 mid-sync, then restart
- 排查过程 / Investigation — two wrong turns: suspected disk space (no), suspected the GPU driver
  (no). Both cost an hour. Neither needs to cost the next one.
- 修复 / Fix — delete the stale lock. **Revert**: nothing to revert; the daemon recreates it
- 未解决 / Open — it still cannot tell a stale lock from a live one

Three things in there that a chat log would not have: the confidence tags, the two dead ends, and the revert procedure. **The dead ends are the part that saves the next agent an afternoon.**

Who this is for

• You run a coding agent on a machine you actually care about — Code