kvm-pilot: Skill for Claude Code

/checkpoint — end-of-session safety sweep

A checkpoint asks one question: is everything I did this session durably captured, or is something quietly at risk of being lost? The next session is often a fresh agent with zero memory of this one — a gap here means redone work or a decision shipped on a false assumption. This replaces "I think I saved everything" with a reproducible sweep ending in SAFE TO STOP or N items need attention, each with the exact fix.

It is the mirror of a morning sweep (which asks what happened while I was away?). Stay in that lane: checkpoint does not score releasability (a separate concern) and does not re-run cloud/telemetry health checks.

Procedure

1. Run the read-only scan (it mutates nothing — that's the safety property):

   bash .claude/skills/checkpoint/scan.sh
   

2. Interpret each surface against the rubric below — the judgement calls the script can't make. Do not skip a surface because it "looked fine"; read the facts.
3. Emit the risk-ordered table + one-line verdict (format at the bottom). Highest-risk first.
4. Always produce the resume handoff (§ Mandatory handoff) — this happens even on an all-✅ scan.
5. Offer the exact fix for every ⚠️; run only the low-risk ones on the operator's go (commit-by-name, save-memory, fix-a-doc). Never promote/release; never blind-add.
6. Meta-reflect and self-improve (§ Self-improvement) — the last thing you do. Did the scan raise a false positive (flagged something fine), miss something you only caught by hand, or produce noise? Did the table help? If yes, amend scan.sh/SKILL.md in this same run to encode the fix and add a § Version history row pairing the change with the real incident. This step is why the skill is worth more each run.

Rubric — interpret the scan, surface by surface (risk-ordered)

C1 · Version control & working tree — highest risk: unsaved code is unrecoverable

• UNCOMMITTED > 0 → ⚠️ commit by name: git -C kvm-pilot commit -m "…" -- <paths>. Never git add -A/. (.gitignore covers venv/artifacts/config.toml/.env, but a stray secret or large file under a new name would be committed irreversibly — stage by name).
• UNTRACKED_NOT_IGNORED > 0 → ⚠️ eyeball each: is it source to track, or a stray artifact/secret to ignore/delete? Never bulk-add.
• ON_PROTECTED_BRANCH: yes (main) — normal here, ✅, not a reason to branch. kvm-pilot is issue-per-finding direct-commit-to-main (CLAUDE.md; agent work is not branched). The C1 risk is unsaved work, not the branch — do not advise switch -c. (v0.2.0: the prior version wrongly assumed a feature-branch/PR flow.)
• UNPUSHED (ahead) > 0 → ⚠️ push it (git -C kvm-pilot push). On this project pushing main runs CI and backs the work up; it does not release (a release needs gh release create). Commits left only-local are the actual work-at-risk. (v0.2.0.)
• BEHIND > 0 → ⚠️ note; rebase before more work.
• DEBUG_MARKERS_IN_DIFF > 0 (breakpoint()/pdb/console.log/FIXME/XXX) → ⚠️ remove before commit. Also eyeball the diff for stray print( added to library modules (client.py, health.py, drivers/ — the CLI uses print legitimately, so the scan doesn't flag it).
• SECRET_SCAN reports N possible literal(s) → 🔴 inspect each; if any is a real credential, ROTATE it (it's in the working diff or the last 25 commits — likely already pushed, so removing it doesn't undo the exposure). Heuristic (quoted password=/token= literals, inline IPMI_PASSWORD=<value>, private-key headers; stock defaults like calvin/admin and env/placeholder refs are excluded), so a hit is "eyeball this," not proof — but never dismiss one without looking. Do not put the secret in RESUME.md or any file to "track" it; RESUME.md is local-only (#209) and still must not hold raw secrets. (v0.10.0)