grounding

name: grounding description: > Use when a long working session should pause for a grounding checkpoint — fired by the huhhb grounding hook after the configured interval (default 2h) or on returning from a long break, or run manually to re-engage with in-flight work before it drifts. Triggers on "grounding", "/grounding", "grounding checkpoint", "ground the session", "are we still on track", "checkpoint the session". Opt-in via huhhb; off by default.

Grounding

A periodic, opt-in checkpoint that keeps a long session grounded — four goals: keep track of in-flight work, keep changes aligned to the repo's documented principles, improve code quality, and re-confirm goals/outcomes. Enable via the huhhb hook (~/.claude/grounding.on) or run /grounding. The hook only nudges (best-effort, never forced) — treat a nudge as standing intent and checkpoint at the next pause. Mechanics, report template, rationalization table, and red-flags live in reference.md.

Order of operations

1. First run — if ~/.claude/grounding.on has no interval_min, ask the preferred interval (default 2h) and persist it. Skip once configured; manual runs proceed regardless.
2. Don't hijack — if the user has a pending request (bug, incident, question), handle it first and offer the checkpoint at the next pause.
3. Read state, skip no-ops — gather the work since last_ground (uncommitted changes plus commits since that epoch; see reference.md for the exact range). Nothing meaningful changed? say "still grounded — nothing to do" and stop. Never invent findings.
4. Offer the menu — list the six checks; the user picks (default all, a subset, or "skip"). Run only what they pick; name any skipped — never a silent omission or a fabricated result.

The checks (run the selected, in order)

1. Work snapshot — git status + git diff --stat; surface uncommitted/unpushed work and suggest logical-chunk commits. Never commit yourself.
2. Code review — actually run /simplify then /security-review and report their output, not an eyeballed verdict. Didn't/can't invoke it? say "manual review — command not invoked"; never relabel your own analysis as the tool's output.
3. Test/build/lint — detect and run the project's test/lint command (package.json/pyproject/ Makefile/…), or offer to; report the real result. Never claim "tests pass" unrun.
4. Repo conformance — read CLAUDE.md / AGENT.md / CONTEXT.md / CONTRIBUTING / docs and check the work against each documented rule, not just the salient one. No docs? say so; don't invent rules.
5. Goal/scope — compare the work to the session's stated objective; flag drift.
6. Gaps & next steps — what's missing or assumed; re-confirm the next steps.

Report, recommend, close

Render the report (reference.md template) for the checks that ran, naming any skipped. Synthesize the findings into a short prioritized "Recommended actions" list — concrete, ranked (blockers/security first), each tied to its finding — and offer to carry them out (confirm-first). Recommend only what a real finding supports; a clean checkpoint recommends nothing — never invent work. Propose any edits as a confirm-first diff; never write or commit unprompted. Ask ≤3 questions. When the checkpoint genuinely completes, stamp last_ground to the state-file path the hook named (only you know it finished). "stop grounding" disables, "not now" snoozes, "skip" skips once.