contextlisted

# Related Context Target (may be empty): $ARGUMENTS **Two depths.** The *user* asked "where are we / what's next / continue" with no area named → [HOP0.md](HOP0.md): counts from the plans and ledgers, nothing opened, ≤ 10 lines, stop. Anything else — a target (area, row, debt `id`, file), a "why is X like this?" question, or **you triggered this yourself before a task** (then the target is that task's area; an empty `$ARGUMENTS` is not a reason for hop 0) → hops 1–3 below, then open **only** the docs those records point at and answer with `file:line` quotes. A follow-up question in the same session digs from where the last hop stopped; never re-run hop 0, never fall back to reading `.claude/docs/` whole. Nothing on record → say so; do not reconstruct an answer from the code. Three things decide whether a change is correct: what the customer asked for (`.claude/docs/specs/`), what was already built and why (`.claude/docs/tasks/`, `.claude/docs/decisions/`), and what is known broken/undecided (`.claude/clio/debt.jsonl`). This skill loads all three cheaply, in that order, without reading the whole archive. **Read-only.** Never writes to `index.jsonl`/`debt.jsonl`, never edits a spec, never fixes a stale record — it *reports* contradictions; `/clio:memo` and `/clio:update` are the writers. Staying silent because "nothing was actionable" is the failure mode. ## Walk the hops, in order Each hop feeds the next — don't skip ahead, and don't stop after hop 1 just because it look