Reuse before rerun

Overview

Every expensive run is a claim: "no usable result exists." Prove that claim in one sweep before spending compute. Most numbers a mature project needs were already measured; the failure mode is not finding them and re-running for hours.

Small runs count too: a "quick 500-row sanity run" on a question that already has a full-corpus answer is a rerun, and so is a spy over ~100+ rows. Not needed for: a single-input probe, unit tests, the fast pre-gate.

Wire this in

sweep.sh (run it; it calls _group.py beside it, and you need not read either) needs bash, git and python3 (standard library only). It reads its roots from env vars — set them once:

• RBR_REPO — repo to search (default: git toplevel, else $PWD)
• RBR_LEDGER — the ledger (default: $RBR_REPO/RESULTS-LEDGER.md)
• RBR_EXTRA_ROOTS — colon-separated extra dirs: session scratchpads, a notes/memory dir, an artifact store. Set this — a large share of results land outside the repo.

<skills-dir> in the commands below is wherever this skill was installed: .claude/skills for a project copy, ~/.claude/skills for a user copy, or ${CLAUDE_PLUGIN_ROOT}/skills under the plugin.

Also substitute your expensive commands below, and your notes / memory tool in step 2.

The protocol (in order, no skipping)

Copy this checklist and track your progress:

- [ ] 1 Question named (corpus · size · config · metric · HEAD)
- [ ] 2 sweep.sh run + one notes/memory query
- [ ] 3 REUSE / EXTEND / RERUN picked
- [ ] 4 Verdict block written; user OK if past the expensive threshold
- [ ] 5 Ledger row appended when the run finishes

1. Name the question in one line: corpus + size + engine/config + metric + code state (a HEAD hash, or "any" if the question is about the past).
2. Sweep once — one Bash call, then one query to your notes / memory tool ("<keyword> result"):

   bash <skills-dir>/reuse-before-rerun/sweep.sh <keyword> [since YYYY-MM-DD] [until YYYY-MM-DD]

3. Pick one outcome:
   • REUSE — same corpus and size exists, and git log --oneline --since=<run date> -- <src> is empty, or the question is about the past. Report path, date, HEAD.
   • EXTEND — a subset, or a paused/partial run (shard files in a scratchpad), exists. Run only the delta: new rows, failing rows, or missing shards.
   • RERUN — nothing usable exists. Say which of the five places came up empty.
   • Commits since the run date do not make it RERUN by themselves. Report the last full number plus the measured gain of each shipped fix, and mark it "incremental, not re-measured".
   • Sweep reports rows not shown? Narrow the keyword or add a date window; return to Step 2.
4. Write the verdict block in every reply that reports a number or launches a run, REUSE included, even when nothing is launched, so the reader can see what was searched. Exactly these 3 lines, right after the numbers:

   Existing: <path · date · HEAD>  |  none — 0 hits for "<keyword>" in ledger, repo, scratchpads, notes, memory tool
   Verdict:  REUSE | EXTEND <delta rows> | RERUN because <none exists | corpus changed>
   Cost:     0 (reused)  |  <rows> rows · ~<minutes> min · <cores> cores

   A run past your expensive threshold (tens of thousands of rows, or an hour+) needs the user's OK this session unless pre-approved. Saying "you'll need a fresh run" also needs the verdict block first.
5. Register when it finishes — append one row to the ledger ($RBR_LEDGER) in the same turn. The next sweep then finds it in step 2.

Answering "where is the run from <date>"

Same command with the date window — sweep.sh <keyword> 2026-01-13 2026-01-15 (until is exclusive) — across all roots, not only the repo. "I can't find it" is not "it does not exist" until the ledger, the repo, every scratchpad, your notes and your memory tool all came up empty. Name each one you checked.

Ledger format

Create it once — newest first, append-only, one row per expensive run. Strict: keep date first as YYYY-MM-DD; sweep.sh reads only rows that start | 20.

| date | tag | engine | HEAD | rows | headline | path |
|---|---|---|---|---|---|---|
| 2026-01-14 | <corpus>-<size> | <engine/config> | <short hash> | 500 | <the numbers a reader needs> | <artifact path> |

tag = corpus + size, as your file names spell it; HEAD = the short commit measured (— if unknown); path = the artifact to open. Never delete a row — a missing row is how a run gets repeated.

Red flags — run the sweep first

• The command you are typing is a benchmark / census / measure / shard script, or names a large corpus file, and no verdict block is in the chat.
• You are about to write "needs a fresh measurement" or "not yet measured" with no verdict block.
• "Let me re-measure to be sure." / "A fresh spy is more trustworthy." / "Quick sanity run."
• You cannot name the prior result you looked at.
• The user said "again and again" or "you already have" this session.

Rationalization table