luca-history

luca-history

Kai forgets things, and forgetting is expensive. When she asks a question about something she thinks happened, run luca.history before anything else. It is cheap for the machine to sweep a week across eight synonyms. If it finds the thing, it saves her ~10 minutes of re-litigating a settled question.

When this fires

Any "did we already do this" shaped question:

• "did we land X" / "didn't we already ship X"
• "I thought we did this" / "we just did this"
• "what happened with X" / "where did X end up"

Default to luca.history here. Do not reach for git log or ls first. If the work landed cleanly, git would show it and Kai would not be asking. If she is asking, the work is probably half-done somewhere git cannot see - an unfinished session, a GitHub issue, a Claude project chat.

Procedure

Step 1: expand the query into ~8 synonyms

LUCA does no LLM calls - it can only strip dates and labels and split tokens mechanically. You can do better. Before calling, expand Kai's phrasing into roughly 8 synonyms and variants: the subject in her words, common rewordings, the likely repo or tool name, kebab and snake and dotted forms. Strip dates, times, issue numbers, and bare labels - those are noise, not the subject.

Pass the synonyms as terms. The bare query still goes in for the record.

Step 2: verify repo-recall is reachable

curl -fsS http://127.0.0.1:7777/api/scan-version >/dev/null 2>&1 || echo "repo-recall down"

If down, suggest brew services start coilysiren/tap/repo-recall and stop.

Step 3: run the sweep

Via MCP (preferred - LUCA is lazy-loaded through mcporter):

mcporter call luca.history query="did we land the history endpoint" \
  terms='["luca.history","history endpoint","greedy sweep","did we land","historical lookup","query expansion","history sweep","day walk"]'

Via CLI fallback (--term is repeatable):

coily exec history -- "did we land the history endpoint" \
  --term "luca.history" --term "history endpoint" --term "greedy sweep"

max_days (default 14) is the day-walk ceiling. LUCA searches a 1-day session window first, then widens a day at a time until it finds a hit or hits the ceiling.

Step 4: reason over the result

LUCA returns matching sessions (with a landed-vs-unfinished read) and a wider search sweep over repos, issues, and commits. Build on it: name the actor where derivable (Claude, Kai, a collaborator), say whether the thing looks landed or half-done, and point Kai at the session uuid / issue / commit. If LUCA found nothing in 14 days, say so plainly - the work may predate repo-recall's index, or live in a Claude project chat it cannot see.

v1 / v2

v1 sweeps last_prompt, the recent-sessions table, and recall_search over repos, sessions, and commits. v2 gets much richer once repo-recall#229's full per-turn session text index lands - then the greedy search hits real prompt, output, and thinking text instead of last_prompt as a proxy.

Sibling

luca.history answers "did we already do this". The separate "what is in flight, what is hanging" endpoint (luca#107) is a sibling capability - reach for that when Kai wants a status sweep rather than a did-we-do-it lookup.