thkt/use-cli-recall

use-cli-recall

Triggers (call without deliberation)

| Trigger | Signal | | -------------------- | ------------------------------------------ | | Temporal reference | 「前に」「あの時」 past events / decisions | | Structural echo | Current problem mirrors a past situation | | Repetition | 「また同じ」 recurring mistake | | Vague back-reference | 「あの件」 past work without specifics | | Module first contact | First edit to a file/module this session |

Commands

| Purpose | Command | | ----------------- | ----------------------------------------------------------- | | Search | recall search "query" (or recall "query" as shorthand) | | Last N days | recall search "query" --days N | | Project filter | recall search "query" --project <path> | | Source filter | recall search "query" --source claude or --source codex | | Limit results | recall search "query" --limit N (default 10, max 100) | | Show session | recall show <session-id> | | Status | recall status | | Incremental index | recall index | | Full rebuild | recall index --force |

Query composition

Write bilingual queries upfront (e.g. recall "認証 auth"). FTS5 trigram tokenization cannot match JA terms of 2 chars or fewer (認証/依存 hit 0), and embeddings do not bridge EN⇄JA (thkt/recall#51). Including both languages covers each search path.

Weak-result retry

recall does not expand queries (caller-is-LLM, thkt/recall#25). Hybrid search returns nearest neighbors, so a poor query yields low-relevance results rather than 0 hits. When results are empty or low-relevance, rewrite the query yourself and retry once: synonyms, EN⇄JA variants, related concept terms.

Pairing with code search

Past decisions (use-cli-recall) vs current code state (ugrep / bfs). Run both on these triggers.

| Trigger | recall query | code search | | -------------------- | ----------------------------- | ---------------------------- | | Module first contact | module name, design rationale | module name, key identifiers | | Structural echo | past similar problem | current similar code |