recall-nudge: UserPromptSubmit hook that runs recall when a proper noun appears (Memory-First, mechanized) #65

目的 (Why)

Agents answer questions about a proper noun (a repo, service, person, host, project) from conversation memory alone, even when a prior session already recorded the fact. recall exists but is manual, and the Memory-First rule ("search memory before answering when a proper noun appears") lives only as prose in per-agent instruction files, so it is skipped under load.

Origin: the Memory-First §0.6 rule from an earlier private family design (2026-03, applied to 12 OpenClaw agents as prose, never mechanized). context-kit has no UserPromptSubmit hook today; all six pieces sit at PreToolUse / PostToolUse boundaries.

完了条件 (Done when)

• New hook hooks/recall-nudge.py (+ .sh wrapper) on UserPromptSubmit: detects proper-noun candidates in the prompt (configurable regex set; default covers repo slugs owner/name, #NNN issue refs, CamelCase / ALLCAPS tokens of 3+ chars, katakana runs) and, when at least one is found and recall is on PATH, runs recall --local-only for the top N candidates with a hard timeout.
• Returns hookSpecificOutput.additionalContext with at most CK_RECALL_NUDGE_MAX_HITS (default 5) hits, each with provenance path, or a one-line "no local hits for: ..." so the agent knows the search ran. Never blocks the prompt; any failure degrades to no output.
• Opt-in via an examples/settings.json block (not default-on); env knobs CK_RECALL_NUDGE_TIMEOUT_MS (default 1500), CK_RECALL_NUDGE_MAX_TERMS (default 3), CK_RECALL_NUDGE_PATTERNS (override).
• Tests: detector unit tests (JA/EN samples, false-positive guard for common words), timeout path, recall-missing path, output cap.
• Docs: docs/recall.md section + README pieces table row; measured latency on the implementer's machine recorded in the PR.
• テスト / lint green

触るファイル / モジュール予測

• hooks/recall-nudge.py
• hooks/recall-nudge.sh
• tests/test_recall_nudge.py
• examples/settings.json
• docs/recall.md
• README.md (pieces table, 1 row)

想定スコープ

• やること: detection + bounded recall invocation + context injection, opt-in.
• やらないこと: changing recall itself; Supermemory / Meili layers (local-only by default to keep latency bounded); auto-writing anything to memory; any host other than Claude Code hooks.

前提条件 (Blocked by)

None. Size: S.