feat(agent_loop): add system-prompt guidance for relative time + multi-source asks

v0.18-α1-2 KRRA Conv diagnostic identified two prompt-shaped Qwen3.5-27B
failure patterns on conversational Korean queries:

  1. "올해 예산은?" / "내년도 일정은?" — agent injected literal year
     ("2024"/"2025") into deep_search, throwing away corpus docs from
     other years. Now told to search the topic first and only narrow by
     year if the topic search returns too many candidates AND there's
     evidence the user wants a specific year.

  2. "X 관련 자료 있어?" / "내용 보여줘" — agent terminated after 1
     deep_search returned 1-2 docs and answered "yes, here is one source".
     Now told that "자료/내용/정보" asks for multiple sources, do at
     least one paraphrase pass before concluding.

Both lines are general-purpose, not KRRA-specific, so they apply to any
Korean / English corpus.

The KRRA Conv -23pp regression vs v0.13 GPT-4o-mini (47% vs 70%) is
documented as a known issue: it stems from a recall ceiling on broad
topical queries (예산 / 인권 / 자료) where 5 ground-truth docs share a
vague topic word and deep_search top-K is dominated by phrase nodes.
Verified by direct corpus probe — 0/150 GT IDs reachable as direct node
ids; matching works only via properties.doc_id and depends on the right
docs landing in deep_search top-K. Real fix requires higher recall or
reranker-on-by-default; deferred to v0.19+ track.

45/45 agent + calibration tests pass.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

Author: SonAIengine

CHANGELOG.md

@@ -6,6 +6,26 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
 ## [Unreleased]
 
+### Changed — `agent_loop` system prompt: relative-time + multi-source guidance
+
+Two new tip lines in the agent prompt, learned from the v0.18-α1-2 KRRA
+Conv diagnostic:
+
+- **Relative time references** ("올해" / "내년도" / "this year"): the agent
+  should NOT inject a literal year number. The corpus may span multiple
+  years and a hard `2024` filter throws away matches. Search the topic
+  first, narrow by year only after evidence the user wants one.
+- **"X 관련 자료/내용/정보" type questions** ask for *multiple* sources.
+  The agent should not stop after the first ``deep_search`` returns 1-2
+  docs — at least one paraphrase pass before concluding.
+
+These guidance lines are general-purpose and apply to any corpus, not
+just KRRA Conv. The KRRA Conv −23pp regression itself is documented as
+known issue: it stems from a recall ceiling on broad topical queries
+where 5 GT docs share a vague topic word (예산 / 인권), not from agent
+reasoning. Real fix would require either a higher deep_search top-K cap
+or reranker-on-by-default for broad queries — both v0.19+ items.
+
 ## [0.17.2] - 2026-04-19
 
 Patch release bundling the license switch and the

src/synaptic/agent_loop.py

@@ -101,7 +101,17 @@
   often mix scripts (e.g. "치즈" → 0, "cheese" → many).
 - For "top N" / "most" / "least" questions, prefer ``aggregate_nodes``
   with ``metric=count`` and ``order=desc``.
-- Prefer concrete IDs over titles when chaining tools."""
+- Prefer concrete IDs over titles when chaining tools.
+- **Relative time references** ("올해" / "내년도" / "this year" / "next year"):
+  do NOT inject a specific year number into the search query — the corpus
+  may span multiple years, and a literal "2024" filter throws away all
+  other matches. Search the topic without the year first; only narrow by
+  year if the topic search returns too many candidates AND you have evidence
+  the user wants a specific year.
+- **"X 관련 자료 / 내용 / 정보" type questions** ask for *multiple* sources.
+  After the first ``deep_search`` returns a few hits, do at least one more
+  ``search`` with paraphrased keywords before concluding. A single document
+  is rarely the complete answer to such a request."""
 
 
 AGENT_TOOLS = [