런타임 계약과 학습자 흐름 보강

apps/api/app/services/memory.py

@@ -37,6 +37,21 @@ _COUNSELING_AGREEMENT_WITHDRAWAL_RE = re.compile(
     r"더\s*이상.*(상담|회기).*(안\s*하|하지\s*않|못\s*하)"
 )
 _DIGEST_SPEAKERS = {"counselor", "client"}
+_LLM_DIGEST_MIN_CHARS = 60
+_LLM_DIGEST_INTERNAL_MARKERS = (
+    "rapport_credit",
+    "effective_openness",
+    "end_state",
+    "evaluation",
+    "정답",
+    "점수",
+    "평가 payload",
+    "평가점수",
+    "평가 점수",
+    "core_belief",
+    "CCD",
+)
+_SESSION_DIGEST_PREFIX_RE = re.compile(r"^S(?P<session_no>\d+):")
 
 
 # ════════════════════════════════════════════════════════════════════════════
@@ -139,7 +154,7 @@ class SessionDigestInput:
 
 @dataclass(frozen=True, slots=True)
 class SessionDigestResult:
-    """Digest writer output contract shared by fallback and future LLM worker."""
+    """Digest writer output contract shared by fallback and LLM worker paths."""
 
     session_id: str
     case_id: str | None
@@ -149,6 +164,35 @@ class SessionDigestResult:
     source: Literal["fallback", "llm"] = "fallback"
 
 
+@dataclass(frozen=True, slots=True)
+class DigestQualityAssessment:
+    """Local quality gate result before an LLM digest can replace fallback."""
+
+    accepted: bool
+    reason: Literal[
+        "ok",
+        "empty",
+        "too_short",
+        "forbidden_substring",
+        "internal_marker",
+        "wrong_session_prefix",
+    ]
+    retryable: bool = False
+    details: tuple[str, ...] = ()
+
+
+@dataclass(frozen=True, slots=True)
+class SessionDigestWorkerOutcome:
+    """Accepted LLM digest result or the reason fallback must remain authoritative."""
+
+    result: SessionDigestResult | None
+    quality: DigestQualityAssessment
+
+    @property
+    def fallback_required(self) -> bool:
+        return self.result is None
+
+
 @dataclass(slots=True)
 class CompressionJob:
     """회기종료 narrative 압축 작업(LLM, 비동기 비블로킹). 큐에 적재될 페이로드.
@@ -305,6 +349,101 @@ def build_compression_messages(job: CompressionJob) -> list[dict[str, str]]:
     return [{"role": "system", "content": system}, {"role": "user", "content": user}]
 
 
+def _normalize_digest_text(value: Any) -> str:
+    return " ".join(str(value or "").split())
+
+
+def assess_llm_digest_quality(
+    digest_input: SessionDigestInput,
+    digest: Any,
+    *,
+    forbidden_substrings: tuple[str, ...] = (),
+    min_chars: int = _LLM_DIGEST_MIN_CHARS,
+) -> DigestQualityAssessment:
+    """Validate an LLM digest candidate without using raw text or end-state data."""
+
+    text = _normalize_digest_text(digest)
+    if not text:
+        return DigestQualityAssessment(accepted=False, reason="empty", retryable=True)
+
+    prefix_match = _SESSION_DIGEST_PREFIX_RE.match(text)
+    if prefix_match and int(prefix_match.group("session_no")) != digest_input.session_no:
+        return DigestQualityAssessment(
+            accepted=False,
+            reason="wrong_session_prefix",
+            retryable=True,
+            details=(prefix_match.group(0),),
+        )
+
+    forbidden_hits = tuple(
+        marker
+        for marker in (str(item).strip() for item in forbidden_substrings)
+        if marker and marker in text
+    )
+    if forbidden_hits:
+        return DigestQualityAssessment(
+            accepted=False,
+            reason="forbidden_substring",
+            retryable=True,
+            details=forbidden_hits[:5],
+        )
+
+    lowered = text.lower()
+    marker_hits = tuple(
+        marker
+        for marker in _LLM_DIGEST_INTERNAL_MARKERS
+        if marker.lower() in lowered
+    )
+    if marker_hits:
+        return DigestQualityAssessment(
+            accepted=False,
+            reason="internal_marker",
+            retryable=True,
+            details=marker_hits,
+        )
+
+    if len(text) < max(1, int(min_chars)):
+        return DigestQualityAssessment(accepted=False, reason="too_short", retryable=True)
+
+    return DigestQualityAssessment(accepted=True, reason="ok")
+
+
+def build_llm_digest_worker_outcome(
+    digest_input: SessionDigestInput,
+    digest: Any,
+    *,
+    forbidden_substrings: tuple[str, ...] = (),
+    min_chars: int = _LLM_DIGEST_MIN_CHARS,
+) -> SessionDigestWorkerOutcome:
+    """Coerce a candidate LLM digest into the shared result contract if it passes."""
+
+    quality = assess_llm_digest_quality(
+        digest_input,
+        digest,
+        forbidden_substrings=forbidden_substrings,
+        min_chars=min_chars,
+    )
+    if not quality.accepted:
+        return SessionDigestWorkerOutcome(result=None, quality=quality)
+
+    normalized = _normalize_digest_text(digest)
+    prefix = f"S{digest_input.session_no}:"
+    if not normalized.startswith(prefix):
+        normalized = f"{prefix} {normalized}"
+
+    return SessionDigestWorkerOutcome(
+        result=SessionDigestResult(
+            session_id=digest_input.session_id,
+            case_id=digest_input.case_id,
+            session_no=digest_input.session_no,
+            digest=normalized,
+            open_threads=digest_input.open_threads,
+            source="llm",
+        ),
+        quality=quality,
+    )
+
+
 def _compact(value: Any, *, limit: int = _SESSION_DIGEST_EXCERPT_CHARS) -> str:
     text = " ".join(str(value or "").split())
     if len(text) <= limit:
@@ -518,9 +657,13 @@ __all__ = [
     "MaskedDigestTurn",
     "SessionDigestInput",
     "SessionDigestResult",
+    "DigestQualityAssessment",
+    "SessionDigestWorkerOutcome",
     "make_carry_over",
     "build_session_digest_input",
     "build_compression_messages",
+    "assess_llm_digest_quality",
+    "build_llm_digest_worker_outcome",
     "build_fallback_digest_result",
     "build_fallback_session_digest",
     "PinnedFactCandidate",