feat: normalize durable fact extraction (#1012)

Closes #1012

- add structured session fact extraction with provenance, temporal metadata,
  canonical normalization, and contradiction grouping
- persist structured metadata into holographic memory auto-extraction with
  canonical-key dedupe across repeated ingests
- add fixture-backed transcript tests plus extraction quality evaluation

agent/account_usage.py

@@ -1,326 +0,0 @@
-from __future__ import annotations
-
-from dataclasses import dataclass
-from datetime import datetime, timezone
-from typing import Any, Optional
-
-import httpx
-
-from agent.anthropic_adapter import _is_oauth_token, resolve_anthropic_token
-from hermes_cli.auth import _read_codex_tokens, resolve_codex_runtime_credentials
-from hermes_cli.runtime_provider import resolve_runtime_provider
-
-
-def _utc_now() -> datetime:
-    return datetime.now(timezone.utc)
-
-
-@dataclass(frozen=True)
-class AccountUsageWindow:
-    label: str
-    used_percent: Optional[float] = None
-    reset_at: Optional[datetime] = None
-    detail: Optional[str] = None
-
-
-@dataclass(frozen=True)
-class AccountUsageSnapshot:
-    provider: str
-    source: str
-    fetched_at: datetime
-    title: str = "Account limits"
-    plan: Optional[str] = None
-    windows: tuple[AccountUsageWindow, ...] = ()
-    details: tuple[str, ...] = ()
-    unavailable_reason: Optional[str] = None
-
-    @property
-    def available(self) -> bool:
-        return bool(self.windows or self.details) and not self.unavailable_reason
-
-
-def _title_case_slug(value: Optional[str]) -> Optional[str]:
-    cleaned = str(value or "").strip()
-    if not cleaned:
-        return None
-    return cleaned.replace("_", " ").replace("-", " ").title()
-
-
-def _parse_dt(value: Any) -> Optional[datetime]:
-    if value in (None, ""):
-        return None
-    if isinstance(value, (int, float)):
-        return datetime.fromtimestamp(float(value), tz=timezone.utc)
-    if isinstance(value, str):
-        text = value.strip()
-        if not text:
-            return None
-        if text.endswith("Z"):
-            text = text[:-1] + "+00:00"
-        try:
-            dt = datetime.fromisoformat(text)
-            return dt if dt.tzinfo else dt.replace(tzinfo=timezone.utc)
-        except ValueError:
-            return None
-    return None
-
-
-def _format_reset(dt: Optional[datetime]) -> str:
-    if not dt:
-        return "unknown"
-    local_dt = dt.astimezone()
-    delta = dt - _utc_now()
-    total_seconds = int(delta.total_seconds())
-    if total_seconds <= 0:
-        return f"now ({local_dt.strftime('%Y-%m-%d %H:%M %Z')})"
-    hours, rem = divmod(total_seconds, 3600)
-    minutes = rem // 60
-    if hours >= 24:
-        days, hours = divmod(hours, 24)
-        rel = f"in {days}d {hours}h"
-    elif hours > 0:
-        rel = f"in {hours}h {minutes}m"
-    else:
-        rel = f"in {minutes}m"
-    return f"{rel} ({local_dt.strftime('%Y-%m-%d %H:%M %Z')})"
-
-
-def render_account_usage_lines(snapshot: Optional[AccountUsageSnapshot], *, markdown: bool = False) -> list[str]:
-    if not snapshot:
-        return []
-    header = f"📈 {'**' if markdown else ''}{snapshot.title}{'**' if markdown else ''}"
-    lines = [header]
-    if snapshot.plan:
-        lines.append(f"Provider: {snapshot.provider} ({snapshot.plan})")
-    else:
-        lines.append(f"Provider: {snapshot.provider}")
-    for window in snapshot.windows:
-        if window.used_percent is None:
-            base = f"{window.label}: unavailable"
-        else:
-            remaining = max(0, round(100 - float(window.used_percent)))
-            used = max(0, round(float(window.used_percent)))
-            base = f"{window.label}: {remaining}% remaining ({used}% used)"
-        if window.reset_at:
-            base += f" • resets {_format_reset(window.reset_at)}"
-        elif window.detail:
-            base += f" • {window.detail}"
-        lines.append(base)
-    for detail in snapshot.details:
-        lines.append(detail)
-    if snapshot.unavailable_reason:
-        lines.append(f"Unavailable: {snapshot.unavailable_reason}")
-    return lines
-
-
-def _resolve_codex_usage_url(base_url: str) -> str:
-    normalized = (base_url or "").strip().rstrip("/")
-    if not normalized:
-        normalized = "https://chatgpt.com/backend-api/codex"
-    if normalized.endswith("/codex"):
-        normalized = normalized[: -len("/codex")]
-    if "/backend-api" in normalized:
-        return normalized + "/wham/usage"
-    return normalized + "/api/codex/usage"
-
-
-def _fetch_codex_account_usage() -> Optional[AccountUsageSnapshot]:
-    creds = resolve_codex_runtime_credentials(refresh_if_expiring=True)
-    token_data = _read_codex_tokens()
-    tokens = token_data.get("tokens") or {}
-    account_id = str(tokens.get("account_id", "") or "").strip() or None
-    headers = {
-        "Authorization": f"Bearer {creds['api_key']}",
-        "Accept": "application/json",
-        "User-Agent": "codex-cli",
-    }
-    if account_id:
-        headers["ChatGPT-Account-Id"] = account_id
-    with httpx.Client(timeout=15.0) as client:
-        response = client.get(_resolve_codex_usage_url(creds.get("base_url", "")), headers=headers)
-        response.raise_for_status()
-    payload = response.json() or {}
-    rate_limit = payload.get("rate_limit") or {}
-    windows: list[AccountUsageWindow] = []
-    for key, label in (("primary_window", "Session"), ("secondary_window", "Weekly")):
-        window = rate_limit.get(key) or {}
-        used = window.get("used_percent")
-        if used is None:
-            continue
-        windows.append(
-            AccountUsageWindow(
-                label=label,
-                used_percent=float(used),
-                reset_at=_parse_dt(window.get("reset_at")),
-            )
-        )
-    details: list[str] = []
-    credits = payload.get("credits") or {}
-    if credits.get("has_credits"):
-        balance = credits.get("balance")
-        if isinstance(balance, (int, float)):
-            details.append(f"Credits balance: ${float(balance):.2f}")
-        elif credits.get("unlimited"):
-            details.append("Credits balance: unlimited")
-    return AccountUsageSnapshot(
-        provider="openai-codex",
-        source="usage_api",
-        fetched_at=_utc_now(),
-        plan=_title_case_slug(payload.get("plan_type")),
-        windows=tuple(windows),
-        details=tuple(details),
-    )
-
-
-def _fetch_anthropic_account_usage() -> Optional[AccountUsageSnapshot]:
-    token = (resolve_anthropic_token() or "").strip()
-    if not token:
-        return None
-    if not _is_oauth_token(token):
-        return AccountUsageSnapshot(
-            provider="anthropic",
-            source="oauth_usage_api",
-            fetched_at=_utc_now(),
-            unavailable_reason="Anthropic account limits are only available for OAuth-backed Claude accounts.",
-        )
-    headers = {
-        "Authorization": f"Bearer {token}",
-        "Accept": "application/json",
-        "Content-Type": "application/json",
-        "anthropic-beta": "oauth-2025-04-20",
-        "User-Agent": "claude-code/2.1.0",
-    }
-    with httpx.Client(timeout=15.0) as client:
-        response = client.get("https://api.anthropic.com/api/oauth/usage", headers=headers)
-        response.raise_for_status()
-    payload = response.json() or {}
-    windows: list[AccountUsageWindow] = []
-    mapping = (
-        ("five_hour", "Current session"),
-        ("seven_day", "Current week"),
-        ("seven_day_opus", "Opus week"),
-        ("seven_day_sonnet", "Sonnet week"),
-    )
-    for key, label in mapping:
-        window = payload.get(key) or {}
-        util = window.get("utilization")
-        if util is None:
-            continue
-        used = float(util) * 100 if float(util) <= 1 else float(util)
-        windows.append(
-            AccountUsageWindow(
-                label=label,
-                used_percent=used,
-                reset_at=_parse_dt(window.get("resets_at")),
-            )
-        )
-    details: list[str] = []
-    extra = payload.get("extra_usage") or {}
-    if extra.get("is_enabled"):
-        used_credits = extra.get("used_credits")
-        monthly_limit = extra.get("monthly_limit")
-        currency = extra.get("currency") or "USD"
-        if isinstance(used_credits, (int, float)) and isinstance(monthly_limit, (int, float)):
-            details.append(
-                f"Extra usage: {used_credits:.2f} / {monthly_limit:.2f} {currency}"
-            )
-    return AccountUsageSnapshot(
-        provider="anthropic",
-        source="oauth_usage_api",
-        fetched_at=_utc_now(),
-        windows=tuple(windows),
-        details=tuple(details),
-    )
-
-
-def _fetch_openrouter_account_usage(base_url: Optional[str], api_key: Optional[str]) -> Optional[AccountUsageSnapshot]:
-    runtime = resolve_runtime_provider(
-        requested="openrouter",
-        explicit_base_url=base_url,
-        explicit_api_key=api_key,
-    )
-    token = str(runtime.get("api_key", "") or "").strip()
-    if not token:
-        return None
-    normalized = str(runtime.get("base_url", "") or "").rstrip("/")
-    credits_url = f"{normalized}/credits"
-    key_url = f"{normalized}/key"
-    headers = {
-        "Authorization": f"Bearer {token}",
-        "Accept": "application/json",
-    }
-    with httpx.Client(timeout=10.0) as client:
-        credits_resp = client.get(credits_url, headers=headers)
-        credits_resp.raise_for_status()
-        credits = (credits_resp.json() or {}).get("data") or {}
-        try:
-            key_resp = client.get(key_url, headers=headers)
-            key_resp.raise_for_status()
-            key_data = (key_resp.json() or {}).get("data") or {}
-        except Exception:
-            key_data = {}
-    total_credits = float(credits.get("total_credits") or 0.0)
-    total_usage = float(credits.get("total_usage") or 0.0)
-    details = [f"Credits balance: ${max(0.0, total_credits - total_usage):.2f}"]
-    windows: list[AccountUsageWindow] = []
-    limit = key_data.get("limit")
-    limit_remaining = key_data.get("limit_remaining")
-    limit_reset = str(key_data.get("limit_reset") or "").strip()
-    usage = key_data.get("usage")
-    if (
-        isinstance(limit, (int, float))
-        and float(limit) > 0
-        and isinstance(limit_remaining, (int, float))
-        and 0 <= float(limit_remaining) <= float(limit)
-    ):
-        limit_value = float(limit)
-        remaining_value = float(limit_remaining)
-        used_percent = ((limit_value - remaining_value) / limit_value) * 100
-        detail_parts = [f"${remaining_value:.2f} of ${limit_value:.2f} remaining"]
-        if limit_reset:
-            detail_parts.append(f"resets {limit_reset}")
-        windows.append(
-            AccountUsageWindow(
-                label="API key quota",
-                used_percent=used_percent,
-                detail=" • ".join(detail_parts),
-            )
-        )
-    if isinstance(usage, (int, float)):
-        usage_parts = [f"API key usage: ${float(usage):.2f} total"]
-        for value, label in (
-            (key_data.get("usage_daily"), "today"),
-            (key_data.get("usage_weekly"), "this week"),
-            (key_data.get("usage_monthly"), "this month"),
-        ):
-            if isinstance(value, (int, float)) and float(value) > 0:
-                usage_parts.append(f"${float(value):.2f} {label}")
-        details.append(" • ".join(usage_parts))
-    return AccountUsageSnapshot(
-        provider="openrouter",
-        source="credits_api",
-        fetched_at=_utc_now(),
-        windows=tuple(windows),
-        details=tuple(details),
-    )
-
-
-def fetch_account_usage(
-    provider: Optional[str],
-    *,
-    base_url: Optional[str] = None,
-    api_key: Optional[str] = None,
-) -> Optional[AccountUsageSnapshot]:
-    normalized = str(provider or "").strip().lower()
-    if normalized in {"", "auto", "custom"}:
-        return None
-    try:
-        if normalized == "openai-codex":
-            return _fetch_codex_account_usage()
-        if normalized == "anthropic":
-            return _fetch_anthropic_account_usage()
-        if normalized == "openrouter":
-            return _fetch_openrouter_account_usage(base_url, api_key)
-    except Exception:
-        return None
-    return None

agent/session_compactor.py

@@ -1,197 +1,546 @@
-"""Session compaction with fact extraction.
+"""Session compaction with structured fact extraction.
 
-Before compressing conversation context, extracts durable facts
-(user preferences, corrections, project details) and saves them
-to the fact store so they survive compression.
-
-Usage:
-    from agent.session_compactor import extract_and_save_facts
-    facts = extract_and_save_facts(messages)
+Before compressing conversation context, extract durable facts with enough
+structure to survive retrieval: source/provenance, temporal anchors,
+normalized canonical keys, and contradiction groups.
 """
 
 from __future__ import annotations
 
-import json
 import logging
 import re
 import time
 from dataclasses import dataclass, field
-from typing import Any, Dict, List, Optional, Tuple
+from datetime import datetime, timezone
+from typing import Any, Dict, List, Tuple
 
 logger = logging.getLogger(__name__)
 
 
+_DEPLOY_METHOD_RE = re.compile(r"\bdeploy(?:ing)?\s+(?:via|through|with)\s+([A-Za-z0-9_./+-]+)", re.IGNORECASE)
+_WATCHDOG_CAP_RE = re.compile(
+    r"\b(?:the\s+)?([A-Za-z0-9_-]+(?:\s+watchdog)?)\s+(?:caps|limits)\s+dispatches(?:\s+per\s+cycle)?\s+to\s+([0-9]+)",
+    re.IGNORECASE,
+)
+_PROVIDER_RE = re.compile(
+    r"\bprovider\s+(?:is|should\s+stay|should\s+be|needs\s+to\s+be)\s+([A-Za-z0-9._/-]+)",
+    re.IGNORECASE,
+)
+_MODEL_RE = re.compile(
+    r"\bmodel\s+(?:is|should\s+stay|should\s+be|needs\s+to\s+be)\s+([A-Za-z0-9._:/-]+)",
+    re.IGNORECASE,
+)
+_PORT_RE = re.compile(r"\bport\s+(?:is|should\s+be)\s+([0-9]+)", re.IGNORECASE)
+_PROJECT_USES_RE = re.compile(r"\b(?:the\s+)?project\s+(?:uses|needs|requires)\s+(.+?)(?:[.!?]|$)", re.IGNORECASE)
+_PREFERENCE_RE = re.compile(r"\bI\s+(?:prefer|like|want|need)\s+(.+?)(?:[.!?]|$)", re.IGNORECASE)
+_CONSTRAINT_RE = re.compile(r"\b(?:do\s+not|don't)\s+(?:ever\s+|again\s+)?(.+?)(?:[.!?]|$)", re.IGNORECASE)
+_DECISION_RE = re.compile(r"\b(?:we|the\s+team)\s+(?:decided|agreed|chose)\s+(?:to\s+)?(.+?)(?:[.!?]|$)", re.IGNORECASE)
+
+
 @dataclass
 class ExtractedFact:
-    """A fact extracted from conversation."""
-    category: str       # "user_pref", "correction", "project", "tool_quirk", "general"
-    entity: str         # what the fact is about
-    content: str        # the fact itself
-    confidence: float   # 0.0-1.0
-    source_turn: int    # which message turn it came from
+    """A durable fact extracted from conversation."""
+
+    category: str
+    entity: str
+    content: str
+    confidence: float
+    source_turn: int
     timestamp: float = 0.0
+    source_role: str = "user"
+    source_text: str = ""
+    normalized_content: str = ""
+    canonical_key: str = ""
+    relation: str = "general"
+    contradiction_group: str = ""
+    status: str = "active"
+    provenance: str = ""
+    observed_at: str = ""
+    evidence: List[Dict[str, Any]] = field(default_factory=list)
+    metadata: Dict[str, Any] = field(default_factory=dict)
 
-
-# Patterns that indicate user preferences
-_PREFERENCE_PATTERNS = [
-    (r"(?:I|we) (?:prefer|like|want|need) (.+?)(?:\.|$)", "preference"),
-    (r"(?:always|never) (?:use|do|run|deploy) (.+?)(?:\.|$)", "preference"),
-    (r"(?:my|our) (?:default|preferred|usual) (.+?) (?:is|are) (.+?)(?:\.|$)", "preference"),
-    (r"(?:make sure|ensure|remember) (?:to|that) (.+?)(?:\.|$)", "instruction"),
-    (r"(?:don'?t|do not) (?:ever|ever again) (.+?)(?:\.|$)", "constraint"),
-]
-
-# Patterns that indicate corrections
-_CORRECTION_PATTERNS = [
-    (r"(?:actually|no[, ]|wait[, ]|correction[: ]|sorry[, ]) (.+)", "correction"),
-    (r"(?:I meant|what I meant was|the correct) (.+?)(?:\.|$)", "correction"),
-    (r"(?:it'?s|its) (?:not|shouldn'?t be|wrong) (.+?)(?:\.|$)", "correction"),
-]
-
-# Patterns that indicate project/tool facts
-_PROJECT_PATTERNS = [
-    (r"(?:the |our )?(?:project|repo|codebase|code) (?:is|uses|needs|requires) (.+?)(?:\.|$)", "project"),
-    (r"(?:deploy|push|commit) (?:to|on) (.+?)(?:\.|$)", "project"),
-    (r"(?:this|that|the) (?:server|host|machine|VPS) (?:is|runs|has) (.+?)(?:\.|$)", "infrastructure"),
-    (r"(?:model|provider|engine) (?:is|should be|needs to be) (.+?)(?:\.|$)", "config"),
-]
+    def __post_init__(self) -> None:
+        if not self.timestamp:
+            self.timestamp = time.time()
+        if not self.observed_at:
+            self.observed_at = _iso_from_timestamp(self.timestamp)
+        if not self.normalized_content:
+            self.normalized_content = _normalize_value(self.content)
+        if not self.provenance:
+            self.provenance = f"conversation:{self.source_role}:{self.source_turn}"
+        if not self.canonical_key:
+            self.canonical_key = _canonical_key(self.entity, self.relation, self.normalized_content)
+        if not self.evidence:
+            self.evidence = [
+                {
+                    "source_role": self.source_role,
+                    "source_turn": self.source_turn,
+                    "source_text": self.source_text or self.content,
+                    "observed_at": self.observed_at,
+                    "provenance": self.provenance,
+                }
+            ]
+        self.metadata = dict(self.metadata or {})
+        self.metadata.setdefault("entity", self.entity)
+        self.metadata.setdefault("relation", self.relation)
+        self.metadata.setdefault("value", self.content)
+        self.metadata.setdefault("normalized_value", self.normalized_content)
+        self.metadata.setdefault("provenance", [self.provenance])
+        self.metadata.setdefault("evidence", list(self.evidence))
+        self.metadata.setdefault("observation_count", len(self.evidence))
+        self.metadata.setdefault("duplicate_count", max(0, self.metadata["observation_count"] - 1))
+        if self.contradiction_group:
+            self.metadata.setdefault("contradiction_group", self.contradiction_group)
+        self.metadata.setdefault("status", self.status)
 
 
 def extract_facts_from_messages(messages: List[Dict[str, Any]]) -> List[ExtractedFact]:
     """Extract durable facts from conversation messages.
 
-    Scans user messages for preferences, corrections, project facts,
-    and infrastructure details that should survive compression.
+    Scans conversation turns for preferences, decisions, corrections, and
+    operational state. Raw candidates are normalized into canonical facts so
+    near-duplicates merge and contradictions remain inspectable.
     """
-    facts = []
-    seen_contents = set()
 
+    raw_candidates: list[ExtractedFact] = []
     for turn_idx, msg in enumerate(messages):
         role = msg.get("role", "")
         content = msg.get("content", "")
-
-        # Only scan user messages and assistant responses with corrections
-        if role not in ("user", "assistant"):
+        if role not in {"user", "assistant"}:
             continue
-        if not content or not isinstance(content, str):
-            continue
-        if len(content) < 10:
-            continue
-
-        # Skip tool results and system messages
         if role == "assistant" and msg.get("tool_calls"):
             continue
+        if not isinstance(content, str) or len(content.strip()) < 10:
+            continue
 
-        extracted = _extract_from_text(content, turn_idx, role)
+        timestamp, observed_at = _message_time(msg)
+        raw_candidates.extend(
+            _extract_from_text(
+                content.strip(),
+                turn_idx=turn_idx,
+                role=role,
+                timestamp=timestamp,
+                observed_at=observed_at,
+            )
+        )
 
-        # Deduplicate by content
-        for fact in extracted:
-            key = f"{fact.category}:{fact.content[:100]}"
-            if key not in seen_contents:
-                seen_contents.add(key)
-                facts.append(fact)
+    return _normalize_candidates(raw_candidates)
+
+
+def evaluate_extraction_quality(messages: List[Dict[str, Any]]) -> Dict[str, Any]:
+    """Return before/after metrics for raw vs normalized extraction quality."""
+
+    raw_candidates: list[ExtractedFact] = []
+    for turn_idx, msg in enumerate(messages):
+        role = msg.get("role", "")
+        content = msg.get("content", "")
+        if role not in {"user", "assistant"}:
+            continue
+        if role == "assistant" and msg.get("tool_calls"):
+            continue
+        if not isinstance(content, str) or len(content.strip()) < 10:
+            continue
+        timestamp, observed_at = _message_time(msg)
+        raw_candidates.extend(
+            _extract_from_text(
+                content.strip(),
+                turn_idx=turn_idx,
+                role=role,
+                timestamp=timestamp,
+                observed_at=observed_at,
+            )
+        )
+
+    normalized = _normalize_candidates(raw_candidates)
+    raw_count = len(raw_candidates)
+    normalized_count = len(normalized)
+    contradiction_groups = {
+        fact.contradiction_group
+        for fact in normalized
+        if fact.status == "contradiction" and fact.contradiction_group
+    }
+    duplicate_count = max(0, raw_count - normalized_count)
+    noise_reduction = (duplicate_count / raw_count) if raw_count else 0.0
+
+    return {
+        "raw_candidates": raw_count,
+        "normalized_facts": normalized_count,
+        "duplicates_merged": duplicate_count,
+        "contradiction_groups": len(contradiction_groups),
+        "noise_reduction": round(noise_reduction, 3),
+    }
+
+
+def _extract_from_text(
+    text: str,
+    *,
+    turn_idx: int,
+    role: str,
+    timestamp: float,
+    observed_at: str,
+) -> List[ExtractedFact]:
+    """Extract raw fact candidates from a single text block."""
+
+    facts: list[ExtractedFact] = []
+    if role != "user":
+        return facts
+
+    deploy_match = _DEPLOY_METHOD_RE.search(text)
+    if deploy_match:
+        method = deploy_match.group(1).strip()
+        facts.append(
+            _build_fact(
+                category="project.decision",
+                entity="project",
+                relation="workflow.deploy_method",
+                value=method,
+                content=f"Deploy via {method}",
+                confidence=0.88,
+                source_turn=turn_idx,
+                source_role=role,
+                source_text=text,
+                timestamp=timestamp,
+                observed_at=observed_at,
+                unique_slot=True,
+            )
+        )
+
+    watchdog_match = _WATCHDOG_CAP_RE.search(text)
+    if watchdog_match:
+        watchdog = watchdog_match.group(1).strip()
+        cap = watchdog_match.group(2).strip()
+        facts.append(
+            _build_fact(
+                category="project.operational",
+                entity=_normalize_entity(watchdog),
+                relation="fleet.dispatch_cap",
+                value=cap,
+                content=f"{watchdog} caps dispatches per cycle to {cap}",
+                confidence=0.92,
+                source_turn=turn_idx,
+                source_role=role,
+                source_text=text,
+                timestamp=timestamp,
+                observed_at=observed_at,
+                unique_slot=True,
+            )
+        )
+
+    provider_match = _PROVIDER_RE.search(text)
+    if provider_match:
+        provider = provider_match.group(1).strip()
+        facts.append(
+            _build_fact(
+                category="project.config",
+                entity="project",
+                relation="config.provider",
+                value=provider,
+                content=f"Provider should stay {provider}",
+                confidence=0.91,
+                source_turn=turn_idx,
+                source_role=role,
+                source_text=text,
+                timestamp=timestamp,
+                observed_at=observed_at,
+                unique_slot=True,
+            )
+        )
+
+    model_match = _MODEL_RE.search(text)
+    if model_match:
+        model = model_match.group(1).strip()
+        facts.append(
+            _build_fact(
+                category="project.config",
+                entity="project",
+                relation="config.model",
+                value=model,
+                content=f"Model should stay {model}",
+                confidence=0.9,
+                source_turn=turn_idx,
+                source_role=role,
+                source_text=text,
+                timestamp=timestamp,
+                observed_at=observed_at,
+                unique_slot=True,
+            )
+        )
+
+    port_match = _PORT_RE.search(text)
+    if port_match:
+        port = port_match.group(1).strip()
+        facts.append(
+            _build_fact(
+                category="project.config",
+                entity="project",
+                relation="config.port",
+                value=port,
+                content=f"Port is {port}",
+                confidence=0.9,
+                source_turn=turn_idx,
+                source_role=role,
+                source_text=text,
+                timestamp=timestamp,
+                observed_at=observed_at,
+                unique_slot=True,
+            )
+        )
+
+    project_match = _PROJECT_USES_RE.search(text)
+    if project_match:
+        value = project_match.group(1).strip().rstrip(".")
+        facts.append(
+            _build_fact(
+                category="project.stack",
+                entity="project",
+                relation="project.stack",
+                value=value,
+                content=f"Project uses {value}",
+                confidence=0.74,
+                source_turn=turn_idx,
+                source_role=role,
+                source_text=text,
+                timestamp=timestamp,
+                observed_at=observed_at,
+                unique_slot=False,
+            )
+        )
+
+    preference_match = _PREFERENCE_RE.search(text)
+    if preference_match:
+        value = preference_match.group(1).strip().rstrip(".")
+        facts.append(
+            _build_fact(
+                category="user_pref.preference",
+                entity="user",
+                relation="user.preference",
+                value=value,
+                content=value,
+                confidence=0.72,
+                source_turn=turn_idx,
+                source_role=role,
+                source_text=text,
+                timestamp=timestamp,
+                observed_at=observed_at,
+                unique_slot=False,
+            )
+        )
+
+    constraint_match = _CONSTRAINT_RE.search(text)
+    if constraint_match:
+        value = constraint_match.group(1).strip().rstrip(".")
+        facts.append(
+            _build_fact(
+                category="user_pref.constraint",
+                entity="user",
+                relation="user.constraint",
+                value=value,
+                content=f"Do not {value}",
+                confidence=0.82,
+                source_turn=turn_idx,
+                source_role=role,
+                source_text=text,
+                timestamp=timestamp,
+                observed_at=observed_at,
+                unique_slot=False,
+            )
+        )
+
+    decision_match = _DECISION_RE.search(text)
+    if decision_match:
+        value = decision_match.group(1).strip().rstrip(".")
+        facts.append(
+            _build_fact(
+                category="project.decision",
+                entity="project",
+                relation="project.decision",
+                value=value,
+                content=f"Decision: {value}",
+                confidence=0.79,
+                source_turn=turn_idx,
+                source_role=role,
+                source_text=text,
+                timestamp=timestamp,
+                observed_at=observed_at,
+                unique_slot=False,
+            )
+        )
 
     return facts
 
 
-def _extract_from_text(text: str, turn_idx: int, role: str) -> List[ExtractedFact]:
-    """Extract facts from a single text block."""
-    facts = []
-    timestamp = time.time()
+def _build_fact(
+    *,
+    category: str,
+    entity: str,
+    relation: str,
+    value: str,
+    content: str,
+    confidence: float,
+    source_turn: int,
+    source_role: str,
+    source_text: str,
+    timestamp: float,
+    observed_at: str,
+    unique_slot: bool,
+) -> ExtractedFact:
+    normalized_value = _normalize_value(value.rstrip(".!?"))
+    value = value.rstrip(".!?")
+    content = content.rstrip(".!?")
+    provenance = f"conversation:{source_role}:{source_turn}"
+    contradiction_group = relation if unique_slot else ""
+    evidence = [
+        {
+            "source_role": source_role,
+            "source_turn": source_turn,
+            "source_text": source_text,
+            "observed_at": observed_at,
+            "provenance": provenance,
+        }
+    ]
+    metadata = {
+        "entity": entity,
+        "relation": relation,
+        "value": value,
+        "normalized_value": normalized_value,
+        "provenance": [provenance],
+        "evidence": list(evidence),
+        "observation_count": 1,
+        "duplicate_count": 0,
+        "status": "active",
+    }
+    if contradiction_group:
+        metadata["contradiction_group"] = contradiction_group
+    return ExtractedFact(
+        category=category,
+        entity=entity,
+        content=content,
+        confidence=confidence,
+        source_turn=source_turn,
+        timestamp=timestamp,
+        source_role=source_role,
+        source_text=source_text,
+        normalized_content=normalized_value,
+        canonical_key=_canonical_key(entity, relation, normalized_value),
+        relation=relation,
+        contradiction_group=contradiction_group,
+        status="active",
+        provenance=provenance,
+        observed_at=observed_at,
+        evidence=evidence,
+        metadata=metadata,
+    )
 
-    # Clean text for pattern matching
-    clean = text.strip()
 
-    # User preference patterns (from user messages)
-    if role == "user":
-        for pattern, subcategory in _PREFERENCE_PATTERNS:
-            for match in re.finditer(pattern, clean, re.IGNORECASE):
-                content = match.group(1).strip() if match.lastindex else match.group(0).strip()
-                if len(content) > 5:
-                    facts.append(ExtractedFact(
-                        category=f"user_pref.{subcategory}",
-                        entity="user",
-                        content=content[:200],
-                        confidence=0.7,
-                        source_turn=turn_idx,
-                        timestamp=timestamp,
-                    ))
+def _normalize_candidates(candidates: List[ExtractedFact]) -> List[ExtractedFact]:
+    """Merge duplicates and mark contradictions while preserving evidence."""
 
-    # Correction patterns (from user messages)
-    if role == "user":
-        for pattern, subcategory in _CORRECTION_PATTERNS:
-            for match in re.finditer(pattern, clean, re.IGNORECASE):
-                content = match.group(1).strip() if match.lastindex else match.group(0).strip()
-                if len(content) > 5:
-                    facts.append(ExtractedFact(
-                        category=f"correction.{subcategory}",
-                        entity="user",
-                        content=content[:200],
-                        confidence=0.8,
-                        source_turn=turn_idx,
-                        timestamp=timestamp,
-                    ))
+    by_key: dict[str, ExtractedFact] = {}
+    contradiction_groups: dict[str, list[ExtractedFact]] = {}
 
-    # Project/infrastructure patterns (from both user and assistant)
-    for pattern, subcategory in _PROJECT_PATTERNS:
-        for match in re.finditer(pattern, clean, re.IGNORECASE):
-            content = match.group(1).strip() if match.lastindex else match.group(0).strip()
-            if len(content) > 5:
-                facts.append(ExtractedFact(
-                    category=f"project.{subcategory}",
-                    entity=subcategory,
-                    content=content[:200],
-                    confidence=0.6,
-                    source_turn=turn_idx,
-                    timestamp=timestamp,
-                ))
+    for candidate in candidates:
+        existing = by_key.get(candidate.canonical_key)
+        if existing is not None:
+            by_key[candidate.canonical_key] = _merge_fact(existing, candidate)
+            continue
 
-    return facts
+        by_key[candidate.canonical_key] = candidate
+        if candidate.contradiction_group:
+            contradiction_groups.setdefault(candidate.contradiction_group, []).append(candidate)
+
+    for group, facts in contradiction_groups.items():
+        canonical_keys = {fact.canonical_key for fact in facts}
+        if len(canonical_keys) <= 1:
+            continue
+        for fact in facts:
+            fact.status = "contradiction"
+            fact.metadata["status"] = "contradiction"
+            fact.metadata["contradiction_group"] = group
+            fact.metadata["contradiction_keys"] = sorted(canonical_keys - {fact.canonical_key})
+
+    return sorted(by_key.values(), key=lambda fact: (fact.source_turn, fact.timestamp, fact.canonical_key))
+
+
+def _merge_fact(existing: ExtractedFact, incoming: ExtractedFact) -> ExtractedFact:
+    existing.confidence = max(existing.confidence, incoming.confidence)
+    existing.timestamp = min(existing.timestamp, incoming.timestamp)
+    existing.source_turn = min(existing.source_turn, incoming.source_turn)
+    if not existing.observed_at or (incoming.observed_at and incoming.observed_at < existing.observed_at):
+        existing.observed_at = incoming.observed_at
+    existing.provenance = min(existing.provenance, incoming.provenance)
+
+    provenance = _ordered_unique(existing.metadata.get("provenance", []), incoming.metadata.get("provenance", []))
+    evidence = _merge_evidence(existing.metadata.get("evidence", []), incoming.metadata.get("evidence", []))
+    observation_count = int(existing.metadata.get("observation_count", len(existing.evidence) or 1))
+    observation_count += int(incoming.metadata.get("observation_count", len(incoming.evidence) or 1))
+
+    existing.evidence = evidence
+    existing.metadata["provenance"] = provenance
+    existing.metadata["evidence"] = evidence
+    existing.metadata["observation_count"] = observation_count
+    existing.metadata["duplicate_count"] = max(0, observation_count - 1)
+    existing.metadata["status"] = existing.status
+    return existing
 
 
 def save_facts_to_store(facts: List[ExtractedFact], fact_store_fn=None) -> int:
     """Save extracted facts to the fact store.
 
-    Args:
-        facts: List of extracted facts.
-        fact_store_fn: Optional callable(category, entity, content, trust).
-            If None, uses the holographic fact store if available.
-
-    Returns:
-        Number of facts saved.
+    If a callback is supplied, prefer the structured signature but fall back to
+    the legacy four-argument callback for compatibility.
     """
-    saved = 0
 
-    if fact_store_fn:
-        for fact in facts:
+    saved = 0
+    for fact in facts:
+        payload = {
+            "category": _store_category(fact.category),
+            "entity": fact.entity,
+            "content": fact.content,
+            "trust": fact.confidence,
+            "metadata": dict(fact.metadata),
+            "canonical_key": fact.canonical_key,
+            "observed_at": fact.observed_at,
+            "source_role": fact.source_role,
+            "source_turn": fact.source_turn,
+            "contradiction_group": fact.contradiction_group,
+            "status": fact.status,
+            "relation": fact.relation,
+        }
+
+        if fact_store_fn:
             try:
-                fact_store_fn(
-                    category=fact.category,
-                    entity=fact.entity,
-                    content=fact.content,
-                    trust=fact.confidence,
-                )
+                fact_store_fn(**payload)
                 saved += 1
-            except Exception as e:
-                logger.debug("Failed to save fact: %s", e)
-    else:
-        # Try holographic fact store
+                continue
+            except TypeError:
+                try:
+                    fact_store_fn(payload["category"], payload["entity"], payload["content"], payload["trust"])
+                    saved += 1
+                    continue
+                except Exception as exc:
+                    logger.debug("Failed to save fact via callback: %s", exc)
+                    continue
+            except Exception as exc:
+                logger.debug("Failed to save fact via callback: %s", exc)
+                continue
+
         try:
             from fact_store import fact_store as _fs
-            for fact in facts:
-                try:
-                    _fs(
-                        action="add",
-                        content=fact.content,
-                        category=fact.category,
-                        tags=fact.entity,
-                        trust_delta=fact.confidence - 0.5,
-                    )
-                    saved += 1
-                except Exception as e:
-                    logger.debug("Failed to save fact via fact_store: %s", e)
+
+            tags = ",".join(filter(None, [fact.entity, fact.relation, fact.status]))
+            _fs(
+                action="add",
+                content=fact.content,
+                category=_store_category(fact.category),
+                tags=tags,
+                trust_delta=fact.confidence - 0.5,
+            )
+            saved += 1
         except ImportError:
             logger.debug("fact_store not available — facts not persisted")
+            break
+        except Exception as exc:
+            logger.debug("Failed to save fact via fact_store: %s", exc)
 
     return saved
 
@@ -204,9 +553,10 @@ def extract_and_save_facts(
 
     Returns (extracted_facts, saved_count).
     """
+
     facts = extract_facts_from_messages(messages)
     if facts:
-        logger.info("Extracted %d facts from conversation", len(facts))
+        logger.info("Extracted %d normalized facts from conversation", len(facts))
         saved = save_facts_to_store(facts, fact_store_fn)
         logger.info("Saved %d/%d facts to store", saved, len(facts))
     else:
@@ -216,16 +566,105 @@ def extract_and_save_facts(
 
 def format_facts_summary(facts: List[ExtractedFact]) -> str:
     """Format extracted facts as a readable summary."""
+
     if not facts:
         return "No facts extracted."
 
-    by_category = {}
-    for f in facts:
-        by_category.setdefault(f.category, []).append(f)
+    by_category: dict[str, list[ExtractedFact]] = {}
+    for fact in facts:
+        by_category.setdefault(fact.category, []).append(fact)
 
     lines = [f"Extracted {len(facts)} facts:", ""]
-    for cat, cat_facts in sorted(by_category.items()):
-        lines.append(f"  {cat}:")
-        for f in cat_facts:
-            lines.append(f"    - {f.content[:80]}")
+    for category, category_facts in sorted(by_category.items()):
+        lines.append(f"  {category}:")
+        for fact in category_facts:
+            suffix = f" [{fact.status}]" if fact.status != "active" else ""
+            lines.append(f"    - {fact.content[:80]}{suffix}")
     return "\n".join(lines)
+
+
+def _store_category(category: str) -> str:
+    if category.startswith("user_pref"):
+        return "user_pref"
+    if category.startswith("project"):
+        return "project"
+    if category.startswith("tool"):
+        return "tool"
+    return "general"
+
+
+def _message_time(msg: Dict[str, Any]) -> Tuple[float, str]:
+    for key in ("created_at", "timestamp", "time"):
+        value = msg.get(key)
+        if value is None:
+            continue
+        if isinstance(value, (int, float)):
+            ts = float(value)
+            return ts, _iso_from_timestamp(ts)
+        if isinstance(value, str):
+            parsed = _parse_time_string(value)
+            if parsed is not None:
+                return parsed, _iso_from_timestamp(parsed) if "T" not in value else value.replace("+00:00", "Z")
+            return time.time(), value
+    now = time.time()
+    return now, _iso_from_timestamp(now)
+
+
+def _parse_time_string(value: str) -> float | None:
+    text = value.strip()
+    if not text:
+        return None
+    try:
+        return float(text)
+    except ValueError:
+        pass
+    try:
+        normalized = text[:-1] + "+00:00" if text.endswith("Z") else text
+        return datetime.fromisoformat(normalized).timestamp()
+    except ValueError:
+        return None
+
+
+def _iso_from_timestamp(value: float) -> str:
+    return datetime.fromtimestamp(value, tz=timezone.utc).isoformat().replace("+00:00", "Z")
+
+
+def _normalize_value(value: str) -> str:
+    normalized = re.sub(r"[^a-z0-9]+", " ", value.lower())
+    normalized = re.sub(r"\s+", " ", normalized).strip()
+    return normalized
+
+
+def _normalize_entity(value: str) -> str:
+    return _normalize_value(value).replace(" ", "_") or "entity"
+
+
+def _canonical_key(entity: str, relation: str, normalized_value: str) -> str:
+    return f"{entity}|{relation}|{normalized_value}"
+
+
+def _ordered_unique(*groups: List[str]) -> List[str]:
+    seen: set[str] = set()
+    ordered: list[str] = []
+    for group in groups:
+        for item in group:
+            if item and item not in seen:
+                seen.add(item)
+                ordered.append(item)
+    return ordered
+
+
+def _merge_evidence(existing: List[Dict[str, Any]], incoming: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
+    seen: set[tuple[str, str, str]] = set()
+    merged: list[dict[str, Any]] = []
+    for item in list(existing) + list(incoming):
+        key = (
+            str(item.get("provenance", "")),
+            str(item.get("observed_at", "")),
+            str(item.get("source_text", "")),
+        )
+        if key in seen:
+            continue
+        seen.add(key)
+        merged.append(dict(item))
+    return merged

cli.py

@@ -13,7 +13,6 @@ Usage:
     python cli.py --list-tools             # List available tools and exit
 """
 
-import concurrent.futures
 import logging
 import os
 import shutil
@@ -64,7 +63,6 @@ from agent.usage_pricing import (
     format_duration_compact,
     format_token_count_compact,
 )
-from agent.account_usage import fetch_account_usage, render_account_usage_lines
 from hermes_cli.banner import _format_context_length, format_banner_version_label
 
 _COMMAND_SPINNER_FRAMES = ("⠋", "⠙", "⠹", "⠸", "⠼", "⠴", "⠦", "⠧", "⠇", "⠏")
@@ -6473,29 +6471,6 @@ class HermesCLI:
         if cost_result.status == "unknown":
             print(f"  Note:             Pricing unknown for {agent.model}")
 
-        # Account limits -- fetched off-thread with a hard timeout so slow
-        # provider APIs don't hang the prompt.
-        provider = getattr(agent, "provider", None) or getattr(self, "provider", None)
-        base_url = getattr(agent, "base_url", None) or getattr(self, "base_url", None)
-        api_key = getattr(agent, "api_key", None) or getattr(self, "api_key", None)
-        account_snapshot = None
-        if provider:
-            with concurrent.futures.ThreadPoolExecutor(max_workers=1) as _pool:
-                try:
-                    account_snapshot = _pool.submit(
-                        fetch_account_usage,
-                        provider,
-                        base_url=base_url,
-                        api_key=api_key,
-                    ).result(timeout=10.0)
-                except (concurrent.futures.TimeoutError, Exception):
-                    account_snapshot = None
-        account_lines = [f"  {line}" for line in render_account_usage_lines(account_snapshot)]
-        if account_lines:
-            print()
-            for line in account_lines:
-                print(line)
-
         if self.verbose:
             logging.getLogger().setLevel(logging.DEBUG)
             for noisy in ('openai', 'openai._base_client', 'httpx', 'httpcore', 'asyncio', 'hpack', 'grpc', 'modal'):

gateway/run.py

@@ -28,8 +28,6 @@ from pathlib import Path
 from datetime import datetime
 from typing import Dict, Optional, Any, List
 
-from agent.account_usage import fetch_account_usage, render_account_usage_lines
-
 # ---------------------------------------------------------------------------
 # SSL certificate auto-detection for NixOS and other non-standard systems.
 # Must run BEFORE any HTTP library (discord, aiohttp, etc.) is imported.
@@ -6483,38 +6481,6 @@ class GatewayRunner:
                     if cached:
                         agent = cached[0]
 
-        # Resolve provider/base_url/api_key for the account-usage fetch.
-        # Prefer the live agent; fall back to persisted billing data on the
-        # SessionDB row so `/usage` still returns account info between turns
-        # when no agent is resident.
-        provider = getattr(agent, "provider", None) if agent and agent is not _AGENT_PENDING_SENTINEL else None
-        base_url = getattr(agent, "base_url", None) if agent and agent is not _AGENT_PENDING_SENTINEL else None
-        api_key = getattr(agent, "api_key", None) if agent and agent is not _AGENT_PENDING_SENTINEL else None
-        if not provider and getattr(self, "_session_db", None) is not None:
-            try:
-                _entry_for_billing = self.session_store.get_or_create_session(source)
-                persisted = self._session_db.get_session(_entry_for_billing.session_id) or {}
-            except Exception:
-                persisted = {}
-            provider = provider or persisted.get("billing_provider")
-            base_url = base_url or persisted.get("billing_base_url")
-
-        # Fetch account usage off the event loop so slow provider APIs don't
-        # block the gateway. Failures are non-fatal -- account_lines stays [].
-        account_lines: list[str] = []
-        if provider:
-            try:
-                account_snapshot = await asyncio.to_thread(
-                    fetch_account_usage,
-                    provider,
-                    base_url=base_url,
-                    api_key=api_key,
-                )
-            except Exception:
-                account_snapshot = None
-            if account_snapshot:
-                account_lines = render_account_usage_lines(account_snapshot, markdown=True)
-
         if agent and hasattr(agent, "session_total_tokens") and agent.session_api_calls > 0:
             lines = []
 
@@ -6572,10 +6538,6 @@ class GatewayRunner:
             if ctx.compression_count:
                 lines.append(f"Compressions: {ctx.compression_count}")
 
-            if account_lines:
-                lines.append("")
-                lines.extend(account_lines)
-
             return "\n".join(lines)
 
         # No agent at all -- check session history for a rough count
@@ -6585,18 +6547,12 @@ class GatewayRunner:
             from agent.model_metadata import estimate_messages_tokens_rough
             msgs = [m for m in history if m.get("role") in ("user", "assistant") and m.get("content")]
             approx = estimate_messages_tokens_rough(msgs)
-            lines = [
-                "📊 **Session Info**",
-                f"Messages: {len(msgs)}",
-                f"Estimated context: ~{approx:,} tokens",
-                "_(Detailed usage available after the first agent response)_",
-            ]
-            if account_lines:
-                lines.append("")
-                lines.extend(account_lines)
-            return "\n".join(lines)
-        if account_lines:
-            return "\n".join(account_lines)
+            return (
+                f"📊 **Session Info**\n"
+                f"Messages: {len(msgs)}\n"
+                f"Estimated context: ~{approx:,} tokens\n"
+                f"_(Detailed usage available after the first agent response)_"
+            )
         return "No usage data available for this session."
 
     async def _handle_insights_command(self, event: MessageEvent) -> str:

plugins/memory/holographic/__init__.py

@@ -356,44 +356,57 @@ class HolographicMemoryProvider(MemoryProvider):
     # -- Auto-extraction (on_session_end) ------------------------------------
 
     def _auto_extract_facts(self, messages: list) -> None:
-        _PREF_PATTERNS = [
-            re.compile(r'\bI\s+(?:prefer|like|love|use|want|need)\s+(.+)', re.IGNORECASE),
-            re.compile(r'\bmy\s+(?:favorite|preferred|default)\s+\w+\s+is\s+(.+)', re.IGNORECASE),
-            re.compile(r'\bI\s+(?:always|never|usually)\s+(.+)', re.IGNORECASE),
-        ]
-        _DECISION_PATTERNS = [
-            re.compile(r'\bwe\s+(?:decided|agreed|chose)\s+(?:to\s+)?(.+)', re.IGNORECASE),
-            re.compile(r'\bthe\s+project\s+(?:uses|needs|requires)\s+(.+)', re.IGNORECASE),
-        ]
+        from agent.session_compactor import evaluate_extraction_quality, extract_facts_from_messages
+
+        def _store_category(category: str) -> str:
+            if category.startswith("user_pref"):
+                return "user_pref"
+            if category.startswith("project"):
+                return "project"
+            if category.startswith("tool"):
+                return "tool"
+            return "general"
+
+        facts = extract_facts_from_messages(messages)
+        if not facts:
+            return
 
         extracted = 0
-        for msg in messages:
-            if msg.get("role") != "user":
-                continue
-            content = msg.get("content", "")
-            if not isinstance(content, str) or len(content) < 10:
-                continue
-
-            for pattern in _PREF_PATTERNS:
-                if pattern.search(content):
-                    try:
-                        self._store.add_fact(content[:400], category="user_pref")
-                        extracted += 1
-                    except Exception:
-                        pass
-                    break
-
-            for pattern in _DECISION_PATTERNS:
-                if pattern.search(content):
-                    try:
-                        self._store.add_fact(content[:400], category="project")
-                        extracted += 1
-                    except Exception:
-                        pass
-                    break
+        for fact in facts:
+            try:
+                metadata = dict(fact.metadata)
+                metadata.setdefault("relation", fact.relation)
+                metadata.setdefault("value", fact.content)
+                metadata.setdefault("provenance", [fact.provenance])
+                metadata.setdefault("evidence", list(fact.evidence))
+                metadata.setdefault("observation_count", len(fact.evidence))
+                metadata.setdefault("duplicate_count", max(0, len(fact.evidence) - 1))
+                self._store.add_fact(
+                    fact.content,
+                    category=_store_category(fact.category),
+                    tags=",".join(filter(None, [fact.entity, fact.relation, fact.status])),
+                    canonical_key=fact.canonical_key,
+                    metadata=metadata,
+                    confidence=fact.confidence,
+                    source_role=fact.source_role,
+                    source_turn=fact.source_turn,
+                    observed_at=fact.observed_at,
+                    contradiction_group=fact.contradiction_group,
+                    status=fact.status,
+                )
+                extracted += 1
+            except Exception as exc:
+                logger.debug("Structured auto-extract failed for %s: %s", fact.canonical_key, exc)
 
         if extracted:
-            logger.info("Auto-extracted %d facts from conversation", extracted)
+            metrics = evaluate_extraction_quality(messages)
+            logger.info(
+                "Auto-extracted %d structured facts from conversation (raw=%d normalized=%d contradictions=%d)",
+                extracted,
+                metrics["raw_candidates"],
+                metrics["normalized_facts"],
+                metrics["contradiction_groups"],
+            )
 
 
 # ---------------------------------------------------------------------------

plugins/memory/holographic/store.py

@@ -3,6 +3,7 @@ SQLite-backed fact store with entity resolution and trust scoring.
 Single-user Hermes memory store plugin.
 """
 
+import json
 import re
 import sqlite3
 import threading
@@ -15,16 +16,24 @@ except ImportError:
 
 _SCHEMA = """
 CREATE TABLE IF NOT EXISTS facts (
-    fact_id         INTEGER PRIMARY KEY AUTOINCREMENT,
-    content         TEXT NOT NULL UNIQUE,
-    category        TEXT DEFAULT 'general',
-    tags            TEXT DEFAULT '',
-    trust_score     REAL DEFAULT 0.5,
-    retrieval_count INTEGER DEFAULT 0,
-    helpful_count   INTEGER DEFAULT 0,
-    created_at      TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
-    updated_at      TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
-    hrr_vector      BLOB
+    fact_id             INTEGER PRIMARY KEY AUTOINCREMENT,
+    content             TEXT NOT NULL UNIQUE,
+    category            TEXT DEFAULT 'general',
+    tags                TEXT DEFAULT '',
+    trust_score         REAL DEFAULT 0.5,
+    retrieval_count     INTEGER DEFAULT 0,
+    helpful_count       INTEGER DEFAULT 0,
+    canonical_key       TEXT DEFAULT '',
+    metadata_json       TEXT DEFAULT '{}',
+    confidence          REAL DEFAULT 0.5,
+    source_role         TEXT DEFAULT '',
+    source_turn         INTEGER DEFAULT -1,
+    observed_at         TEXT DEFAULT '',
+    contradiction_group TEXT DEFAULT '',
+    status              TEXT DEFAULT 'active',
+    created_at          TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
+    updated_at          TIMESTAMP DEFAULT CURRENT_TIMESTAMP,
+    hrr_vector          BLOB
 );
 
 CREATE TABLE IF NOT EXISTS entities (
@@ -41,9 +50,11 @@ CREATE TABLE IF NOT EXISTS fact_entities (
     PRIMARY KEY (fact_id, entity_id)
 );
 
-CREATE INDEX IF NOT EXISTS idx_facts_trust    ON facts(trust_score DESC);
-CREATE INDEX IF NOT EXISTS idx_facts_category ON facts(category);
-CREATE INDEX IF NOT EXISTS idx_entities_name  ON entities(name);
+CREATE INDEX IF NOT EXISTS idx_facts_trust               ON facts(trust_score DESC);
+CREATE INDEX IF NOT EXISTS idx_facts_category            ON facts(category);
+CREATE INDEX IF NOT EXISTS idx_facts_canonical_key       ON facts(canonical_key);
+CREATE INDEX IF NOT EXISTS idx_facts_contradiction_group ON facts(contradiction_group);
+CREATE INDEX IF NOT EXISTS idx_entities_name             ON entities(name);
 
 CREATE VIRTUAL TABLE IF NOT EXISTS facts_fts
     USING fts5(content, tags, content=facts, content_rowid=fact_id);
@@ -129,10 +140,23 @@ class MemoryStore:
         """Create tables, indexes, and triggers if they do not exist. Enable WAL mode."""
         self._conn.execute("PRAGMA journal_mode=WAL")
         self._conn.executescript(_SCHEMA)
-        # Migrate: add hrr_vector column if missing (safe for existing databases)
         columns = {row[1] for row in self._conn.execute("PRAGMA table_info(facts)").fetchall()}
-        if "hrr_vector" not in columns:
-            self._conn.execute("ALTER TABLE facts ADD COLUMN hrr_vector BLOB")
+        migrations = {
+            "hrr_vector": "ALTER TABLE facts ADD COLUMN hrr_vector BLOB",
+            "canonical_key": "ALTER TABLE facts ADD COLUMN canonical_key TEXT DEFAULT ''",
+            "metadata_json": "ALTER TABLE facts ADD COLUMN metadata_json TEXT DEFAULT '{}'",
+            "confidence": "ALTER TABLE facts ADD COLUMN confidence REAL DEFAULT 0.5",
+            "source_role": "ALTER TABLE facts ADD COLUMN source_role TEXT DEFAULT ''",
+            "source_turn": "ALTER TABLE facts ADD COLUMN source_turn INTEGER DEFAULT -1",
+            "observed_at": "ALTER TABLE facts ADD COLUMN observed_at TEXT DEFAULT ''",
+            "contradiction_group": "ALTER TABLE facts ADD COLUMN contradiction_group TEXT DEFAULT ''",
+            "status": "ALTER TABLE facts ADD COLUMN status TEXT DEFAULT 'active'",
+        }
+        for column, ddl in migrations.items():
+            if column not in columns:
+                self._conn.execute(ddl)
+        self._conn.execute("CREATE INDEX IF NOT EXISTS idx_facts_canonical_key ON facts(canonical_key)")
+        self._conn.execute("CREATE INDEX IF NOT EXISTS idx_facts_contradiction_group ON facts(contradiction_group)")
         self._conn.commit()
 
     # ------------------------------------------------------------------
@@ -144,41 +168,148 @@ class MemoryStore:
         content: str,
         category: str = "general",
         tags: str = "",
+        *,
+        canonical_key: str = "",
+        metadata: dict | None = None,
+        confidence: float | None = None,
+        source_role: str = "",
+        source_turn: int = -1,
+        observed_at: str = "",
+        contradiction_group: str = "",
+        status: str = "active",
     ) -> int:
         """Insert a fact and return its fact_id.
 
-        Deduplicates by content (UNIQUE constraint). On duplicate, returns
-        the existing fact_id without modifying the row. Extracts entities from
-        the content and links them to the fact.
+        Exact duplicates are deduplicated by content. Near-duplicates are
+        normalized by canonical_key, with provenance/evidence merged into the
+        existing row. Contradictions sharing the same contradiction_group remain
+        stored as separate rows and are marked inspectably.
         """
         with self._lock:
             content = content.strip()
             if not content:
                 raise ValueError("content must not be empty")
 
+            metadata = dict(metadata or {})
+            canonical_key = canonical_key.strip()
+            contradiction_group = contradiction_group.strip()
+            observed_at = observed_at.strip()
+            status = status or "active"
+            trust_score = self.default_trust if confidence is None else _clamp_trust(confidence)
+            metadata_json = json.dumps(metadata, sort_keys=True)
+
+            if canonical_key:
+                existing = self._conn.execute(
+                    "SELECT fact_id, metadata_json, trust_score, confidence, observed_at FROM facts WHERE canonical_key = ?",
+                    (canonical_key,),
+                ).fetchone()
+                if existing is not None:
+                    merged_metadata = self._merge_metadata(existing["metadata_json"], metadata)
+                    merged_trust = max(float(existing["trust_score"]), trust_score)
+                    merged_observed_at = existing["observed_at"] or observed_at
+                    if observed_at and merged_observed_at:
+                        merged_observed_at = min(merged_observed_at, observed_at)
+                    elif observed_at:
+                        merged_observed_at = observed_at
+                    self._conn.execute(
+                        """
+                        UPDATE facts
+                        SET metadata_json = ?,
+                            trust_score = ?,
+                            confidence = ?,
+                            observed_at = ?,
+                            updated_at = CURRENT_TIMESTAMP
+                        WHERE fact_id = ?
+                        """,
+                        (
+                            json.dumps(merged_metadata, sort_keys=True),
+                            merged_trust,
+                            max(float(existing["confidence"] or 0.0), confidence or trust_score),
+                            merged_observed_at,
+                            existing["fact_id"],
+                        ),
+                    )
+                    self._conn.commit()
+                    return int(existing["fact_id"])
+
+            contradiction_rows = []
+            if contradiction_group:
+                contradiction_rows = self._conn.execute(
+                    """
+                    SELECT fact_id, canonical_key, metadata_json
+                    FROM facts
+                    WHERE contradiction_group = ?
+                      AND canonical_key != ?
+                    """,
+                    (contradiction_group, canonical_key),
+                ).fetchall()
+                if contradiction_rows:
+                    status = "contradiction"
+                    metadata = dict(metadata)
+                    metadata["status"] = "contradiction"
+                    metadata["contradiction_group"] = contradiction_group
+                    metadata["contradiction_keys"] = sorted(
+                        {
+                            canonical_key,
+                            *[str(row["canonical_key"]) for row in contradiction_rows if row["canonical_key"]],
+                        }
+                        - {""}
+                    )
+                    metadata_json = json.dumps(metadata, sort_keys=True)
+
             try:
                 cur = self._conn.execute(
                     """
-                    INSERT INTO facts (content, category, tags, trust_score)
-                    VALUES (?, ?, ?, ?)
+                    INSERT INTO facts (
+                        content,
+                        category,
+                        tags,
+                        trust_score,
+                        canonical_key,
+                        metadata_json,
+                        confidence,
+                        source_role,
+                        source_turn,
+                        observed_at,
+                        contradiction_group,
+                        status
+                    )
+                    VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
                     """,
-                    (content, category, tags, self.default_trust),
+                    (
+                        content,
+                        category,
+                        tags,
+                        trust_score,
+                        canonical_key,
+                        metadata_json,
+                        confidence if confidence is not None else trust_score,
+                        source_role,
+                        source_turn,
+                        observed_at,
+                        contradiction_group,
+                        status,
+                    ),
                 )
                 self._conn.commit()
                 fact_id: int = cur.lastrowid  # type: ignore[assignment]
             except sqlite3.IntegrityError:
-                # Duplicate content — return existing id
                 row = self._conn.execute(
                     "SELECT fact_id FROM facts WHERE content = ?", (content,)
                 ).fetchone()
                 return int(row["fact_id"])
 
-            # Entity extraction and linking
+            if contradiction_rows:
+                self._mark_contradictions(
+                    contradiction_group=contradiction_group,
+                    new_canonical_key=canonical_key,
+                    existing_rows=contradiction_rows,
+                )
+
             for name in self._extract_entities(content):
                 entity_id = self._resolve_entity(name)
                 self._link_fact_entity(fact_id, entity_id)
 
-            # Compute HRR vector after entity linking
             self._compute_hrr_vector(fact_id, content)
             self._rebuild_bank(category)
 
@@ -211,6 +342,9 @@ class MemoryStore:
             sql = f"""
                 SELECT f.fact_id, f.content, f.category, f.tags,
                        f.trust_score, f.retrieval_count, f.helpful_count,
+                       f.canonical_key, f.metadata_json, f.confidence,
+                       f.source_role, f.source_turn, f.observed_at,
+                       f.contradiction_group, f.status,
                        f.created_at, f.updated_at
                 FROM facts f
                 JOIN facts_fts fts ON fts.rowid = f.fact_id
@@ -336,7 +470,11 @@ class MemoryStore:
 
             sql = f"""
                 SELECT fact_id, content, category, tags, trust_score,
-                       retrieval_count, helpful_count, created_at, updated_at
+                       retrieval_count, helpful_count,
+                       canonical_key, metadata_json, confidence,
+                       source_role, source_turn, observed_at,
+                       contradiction_group, status,
+                       created_at, updated_at
                 FROM facts
                 WHERE trust_score >= ?
                   {category_clause}
@@ -387,6 +525,89 @@ class MemoryStore:
                 "helpful_count": row["helpful_count"] + helpful_increment,
             }
 
+    # ------------------------------------------------------------------
+    # Metadata helpers
+    # ------------------------------------------------------------------
+
+    def _load_metadata(self, metadata_json: str | None) -> dict:
+        if not metadata_json:
+            return {}
+        try:
+            data = json.loads(metadata_json)
+            return data if isinstance(data, dict) else {}
+        except Exception:
+            return {}
+
+    def _merge_metadata(self, existing_json: str | None, incoming: dict | None) -> dict:
+        existing = self._load_metadata(existing_json)
+        incoming = dict(incoming or {})
+        merged = dict(existing)
+        merged.update({k: v for k, v in incoming.items() if k not in {"provenance", "evidence", "observation_count", "duplicate_count", "contradiction_keys"}})
+
+        provenance = []
+        seen_provenance: set[str] = set()
+        for item in list(existing.get("provenance", [])) + list(incoming.get("provenance", [])):
+            if item and item not in seen_provenance:
+                seen_provenance.add(item)
+                provenance.append(item)
+
+        evidence = []
+        seen_evidence: set[tuple[str, str, str]] = set()
+        for item in list(existing.get("evidence", [])) + list(incoming.get("evidence", [])):
+            if not isinstance(item, dict):
+                continue
+            key = (
+                str(item.get("provenance", "")),
+                str(item.get("observed_at", "")),
+                str(item.get("source_text", "")),
+            )
+            if key in seen_evidence:
+                continue
+            seen_evidence.add(key)
+            evidence.append(dict(item))
+
+        observation_count = int(existing.get("observation_count", max(1, len(existing.get("evidence", [])) or 1)))
+        observation_count += int(incoming.get("observation_count", max(1, len(incoming.get("evidence", [])) or 1)))
+
+        contradiction_keys = []
+        seen_keys: set[str] = set()
+        for item in list(existing.get("contradiction_keys", [])) + list(incoming.get("contradiction_keys", [])):
+            if item and item not in seen_keys:
+                seen_keys.add(item)
+                contradiction_keys.append(item)
+
+        merged["provenance"] = provenance
+        merged["evidence"] = evidence
+        merged["observation_count"] = observation_count
+        merged["duplicate_count"] = max(0, observation_count - 1)
+        if contradiction_keys:
+            merged["contradiction_keys"] = contradiction_keys
+        return merged
+
+    def _mark_contradictions(self, contradiction_group: str, new_canonical_key: str, existing_rows: list[sqlite3.Row]) -> None:
+        for row in existing_rows:
+            metadata = self._load_metadata(row["metadata_json"])
+            keys = []
+            seen: set[str] = set()
+            for item in list(metadata.get("contradiction_keys", [])) + [new_canonical_key]:
+                if item and item not in seen:
+                    seen.add(item)
+                    keys.append(item)
+            metadata["status"] = "contradiction"
+            metadata["contradiction_group"] = contradiction_group
+            metadata["contradiction_keys"] = keys
+            self._conn.execute(
+                """
+                UPDATE facts
+                SET status = 'contradiction',
+                    metadata_json = ?,
+                    updated_at = CURRENT_TIMESTAMP
+                WHERE fact_id = ?
+                """,
+                (json.dumps(metadata, sort_keys=True), row["fact_id"]),
+            )
+        self._conn.commit()
+
     # ------------------------------------------------------------------
     # Entity helpers
     # ------------------------------------------------------------------
@@ -560,8 +781,14 @@ class MemoryStore:
     # ------------------------------------------------------------------
 
     def _row_to_dict(self, row: sqlite3.Row) -> dict:
-        """Convert a sqlite3.Row to a plain dict."""
-        return dict(row)
+        """Convert a sqlite3.Row to a plain dict with decoded metadata."""
+        data = dict(row)
+        metadata = self._load_metadata(data.get("metadata_json"))
+        if metadata:
+            data["metadata"] = metadata
+            data.setdefault("relation", metadata.get("relation"))
+        data.pop("metadata_json", None)
+        return data
 
     def close(self) -> None:
         """Close the database connection."""

tests/fixtures/memory_extraction_fragments.json

@@ -0,0 +1,63 @@
+{
+  "preferences_and_duplicates": [
+    {
+      "role": "user",
+      "content": "Deploy via Ansible for production changes.",
+      "created_at": "2026-04-22T10:00:00Z"
+    },
+    {
+      "role": "user",
+      "content": "We deploy through Ansible on this repo.",
+      "created_at": "2026-04-22T10:01:00Z"
+    },
+    {
+      "role": "user",
+      "content": "Gitea-first for repository work.",
+      "created_at": "2026-04-22T10:02:00Z"
+    }
+  ],
+  "operational_and_contradictions": [
+    {
+      "role": "user",
+      "content": "The BURN watchdog caps dispatches per cycle to 6.",
+      "created_at": "2026-04-22T11:00:00Z"
+    },
+    {
+      "role": "user",
+      "content": "The provider should stay openai-codex/gpt-5.4.",
+      "created_at": "2026-04-22T11:01:00Z"
+    },
+    {
+      "role": "user",
+      "content": "Correction: the provider should stay mimo-v2-pro.",
+      "created_at": "2026-04-22T11:02:00Z"
+    }
+  ],
+  "mixed_transcript": [
+    {
+      "role": "user",
+      "content": "Deploy via Ansible for production changes.",
+      "created_at": "2026-04-22T10:00:00Z"
+    },
+    {
+      "role": "user",
+      "content": "We deploy through Ansible on this repo.",
+      "created_at": "2026-04-22T10:01:00Z"
+    },
+    {
+      "role": "user",
+      "content": "The BURN watchdog caps dispatches per cycle to 6.",
+      "created_at": "2026-04-22T11:00:00Z"
+    },
+    {
+      "role": "user",
+      "content": "The provider should stay openai-codex/gpt-5.4.",
+      "created_at": "2026-04-22T11:01:00Z"
+    },
+    {
+      "role": "user",
+      "content": "Correction: the provider should stay mimo-v2-pro.",
+      "created_at": "2026-04-22T11:02:00Z"
+    }
+  ]
+}

tests/gateway/test_usage_command.py

@@ -175,79 +175,3 @@ class TestUsageCachedAgent:
             result = await runner._handle_usage_command(event)
 
         assert "Cost: included" in result
-
-
-class TestUsageAccountSection:
-    """Account-limits section appended to /usage output."""
-
-    @pytest.mark.asyncio
-    async def test_usage_command_includes_account_section(self, monkeypatch):
-        agent = _make_mock_agent(provider="openai-codex")
-        agent.base_url = "https://chatgpt.com/backend-api/codex"
-        agent.api_key = "unused"
-        runner = _make_runner(SK, cached_agent=agent)
-        event = MagicMock()
-
-        monkeypatch.setattr(
-            "gateway.run.fetch_account_usage",
-            lambda provider, base_url=None, api_key=None: object(),
-        )
-        monkeypatch.setattr(
-            "gateway.run.render_account_usage_lines",
-            lambda snapshot, markdown=False: [
-                "📈 **Account limits**",
-                "Provider: openai-codex (Pro)",
-                "Session: 85% remaining (15% used)",
-            ],
-        )
-        with patch("agent.rate_limit_tracker.format_rate_limit_compact", return_value="RPM: 50/60"), \
-             patch("agent.usage_pricing.estimate_usage_cost") as mock_cost:
-            mock_cost.return_value = MagicMock(amount_usd=None, status="included")
-            result = await runner._handle_usage_command(event)
-
-        assert "📊 **Session Token Usage**" in result
-        assert "📈 **Account limits**" in result
-        assert "Provider: openai-codex (Pro)" in result
-
-    @pytest.mark.asyncio
-    async def test_usage_command_uses_persisted_provider_when_agent_not_running(self, monkeypatch):
-        runner = _make_runner(SK)
-        runner._session_db = MagicMock()
-        runner._session_db.get_session.return_value = {
-            "billing_provider": "openai-codex",
-            "billing_base_url": "https://chatgpt.com/backend-api/codex",
-        }
-        session_entry = MagicMock()
-        session_entry.session_id = "sess-1"
-        runner.session_store.get_or_create_session.return_value = session_entry
-        runner.session_store.load_transcript.return_value = [
-            {"role": "user", "content": "earlier"},
-        ]
-
-        calls = {}
-
-        async def _fake_to_thread(fn, *args, **kwargs):
-            calls["args"] = args
-            calls["kwargs"] = kwargs
-            return fn(*args, **kwargs)
-
-        monkeypatch.setattr("gateway.run.asyncio.to_thread", _fake_to_thread)
-        monkeypatch.setattr(
-            "gateway.run.fetch_account_usage",
-            lambda provider, base_url=None, api_key=None: object(),
-        )
-        monkeypatch.setattr(
-            "gateway.run.render_account_usage_lines",
-            lambda snapshot, markdown=False: [
-                "📈 **Account limits**",
-                "Provider: openai-codex (Pro)",
-            ],
-        )
-
-        event = MagicMock()
-        result = await runner._handle_usage_command(event)
-
-        assert calls["args"] == ("openai-codex",)
-        assert calls["kwargs"]["base_url"] == "https://chatgpt.com/backend-api/codex"
-        assert "📊 **Session Info**" in result
-        assert "📈 **Account limits**" in result

tests/plugins/memory/test_holographic_auto_extract.py

@@ -0,0 +1,50 @@
+"""Integration tests for holographic auto-extraction with structured fact persistence."""
+
+import json
+import sys
+from pathlib import Path
+
+sys.path.insert(0, str(Path(__file__).resolve().parents[3]))
+
+from plugins.memory.holographic import HolographicMemoryProvider
+
+_FIXTURE_PATH = Path(__file__).resolve().parents[2] / "fixtures" / "memory_extraction_fragments.json"
+
+
+def _load_fixture(name: str):
+    return json.loads(_FIXTURE_PATH.read_text())[name]
+
+
+class TestHolographicAutoExtract:
+    def test_auto_extract_persists_structured_metadata_and_normalizes_duplicates(self, tmp_path):
+        provider = HolographicMemoryProvider(
+            config={
+                "db_path": str(tmp_path / "memory_store.db"),
+                "auto_extract": True,
+                "default_trust": 0.5,
+            }
+        )
+        provider.initialize("test-session")
+
+        messages = _load_fixture("mixed_transcript")
+        provider.on_session_end(messages)
+        provider.on_session_end(messages)
+
+        facts = provider._store.list_facts(min_trust=0.0, limit=20)
+        deploy_facts = [f for f in facts if f.get("relation") == "workflow.deploy_method"]
+        provider_facts = [f for f in facts if f.get("contradiction_group") == "config.provider"]
+
+        assert len(deploy_facts) == 1
+        assert deploy_facts[0]["metadata"]["duplicate_count"] >= 3
+        assert deploy_facts[0]["observed_at"] == "2026-04-22T10:00:00Z"
+        assert deploy_facts[0]["metadata"]["provenance"] == [
+            "conversation:user:0",
+            "conversation:user:1",
+        ]
+
+        assert len(provider_facts) == 2
+        assert {f["status"] for f in provider_facts} == {"contradiction"}
+        assert {f["metadata"]["value"] for f in provider_facts} == {
+            "openai-codex/gpt-5.4",
+            "mimo-v2-pro",
+        }

tests/test_account_usage.py

@@ -1,203 +0,0 @@
-from datetime import datetime, timezone
-
-from agent.account_usage import (
-    AccountUsageSnapshot,
-    AccountUsageWindow,
-    fetch_account_usage,
-    render_account_usage_lines,
-)
-
-
-class _Response:
-    def __init__(self, payload, status_code=200):
-        self._payload = payload
-        self.status_code = status_code
-
-    def raise_for_status(self):
-        if self.status_code >= 400:
-            raise RuntimeError(f"HTTP {self.status_code}")
-
-    def json(self):
-        return self._payload
-
-
-class _Client:
-    def __init__(self, payload):
-        self._payload = payload
-
-    def __enter__(self):
-        return self
-
-    def __exit__(self, exc_type, exc, tb):
-        return False
-
-    def get(self, url, headers=None):
-        return _Response(self._payload)
-
-
-class _RoutingClient:
-    def __init__(self, payloads):
-        self._payloads = payloads
-
-    def __enter__(self):
-        return self
-
-    def __exit__(self, exc_type, exc, tb):
-        return False
-
-    def get(self, url, headers=None):
-        return _Response(self._payloads[url])
-
-
-def test_fetch_account_usage_codex(monkeypatch):
-    monkeypatch.setattr(
-        "agent.account_usage.resolve_codex_runtime_credentials",
-        lambda refresh_if_expiring=True: {
-            "provider": "openai-codex",
-            "base_url": "https://chatgpt.com/backend-api/codex",
-            "api_key": "***",
-        },
-    )
-    monkeypatch.setattr(
-        "agent.account_usage._read_codex_tokens",
-        lambda: {"tokens": {"account_id": "acct_123"}},
-    )
-    monkeypatch.setattr(
-        "agent.account_usage.httpx.Client",
-        lambda timeout=15.0: _Client(
-            {
-                "plan_type": "pro",
-                "rate_limit": {
-                    "primary_window": {
-                        "used_percent": 15,
-                        "reset_at": 1_900_000_000,
-                        "limit_window_seconds": 18000,
-                    },
-                    "secondary_window": {
-                        "used_percent": 40,
-                        "reset_at": 1_900_500_000,
-                        "limit_window_seconds": 604800,
-                    },
-                },
-                "credits": {"has_credits": True, "balance": 12.5},
-            }
-        ),
-    )
-
-    snapshot = fetch_account_usage("openai-codex")
-
-    assert snapshot is not None
-    assert snapshot.plan == "Pro"
-    assert len(snapshot.windows) == 2
-    assert snapshot.windows[0].label == "Session"
-    assert snapshot.windows[0].used_percent == 15.0
-    assert snapshot.windows[0].reset_at == datetime.fromtimestamp(1_900_000_000, tz=timezone.utc)
-    assert "Credits balance: $12.50" in snapshot.details
-
-
-def test_render_account_usage_lines_includes_reset_and_provider():
-    snapshot = AccountUsageSnapshot(
-        provider="openai-codex",
-        source="usage_api",
-        fetched_at=datetime.now(timezone.utc),
-        plan="Pro",
-        windows=(
-            AccountUsageWindow(
-                label="Session",
-                used_percent=25,
-                reset_at=datetime.now(timezone.utc),
-            ),
-        ),
-        details=("Credits balance: $9.99",),
-    )
-    lines = render_account_usage_lines(snapshot)
-
-    assert lines[0] == "📈 Account limits"
-    assert "openai-codex (Pro)" in lines[1]
-    assert "Session: 75% remaining (25% used)" in lines[2]
-    assert "Credits balance: $9.99" in lines[3]
-
-
-def test_fetch_account_usage_openrouter_uses_limit_remaining_and_ignores_deprecated_rate_limit(monkeypatch):
-    monkeypatch.setattr(
-        "agent.account_usage.resolve_runtime_provider",
-        lambda requested, explicit_base_url=None, explicit_api_key=None: {
-            "provider": "openrouter",
-            "base_url": "https://openrouter.ai/api/v1",
-            "api_key": "***",
-        },
-    )
-    monkeypatch.setattr(
-        "agent.account_usage.httpx.Client",
-        lambda timeout=10.0: _RoutingClient(
-            {
-                "https://openrouter.ai/api/v1/credits": {
-                    "data": {"total_credits": 300.0, "total_usage": 10.92}
-                },
-                "https://openrouter.ai/api/v1/key": {
-                    "data": {
-                        "limit": 100.0,
-                        "limit_remaining": 70.0,
-                        "limit_reset": "monthly",
-                        "usage": 12.5,
-                        "usage_daily": 0.5,
-                        "usage_weekly": 2.0,
-                        "usage_monthly": 8.0,
-                        "rate_limit": {"requests": -1, "interval": "10s"},
-                    }
-                },
-            }
-        ),
-    )
-
-    snapshot = fetch_account_usage("openrouter")
-
-    assert snapshot is not None
-    assert snapshot.windows == (
-        AccountUsageWindow(
-            label="API key quota",
-            used_percent=30.0,
-            detail="$70.00 of $100.00 remaining • resets monthly",
-        ),
-    )
-    assert "Credits balance: $289.08" in snapshot.details
-    assert "API key usage: $12.50 total • $0.50 today • $2.00 this week • $8.00 this month" in snapshot.details
-    assert all("-1 requests / 10s" not in line for line in render_account_usage_lines(snapshot))
-
-
-def test_fetch_account_usage_openrouter_omits_quota_window_when_key_has_no_limit(monkeypatch):
-    monkeypatch.setattr(
-        "agent.account_usage.resolve_runtime_provider",
-        lambda requested, explicit_base_url=None, explicit_api_key=None: {
-            "provider": "openrouter",
-            "base_url": "https://openrouter.ai/api/v1",
-            "api_key": "***",
-        },
-    )
-    monkeypatch.setattr(
-        "agent.account_usage.httpx.Client",
-        lambda timeout=10.0: _RoutingClient(
-            {
-                "https://openrouter.ai/api/v1/credits": {
-                    "data": {"total_credits": 100.0, "total_usage": 25.5}
-                },
-                "https://openrouter.ai/api/v1/key": {
-                    "data": {
-                        "limit": None,
-                        "limit_remaining": None,
-                        "usage": 25.5,
-                        "usage_daily": 1.25,
-                        "usage_weekly": 4.5,
-                        "usage_monthly": 18.0,
-                    }
-                },
-            }
-        ),
-    )
-
-    snapshot = fetch_account_usage("openrouter")
-
-    assert snapshot is not None
-    assert snapshot.windows == ()
-    assert "Credits balance: $74.50" in snapshot.details
-    assert "API key usage: $25.50 total • $1.25 today • $4.50 this week • $18.00 this month" in snapshot.details

tests/test_session_compactor.py

@@ -1,6 +1,6 @@
 """Tests for session compaction with fact extraction."""
 
-import pytest
+import json
 import sys
 from pathlib import Path
 
@@ -8,12 +8,19 @@ sys.path.insert(0, str(Path(__file__).resolve().parent.parent))
 
 from agent.session_compactor import (
     ExtractedFact,
-    extract_facts_from_messages,
-    save_facts_to_store,
+    evaluate_extraction_quality,
     extract_and_save_facts,
+    extract_facts_from_messages,
     format_facts_summary,
+    save_facts_to_store,
 )
 
+_FIXTURE_PATH = Path(__file__).resolve().parent / "fixtures" / "memory_extraction_fragments.json"
+
+
+def _load_fixture(name: str):
+    return json.loads(_FIXTURE_PATH.read_text())[name]
+
 
 class TestFactExtraction:
     def test_extract_preference(self):
@@ -60,14 +67,48 @@ class TestFactExtraction:
             {"role": "user", "content": "I prefer Python."},
         ]
         facts = extract_facts_from_messages(messages)
-        # Should deduplicate
         python_facts = [f for f in facts if "Python" in f.content]
         assert len(python_facts) == 1
 
+    def test_structured_fact_preserves_provenance_and_temporal_metadata(self):
+        facts = extract_facts_from_messages(_load_fixture("preferences_and_duplicates"))
+        deploy_fact = next(f for f in facts if f.relation == "workflow.deploy_method")
+        assert deploy_fact.source_role == "user"
+        assert deploy_fact.source_turn == 0
+        assert deploy_fact.observed_at == "2026-04-22T10:00:00Z"
+        assert deploy_fact.provenance == "conversation:user:0"
+        assert deploy_fact.canonical_key
+        assert deploy_fact.evidence
+        assert deploy_fact.evidence[0]["source_text"].startswith("Deploy via Ansible")
+
+    def test_near_duplicate_facts_are_normalized_into_one_canonical_fact(self):
+        facts = extract_facts_from_messages(_load_fixture("preferences_and_duplicates"))
+        deploy_facts = [f for f in facts if f.relation == "workflow.deploy_method"]
+        assert len(deploy_facts) == 1
+        assert len(deploy_facts[0].evidence) == 2
+        assert deploy_facts[0].metadata["duplicate_count"] == 1
+
+    def test_contradictory_facts_are_preserved_for_unique_slots(self):
+        facts = extract_facts_from_messages(_load_fixture("operational_and_contradictions"))
+        provider_facts = [f for f in facts if f.contradiction_group == "config.provider"]
+        assert len(provider_facts) == 2
+        assert {f.status for f in provider_facts} == {"contradiction"}
+        assert {f.normalized_content for f in provider_facts} == {
+            "openai codex gpt 5 4",
+            "mimo v2 pro",
+        }
+
+    def test_quality_evaluation_reports_noise_reduction(self):
+        metrics = evaluate_extraction_quality(_load_fixture("mixed_transcript"))
+        assert metrics["raw_candidates"] > metrics["normalized_facts"]
+        assert metrics["noise_reduction"] > 0
+        assert metrics["contradiction_groups"] == 1
+
 
 class TestSaveFacts:
     def test_save_with_callback(self):
         saved = []
+
         def mock_save(category, entity, content, trust):
             saved.append({"category": category, "content": content})
 
@@ -76,6 +117,38 @@ class TestSaveFacts:
         assert count == 1
         assert len(saved) == 1
 
+    def test_save_with_extended_callback_metadata(self):
+        saved = []
+
+        def mock_save(category, entity, content, trust, **kwargs):
+            saved.append({
+                "category": category,
+                "entity": entity,
+                "content": content,
+                "trust": trust,
+                **kwargs,
+            })
+
+        fact = ExtractedFact(
+            "project.operational",
+            "watchdog",
+            "BURN watchdog caps dispatches per cycle to 6",
+            0.9,
+            2,
+            source_role="user",
+            observed_at="2026-04-22T11:00:00Z",
+            provenance="conversation:user:2",
+            canonical_key="project.operational|watchdog|dispatch_cap|6",
+            relation="fleet.dispatch_cap",
+            contradiction_group="fleet.dispatch_cap",
+            metadata={"duplicate_count": 0},
+        )
+        count = save_facts_to_store([fact], fact_store_fn=mock_save)
+        assert count == 1
+        assert saved[0]["canonical_key"] == fact.canonical_key
+        assert saved[0]["observed_at"] == "2026-04-22T11:00:00Z"
+        assert saved[0]["metadata"]["duplicate_count"] == 0
+
 
 class TestFormatSummary:
     def test_empty(self):