[BURN #197] feat: provenance chain — add source_session, source_model, source_provider, timestamp to facts

- Adds schemas/provenance.json defining provenance object schema
- Updates harvester.py write_knowledge() to attach provenance to every fact
- Adds extract_provider() helper to infer provider from API base URL
- Updates SCHEMA.md documenting provenance field and object
- Provenance fields: source_session, source_model, source_provider,
  timestamp (harvested_at), extraction_method, confidence, verified
- Does NOT retroactively modify existing facts in index.json
Closes #197

knowledge/SCHEMA.md

@@ -43,9 +43,26 @@ The harvester writes to both. The bootstrapper reads from index.json. Humans edi
 | `last_confirmed` | date | no | ISO-8601 date last seen in a session |
 | `expires` | date | no | Optional. After this date, fact is stale |
 | `related` | string[] | no | IDs of related facts |
+| `provenance` | object | no | Provenance metadata — see Provenance Object section below |
 
 ### ID Format: `{domain}:{category}:{sequence}`
 
+
+
+### Provenance Object
+
+Every fact may include a [`provenance`](#fact-object) field that tracks its origin.
+
+| Field | Type | Required | Description |
+|-------|------|----------|-------------|
+| `source_session` | string | yes | Session ID / file path where this fact was extracted |
+| `source_model` | string | yes | Model name used for extraction (e.g., `xiaomi/mimo-v2-pro`) |
+| `source_provider` | string | yes | Provider name (`nous`, `openrouter`, `anthropic`, `openai`, etc.) |
+| `timestamp` | date-time | yes | Extraction timestamp (ISO-8601 UTC) |
+| `extraction_method` | enum | yes | `llm_extraction`, `manual`, or `retroactive_harvest` |
+| `confidence` | float | yes | Confidence at extraction time (0.0–1.0) |
+| `verified` | boolean | yes | `true` if fact has been manually reviewed, else `false` |
+
 ### Categories
 
 | Category | Definition |
@@ -85,6 +102,35 @@ knowledge/
     └── {agent-type}.yaml
 ```
 
+
+
+### Provenance Object (added via `write_knowledge()` and harvester)
+
+```json
+{
+  "source_session": "string — session ID or file path",
+  "source_model": "string — model used for extraction",
+  "source_provider": "string — provider name (nous, openrouter, etc.)",
+  "timestamp": "string — ISO-8601 UTC extraction time",
+  "extraction_method": "string — llm_extraction|manual|retroactive_harvest",
+  "confidence": "float — 0.0–1.0 confidence from extraction",
+  "verified": "boolean — whether fact has been manually verified"
+}
+```
+
+The `provenance` field is attached to every fact harvested via `write_knowledge()`. It provides traceability: which session produced this fact, which model/provider extracted it, when, and with what confidence.
+
+| Provenance Field | Type | Required | Description |
+|------------------|------|----------|-------------|
+| `source_session` | string | yes | Session ID / file path where extracted |
+| `source_model` | string | yes | Model name (e.g., `xiaomi/mimo-v2-pro`) |
+| `source_provider` | string | yes | Provider (`nous`, `openrouter`, `anthropic`, `openai`) |
+| `timestamp` | date-time | yes | Extraction timestamp (ISO-8601) |
+| `extraction_method` | enum | yes | `llm_extraction`, `manual`, or `retroactive_harvest` |
+| `confidence` | float | yes | Confidence score (0.0–1.0) at extraction time |
+| `verified` | boolean | yes | `true` if manually reviewed, else `false` |
+
+
 ## YAML File Format
 
 YAML files use frontmatter for metadata, then markdown sections with fact entries:

schemas/provenance.json

@@ -0,0 +1,52 @@
+{
+  "$schema": "http://json-schema.org/draft-07/schema#",
+  "title": "Knowledge Provenance",
+  "description": "Provenance metadata attached to every knowledge fact",
+  "type": "object",
+  "required": [
+    "source_session",
+    "source_model",
+    "source_provider",
+    "timestamp"
+  ],
+  "properties": {
+    "source_session": {
+      "type": "string",
+      "description": "Session ID or file path where this fact was extracted"
+    },
+    "source_model": {
+      "type": "string",
+      "description": "Model used for extraction (e.g., 'xiaomi/mimo-v2-pro')"
+    },
+    "source_provider": {
+      "type": "string",
+      "description": "Provider name (nous, openrouter, anthropic, etc.)"
+    },
+    "timestamp": {
+      "type": "string",
+      "format": "date-time",
+      "description": "UTC ISO-8601 timestamp when this fact was extracted"
+    },
+    "extraction_method": {
+      "type": "string",
+      "description": "How the fact was extracted (llm_extraction, manual, retroactive_harvest)",
+      "enum": [
+        "llm_extraction",
+        "manual",
+        "retroactive_harvest"
+      ],
+      "default": "llm_extraction"
+    },
+    "confidence": {
+      "type": "number",
+      "minimum": 0,
+      "maximum": 1,
+      "description": "Confidence assigned during extraction (copied from top-level fact)"
+    },
+    "verified": {
+      "type": "boolean",
+      "description": "Whether this fact has been manually verified",
+      "default": false
+    }
+  }
+}

scripts/harvester.py

@@ -27,6 +27,22 @@ sys.path.insert(0, str(SCRIPT_DIR))
 
 from session_reader import read_session, extract_conversation, truncate_for_context, messages_to_text
 
+def extract_provider(api_base: str) -> str:
+    """Infer provider name from API base URL."""
+    url = api_base.lower()
+    if 'nousresearch' in url or 'nous' in url:
+        return 'nous'
+    if 'openrouter' in url:
+        return 'openrouter'
+    if 'anthropic' in url:
+        return 'anthropic'
+    if 'openai' in url:
+        return 'openai'
+    # Fallback: try to extract hostname
+    from urllib.parse import urlparse
+    host = urlparse(api_base).netloc
+    return host.split('.')[0] if host else 'unknown' 
+
 # --- Configuration ---
 
 DEFAULT_API_BASE = os.environ.get("HARVESTER_API_BASE", "https://api.nousresearch.com/v1")
@@ -229,15 +245,34 @@ def validate_fact(fact: dict) -> bool:
     return True
 
 
-def write_knowledge(index: dict, new_facts: list[dict], knowledge_dir: str, source_session: str = ""):
-    """Write new facts to the knowledge store."""
+def write_knowledge(index: dict, new_facts: list[dict], knowledge_dir: str, source_session: str = "", model: str = "", provider: str = ""):
+    """Write new facts to the knowledge store.
+    
+    Adds provenance metadata to each fact. If model/provider are empty, tries to
+    infer from environment or defaults.
+    """
     kdir = Path(knowledge_dir)
     kdir.mkdir(parents=True, exist_ok=True)
     
-    # Add source tracking to each fact
+    # Determine model/provider defaults if not provided
+    model = model or os.environ.get("HARVESTER_MODEL", "xiaomi/mimo-v2-pro")
+    provider = provider or os.environ.get("HARVESTER_PROVIDER", "nous")
+    
+    timestamp = datetime.now(timezone.utc).isoformat()
+    
+    # Add provenance to each fact
     for fact in new_facts:
-        fact['source_session'] = source_session
-        fact['harvested_at'] = datetime.now(timezone.utc).isoformat()
+        provenance = {
+            'source_session': source_session,
+            'source_model': model,
+            'source_provider': provider,
+            'timestamp': timestamp,
+            'extraction_method': 'llm_extraction',
+            'confidence': fact.get('confidence', 0.5),
+            'verified': False
+        }
+        fact['provenance'] = provenance
+        fact['harvested_at'] = timestamp
     
     # Update index
     index['facts'].extend(new_facts)
@@ -330,7 +365,7 @@ def harvest_session(session_path: str, knowledge_dir: str, api_base: str, api_ke
         
         # 8. Write (unless dry run)
         if new_facts and not dry_run:
-            write_knowledge(existing_index, new_facts, knowledge_dir, source_session=session_path)
+            write_knowledge(existing_index, new_facts, knowledge_dir, source_session=session_path, model=model, provider=extract_provider(api_base))
         
         stats['elapsed_seconds'] = round(time.time() - start_time, 2)
         return stats