Add per-agent model assignments (agent_models table)

Festinger now reads X-Agent-Name from every intercepted request and
resolves the utility LLM model in priority order:
  1. agent_models table  — agent-specific (e.g. gunnar → claude, rind → qwen)
  2. write_model_id config — global default
  3. Request mirror       — same provider/model Agent0 is currently using

New API: GET/PUT/DELETE /agent-models
New admin UI: "Agent models" section with assignment form and table.

Agent0 side: add a custom header X-Agent-Name: <name> in the LLM
provider config per agent container (AGENT_NAME env var can drive this).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

plugins/festinger/db/schema.sql

@@ -111,3 +111,14 @@ CREATE TABLE IF NOT EXISTS kg_write_log (
 CREATE INDEX IF NOT EXISTS kwl_created_idx ON kg_write_log (created_at DESC);
 CREATE INDEX IF NOT EXISTS kwl_concept_idx ON kg_write_log (concept_id);
 CREATE INDEX IF NOT EXISTS kwl_op_idx      ON kg_write_log (op);
+
+-- ---------------------------------------------------------------------------
+-- agent_models — per-agent LLM model assignments
+-- Maps an agent identity (from X-Agent-Name header) to a specific model.
+-- Priority over write_model_id (global default) when agent_name is present.
+-- ---------------------------------------------------------------------------
+CREATE TABLE IF NOT EXISTS agent_models (
+    agent_name  TEXT PRIMARY KEY,              -- normalised lowercase, e.g. 'gunnar', 'rind'
+    model_id    INT  NOT NULL REFERENCES models(id) ON DELETE CASCADE,
+    created_at  TIMESTAMPTZ NOT NULL DEFAULT now()
+);

plugins/festinger/festinger/db.py

@@ -60,6 +60,16 @@ async def init_schema(pool: asyncpg.Pool) -> None:
         await conn.execute(
             "ALTER TABLE soas ADD COLUMN IF NOT EXISTS first_seen_context TEXT NOT NULL DEFAULT ''"
         )
+        # Migration: per-agent model assignments (agent_name → model_id)
+        await conn.execute(
+            """
+            CREATE TABLE IF NOT EXISTS agent_models (
+                agent_name  TEXT PRIMARY KEY,
+                model_id    INT  NOT NULL REFERENCES models(id) ON DELETE CASCADE,
+                created_at  TIMESTAMPTZ NOT NULL DEFAULT now()
+            )
+            """
+        )
     log.info("schema applied")
 
 

plugins/festinger/festinger/main.py

@@ -578,6 +578,11 @@ def _extract_request_model_config(
 # Saliency + recollection pipeline
 # ---------------------------------------------------------------------------
 
+def _agent_name_from_headers(headers: dict) -> str:
+    """Extract and normalise the agent identity from X-Agent-Name header."""
+    return headers.get("x-agent-name", "").strip().lower()
+
+
 async def process_prompt(body: dict, path: str, pool, cfg: dict, request_headers: dict | None = None) -> dict:
     """
     Run the saliency + recollection pipeline over the prompt.
@@ -587,9 +592,11 @@ async def process_prompt(body: dict, path: str, pool, cfg: dict, request_headers
     conf_floor = float(await get_config(pool, "recollection_confidence_floor", "0.6"))
     recency_days = int(await get_config(pool, "recollection_recency_days", "90"))
 
+    hdrs = request_headers or {}
     # Derive a ModelConfig from the intercepted request so context discovery can
     # mirror Agent0's current model without a separate write_model_id config.
-    request_model = _extract_request_model_config(path, body, request_headers or {}, cfg)
+    request_model = _extract_request_model_config(path, body, hdrs, cfg)
+    agent_name = _agent_name_from_headers(hdrs)
 
     # Extract only the last user message — agent responses and reasoning traces
     # are noise for both cue scanning and concept discovery.
@@ -651,7 +658,11 @@ async def process_prompt(body: dict, path: str, pool, cfg: dict, request_headers
 
     # 3. Enqueue for LLM-driven discovery if there are candidates to evaluate.
     if novel_candidates and len(user_text) >= 20:
-        await enqueue_context_discover(user_text, novel_candidates, fallback_model=request_model)
+        await enqueue_context_discover(
+            user_text, novel_candidates,
+            agent_name=agent_name,
+            fallback_model=request_model,
+        )
 
     if not salient_for_read:
         return body
@@ -964,6 +975,81 @@ async def delete_model(model_id: int, request: Request) -> dict:
     return {"status": "ok", "deleted": model_id}
 
 
+# ---------------------------------------------------------------------------
+# /agent-models — per-agent model assignments
+# ---------------------------------------------------------------------------
+
+@app.get("/agent-models")
+async def list_agent_models(request: Request) -> dict:
+    pool = request.app.state.pool
+    async with pool.acquire() as conn:
+        rows = await conn.fetch(
+            """
+            SELECT am.agent_name, am.model_id, am.created_at,
+                   m.provider, m.model_name, m.base_url
+            FROM agent_models am
+            JOIN models m ON m.id = am.model_id
+            ORDER BY am.agent_name
+            """
+        )
+    return {"agent_models": [
+        {
+            "agent_name": r["agent_name"],
+            "model_id": r["model_id"],
+            "provider": r["provider"],
+            "model_name": r["model_name"],
+            "base_url": r["base_url"] or "",
+            "created_at": r["created_at"].isoformat(),
+        }
+        for r in rows
+    ]}
+
+
+@app.put("/agent-models/{agent_name}")
+async def set_agent_model(agent_name: str, request: Request) -> dict:
+    """
+    Assign a model to an agent.  agent_name is normalised to lowercase.
+    Body: {"model_id": 3}
+    """
+    pool = request.app.state.pool
+    data = await request.json()
+    model_id = data.get("model_id")
+    if not model_id:
+        return {"error": "model_id is required"}
+    name = agent_name.strip().lower()
+    if not name:
+        return {"error": "agent_name must not be empty"}
+    async with pool.acquire() as conn:
+        # Verify the model exists
+        row = await conn.fetchrow("SELECT id, provider, model_name FROM models WHERE id=$1", int(model_id))
+        if not row:
+            return {"error": f"model {model_id} not found"}
+        await conn.execute(
+            """
+            INSERT INTO agent_models (agent_name, model_id)
+            VALUES ($1, $2)
+            ON CONFLICT (agent_name) DO UPDATE SET model_id = EXCLUDED.model_id, created_at = now()
+            """,
+            name, int(model_id),
+        )
+    log.info("agent_model set  agent=%s model_id=%s provider=%s model=%s",
+             name, model_id, row["provider"], row["model_name"])
+    return {"status": "ok", "agent_name": name, "model_id": int(model_id)}
+
+
+@app.delete("/agent-models/{agent_name}")
+async def delete_agent_model(agent_name: str, request: Request) -> dict:
+    pool = request.app.state.pool
+    name = agent_name.strip().lower()
+    async with pool.acquire() as conn:
+        result = await conn.execute("DELETE FROM agent_models WHERE agent_name=$1", name)
+    deleted = int(result.split()[-1]) if result else 0
+    if not deleted:
+        return {"error": f"no assignment found for agent '{name}'"}
+    log.info("agent_model deleted  agent=%s", name)
+    return {"status": "ok", "deleted": name}
+
+
 @app.get("/models/discover")
 async def discover_models(base_url: str = "http://host.docker.internal:1234") -> dict:
     """
@@ -1935,6 +2021,31 @@ ADMIN_HTML = """<!DOCTYPE html>
     </details>
   </div>
 
+  <h2>Agent models</h2>
+  <p style="font-size:0.83em;color:#666;margin-bottom:0.8em">
+    Assign a model per agent. Agent0 must send <code>X-Agent-Name: &lt;name&gt;</code> on every request.
+    Takes priority over the global write model.
+  </p>
+  <table id="agent-models-table" style="margin-bottom:0.8em">
+    <thead><tr><th>Agent</th><th>Model ID</th><th>Provider</th><th>Model name</th><th></th></tr></thead>
+    <tbody id="agent-models-tbody"><tr><td colspan="5">Loading…</td></tr></tbody>
+  </table>
+  <details style="margin-bottom:1em">
+    <summary style="cursor:pointer;font-size:0.9em;color:#555">Assign model to agent…</summary>
+    <div style="margin-top:0.6em;display:flex;gap:0.7em;flex-wrap:wrap;align-items:flex-end">
+      <label style="font-size:0.85em">Agent name (e.g. gunnar)
+        <input id="am-agent" type="text" placeholder="gunnar"
+               style="font-family:monospace;padding:5px 8px;border:1px solid #ccc;border-radius:3px;display:block;margin-top:2px;width:140px">
+      </label>
+      <label style="font-size:0.85em">Model
+        <select id="am-model" style="font-family:monospace;padding:5px 8px;border:1px solid #ccc;border-radius:3px;display:block;margin-top:2px">
+          <option value="">— select —</option>
+        </select>
+      </label>
+      <button onclick="assignAgentModel(this)" style="height:32px">Assign</button>
+    </div>
+  </details>
+
   <h2>Actions</h2>
   <div class="actions">
     <button class="primary" onclick="runResolution(this)">Run conflict resolution now</button>
@@ -2011,6 +2122,11 @@ ADMIN_HTML = """<!DOCTYPE html>
           <td><button onclick="setConfig('write_model_id','${{m.id}}')" style="padding:2px 8px;font-size:0.8em;${{writeId==String(m.id)?'background:#2a7a2a;color:#fff;border-color:#2a7a2a':''}}">${{writeId==String(m.id)?'✓ active':'set'}}</button></td>
           <td><button onclick="deleteModel(${{m.id}},this)" style="padding:2px 8px;font-size:0.8em;color:#b00;border-color:#b00">✕</button></td>
         </tr>`).join('');
+
+      // Populate model dropdown in agent-models assignment form
+      const sel = document.getElementById('am-model');
+      sel.innerHTML = '<option value="">— select —</option>' +
+        md.models.map(m => `<option value="${{m.id}}">${{m.id}} — ${{m.provider}} / ${{m.model_name}}</option>`).join('');
     }}
 
     async function addModel(btn) {{
@@ -2052,6 +2168,57 @@ ADMIN_HTML = """<!DOCTYPE html>
       await loadModels();
     }}
 
+    async function loadAgentModels() {{
+      const r = await fetch('/agent-models');
+      const d = await r.json();
+      const tbody = document.getElementById('agent-models-tbody');
+      if (!d.agent_models.length) {{
+        tbody.innerHTML = '<tr><td colspan="5" style="color:#999">No assignments yet.</td></tr>';
+        return;
+      }}
+      tbody.innerHTML = d.agent_models.map(a => `
+        <tr>
+          <td><strong>${{a.agent_name}}</strong></td>
+          <td>${{a.model_id}}</td>
+          <td>${{a.provider}}</td>
+          <td>${{a.model_name}}${{a.base_url ? ' <span style="color:#999;font-size:0.85em">(' + a.base_url + ')</span>' : ''}}</td>
+          <td><button onclick="removeAgentModel('${{a.agent_name}}',this)" style="padding:2px 8px;font-size:0.8em;color:#b00;border-color:#b00">✕</button></td>
+        </tr>`).join('');
+    }}
+
+    async function assignAgentModel(btn) {{
+      const agent = document.getElementById('am-agent').value.trim().toLowerCase();
+      const model_id = document.getElementById('am-model').value;
+      if (!agent) {{ alert('Enter an agent name.'); return; }}
+      if (!model_id) {{ alert('Select a model.'); return; }}
+      btn.disabled = true;
+      try {{
+        const r = await fetch('/agent-models/' + encodeURIComponent(agent), {{
+          method: 'PUT',
+          headers: {{'Content-Type': 'application/json'}},
+          body: JSON.stringify({{model_id: parseInt(model_id)}})
+        }});
+        const d = await r.json();
+        if (d.error) {{ showResult('Error: ' + d.error, false); return; }}
+        showResult('Assigned model ' + model_id + ' to agent "' + agent + '".', true);
+        document.getElementById('am-agent').value = '';
+        await loadAgentModels();
+      }} catch(e) {{ showResult('Error: ' + e.message, false); }}
+      finally {{ btn.disabled = false; }}
+    }}
+
+    async function removeAgentModel(agent, btn) {{
+      if (!confirm('Remove model assignment for "' + agent + '"?')) return;
+      btn.disabled = true;
+      try {{
+        const r = await fetch('/agent-models/' + encodeURIComponent(agent), {{method: 'DELETE'}});
+        const d = await r.json();
+        if (d.error) {{ showResult('Error: ' + d.error, false); return; }}
+        await loadAgentModels();
+      }} catch(e) {{ showResult('Error: ' + e.message, false); }}
+      finally {{ btn.disabled = false; }}
+    }}
+
     async function loadConflicts() {{
       const r = await fetch('/conflicts');
       const d = await r.json();
@@ -2200,6 +2367,7 @@ ADMIN_HTML = """<!DOCTYPE html>
     loadConflicts();
     loadLog(0);
     loadModels();
+    loadAgentModels();
   </script>
 </body>
 </html>

plugins/festinger/festinger/write_queue.py

@@ -32,15 +32,18 @@ class ContextDiscoverRequest:
     User message text plus candidate tokens (words absent from the English
     dictionary) submitted for LLM-driven concept discovery.
 
-    candidate_tokens are hints for the LLM — it decides which are real domain
+    candidate_tokens: hints for the LLM — it decides which are real domain
-    concepts vs typos/noise, and extracts relationship triples from the text.
+      concepts vs typos/noise, and extracts relationship triples from the text.
 
-    fallback_model: if set, used when write_model_id is not configured in the
+    agent_name: normalised agent identity from X-Agent-Name header (lowercase).
-    DB.  Allows Festinger to mirror the model Agent0 is currently using without
+      Used to look up an agent-specific model in the agent_models table.
-    any extra configuration.
+
+    fallback_model: last-resort ModelConfig derived from the intercepted request.
+      Used when neither an agent-specific model nor write_model_id is configured.
     """
     user_text: str
     candidate_tokens: list[str]
+    agent_name: str = ""
     fallback_model: Optional[ModelConfig] = None
 
 
@@ -62,19 +65,23 @@ _llm_semaphore: asyncio.Semaphore | None = None
 async def enqueue_context_discover(
     user_text: str,
     candidate_tokens: list[str],
+    agent_name: str = "",
     fallback_model: Optional[ModelConfig] = None,
 ) -> None:
     """
     Enqueue a user message for LLM-driven concept discovery and relation extraction.
 
-    fallback_model: ModelConfig derived from the intercepted request.  Used when
+    agent_name: normalised agent identity (from X-Agent-Name header).
-    write_model_id is not configured — lets Festinger mirror Agent0's model
+      Festinger looks up an agent-specific model in the agent_models table.
-    without any extra setup.
+
+    fallback_model: last-resort ModelConfig from the intercepted request.
+      Used when neither an agent-specific model nor write_model_id is configured.
     """
     try:
         _queue.put_nowait(ContextDiscoverRequest(
             user_text=user_text,
             candidate_tokens=candidate_tokens,
+            agent_name=agent_name,
             fallback_model=fallback_model,
         ))
     except asyncio.QueueFull:
@@ -119,7 +126,8 @@ async def _worker(pool: asyncpg.Pool) -> None:
                 # Slow path: fire off without awaiting so the worker stays free.
                 asyncio.create_task(
                     _process_context_discover_guarded(
-                        pool, item.user_text, item.candidate_tokens, item.fallback_model
+                        pool, item.user_text, item.candidate_tokens,
+                        item.agent_name, item.fallback_model,
                     )
                 )
         except Exception as e:
@@ -132,13 +140,14 @@ async def _process_context_discover_guarded(
     pool: asyncpg.Pool,
     user_text: str,
     candidate_tokens: list[str],
+    agent_name: str = "",
     fallback_model: Optional[ModelConfig] = None,
 ) -> None:
     """Wrapper that acquires the LLM semaphore before concept discovery."""
     assert _llm_semaphore is not None
     async with _llm_semaphore:
         try:
-            await _process_context_discover(pool, user_text, candidate_tokens, fallback_model)
+            await _process_context_discover(pool, user_text, candidate_tokens, agent_name, fallback_model)
         except Exception as e:
             log.exception("context discover task error: %s", e)
 
@@ -171,20 +180,74 @@ async def _process_cue(pool: asyncpg.Pool, triple: CueTriple) -> None:
         log.info("cue triple collision: %s", collision)
 
 
+async def _resolve_discover_model(
+    pool: asyncpg.Pool,
+    agent_name: str,
+    fallback_model: Optional[ModelConfig],
+) -> Optional[ModelConfig]:
+    """
+    Resolve which LLM to use for context discovery.
+
+    Priority:
+      1. Agent-specific model  — agent_models table, keyed by agent_name
+      2. Global default        — write_model_id config key
+      3. Request mirror        — fallback_model (same provider/model Agent0 used)
+    """
+    if agent_name:
+        async with pool.acquire() as conn:
+            row = await conn.fetchrow(
+                """
+                SELECT m.provider, m.model_name, m.api_key, m.base_url
+                FROM agent_models am
+                JOIN models m ON m.id = am.model_id
+                WHERE am.agent_name = $1
+                """,
+                agent_name,
+            )
+        if row:
+            log.debug(
+                "context discover: agent=%s → provider=%s model=%s",
+                agent_name, row["provider"], row["model_name"],
+            )
+            return ModelConfig(
+                provider=row["provider"],
+                model_name=row["model_name"],
+                api_key=row["api_key"],
+                base_url=row["base_url"] or "",
+            )
+
+    write_model_id = await get_config(pool, "write_model_id")
+    if write_model_id:
+        m = await get_model_config(pool, write_model_id)
+        if m:
+            return m
+        log.warning("write_model_id=%s not found in models table", write_model_id)
+
+    if fallback_model:
+        log.debug(
+            "context discover: mirroring request model provider=%s model=%s",
+            fallback_model.provider, fallback_model.model_name,
+        )
+        return fallback_model
+
+    return None
+
+
 async def _process_context_discover(
     pool: asyncpg.Pool,
     user_text: str,
     candidate_tokens: list[str],
+    agent_name: str = "",
     fallback_model: Optional[ModelConfig] = None,
 ) -> None:
     """
     Ask the LLM to evaluate candidate tokens (dictionary misses) and extract
     relationship triples for those it judges to be real domain concepts.
 
-    Model selection priority:
+    Model selection priority (see _resolve_discover_model):
-      1. write_model_id config key → explicit utility model override (DB models table)
+      1. Agent-specific model (agent_models table)
-      2. fallback_model              → mirrors the model Agent0 used for this request
+      2. write_model_id config (global default)
-      3. Neither set                 → skip (log debug)
+      3. Request mirror (fallback_model)
 
     Saliency feedback loop:
       - Confirmed concepts (triples inserted) → saliency raised to NOVEL_CONFIRMED_SALIENCY
@@ -192,20 +255,9 @@ async def _process_context_discover(
       - Rejected concepts (no triples) → saliency stays low (NOVEL_INITIAL_SALIENCY),
         effectively hiding typos and noise from the recollection engine.
     """
-    write_model_id = await get_config(pool, "write_model_id")
+    model = await _resolve_discover_model(pool, agent_name, fallback_model)
-    if write_model_id:
+    if not model:
-        model = await get_model_config(pool, write_model_id)
+        log.debug("no model resolved for context discover (agent=%r) — skipping", agent_name)
-        if not model:
-            log.warning("write_model_id=%s not found in models table", write_model_id)
-            return
-    elif fallback_model:
-        model = fallback_model
-        log.debug(
-            "context discover: using request model provider=%s model=%s",
-            model.provider, model.model_name,
-        )
-    else:
-        log.debug("no write_model_id configured and no request model available — skipping context discover")
         return
 
     seed_dims = ["type", "membership", "runs-on", "tech", "owned-by", "geography"]