judges use claude-opus-4-8; everything else stays on sonnet-4-6

Both LLM-as-judge sites — the door governor (Runtime::judge) and the topology
comparison scorer (JudgeScorer) — now use the judge model (default
claude-opus-4-8, override CLAWMATES_JUDGE_MODEL). Execution/reasoning turns keep
using the configured default model (claude-sonnet-4-6).

Co-Authored-By: Claude Opus 4.8 <[email protected]>

crates/cm-api/src/routes/topology.rs

@@ -91,10 +91,14 @@ pub async fn compare_topologies(
     Json(req): Json<CompareRequest>,
 ) -> Result<Json<Comparison>, ApiError> {
     let provider = state.runtime.provider();
+    // Execution turns run on the configured default model (sonnet); the judge
+    // uses the stronger judge model (default claude-opus-4-8).
     let model = state.runtime.model().to_string();
+    let judge_model =
+        std::env::var("CLAWMATES_JUDGE_MODEL").unwrap_or_else(|_| "claude-opus-4-8".to_string());
     let executor =
-        ProviderExecutor::new(provider.clone(), model.clone(), state.runtime.max_tokens());
+        ProviderExecutor::new(provider.clone(), model, state.runtime.max_tokens());
-    let scorer = JudgeScorer::new(provider, model, 16);
+    let scorer = JudgeScorer::new(provider, judge_model, 16);
     let cmp = compare(&req.graphs, &req.task, &executor, &scorer)
         .await
         .map_err(|_| ApiError::Internal)?;

crates/cm-runtime/src/runtime.rs

@@ -26,6 +26,14 @@ use crate::tools::{ToolContext, ToolRegistry};
 /// How long a pending approval stays decidable before it expires.
 const APPROVAL_TTL: time::Duration = time::Duration::hours(24);
 
+/// Model used for LLM-as-judge roles (the door governor and the topology
+/// comparison scorer). Judges use the strongest model — default
+/// `claude-opus-4-8` — while everything else runs on the configured default
+/// model (`claude-sonnet-4-6`). Override with `CLAWMATES_JUDGE_MODEL`.
+pub fn judge_model() -> String {
+    std::env::var("CLAWMATES_JUDGE_MODEL").unwrap_or_else(|_| "claude-opus-4-8".to_string())
+}
+
 #[derive(Debug, Clone)]
 pub struct RuntimeConfig {
     pub model: String,
@@ -235,21 +243,23 @@ impl Runtime {
         }
     }
 
-    /// Governor agent: ask the configured model to judge an action. Returns
+    /// Governor agent: ask the **judge model** to judge an action. Returns
     /// `(allow, reason)`. The verdict is the first token (`ALLOW`/`DENY`) of the
     /// model's reply. Best-effort and **fail-open** — if the model is
     /// unreachable it returns `(true, …)` so a governor outage doesn't halt
     /// autonomous agents (the governor is an extra soft check atop deterministic
-    /// policy, not the only gate).
+    /// policy, not the only gate). Judges use the strongest model
+    /// ([`judge_model`], default `claude-opus-4-8`); everything else runs on the
+    /// configured default model.
     pub async fn judge(&self, system: &str, user: &str) -> (bool, String) {
         let request = ChatRequest {
             system: system.to_string(),
+            model: judge_model(),
             messages: vec![ChatMessage {
                 role: ChatRole::User,
                 parts: vec![ContentPart::text(user)],
             }],
             tools: vec![],
-            model: self.inner.config.model.clone(),
             max_tokens: 256,
         };
         let mut text = String::new();