diff --git a/agents/ollama.ts b/agents/ollama.ts
index 1cf8210..51baa55 100644
--- a/agents/ollama.ts
+++ b/agents/ollama.ts
@@ -87,14 +87,28 @@ function pruneToolMessages(messages: Message[], keepRecent = 6): Message[] {
   let pruned = 0;
   const result = messages.map((msg, i) => {
     if (!toPrune.has(i)) return msg;
-    const originalLen = typeof msg.content === "string" ? msg.content.length : 0;
+    const originalLen =
+      typeof msg.content === "string" ? msg.content.length : 0;
     pruned++;
-    return { ...msg, content: `[pruned: was ${originalLen} chars — context limit approached]` };
+    return {
+      ...msg,
+      content: `[pruned: was ${originalLen} chars — context limit approached]`,
+    };
   });
   log.info(`» pruned ${pruned} old tool message(s) to reduce context`);
   return result;
 }
 
+async function unloadModel(ollama: Ollama, model: string): Promise<void> {
+  try {
+    await ollama.generate({ model, keep_alive: 0, prompt: "" });
+    log.info(`» unloaded model ${model} from Ollama`);
+  } catch (err) {
+    const msg = err instanceof Error ? err.message : String(err);
+    log.warning(`» failed to unload model ${model}: ${msg}`);
+  }
+}
+
 async function runOllamaLoop(ctx: AgentRunContext): Promise<AgentResult> {
   const ollamaHost = process.env.OLLAMA_HOST ?? "";
   if (!ollamaHost) {
@@ -162,6 +176,8 @@ async function runOllamaLoop(ctx: AgentRunContext): Promise<AgentResult> {
         },
       );
     } catch (err) {
+      await unloadModel(ollama, model);
+
       const lastError = err instanceof Error ? err.message : String(err);
       log.error(`Ollama error: ${lastError}`);
       return { success: false, error: `Ollama request failed: ${lastError}` };
@@ -208,21 +224,14 @@ async function runOllamaLoop(ctx: AgentRunContext): Promise<AgentResult> {
         continue;
       }
 
+      await unloadModel(ollama, model);
+
       if (pendingModeNudge) {
         log.info("» agent finished after mode nudge (no tool calls)");
       } else {
         log.info("» agent finished (no tool calls)");
       }
 
-      // unload the model from memory immediately since we're done with it, to free up resources for future runs. see #829.
-      try {
-        await ollama.generate({ model, keep_alive: 0, prompt: "" });
-        log.info(`» unloaded model ${model} from Ollama`);
-      } catch (err) {
-        const msg = err instanceof Error ? err.message : String(err);
-        log.warning(`» failed to unload model ${model}: ${msg}`);
-      }
-
       return {
         success: true,
         output: assistantMessage.content || undefined,
@@ -299,16 +308,9 @@ async function runOllamaLoop(ctx: AgentRunContext): Promise<AgentResult> {
     }
   }
 
-  log.warning(`» agent hit max iterations (${MAX_ITERATIONS})`);
+  await unloadModel(ollama, model);
 
-  // unload the model from memory immediately since we're done with it, to free up resources for future runs. see #829.
-  try {
-    await ollama.generate({ model, keep_alive: 0, prompt: "" });
-    log.info(`» unloaded model ${model} from Ollama`);
-  } catch (err) {
-    const msg = err instanceof Error ? err.message : String(err);
-    log.warning(`» failed to unload model ${model}: ${msg}`);
-  }
+  log.warning(`» agent hit max iterations (${MAX_ITERATIONS})`);
 
   return {
     success: false,