chore: some more improvements to try and get as close to original as possible

agents/ollama.ts

@@ -69,6 +69,32 @@ async function callMcpTool(
   }
 }
 
+/**
+ * When context approaches the limit, truncate the content of old tool-result
+ * messages to free space. Keeps the most recent N tool results intact so the
+ * model still has fresh context; replaces earlier ones with a size notice.
+ * Never touches system/user/assistant messages — only tool messages.
+ */
+function pruneToolMessages(messages: Message[], keepRecent = 6): Message[] {
+  const toolIndices: number[] = [];
+  for (let i = 0; i < messages.length; i++) {
+    if (messages[i].role === "tool") toolIndices.push(i);
+  }
+  const pruneCount = Math.max(0, toolIndices.length - keepRecent);
+  if (pruneCount === 0) return messages;
+
+  const toPrune = new Set(toolIndices.slice(0, pruneCount));
+  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;
+    pruned++;
+    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 runOllamaLoop(ctx: AgentRunContext): Promise<AgentResult> {
   const ollamaHost = process.env.OLLAMA_HOST ?? "";
   if (!ollamaHost) {
@@ -89,7 +115,7 @@ async function runOllamaLoop(ctx: AgentRunContext): Promise<AgentResult> {
   const tools = await getOllamaTools(mcpClient);
   log.info(`» ${tools.length} tools available`);
 
-  const messages: Message[] = [
+  let messages: Message[] = [
     {
       role: "user",
       content: ctx.instructions.full,
@@ -149,6 +175,9 @@ async function runOllamaLoop(ctx: AgentRunContext): Promise<AgentResult> {
       log.info(
         `  context: ${promptTokens} prompt + ${evalTokens ?? 0} eval = ${total} tokens (${pct}% of limit)`,
       );
+      if (promptTokens > 200_000) {
+        messages = pruneToolMessages(messages);
+      }
     }
 
     const assistantMessage = response.message;