chore: some retry logic

agents/ollama.ts

@@ -2,6 +2,7 @@ import { Client } from "@modelcontextprotocol/sdk/client/index.js";
 import { StreamableHTTPClientTransport } from "@modelcontextprotocol/sdk/client/streamableHttp.js";
 import { Ollama, type Message, type ToolCall } from "ollama";
 import { log } from "../utils/cli.ts";
+import { retry } from "../utils/retry.ts";
 import { agent, type AgentResult, type AgentRunContext } from "./shared.ts";
 
 const DEFAULT_MODEL = "qwen3.6:35b";
@@ -113,20 +114,38 @@ async function runOllamaLoop(ctx: AgentRunContext): Promise<AgentResult> {
 
     let response: Awaited<ReturnType<typeof ollama.chat>>;
     try {
-      response = await ollama.chat({
-        model,
-        messages,
-        tools,
-        keep_alive: -1,
-        think: false,
-        options: { num_ctx: 262144 },
-      });
+      response = await retry(
+        () => ollama.chat({
+          model,
+          messages,
+          tools,
+          keep_alive: -1,
+          think: false,
+          options: { num_ctx: 262144 },
+        }),
+        {
+          delaysMs: [3_000, 8_000],
+          shouldRetry: (err) => {
+            const msg = err instanceof Error ? err.message : String(err);
+            return /unexpected EOF|XML syntax error|ECONNRESET|ETIMEDOUT|fetch failed/i.test(msg);
+          },
+          label: `Ollama turn ${iterations}`,
+        },
+      );
     } catch (err) {
       const lastError = err instanceof Error ? err.message : String(err);
       log.error(`Ollama error: ${lastError}`);
       return { success: false, error: `Ollama request failed: ${lastError}` };
     }
 
+    const promptTokens = (response as Record<string, unknown>).prompt_eval_count as number | undefined;
+    const evalTokens = (response as Record<string, unknown>).eval_count as number | undefined;
+    if (promptTokens !== undefined) {
+      const total = promptTokens + (evalTokens ?? 0);
+      const pct = Math.round((total / 262144) * 100);
+      log.info(`  context: ${promptTokens} prompt + ${evalTokens ?? 0} eval = ${total} tokens (${pct}% of limit)`);
+    }
+
     const assistantMessage = response.message;
     messages.push(assistantMessage);