fix(provider): retry truncated streams without terminal marker

packages/moss-agent/src/core/llm/llm-error-classifier.ts

@@ -59,7 +59,7 @@ function isAbortLike(message: string): boolean {
 }
 
 function isPrematureClose(message: string): boolean {
-  return /err_stream_premature_close|premature close|stream closed prematurely|other side closed|stream.*terminated|^(?:llm\s+stream\s+error:\s*)?terminated\.?$/i.test(
+  return /err_stream_premature_close|premature close|stream closed prematurely|other side closed|stream.*terminated|stream (?:ended|terminated) without (?:\[done\]\s*(?:or\s*)?)?finish_reason|^(?:llm\s+stream\s+error:\s*)?terminated\.?$/i.test(
     message
   );
 }

packages/moss-agent/src/provider/error-classify.ts

@@ -221,6 +221,7 @@ function matchOpaqueStreamConnectionDrop(msg: string): boolean {
     m === 'connection error.' ||
     /^(?:llm\s+stream\s+error:\s*)?terminated\.?$/i.test(msg.trim()) ||
     /^(?:llm\s+stream\s+error:\s*)?connection error\.?$/i.test(msg.trim()) ||
+    /stream (?:ended|terminated) without (?:\[done\]\s*(?:or\s*)?)?finish_reason/i.test(msg) ||
     /terminated.*other side closed|other side closed|stream.*terminated/i.test(m)
   );
 }

packages/moss-agent/test/error-classify-500.spec.mjs

@@ -29,6 +29,17 @@ for (const status of [500, 502, 503, 504, 529]) {
   assert.equal(r.retryable, true, `HTTP ${status} is retryable`);
 }
 
+// ─── truncated successful HTTP streams are also transient ─────────────────
+
+for (const errorMessage of [
+  'LLM stream incomplete: Stream ended without finish_reason',
+  'OpenAI provider: stream terminated without [DONE] or finish_reason',
+]) {
+  const r = classifyProviderError({ errorMessage });
+  assert.equal(r.category, 'service_unavailable', 'missing terminal marker is a gateway interruption');
+  assert.equal(r.retryable, true, 'missing terminal marker is retried');
+}
+
 // ─── non-retryable errors stay non-retryable (no false regression) ─────────
 
 {

packages/moss-agent/test/provider-retry-contract.spec.mjs

@@ -99,4 +99,32 @@ function makeAgent(provider, maxLLMRetries = 1) {
   assert.equal(calls, 2, '429 gets one bounded retry, not nested retry loops');
 }
 
+{
+  let calls = 0;
+  const provider = {
+    id: 'missing-terminal-marker', displayName: 'missing-terminal-marker',
+    async complete() { throw new Error('not used'); },
+    async stream(_options, onEvent) {
+      calls++;
+      if (calls === 1) {
+        throw new Error('LLM stream incomplete: Stream ended without finish_reason');
+      }
+      onEvent({ type: 'message_start' });
+      onEvent({ type: 'content_block_start' });
+      onEvent({ type: 'content_block_delta', text: 'recovered after truncated stream' });
+      onEvent({ type: 'content_block_stop' });
+      onEvent({ type: 'message_delta', stopReason: 'end_turn' });
+      onEvent({ type: 'message_stop' });
+      return {
+        stopReason: 'end_turn',
+        content: [{ type: 'text', text: 'recovered after truncated stream' }],
+      };
+    },
+  };
+  const agent = makeAgent(provider);
+  const result = await agent.chat('missing-terminal-marker', 'hello');
+  assert.equal(result.response, 'recovered after truncated stream');
+  assert.equal(calls, 2, 'a stream missing finish_reason is retried inside the same agent run');
+}
+
 console.log('[PASS] provider auth and rate-limit retry contracts');