feat(agent): auto-continue on silent-fail tool calls too

Promissory-pattern catch covered 'let me try another approach' but
missed the related failure: agent gave a confident final answer
('There are no orphan nodes') based on a tool call that exited
non-zero with empty stdout (e.g. 'graph json | grep -q ...').

Add a second auto-continue trigger: last tool exit != 0 AND last
tool stdout is empty AND model wrote a non-empty final answer.
Forces the model to re-derive its conclusion via a different query
(e.g. run-js with explicit return) before answering.

Author: snomiao

src/agent/llm/session.ts

@@ -589,34 +589,51 @@ export async function streamSession(
     combinedText += finalText
     const toolsDuringTurn = toolCalls.length - before
 
-    // Auto-continue heuristic: the model wrote a promissory phrase
-    // ("let me try another approach", "I'll investigate further", …)
-    // but stopped emitting tool calls. Inject a continue prompt and
-    // re-stream — up to MAX_AUTOCONTINUE times.
+    // Auto-continue heuristics — force another tool round when:
+    //   (a) the model wrote a promissory phrase but stopped, OR
+    //   (b) the model gave a confident final answer that rests on a
+    //       tool call which produced no stdout and a non-zero exit
+    //       (the silent-grep / empty-pipe failure mode).
+    const lastTool = toolCalls[toolCalls.length - 1]
+    const silentFail =
+      lastTool !== undefined &&
+      lastTool.exitCode !== 0 &&
+      lastTool.stdout.trim().length === 0
+    const promissory =
+      finalText.length > 0 && PROMISSORY_PATTERN.test(finalText)
     const stalled =
-      finalText.length > 0 &&
-      PROMISSORY_PATTERN.test(finalText) &&
-      // If the model just made a tool call this turn AND ended with the
-      // promise, it's mid-iteration — give it one more push. If it
-      // didn't make any tool call this turn, it's definitely stuck.
-      autoContinues < MAX_AUTOCONTINUE
+      autoContinues < MAX_AUTOCONTINUE &&
+      (promissory || (silentFail && finalText.length > 0))
 
     if (!stalled) break
 
     autoContinues++
+    let nudge: string
+    if (silentFail && !promissory) {
+      nudge =
+        '[auto-continue] Your last tool call exited ' +
+        String(lastTool?.exitCode) +
+        ' with empty stdout — your final answer rests on a query that ' +
+        'returned nothing. Verify your conclusion with a different ' +
+        'approach (e.g. run-js with a return statement) before answering. ' +
+        'Do not narrate now — run the verifying tool call.'
+    } else if (toolsDuringTurn === 0) {
+      nudge =
+        '[auto-continue] You promised follow-up action but stopped without ' +
+        'executing it. Run the next tool call now — do not narrate, act.'
+    } else {
+      nudge =
+        '[auto-continue] Your last turn surfaced something that needs ' +
+        'another step. Take that step now via run_shell — no preamble.'
+    }
     messages = [
       ...messages,
       { role: 'assistant', content: finalText },
-      {
-        role: 'user',
-        content:
-          toolsDuringTurn === 0
-            ? '[auto-continue] You promised follow-up action but stopped without executing it. Run the next tool call now — do not narrate, act.'
-            : '[auto-continue] Your last turn surfaced something that needs another step. Take that step now via run_shell — no preamble.'
-      }
+      { role: 'user', content: nudge }
     ]
+    const reason = silentFail && !promissory ? 'silent-fail' : 'promissory'
     onDelta(
-      `\n\n[agent: auto-continuing — promissory stop detected, attempt ${autoContinues}/${MAX_AUTOCONTINUE}]\n\n`
+      `\n\n[agent: auto-continuing — ${reason} detected, attempt ${autoContinues}/${MAX_AUTOCONTINUE}]\n\n`
     )
   }