fix: address agent pipeline process failure when no tools are used

1. presenter.ts: Stop echoing task goals when cleanOutput() strips everything and no files were changed; properly delegate to presentFailure().

2. repairLoop.ts: If QC flags TOOL_MISSING, inject a strong directive in the repair task telling the agent it MUST call tools.

3. ReactAgentAdapter.ts: Relax the isSubstantive filter to not discard multi-paragraph text that starts with a thinking opener.

Author: junkyard22

apps/desktop/src/agents/ReactAgentAdapter.ts

@@ -595,8 +595,14 @@ export class ReactAgentAdapter implements AgentAdapter {
           // would surface internal monologue to the user if the loop later exits
           // at max_iterations.
           const extractedText = stripThoughtBlocks(cleanedOutput);
-          const isSubstantive = extractedText.trim().length > 0
-            && !/^(Good|OK|Alright|Let me|I('ll| will| need| want| should)|Now |First)/im.test(extractedText.trim());
+          // Only reject short single-sentence thinking preambles.  Multi-paragraph
+          // text that starts with a thinking opener may contain real content after
+          // the first sentence — don't discard it.
+          const trimmed = extractedText.trim();
+          const looksLikePureThinking = trimmed.length < 500
+            && !/\n/.test(trimmed)
+            && /^(Good|OK|Alright|Let me|I('ll| will| need| want| should)|Now |First)/im.test(trimmed);
+          const isSubstantive = trimmed.length > 0 && !looksLikePureThinking;
           if (isSubstantive) {
             currentOutputText = extractedText;
           }

packages/benson-core/src/presenter.ts

@@ -249,12 +249,11 @@ function presentSuccess(result: ExecutionResult, task: TaskSpec): string {
     return `Done. Created or updated:\n\n${fileList}`;
   }
 
-  // Last resort — construct a concise confirmation from the task goals.
-  const summary = task.goals.length === 1
-    ? task.goals[0]
-    : task.goals.map((g, i) => `${i + 1}. ${g}`).join("\n");
-
-  return `Done. Here is what was completed:\n\n${summary}`;
+  // Last resort — cleanOutput stripped everything and no files were changed.
+  // This means the agent produced no useful work.  Do NOT echo the task goals
+  // back ("Done. Here is what was completed:") — that's deceptive.  Instead
+  // delegate to the failure presenter so the user gets an honest answer.
+  return presentFailure(result, task);
 }
 
 function presentWarn(result: ExecutionResult, task: TaskSpec): string {

packages/orca-core/src/repairLoop.ts

@@ -74,8 +74,22 @@ export async function handleRepairLoop(
     // Prefer the AHP-based targeted repair prompt (set by Pappy's verifyAHPPacket)
     // over the legacy generic repairTask string when available.
     const repairInstruction = ahpPacket?.repairPrompt ?? currentQC.repairTask!;
+
+    // ── Tool-missing escalation ─────────────────────────────────────────────
+    // If QC flagged TOOL_MISSING, the agent ran but never called tools.
+    // Inject a strong directive — the agent MUST use its available tools.
+    const hasToolMissing = currentQC.issues.some(
+      (i) => i.code === 'TOOL_MISSING' || /tool.*(missing|not.*call)/i.test(i.message ?? i.description ?? ''),
+    );
+    const toolNudge = hasToolMissing
+      ? `\n\nCRITICAL: The previous attempt completed WITHOUT calling any tools. ` +
+        `You MUST use your available tools (read_file, write_file, list_directory, etc.) ` +
+        `to actually perform the work. Do NOT just describe what you would do — ` +
+        `call the tools and do it. Use <tool_call> blocks to invoke tools.`
+      : '';
+
     const repairSpec: OrcaTaskSpec = {
-      originalUserMessage: `${repairInstruction}${roleHint}`,
+      originalUserMessage: `${repairInstruction}${roleHint}${toolNudge}`,
       intent: "repair",
       goals: [
         ...originalTask.goals,