Fix 'ok' short predictions: add explicit instruction to agent_input

scripts/conversation-context.mjs

@@ -569,5 +569,18 @@ Clone threshold: ${safeThreshold}
 ${historyBlock}
 
 ### assistant (current iter ${safeIteration}):
-${currentAssistantBlock}`
+${currentAssistantBlock}
+
+---
+Given the conversation and the assistant's work above, predict the most
+likely next prompt the user would send to continue this task.
+
+Rules:
+- Write as if you are the user typing their next message.
+- Be specific: reference concrete artifacts, file names, test results, or
+  next steps that follow naturally from what the assistant just did.
+- Minimum 1 full sentence. No single words, no "ok", no "looks good",
+  no bare affirmations.
+- If the current work is complete, the prompt should ask for the next
+  logical feature, verification step, or follow-up action.`
 }

tests/stop-hook-v2.test.mjs

@@ -210,6 +210,9 @@ describe('Clone Loop v2 stop hook', () => {
         assert.equal(calls[1].params.arguments.agent, 'Claude Code Clone Loop')
         assert.match(calls[1].params.arguments.agent_input, /Fix the bug and run tests/)
         assert.match(calls[1].params.arguments.agent_input, /Tests passed\. What next\?/)
+        assert.match(calls[1].params.arguments.agent_input, /predict the most/)
+        assert.match(calls[1].params.arguments.agent_input, /No single words, no "ok"/)
+
         assert.equal(calls[1].params.arguments.threshold, 0.6)
         assert.equal(calls[1].params.arguments.session_id, 'session-123')