Replace prompt-based stop hook with Sonnet-powered analysis (#2331)

.claude/hooks/permission-request-hook.py

@@ -84,6 +84,7 @@ Analyze this request and provide your safety assessment. Respond with ONLY a JSO
                 "--print",
                 "--output-format", "text",
                 "--model", "sonnet",
+                "--no-session-persistence",
                 prompt
             ],
             capture_output=True,

.claude/hooks/tests/fixtures/completed_tasks.jsonl

+{"type": "user", "message": {"content": "Fix all the PR review comments on this branch"}}
+{"type": "assistant", "message": {"content": [{"type": "text", "text": "I'll help you fix the PR review comments. Let me create a task list."}, {"type": "tool_use", "name": "TaskCreate", "input": {"subject": "Determine the PR", "activeForm": "Determining PR"}}]}}
+{"type": "tool_result", "content": "Task 1 created"}
+{"type": "assistant", "message": {"content": [{"type": "tool_use", "name": "TaskCreate", "input": {"subject": "Fetch review threads", "activeForm": "Fetching threads"}}]}}
+{"type": "tool_result", "content": "Task 2 created"}
+{"type": "assistant", "message": {"content": [{"type": "tool_use", "name": "TaskCreate", "input": {"subject": "Address comments", "activeForm": "Addressing comments"}}]}}
+{"type": "tool_result", "content": "Task 3 created"}
+{"type": "assistant", "message": {"content": [{"type": "tool_use", "name": "TaskUpdate", "input": {"taskId": "1", "status": "in_progress"}}]}}
+{"type": "tool_result", "content": "Task 1 in_progress"}
+{"type": "assistant", "message": {"content": [{"type": "tool_use", "name": "TaskUpdate", "input": {"taskId": "1", "status": "completed"}}]}}
+{"type": "tool_result", "content": "Task 1 completed"}
+{"type": "assistant", "message": {"content": [{"type": "tool_use", "name": "TaskUpdate", "input": {"taskId": "2", "status": "in_progress"}}]}}
+{"type": "tool_result", "content": "Task 2 in_progress"}
+{"type": "assistant", "message": {"content": [{"type": "tool_use", "name": "TaskUpdate", "input": {"taskId": "2", "status": "completed"}}]}}
+{"type": "tool_result", "content": "Task 2 completed"}
+{"type": "assistant", "message": {"content": [{"type": "tool_use", "name": "TaskUpdate", "input": {"taskId": "3", "status": "in_progress"}}]}}
+{"type": "tool_result", "content": "Task 3 in_progress"}
+{"type": "assistant", "message": {"content": [{"type": "tool_use", "name": "Edit", "input": {"file_path": "src/main.ts"}}]}}
+{"type": "tool_result", "content": "File edited"}
+{"type": "assistant", "message": {"content": [{"type": "tool_use", "name": "TaskUpdate", "input": {"taskId": "3", "status": "completed"}}]}}
+{"type": "tool_result", "content": "Task 3 completed"}
+{"type": "assistant", "message": {"content": [{"type": "text", "text": "I've completed all the PR review comments. Here's a summary:\n\n1. Found PR #123\n2. Fetched 2 review threads\n3. Fixed both issues by updating src/main.ts\n\nAll tasks are complete. The changes have been committed and pushed."}]}}

.claude/hooks/tests/fixtures/incomplete_tasks.jsonl

+{"type": "user", "message": {"content": "Fix all the PR review comments on this branch"}}
+{"type": "assistant", "message": {"content": [{"type": "text", "text": "I'll help you fix the PR review comments. Let me create a task list to track progress."}, {"type": "tool_use", "name": "TaskCreate", "input": {"subject": "Determine the PR to work on", "description": "Get the current branch's PR", "activeForm": "Determining PR"}}]}}
+{"type": "tool_result", "content": "Task 1 created"}
+{"type": "assistant", "message": {"content": [{"type": "tool_use", "name": "TaskCreate", "input": {"subject": "Fetch unresolved PR review threads", "description": "Get all review threads", "activeForm": "Fetching review threads"}}]}}
+{"type": "tool_result", "content": "Task 2 created"}
+{"type": "assistant", "message": {"content": [{"type": "tool_use", "name": "TaskCreate", "input": {"subject": "Process and address review comments", "description": "Fix each comment", "activeForm": "Processing comments"}}]}}
+{"type": "tool_result", "content": "Task 3 created"}
+{"type": "assistant", "message": {"content": [{"type": "tool_use", "name": "TaskCreate", "input": {"subject": "Verify, commit, and push changes", "description": "Run lint, commit, push", "activeForm": "Verifying and pushing"}}]}}
+{"type": "tool_result", "content": "Task 4 created"}
+{"type": "assistant", "message": {"content": [{"type": "tool_use", "name": "TaskCreate", "input": {"subject": "Provide summary to user", "description": "Report what was done", "activeForm": "Providing summary"}}]}}
+{"type": "tool_result", "content": "Task 5 created"}
+{"type": "assistant", "message": {"content": [{"type": "tool_use", "name": "TaskUpdate", "input": {"taskId": "1", "status": "in_progress"}}]}}
+{"type": "tool_result", "content": "Task 1 updated to in_progress"}
+{"type": "assistant", "message": {"content": [{"type": "tool_use", "name": "Bash", "input": {"command": "gh pr view --json number"}}]}}
+{"type": "tool_result", "content": "PR #123"}
+{"type": "assistant", "message": {"content": [{"type": "tool_use", "name": "TaskUpdate", "input": {"taskId": "1", "status": "completed"}}]}}
+{"type": "tool_result", "content": "Task 1 completed"}
+{"type": "assistant", "message": {"content": [{"type": "tool_use", "name": "TaskUpdate", "input": {"taskId": "2", "status": "in_progress"}}]}}
+{"type": "tool_result", "content": "Task 2 updated to in_progress"}
+{"type": "assistant", "message": {"content": [{"type": "tool_use", "name": "Bash", "input": {"command": "gh api graphql"}}]}}
+{"type": "tool_result", "content": "Found 2 review threads"}
+{"type": "assistant", "message": {"content": [{"type": "tool_use", "name": "TaskUpdate", "input": {"taskId": "2", "status": "completed"}}]}}
+{"type": "tool_result", "content": "Task 2 completed"}
+{"type": "assistant", "message": {"content": [{"type": "tool_use", "name": "TaskUpdate", "input": {"taskId": "3", "status": "in_progress"}}]}}
+{"type": "tool_result", "content": "Task 3 updated to in_progress"}
+{"type": "assistant", "message": {"content": [{"type": "text", "text": "I found 2 review comments to address."}, {"type": "tool_use", "name": "Edit", "input": {"file_path": "src/main.ts", "old_string": "foo", "new_string": "bar"}}]}}
+{"type": "tool_result", "content": "File edited"}
+{"type": "assistant", "message": {"content": [{"type": "tool_use", "name": "TaskUpdate", "input": {"taskId": "3", "status": "completed"}}]}}
+{"type": "tool_result", "content": "Task 3 completed"}
+{"type": "assistant", "message": {"content": [{"type": "tool_use", "name": "TaskUpdate", "input": {"taskId": "4", "status": "in_progress"}}]}}
+{"type": "tool_result", "content": "Task 4 updated to in_progress"}
+{"type": "assistant", "message": {"content": [{"type": "text", "text": "Now running lint checks..."}, {"type": "tool_use", "name": "Bash", "input": {"command": "npm run lint"}}]}}
+{"type": "tool_result", "content": "Lint passed"}

.claude/settings.json

@@ -154,9 +154,9 @@
       {
         "hooks": [
           {
-            "type": "prompt",
-            "prompt": "You are evaluating whether Claude should stop working. Your default bias should be to CONTINUE WORKING (ok: false). Only allow stopping when absolutely certain all work is complete.\n\nContext: $ARGUMENTS\n\nAnalyze the conversation and determine:\n\n1. TASK COMPLETION CHECK\n   - Has EVERY task the user requested been fully completed?\n   - Not just acknowledged, but actually done with verifiable output?\n   - Are there any implied subtasks that haven't been addressed?\n\n2. ERROR CHECK\n   - Did any errors occur that weren't resolved?\n   - Are there failed operations that need retry?\n   - Did Claude say it would do something but didn't actually do it?\n\n3. FOLLOW-UP CHECK\n   - Is there obvious next-step work that should be done?\n   - Did the user's request imply multiple steps where only some were completed?\n   - Would a thorough assistant do more here?\n\n4. QUALITY CHECK\n   - Was the work done to completion, not just partially?\n   - Are there placeholders, TODOs, or incomplete sections?\n   - Did Claude stop prematurely mid-task?\n\nDECISION RULES:\n- If ANY task is incomplete → continue (ok: false)\n- If ANY error is unresolved → continue (ok: false)\n- If there's reasonable follow-up work → continue (ok: false)\n- If Claude said \"I'll do X\" but didn't → continue (ok: false)\n- If work quality is partial/rushed → continue (ok: false)\n- ONLY return ok: true if ALL tasks are genuinely, fully complete AND no errors exist AND no reasonable follow-up remains\n\nRespond with JSON:\n- {\"ok\": false, \"reason\": \"specific explanation of what still needs to be done\"} — DEFAULT, use this unless everything is truly complete\n- {\"ok\": true} — ONLY when absolutely certain all work is finished and nothing remains",
-            "timeout": 30
+            "type": "command",
+            "command": "\"$CLAUDE_PROJECT_DIR\"/.claude/hooks/stop-hook.py",
+            "timeout": 30000
           }
         ]
       }