[FEATURE] PostThinking Hook

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Claude Code's hook system has a visibility gap. Currently available hooks:

• UserPromptSubmit: Sees user's raw input before Claude processes it
• PreToolUse: Fires after Claude has already committed to an action

Missing: A hook that fires after Claude has interpreted the request and formed a plan, but before execution begins.

Having insight into Claude intent might be a really beneficial piece of information for the hooks.

Proposed Solution

Add a PostThinking hook that fires once per turn, after Claude's thinking phase completes but before the first tool call.

Hook input:

 {                                                                                                                                                                                                                                                                  
   "session_id": "abc-123",                                                                                                                                                                                                                                         
   "user_prompt": "help me clean up this authentication code",                                                                                                                                                                                                      
   "thinking": "The user wants to simplify their auth code. I'll use the simplify skill...",                                                                                                                                                                        
   "planned_tools": ["Read", "Skill"],                                                                                                                                                                                                                              
   "first_tool": {"name": "Read", "input": {"file_path": "/src/auth.ts"}}                                                                                                                                                                                           
 } 

Hook output capabilities:

• Inject context: {"additionalContext": "..."}
• Block execution: {"block": true, "message": "..."}
• Modify first tool (optional): {"modifyFirstTool": {...}}

Alternative Solutions

Current workaround: Two-phase flag mechanism

1. UserPromptSubmit creates a flag file: .skill-check-pending
2. PreToolUse checks flag, parses transcript JSONL to extract thinking, then deletes flag

Priority

Critical - Blocking my work

Feature Category

CLI commands and flags

Use Case Example

• Intelligent skill activation - Match skills based on Claude's plan ("I'll use simplify...") instead of guessing from ambiguous user input. Current approach requires API calls with ~60% confidence; PostThinking enables ~95% confidence from direct intent.
• Dangerous operation prevention - Detect rm -rf, git push --force, or DROP TABLE in thinking and block before execution starts. Currently can only warn after Claude has committed to the action.
• Cost control - Limit expensive operations. If thinking mentions "multiple web searches" or "analyze all files," inject limits before they happen.

Additional Context

_No response_