Claude Code Undocumented Features: Hooks, Memory, YOLO Classifier & More

André Figueira read the Claude Code source code (v2.1.87) and found dozens of undocumented configuration options. The official docs cover basics, but the actual package—freely inspectable in node_modules—exposes fields for hooks, agent memory, auto-mode rules, and more. Here's what you can actually configure right now.

YOLO Classifier: Auto-approve Actions with Plain English

Internally called the "YOLO Classifier" (yoloClassifier.ts), this system lets you describe your environment in plain English to decide when Claude can auto-approve operations. In settings.json, add a description like "this is a staging server, destructive operations are acceptable" to allow certain actions without prompting.

Hook Return Values That Change Behavior Mid-Flight

Hooks receive JSON on stdin and can return event-specific JSON on stdout. Here are the key return fields:

• PreToolUse: updatedInput (rewrite tool input before execution), permissionDecision (force allow/deny), permissionDecisionReason, additionalContext
• SessionStart: watchPaths (automatic file watching triggering FileChanged events), initialUserMessage (prepend content to first user message), additionalContext
• PostToolUse: updatedMCPToolOutput, additionalContext
• PermissionRequest: decision (programmatic allow/deny with updatedInput or updatedPermissions)

Example: a PreToolUse hook that adds --dry-run to every git push command. In .claude/settings.json:

{
  "hooks": {
    "PreToolUse": [{
      "matcher": "Bash",
      "hooks": [{
        "type": "command",
        "command": "~/.claude/hooks/dry-run-pushes.sh"
      }]
    }]
  }
}

Script at ~/.claude/hooks/dry-run-pushes.sh:

#!/bin/bash
INPUT=$(jq -r '.tool_input.command' 
This rewrites the command before execution, an undocumented field.

Session Context & File Watching

SessionStart hooks can set watchPaths to monitor files like package.json, .env, and tsconfig.json for changes, triggering FileChanged events. They can also inject initialUserMessage to prepend context, e.g., current git branch and uncommitted changes count.

Example script returning watch paths and initial message:

#!/bin/bash
BRANCH=$(git branch --show-current 2>/dev/null)
CHANGES=$(git status --porcelain 2>/dev/null | wc -l | tr -d ' ')
jq -n \
  --arg branch "$BRANCH" \
  --arg changes "$CHANGES" \
  '{
    "watchPaths": ["package.json", ".env", "tsconfig.json"],
    "initialUserMessage": "Current branch: \($branch), uncommitted changes: \($changes)"
  }'

Persistent Agent Memory & Auto-Mode Rules

Agent memory can be configured to persist across sessions using fields in the ~/.claude/agents/ directory. Auto-mode rules can be described in plain English files, telling Claude when to approve actions automatically. The source also hints at "self-improving dream loops" where Claude can refine its own configs.

File Locations

• Settings: ~/.claude/settings.json (personal) or .claude/settings.json (project, shareable via git)
• Skills: ~/.claude/skills/<name>/SKILL.md (personal) or .claude/skills/<name>/SKILL.md (project)
• Agents: ~/.claude/agents/<name>.md (personal) or .claude/agents/<name>.md (project)
• Hook scripts: Conventionally in ~/.claude/hooks/, must be executable (chmod +x)

These are practical for developers who want to customize Claude Code beyond the documented APIs. Expect these features to change between releases.