ClaudeCodeUndocumentedConfiguration
title: Claude Code Undocumented Configuration type: technique created: 2026-06-25 last_updated: 2026-06-25 related: ["radar/techniques/BuildingSkillsForClaude"] sources: ["https://buildingbetter.tech/p/i-read-the-claude-code-source-code"] radar_quadrant: Techniques radar_ring: Assess radar_position: inner
A body of configuration options discovered by reading the Claude Code source directly, covering hook response fields, skill frontmatter extensions, agent config, and auto-mode settings not present in the official documentation as of mid-2026.
PreToolUse hooks can return structured fields beyond what the docs describe:
| Field | Effect |
|---|---|
updatedInput |
Rewrites the tool's input before execution |
permissionDecision |
Forces allow or deny on a tool call |
permissionDecisionReason |
Explanation shown in the UI |
additionalContext |
Injects text into the conversation |
SessionStart hooks can return watchPaths, initialUserMessage, and additionalContext. PostToolUse hooks can return updatedMCPToolOutput and additionalContext.
Three hook-level flags with no official documentation:
-
once: true-- fires exactly once, then auto-removes -
async: true-- runs in background without blocking the agent -
asyncRewake: true-- non-blocking normally, but blocks if exit code is 2
Six frontmatter fields beyond the documented set:
| Field | Effect |
|---|---|
model |
Overrides which model runs the skill (haiku/opus) |
effort |
Controls thinking depth (low/medium/high/max) |
hooks |
Skill-scoped hooks that register/deregister with the skill |
agent |
Delegates skill execution to a named custom agent |
disable-model-invocation: true |
Prevents auto-invocation of the skill |
shell |
Specifies the execution shell |
-
color-- UI color designation -
memory-- persistent memory scope (user/project/local) -
omitClaudeMd: true-- skips the project instruction hierarchy for this agent -
criticalSystemReminder_EXPERIMENTAL-- re-injected at every turn -
requiredMcpServers-- declares required MCP server dependencies
autoMode.environment accepts plain English context strings describing the setup, used to improve Claude Code's safety decisions in autonomous mode.
Placed in Assess at inner position. These configuration surfaces materially expand what is achievable with Claude Code hooks and skills -- particularly updatedInput for pre-processing tool calls, once: true for single-fire hooks, and effort for per-skill thinking depth. Complements radar/techniques/BuildingSkillsForClaude, which covers the SKILL.md methodology. Trial gate: use at least two undocumented fields (e.g. updatedInput + once: true) in a production hook setup and validate behaviour.
