🌐 日本語

Lifecycle × Configuration Map

NOTE

Shows which configuration layers are active at each phase of Claude Code's task flow.
Configuration names are representative examples. Layers from Parts 3–7 are cut across the lifecycle.

Related Issue: #21

Lifecycle Flow

Configuration Effective at Each Phase

Session Start (SessionStart → CLAUDE.md Load)

Configuration Layer	What Happens	Related Part
CLAUDE.md (all levels)	Merged in order: Global → Project → Local, injected as resident context	Part 3
settings.json	Runtime settings (permissions, environment variables, thinking mode, etc.) applied. Not injected to LLM context	Part 7
MCP tool definitions	Tool definitions from connected MCP servers resident-injected to context (index-only when Tool Search enabled)	Part 6
Hook: SessionStart	Environment check, log initialization. stdout added to context	Part 7
Hook: InstructionsLoaded	Fires when CLAUDE.md / rules files are loaded. Audit log, compliance tracking	Part 7

On User Input (UserPromptSubmit)

Configuration Layer	What Happens	Related Part
Hook: UserPromptSubmit	Input validation, additional context injection. stdout added to context. Can block prompt with exit 2	Part 7
Prompt Hook	Evaluate input using LLM with type: "prompt"	Part 7

Before Tool Execution (PreToolUse)

Configuration Layer	What Happens	Related Part
.claude/rules/	Rules matching glob pattern of target file injected to context	Part 4
Skills	Task-specific procedures deployed to context via automatic LLM judgment or / invocation	Part 5
settings.json (permissions)	allow / deny rules permit/deny tool use. Hook's deny rule takes priority over allow	Part 7
Hook: PreToolUse	Block dangerous commands, rewrite input (updatedInput). Control allow / deny / ask with permissionDecision	Part 7
Hook: PermissionRequest	Fires when permission dialog shown. Auto-approve/deny possible	Part 7

After Tool Execution (PostToolUse / PostToolUseFailure)

Configuration Layer	What Happens	Related Part
Hook: PostToolUse	Auto-format (prettier, etc.), lint execution, logging. Tools already executed, cannot be reverted	Part 7
Hook: PostToolUseFailure	Error log on tool failure, retry decision	Part 7

Sub-Agent and Tasks (SubagentStart/Stop, TaskCreated/Completed)

Configuration Layer	What Happens	Related Part
Agents	Execute in independent context window. Distill results back to main	Part 5
Hook: SubagentStart	Inject context when sub-agent generated	Part 7
Hook: SubagentStop	Validate results on sub-agent completion. Can block with exit 2	Part 7
Hook: TaskCreated	Enforce naming conventions, validate on task creation	Part 7
Hook: TaskCompleted	Validate task completion conditions	Part 7

Response Complete (Stop / StopFailure)

Configuration Layer	What Happens	Related Part
Hook: Stop	Quality gate, continuation decision. Exit 2 or decision: "block" prevents stop and continues work	Part 7
Hook: StopFailure	Error log on API error, send alerts	Part 7
Agent Hook	Sub-agent multi-turn verification (test execution, etc.) with type: "agent"	Part 7

Context Compression (PreCompact / PostCompact)

Configuration Layer	What Happens	Related Part
Hook: PreCompact	Validate before compression. Backup critical information, etc.	Part 7
Hook: PostCompact	Validate after compression	Part 7
Hook: SessionStart (matcher: compact)	After compression, context re-injection possible on session restart	Part 7

Async Events (Fire Parallel to Loop)

Configuration Layer	What Happens	Related Part
Hook: CwdChanged	Reload environment variables on working directory change (direnv, etc.)	Part 7
Hook: FileChanged	Detect watched file changes. Specify filename with matcher	Part 7
Hook: ConfigChange	Security audit on config file change. Can block with exit 2	Part 7
Hook: Notification	Desktop notification, alert on notification event	Part 7

Configuration Layer Timing Summary

Configuration Layer	Active Timing	Context Consumption
CLAUDE.md	Injected on session start, resident across all turns	Always
.claude/rules/	Injected when glob matches file operation	Conditional only
Skills	Injected on / invocation or LLM auto-judgment	On invocation only
Agents	Started with Agent() / Task(). Independent context	Main doesn't consume
MCP tool definitions	Injected on session start (index-only with Tool Search)	Always (or on search)
settings.json	Always applied at runtime	None
Hooks	Fire at each lifecycle event	None (except Prompt Hook)

TIP

While problem-countermeasure-map.md shows "structural problem → which configuration solves it," this page shows "which phase of the lifecycle → which configuration is active." Reading both together gives a three-dimensional understanding of the configuration landscape.

---

NOTE

For detailed Hook events (JSON input/output schema, matcher specs, etc.), see the official reference:
Hooks reference | Hooks guide