harness-engineering

Engineering reference for the Claude Code agent harness — hooks, permissions, MCP, subagents, context, skills, output styles, settings. Captures the stable method; version-specific facts are marked VOLATILE.

Triggers

Apply this skill when the task involves the agent's operating environment, not the product code:

• Writing or debugging a hook (PreToolUse, PostToolUse, SessionStart, Stop, …)
• Tuning permissions / tool routing (allow/ask/deny, permission modes)
• Adding or configuring an MCP server, or diagnosing deferred-tool / ToolSearch issues
• Orchestrating subagents (delegation, context isolation, worktree, fan-out)
• Context management — budgeting, compaction, what survives, memory/CLAUDE.md loading
• Authoring a skill (SKILL.md structure, triggers, progressive disclosure) or an output style
• Editing settings.json or reasoning about config precedence
• Deciding where a rule belongs — hook vs. CLAUDE.md vs. permission vs. skill vs. style

Do NOT apply for ordinary application-code questions, single-file edits, or product-domain logic — this skill is about the harness, not what you build with it.