feature-guide

Claude Code Native Architecture

Quick reference for CC's native features. Source: code.claude.com/docs/en/. Updated by docs-monitor agent. Last updated: 2026-06-18

Workflow rule (always-on while this skill is loaded)

Always use the Guide-and-Verify workflow when designing, editing, or generating any Claude Code configuration artifact. Skipping it is a regression in process, not a shortcut.

GUIDE Phase — Before generating code or producing a plan

Before generating or editing a Claude Code config artifact, AND before producing an implementation plan or design document you MUST:

1. Read the single matching reference file for the artifact type from the table below — not the whole skill. Mapping:
   • agents/ → references/agents.md
   • skills/ → references/skills.md
   • hooks/, settings.json (hooks block) → references/hooks.md
   • commands/ → references/tools-and-scheduling.md
   • output-styles/ → references/skills.md (output style section)
   • schedules/ → references/tools-and-scheduling.md
   • .mcp.json, plugin manifests → references/mcp-and-plugins.md
   • permission modes (auto, plan, acceptEdits, bypassPermissions) → references/modes-and-permissions.md
   • memory, context, transcripts → references/memory-and-context.md
   • any field newly available or recently changed → references/changelog.md
2. Plan-phase Guide is non-optional. A plan that proposes a deprecated feature, an incorrect hook event, or a forbidden frontmatter field must be caught before any code is written. If you are entering plan mode or are a sub-agent producing a design, read the matching reference BEFORE drafting the plan, not after.

When the work crosses plugin boundaries (modifying a plugin manifest, adding a marketplace entry, changing version, adding hooks/skills/agents to a plugin) you MUST also:

• Re-read references/mcp-and-plugins.md for plugin agent constraints (permissionMode/hooks/mcpServers are forbidden in plugin agent frontmatter).
• Verify version-bump implications: changes only ship to users when version in plugin.json is bumped.

VERIFY Phase — After generating code

The cc-native plugin runs two automated verifications:

1. Deterministic lint (hooks/cc-native-verify.py, PostToolUse) — exits 0 pass, 1 warnings, 2 hard fail.
2. Semantic audit (agents/auditor.md, invoked via Stop hook directive as cc-native:auditor) — returns per-file verdicts: block / warn / pass.

You MUST:

1. Read phase — re-read the modified files (the verify hook may have already flagged them).
2. Lint check — if cc-native-verify exited with code 2, fix the errors before continuing. Code 1 warnings should also be addressed unless explicitly out of scope.
3. Audit check — when the Stop hook injects an additionalContext directive instructing audit, invoke the cc-native:auditor subagent via the Task tool with the listed file paths. Do not skip this step on the grounds that the lint passed — semantic issues (wrong hook event for the goal, over-broad tools, dangling cross-references) are exactly what the lint cannot catch.
4. Block on findings — you are prohibited from declaring done if:
   • any cc-native-verify exit-2 errors remain
   • the cc-native:auditor returned any block-severity finding
5. Fix in code — based on the rule's rationale (verifier message or auditor finding text), edit the file. Do not silence the warning by changing the lint config.
6. Re-verify — after fixes, the next Edit/Write triggers cc-native-verify again automatically; for the auditor, re-invoke with the same file list.

Why this exists

Claude Code feature surface changes weekly. Without these guard rails, agents drift: they author a hook with an event name that was renamed three releases ago, frontmatter that's silently ignored by the plugin loader, or a skill description that never triggers. The lint catches mechanical errors; the auditor catches intent errors; this skill provides the always-current ground truth. This rule forces the loop to close.

Agentic Loop (always-loaded summary)

Three phases: gather context (Read/Grep/Glob) → take action (Edit/Write/Bash) → verify results (run tests, re-read). Phases blend together. Never skip verify. Claude decides what each step requires based on what it learned from the previous step.

How to use this skill

Read only the reference file that matches the question — do not read all of them. Each file is self-contained and stands alone.

Topic	Reference file
Subagents, custom agents, agent teams, worktrees	references/agents.md
Skills (frontmatter, progressive disclosure, lifecycle, invocation control)	references/skills.md
Hooks (events, types, matchers, exit codes, JSON output)	references/hooks.md
MCP servers, plugins, channels	references/mcp-and-plugins.md
Plan mode, permission modes, auto mode, checkpointing	references/modes-and-permissions.md
Memory hierarchy, CLAUDE.md, auto memory, context management	references/memory-and-context.md
Tools reference, scheduled tasks (CronCreate, /schedule, /loop), session management, effort levels	references/tools-and-scheduling.md
Notable settings keys	references/settings.md
New commands and recent version notes (v2.1.116 → v2.1.181)	references/changelog.md

Routing tips

• "How does X work in CC?" → look up X in the table, read that one file.
• "What changed in version Y?" → references/changelog.md.
• "Which setting controls Z?" → references/settings.md, then maybe a topic file.
• "How do I structure a skill?" → references/skills.md (covers progressive disclosure).
• Multi-topic questions: read the matching files, do not preload others.