This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

bun install          # NOT npm/yarn — this project uses Bun exclusively
bun run build        # bundles to dist/cli.js (~23MB single file)
bun dist/cli.js      # run the CLI

There are no tests. There is no linter. The build script is build.ts (not a bundler config file).

Rebuilt Claude Code CLI from Anthropic's leaked source (1,929 files extracted from a source map). The src/ directory is the original source. The stubs/ directory contains proprietary code extracted from npm packages and source maps. This is not a fork — it's a reconstruction.

• Feature flags: All feature('FLAG_NAME') calls return false via the shim at shims/bun-bundle.ts. This is intentional — it disables unreleased/internal features. Do not change this.
• MACRO.VERSION is hardcoded to 2.1.88 in build.ts. Do not change this.
• Auto-updater is patched: src/utils/autoUpdater.ts line 72 has an early return to disable remote version checks. Do not remove this.
• [STUB] files contain minimal placeholder code. They exist because the source was missing (not in the leak), requires native binaries (Swift/Rust), or is Anthropic-internal (USER_TYPE === 'ant'). Do not try to "fix" or "complete" stubs unless explicitly asked.
• stubs/ files are extracted source — treat them as read-only reference material, not code to modify.

• Entrypoint: src/main.tsx — Commander.js CLI parser. Bootstrap is src/entrypoints/cli.tsx.
• UI: Custom Ink fork at src/ink/ — this is NOT the npm ink package. It's a full React-based terminal renderer with custom flexbox layout (src/native-ts/yoga-layout/).
• System prompt: src/constants/prompts.ts — the actual instructions sent to Claude. This is the single most important file for understanding Claude Code's behavior.
• Tools live in src/tools/ with each tool in its own directory (BashTool/, FileEditTool/, etc.).
• Permissions classifier: src/utils/permissions/yoloClassifier.ts (52KB) — auto-mode logic. The 3 prompt .txt files it references were dead-code-eliminated and are missing.
• Plugin system: src/utils/plugins/ (65+ files) — fully working. Loads from .claude/plugins/.
• Hook system: src/utils/hooks/ (155 files) — fully working. Config in settings.json.
• CLAUDE.md loader: src/utils/claudemd.ts — how this very file gets loaded. Supports @path includes, frontmatter with paths: globs, HTML comment stripping.

The base source is v2.1.88. The following features and fixes from the official 2.1.89–2.1.91 changelogs have been backported:

• disableSkillShellExecution setting — disables !cmd`` and ````!``` shell execution in skills/commands/plugins (types.ts, `promptShellExecution.ts`)
• MCP _meta["anthropic/maxResultSizeChars"] override — MCP tools can request up to 500K chars for large results like DB schemas (mcpValidation.ts, client.ts)
• Multi-line prompts in deep links — encoded newlines %0A no longer rejected in claude-cli://open?q= (parseDeepLink.ts)
• Plugin bin/ executables — plugins can ship executables under bin/ that are prepended to PATH for Bash tool (plugin.ts, pluginLoader.ts, subprocessEnv.ts)
• Faster stripAnsi via Bun.stripANSI — native Bun path when available, fallback to strip-ansi package (src/utils/stripAnsi.ts + 12 import sites)

• .husky added to protected directories — prevents auto-editing in acceptEdits mode (filesystem.ts)
• DNS commands removed from auto-allow — Get-DnsClientCache fully removed, ipconfig /displaydns removed for DNS cache privacy (readOnlyValidation.ts)
• CLAUDE_CODE_PLUGIN_KEEP_MARKETPLACE_ON_FAILURE env var — keeps existing marketplace cache when git pull fails, useful in offline environments (marketplaceManager.ts)

• cleanupPeriodDays: 0 validation — now rejected with actionable error pointing to --no-session-persistence (types.ts)
• defer permission decision for PreToolUse hooks — headless sessions can pause and resume via --resume (PermissionRule.ts, coreSchemas.ts, hooks.ts)
• MCP_CONNECTION_NONBLOCKING=true env var — bounds MCP server connections at 5s instead of 30s (client.ts)
• Autocompact thrash loop — actionable error message when circuit breaker trips after 3 consecutive failures (autoCompact.ts)
• MCP tool errors: all content blocks — multi-element error content no longer truncated to first block (client.ts)
• Edit works on files viewed via Bash — cat, sed -n, head, tail output tracked in readFileState so Edit doesn't require separate Read (BashTool.tsx)
• Hook output >50K saved to disk — large hook output written to temp file with path + preview instead of injecting into context (hooks.ts)

• PermissionDenied hook event, CLAUDE_CODE_NO_FLICKER=1, showThinkingSummaries setting, MAX_EDIT_FILE_SIZE (1 GiB OOM protection), BASH_LIST_COMMANDS (ls/tree/du summary), paste !command bash mode

import { query } from "@anthropic-ai/claude-agent-sdk";
const response = query({
  prompt: "your prompt",
  options: {
    pathToClaudeCodeExecutable: "/path/to/claude-code/dist/cli.js",
  },
});

SDK source is at stubs/downloads/claude-agent-sdk/.

• Computer Use — logic extracted (stubs/@ant/computer-use-mcp/src/toolCalls.ts, 137KB) but needs native Swift/Rust binaries for screen capture/input.
• Auto-mode classifier prompts — 3 .txt files were DCE'd by the TRANSCRIPT_CLASSIFIER flag.
• Feature-flagged features (voice, coordinator, ultraplan, bridge) — disabled via shim, many need backend infra.
• Ant-only tools (Tungsten, REPL, SuggestBackgroundPR) — internal Anthropic tools, never available externally.

Files we created (not from the original leak) to satisfy imports. Each contains minimal placeholder code.

• Always rebuild after source changes: bun run build
• The build produces a single dist/cli.js file — there's no dev server or watch mode
• If adding new stubs, they go in stubs/ and need corresponding shim entries if they're imported as packages
• SDK types can be regenerated: bun scripts/generate-sdk-types.ts