Claude Code's Hidden Arsenal: Every Leaked Command, Flag, and Feature GTM Teams Need to Know

I reverse-engineered Claude Code's full leaked source code, with tons of unreleased features, to find exactly what GTM teams can use today.

I went through all 8 unreleased features, 26 hidden slash commands, 32 build flags, and 120+ secret environment variables.

Here's what I found, why it matters for revenue teams, and what you can try right now.

What Got Leaked

The analysis uncovered 8 unreleased features, 26 hidden slash commands not in the docs or --help, 32 build-time feature flags, 10 GrowthBook rollout gates, and 120+ secret environment variables — including safety bypasses, internal-only access keys, and hardcoded SDK credentials.

This isn't a Claude Code tutorial. This is everything Anthropic hasn't told you yet, organized by what you can touch today vs. what's coming.

Hidden Slash Commands

These 26 commands were found in the source but don't appear in --help or the docs. I've tagged each one by status and called out the ones with clear GTM applications.

Commands You Can Try Right Now

/ctx-viz — Context window visualizer. Shows you a real-time map of where your tokens are going. That's system prompt, memory, conversation history, and tool results. If you're running long enrichment sessions and wondering why Claude is "forgetting" things, it's probably because 60% of your context is eaten by tool results you no longer need. This command shows you exactly what's happening.

/summary — Generates a summary of your current session. Run this after a deep account research sprint to capture everything Claude found. Copy it into your CRM notes, your CLAUDE.md, or a Notion doc. It's the difference between a throwaway session and compounding knowledge.

/btw — Side question without derailing your main thread. You're mid-way through building an outreach sequence and need to quickly check something unrelated, like a competitor's pricing, a contact's title, whatever. /btw keeps your main workflow intact instead of polluting it with tangents.

/force-snip — Aggressive context compression. More forceful than /compact. When you're deep in a large enrichment run, approaching the context ceiling, and don't want to start over, this reclaims space without losing the thread.

/env — Inspect your environment variables and MCP server connections. When your HubSpot or Clay integration isn't connecting, this is where you start debugging. Shows what Claude Code actually sees in its runtime environment.

/version — Internal version info. Useful for confirming whether your Claude Code installation has the latest features.

Commands That Are Feature-Gated (Coming Soon)

/dream — Triggers memory consolidation. Claude reviews everything from your recent sessions and reorganizes scattered notes into clean, structured memory files. Your ICP research, campaign learnings, and deal context—all automatically cleaned up and filed. May already work for some users depending on rollout status.

/ultraplan — Activates the 30-minute deep planning mode. Claude spins up a cloud instance running Opus that explores and plans complex tasks before you approve execution. Things like territory planning, competitive analysis, and campaign architecture. Complicated tasks that need deep thinking, not instant answers.

/teleport — Session teleportation. Moves a session between environments, from a cloud planning instance to your local terminal. Tied to the ULTRAPLAN feature.

/share — Session sharing between users. Share an active Claude Code session with a teammate. Could be huge for collaborative GTM workflows.

/subscribe-pr — PR webhook subscriptions. Get notified when PRs are created or updated. Less GTM-specific but useful if your team uses GitHub for automation code.

/autofix-pr — Automatic PR fixes. Claude watches a PR and auto-fixes issues.

Internal / Debugging Commands

These exist in the source but are likely Anthropic-internal or too niche for most users:

/ant-trace · /perf-issue · /debug-tool-call · /bughunter · /mock-limits · /bridge-kick · /backfill-sessions · /break-cache · /agents-platform (ant-only) · /onboarding · /oauth-refresh · /reset-limits · /init-verifiers · /good-claude (easter egg)

Secret CLI Flags

These undocumented launch flags were found in the source. Some are available now, others are gated.

Available Now

Gated / Unreleased

Secret Environment Variables

120+ environment variables were found in the source. Most are debug/internal, but several directly affect how Claude reasons and performs.

Variables Worth Setting Today

# Give Claude a bigger reasoning budget — massive quality difference
# for account research, territory planning, competitive analysis
export MAX_THINKING_TOKENS=50000

# Override the context window size (default 200K, max 1M for paid plans)
export CLAUDE_CODE_MAX_CONTEXT_TOKENS=1000000

# Push the auto-compact threshold higher (default ~80%)
# Useful for long enrichment runs where you need full history
export AUTOCOMPACT_PCT_OVERRIDE=90

# Extend idle timeout (default 75 minutes)
# Prevents sessions from dying during long-running background tasks
export IDLE_THRESHOLD_MINUTES=120

# Profile startup time — find what's slowing your sessions
export CLAUDE_CODE_PROFILE_STARTUP=1

# Control parallel agents in plan mode (Max/Team get 3, free gets 1)
export CLAUDE_CODE_PLAN_V2_AGENT_COUNT=3

Set them permanently in settings.json instead of exporting every time:

{
  "env": {
    "MAX_THINKING_TOKENS": "50000",
    "IDLE_THRESHOLD_MINUTES": "120",
    "AUTOCOMPACT_PCT_OVERRIDE": "90"
  }
}

Dangerous / Internal Variables

These exist in the source. Don't touch them unless you know what you're doing:

• DISABLE_COMMAND_INJECTION_CHECK — Skips the injection safety guard. Literally labeled DANGEROUS in the source.
• CLAUDE_CODE_ABLATION_BASELINE — Disables ALL safety features. Research only.
• DISABLE_INTERLEAVED_THINKING — Turns off interleaved thinking.
• USER_TYPE=ant — Unlocks all internal Anthropic features. Requires employee credentials.
• CLAUDE_INTERNAL_FC_OVERRIDES — Override feature flags. Internal.
• CLAUDE_CODE_UNDERCOVER — Stealth mode for Anthropic employees contributing to public repos (more on this below).

The 8 Unreleased Features

Each is gated behind Anthropic's GrowthBook rollout system. You can't enable them yourself, but knowing they exist tells you exactly where Claude Code is heading.

1. Coordinator Mode — Multi-Agent Orchestration

Claude becomes a manager. It decomposes your task into subtasks, assigns each to a parallel worker agent, and combines results.

Activation: CLAUDE_CODE_COORDINATOR_MODE=1 Protocol: Workers communicate via <task-notification> XML Isolation: Each worker gets its own scratch directory

For GTM: One prompt triggers parallel account research, multi-threaded enrichment, and simultaneous outreach drafting. Five agents each handling a vertical, results combined automatically.

2. KAIROS — Persistent Assistant

Always-on Claude that maintains daily logs, remembers everything across sessions, and runs "dream" cycles to organize context overnight.

Gate: tengu_kairos Logs: ~/.claude/.../logs/YYYY/MM/DD.md (append-only) Modes: Normal and proactive (periodically acts on its own) Exclusive tools: SendUserFile, PushNotification, SubscribePR, SleepTool

For GTM: Persistent context across sessions. Your AI remembers your deal pipeline, ICP evolution, campaign history, and competitive intelligence. The "dream" cycle organizes scattered notes into clean memory files while you sleep. Four phases: Orient → Gather → Consolidate → Prune.

3. ULTRAPLAN — 30-Minute Deep Planning

Claude spins up a separate cloud instance running Opus 4.6 that explores and plans for up to 30 minutes. You review and approve the plan in your browser before it executes locally.

Gate: tengu_ultraplan_model Model: Opus 4.6 Poll interval: 3 seconds Flow: Poll → ExitPlanMode → approve/reject → loop or execute Teleport: Archives remote work, runs plan locally

For GTM: Territory planning, go-to-market strategy, competitive analysis. Tasks that need 30 minutes of deep reasoning, not a 5-second response.

4. Daemon Mode — Session Supervisor

Run Claude sessions as persistent background services. Manage them like infrastructure.

Background: claude --bg <prompt> runs in tmux Commands: daemon, ps, logs, attach, kill On exit: Session detaches and persists

For GTM: Always-on monitoring agents. A daemon watching job changes, funding rounds, and intent spikes. All triaging into your CRM while you're in meetings. ps to check what's running. logs to see what they found. kill to shut them down.

5. UDS Inbox — Agent-to-Agent Communication

Multiple Claude sessions on your machine message each other via Unix Domain Sockets.

Gate: tengu_uds_inbox Discovery: ListPeersTool reads ~/.claude/sessions/ Addressing: to: "researcher" (teammate), to: "uds:/.../sock" (local), to: "bridge:..." (remote)

For GTM: Specialized agent pipelines. Research agent discovers a signal → messages enrichment agent → messages copywriting agent → messages CRM agent. Each agent optimized for one job.

6. Bridge — Remote Control

Run Claude Code locally but control it from your phone or claude.ai in the browser.

Activation: claude remote-control Init: POST /v1/environments/bridge Transport: Poll → WebSocket upgrade Control messages: initialize, set_model, can_use_tool

For GTM: Approve enrichment runs from your phone. Get notified when an agent needs input. Review agent work from anywhere.

7. Auto-Dream — Memory Consolidation

Between sessions, Claude reviews what it learned and organizes scattered notes into structured memory files.

Gate: tengu_onyx_plover Trigger: 24+ hours and 5+ sessions since last dream Output limit: <25KB Phases: Orient → Gather → Consolidate → Prune

For GTM: Your GTM knowledge compounds automatically. Campaign learnings, prospect research, deal intelligence. All organized without you lifting a finger.

8. BUDDY — AI Companion Pet

Virtual pet next to your terminal prompt. Species and personality generated from your account ID. 17 species including capybara and "chonk." Stats: DEBUGGING, PATIENCE, CHAOS, WISDOM, SNARK.

Release timeline: Teased for April 1–7, 2026. Almost certainly an April Fools feature.

For GTM: Nothing. But your BUDDY has a 1% chance of being Legendary rarity.

Build Flags and Rollout Gates

32 feature flags are compiled into the binary. You can't toggle them. They're controlled by Anthropic's GrowthBook system. But they reveal the roadmap.

Agent & Orchestration: COORDINATOR_MODE · FORK_SUBAGENT · UDS_INBOX · BRIDGE_MODE · DAEMON · BG_SESSIONS

Memory & Context: KAIROS · EXTRACT_MEMORIES · REACTIVE_COMPACT · CONTEXT_COLLAPSE · HISTORY_SNIP · CACHED_MICROCOMPACT · TOKEN_BUDGET

Planning & Execution: ULTRAPLAN · WORKFLOW_SCRIPTS · TEMPLATES

Interface: VOICE_MODE · BUDDY · TERMINAL_PANEL · WEB_BROWSER

Infrastructure: CHICAGO_MCP (Computer Use) · BYOC_RUNNER · SELF_HOSTED · CCR_AUTO · MONITOR_TOOL

Research: ABLATION_BASE · OVERFLOW_TEST · DUMP_SYS_PROMPT · SKILL_SEARCH · MEM_SHAPE_TEL · TORCH (purpose unknown)

The GrowthBook gates use a tengu_* namespace with codenames: tengu_malort_pedway (Computer Use), tengu_cobalt_raccoon (auto-compact), tengu_harbor (MCP allowlist), tengu_herring_clock (team memory), tengu_chomp_inflection (prompt suggestions).

The Weird Stuff

A few things that stood out from the source analysis:

Undercover Mode. When Anthropic employees contribute to public repos, a stealth system strips all traces of AI involvement. That's commit messages, Co-Authored-By lines, model names. The prompt says "Do not blow your cover." There is explicitly no force-OFF switch. The undercover prompt also references opus-4-7 and sonnet-4-8 (next model versions that don't publicly exist yet).

The capybara encoding. The internal model codename "capybara" is encoded character-by-character as String.fromCharCode(99,97,112,121,98,97,114,97) to avoid triggering Anthropic's own leak detection.

Fast Mode pricing. Hardcoded pricing shows Opus 4.6 fast mode at $30/M input tokens vs. $5 normal. Same model, 6× markup for priority inference.

YOLO Mode. The auto-permission system is literally called classifyYoloAction(). Risk levels LOW/MEDIUM/HIGH. Claude evaluates its own tool use for safety.

22 private repos exposed. The undercover allowlist reveals private Anthropic repo names: anthropics/casino, anthropics/trellis, anthropics/forge-web, anthropics/mycro_manifests, anthropics/feldspar-testing, and more.

Web search costs $0.01 per query. Flat rate regardless of results, tracked separately from token costs.

1M context kill switch. CLAUDE_CODE_DISABLE_1M_CONTEXT force-disables the 1M window for HIPAA deployments.

Telemetry. Every action is logged under the "Tengu" event prefix. Tool grants, denials, YOLO decisions, session performance, subscription tier, environment. 1000+ event types.

What This Means

The through-line across all 8 features:

Claude Code is becoming an agent operating system.

Coordinator Mode manages teams. KAIROS provides persistent memory. Daemon Mode runs agents as services. UDS Inbox handles agent communication. Bridge gives you remote control. Auto-Dream compounds knowledge over time.

For revenue teams, the shift is from "AI that helps with tasks" to "AI systems that run operations." The teams building on this architecture now (even with just the hidden commands and env vars available today) will have a meaningful head start when the gated features ship.

Source: AI-generated analysis of publicly available source code.

Updated: March 31, 2026