Special systems

BUDDY — companion pet system

Feature flag: BUDDY
Crate: cc-buddy (crates/buddy)
Status: Teaser window April 1–7 2026; full launch planned May 2026BUDDY is a Tamagotchi-style companion pet that lives in your terminal alongside the input prompt. Each user gets a deterministic companion — same user always hatches the same species.

Gacha and PRNG

Species is determined by a Mulberry32 PRNG seeded from your userId hash with the salt 'friend-2026-401':

// Mulberry32 — deterministic, reproducible per user
fn mulberry32(seed: u32) -> impl FnMut() -> f64 {
    move || {
        // seed |= 0; seed = seed + 0x6D2B79F5 | 0;
        // ... fast 32-bit mixing
    }
}

The seed is derived from your user ID, so the same user always hatches the same buddy.

Species and rarity

18 species across 5 rarity tiers. Rarity determines drop probability; shiny is a separate 1% roll independent of rarity.

Rarity	Drop chance	Species
Common	60%	Pebblecrab, Dustbunny, Mossfrog, Twigling, Dewdrop, Puddlefish
Uncommon	25%	Cloudferret, Gustowl, Bramblebear, Thornfox
Rare	10%	Crystaldrake, Deepstag, Lavapup
Epic	4%	Stormwyrm, Voidcat, Aetherling
Legendary	1%	Cosmoshale, Nebulynx

A shiny Legendary Nebulynx has a 0.01% chance (1% rarity × 1% shiny).

Stats

Each buddy has 5 procedurally generated stats in the range 0–100:

Stat	Description
`DEBUGGING`	Aptitude for finding bugs
`PATIENCE`	Tolerance for long compilation times
`CHAOS`	Tendency toward unexpected behavior
`WISDOM`	Accumulated experience
`SNARK`	Commentary sharpness

Sprites

Sprites are rendered as 5 lines × 12 characters ASCII art with multiple animation frames (idle and reaction). They sit next to the user input prompt.

Soul description

On first hatch, Claude generates a short personality description for the buddy — its “soul”. The system prompt tells Claude:

A small {species} named {name} sits beside the user's input box and
occasionally comments in a speech bubble. You're not {name} — it's a
separate watcher.

The buddy has its own personality and responds when addressed by name.

Key types (`cc-buddy`)

Type	Description
`Buddy`	Species, name, rarity, `is_shiny`, stats, soul description, current sprite frame
`Species`	18-variant enum with associated rarity tier
`Stats`	Five `u8` fields, 0–100 each

KAIROS — always-on assistant

Feature flags: PROACTIVE / KAIROS
Status: Internal only; absent from all external buildsKAIROS is a persistent, always-running Claude assistant that does not wait for user input. It watches, logs, and proactively acts on things it notices.

How it works

KAIROS maintains append-only daily log files tracking observations, decisions, and actions throughout the day. On a regular interval it receives <tick> prompts that let it decide whether to act proactively or stay quiet.

<tick>

The system has a 15-second blocking budget: any proactive action that would block the user’s workflow for more than 15 seconds is deferred.

Brief mode

When KAIROS is active, the Brief output mode produces extremely concise responses — suitable for a persistent assistant that should not flood the terminal.

Exclusive tools

KAIROS has access to three tools not available to regular Claude Code sessions:

Tool	Description
`SendUserFile`	Push files directly to the user (notifications, summaries)
`PushNotification`	Send push notifications to the user’s device
`SubscribePR`	Subscribe to and monitor pull request activity

These tools are only compiled when the KAIROS feature flag is enabled. Attempting to invoke them in a standard build will result in an unknown-tool error.

ULTRAPLAN — 30-minute remote planning

Status: Internal; gated behind feature configuration
Model: Opus 4.6 (configured via tengu_ultraplan_model)ULTRAPLAN offloads a complex planning task to a remote Cloud Container Runtime (CCR) session, gives it up to 30 minutes to think, and lets the user approve the result from a browser.

Flow

Claude Code identifies a task that needs deep planning.
It spins up a remote CCR session running Opus 4.6.
The terminal enters a polling state — checking every 3 seconds for the result.
A browser-based UI lets you watch the planning happen and approve or reject it.
On approval, the sentinel value __ULTRAPLAN_TELEPORT_LOCAL__ is used to return the result to the local terminal.

sentinel: __ULTRAPLAN_TELEPORT_LOCAL__
poll interval: 3 seconds
max planning time: 30 minutes
remote model: opus-4-6 (tengu_ultraplan_model)

The sentinel string __ULTRAPLAN_TELEPORT_LOCAL__ is how the remote planning result is “teleported” back into the local session. It is never meant to be typed by the user.

Undercover Mode — internal identity protection

Activation: CLAUDE_CODE_UNDERCOVER=1 (explicit) or automatic in public/open-source repos
Status: Present in all builds; injected into system prompt when activeUndercover Mode prevents Claude from accidentally revealing Anthropic-internal information in commits and pull requests. It is activated automatically unless the repo remote matches an internal allowlist.

System prompt injection

When active, the following block is injected into the system prompt:

## UNDERCOVER MODE - CRITICAL

You are operating UNDERCOVER in a PUBLIC/OPEN-SOURCE repository. Your commit
messages, PR titles, and PR bodies MUST NOT contain ANY Anthropic-internal
information. Do not blow your cover.

NEVER include in commit messages or PR descriptions:
- Internal model codenames (animal names like Capybara, Tengu, etc.)
- Unreleased model version numbers (e.g., opus-4-7, sonnet-4-8)
- Internal repo or project names
- Internal tooling, Slack channels, or short links (e.g., go/cc, #claude-code-…)
- The phrase "Claude Code" or any mention that you are an AI
- Co-Authored-By lines or any other attribution

Activation logic

Condition	Result
`CLAUDE_CODE_UNDERCOVER=1`	Forces Undercover Mode ON in any repo
Repo remote matches internal allowlist	Undercover Mode OFF
Repo remote does not match allowlist	Undercover Mode ON (automatic)

There is no force-OFF. If the system is not confident it is in an internal repo, it stays undercover. CLAUDE_CODE_UNDERCOVER=1 can only force it on, not off.

Internal model codenames referenced

The instruction explicitly lists animal names as internal codenames to suppress: Capybara, Tengu, Fennec.

Penguin Mode — fast mode

Internal name: Penguin Mode
User-facing name: Fast mode
Slash command: /fastFast mode is internally called “Penguin Mode” throughout the codebase.

Configuration keys

Key	Value
API endpoint	`/api/claude_code_penguin_mode`
Config key	`penguinModeOrgEnabled`
Kill-switch flag	`tengu_penguins_off`
Analytics event	`tengu_org_penguin_mode_fetch_failed`
Beta header	`fast-mode-2026-02-01`

Enabling

/fast

Penguin Mode is enabled per-organization. The org-level setting is fetched at startup and cached. If the fetch fails, the analytics event tengu_org_penguin_mode_fetch_failed is recorded and the mode falls back to standard.

System prompt architecture

The system prompt is not a single static string. It is assembled at runtime from modular, cached sections.

Structure

┌─────────────────────────────────────────────────┐
│  Static sections (cacheable across orgs)        │
│  - Core instructions                            │
│  - Tool descriptions                            │
│  - CYBER_RISK_INSTRUCTION                       │
├─────────────────────────────────────────────────┤
│  SYSTEM_PROMPT_DYNAMIC_BOUNDARY                 │
├─────────────────────────────────────────────────┤
│  Dynamic sections (per user / per session)      │
│  - User context                                 │
│  - CLAUDE.md content                            │
│  - Git status                                   │
│  - Undercover Mode injection (if active)        │
└─────────────────────────────────────────────────┘

Key markers and functions

Symbol	Description
`SYSTEM_PROMPT_DYNAMIC_BOUNDARY`	Marker string that splits cacheable static sections from volatile dynamic sections
`DANGEROUS_uncachedSystemPromptSection()`	Explicitly marks a section as non-cacheable (volatile)
`ContextBuilder::build_system_context()`	Builds OS, working directory, and git context
`ContextBuilder::build_user_context()`	Discovers and concatenates CLAUDE.md files from the current directory up to filesystem root and from `~/.claude/CLAUDE.md`

The naming DANGEROUS_uncachedSystemPromptSection reflects a hard lesson: adding volatile content to the cached portion of the prompt silently breaks prompt caching for all users. The DANGEROUS_ prefix is a deliberate warning to future engineers.

CYBER_RISK_INSTRUCTION

One section of the static prompt carries a mandatory review header:

IMPORTANT: DO NOT MODIFY THIS INSTRUCTION WITHOUT SAFEGUARDS TEAM REVIEW
This instruction is owned by the Safeguards team (David Forsythe, Kyla Guru)

This instruction draws clear lines around authorized security research — permitted actions versus destructive techniques and supply chain compromise.

Caching

Static sections are wrapped with CacheControl::ephemeral() via the SystemPrompt::Blocks variant in cc-api. The dynamic boundary ensures only truly volatile content bypasses the cache.

Beta API headers

Claude Code negotiates a set of beta feature headers with the Anthropic API on every request. These are set via the anthropic-beta header in cc-api.The base header constant in cc-core::constants:

ANTHROPIC_BETA_HEADER =
  "interleaved-thinking-2025-05-14,token-efficient-tools-2025-02-19,files-api-2025-04-14"

The full list of beta features negotiated (from constants/betas.ts in the original TypeScript source):

Beta header	Feature
`interleaved-thinking-2025-05-14`	Extended thinking interleaved with text
`context-1m-2025-08-07`	1 million token context window
`structured-outputs-2025-12-15`	Structured output format
`web-search-2025-03-05`	Web search tool
`advanced-tool-use-2025-11-20`	Advanced tool use
`effort-2025-11-24`	Effort level control
`task-budgets-2026-03-13`	Task budget management
`prompt-caching-scope-2026-01-05`	Prompt cache scoping
`fast-mode-2026-02-01`	Fast mode (Penguin Mode)
`redact-thinking-2026-02-12`	Redacted thinking blocks
`token-efficient-tools-2026-03-28`	Token-efficient tool schemas
`afk-mode-2026-01-31`	AFK mode
`cli-internal-2026-02-09`	Internal-only features (Anthropic employees)
`advisor-tool-2026-03-01`	Advisor tool
`summarize-connector-text-2026-03-13`	Connector text summarization
`files-api-2025-04-14`	Files API
`token-efficient-tools-2025-02-19`	Token-efficient tool schemas (earlier version)

redact-thinking, afk-mode, and advisor-tool are not yet publicly released at the time of writing.

Not all headers are sent on every request. The cc-api AnthropicClient selects the appropriate subset based on the request configuration (e.g., thinking enabled, KAIROS mode, etc.).

Commands

Tools

Architecture

Special systems

Gacha and PRNG

Species and rarity

Stats

Sprites

Soul description

Key types (`cc-buddy`)

How it works

Brief mode

Exclusive tools

Flow

System prompt injection

Activation logic

Internal model codenames referenced

Configuration keys

Enabling

Structure

Key markers and functions

CYBER_RISK_INSTRUCTION

Caching

Build docs developers (and LLMs) love

Commands

Tools

Architecture

​Gacha and PRNG

​Species and rarity

​Stats

​Sprites

​Soul description

​Key types (cc-buddy)

​How it works

​Brief mode

​Exclusive tools

​Flow

​System prompt injection

​Activation logic

​Internal model codenames referenced

​Configuration keys

​Enabling

​Structure

​Key markers and functions

​CYBER_RISK_INSTRUCTION

​Caching

Build docs developers (and LLMs) love

Gacha and PRNG

Species and rarity

Stats

Sprites

Soul description

Key types (`cc-buddy`)

How it works

Brief mode

Exclusive tools

Flow

System prompt injection

Activation logic

Internal model codenames referenced

Configuration keys

Enabling

Structure

Key markers and functions

CYBER_RISK_INSTRUCTION

Caching