44 feature flags, 20 externally inactive. KAIROS, Auto-Dream, UltraPlan, Coordinator, Bridge, Daemon, UDS Inbox, Buddy, plus anti-distillation and undercover mode.
The most interesting part of the Claude Code source leak is not “what works right now.” It is what does not work yet.
Of 44 feature flags, 20 are externally inactive — roughly 45% of the roadmap is experimental. Behind those 20 disabled flags sit 8 independent feature modules, each pointing toward the future direction of AI coding tools.
Feature Connection Map
Before examining individual features, here is the full picture. The 8 features are not isolated experiments — they connect along four axes.
graph TB
subgraph "Memory Axis"
KAIROS["KAIROS<br/>Proactive Agent"]
AD["Auto-Dream<br/>Memory Consolidation"]
end
subgraph "Background Axis"
DAEMON["Daemon<br/>Background Execution"]
UDS["UDS Inbox<br/>Inter-Session Comms"]
end
subgraph "Multi-Agent Axis"
COORD["Coordinator<br/>Leader-Worker"]
UP["UltraPlan<br/>30-min Opus Planning"]
end
subgraph "Remote Axis"
BRIDGE["Bridge<br/>Remote Control"]
BUDDY["Buddy<br/>Virtual Pet"]
end
KAIROS <-->|"dream mode<br/>cross-session memory"| AD
DAEMON <-->|"tmux + UDS<br/>process isolation"| UDS
COORD <-->|"leader can invoke<br/>UltraPlan"| UP
BRIDGE <-->|"remote + persistence<br/>session continuity"| DAEMON
COORD -->|"work dispatch"| UDS
KAIROS -->|"5-min cron =<br/>background loop"| DAEMON
The Memory axis handles memory that outlives sessions. The Background axis enables execution without a terminal. The Multi-Agent axis provides multi-agent collaboration. The Remote axis expands access pathways.
1. KAIROS: The Agent That Never Sleeps
KAIROS (Knowledge and Autonomous Intelligence with Reliable Operation System) is the most ambitious unreleased feature: a persistent proactive assistant that observes and acts without explicit user commands. Found in src/memdir/ and src/tools/SleepTool/.
| Component | Role | Implementation Status |
|---|---|---|
Dream Mode (/dream skill) | Session-to-session memory consolidation | Skill definition exists, execution incomplete |
| GitHub Webhooks | External event detection (PRs, Issues) | Interface defined, handler stubs |
| Background Daemon | Always-on without terminal | Linked to Daemon module |
| 5-min cron refresh | Periodic context update | Cron logic exists, inactive |
| SleepTool | Enter/exit standby state | Tool registered, feature-gated |
The structure is laid out but the execution layer is incomplete. Architecturally, it requires Daemon + UDS Inbox + Auto-Dream to function — meaning KAIROS is the final puzzle piece that completes once other features stabilize.
2. UltraPlan: 30 Minutes of Thinking
UltraPlan allocates up to 30 minutes of dedicated inference time on a remote cloud Opus-class model for deep planning. Found in src/commands/ultraplan.tsx.
Workflow: user submits a complex planning request -> request sent to remote Opus model -> browser-based approval workflow (cost/time confirmation) -> polling-based execution (local periodically checks progress) -> completed plan teleported back to local.
The browser approval step is pragmatic engineering. It prevents accidental “$50 charge from running 30 minutes of Opus” scenarios.
3. Auto-Dream: Memory Consolidation While You Sleep
Implemented in src/services/autoDream/autoDream.ts, Auto-Dream is a background memory consolidation engine that runs after sessions end. A forked sub-agent operates independently from the main context:
- Reviews the entire just-finished conversation
- Extracts core patterns and learnings
- Organizes them into
memdir/storage - Cleans up stale context
- Strengthens associations between related memories
The process mirrors how the human brain consolidates short-term memory into long-term memory during sleep — hence the name “Dream.” Because it runs as a sub-agent, it does not consume the main session’s context window.
4. Coordinator Mode: The Leader Never Edits Code
Found in src/coordinator/ and src/utils/swarm/, Coordinator Mode is a leader-worker multi-agent orchestration system.
flowchart LR
subgraph "Phase 1: Research"
direction TB
R1["Worker A<br/>Parallel Investigation"]
R2["Worker B<br/>Parallel Investigation"]
end
subgraph "Phase 2: Synthesis"
direction TB
LEADER["Leader<br/>Reads Code Directly<br/>No Delegation"]
end
subgraph "Phase 3: Implement"
direction TB
I1["Worker C<br/>Per-Area Changes"]
I2["Worker D<br/>Per-Area Changes"]
end
subgraph "Phase 4: Verify"
direction TB
V1["Worker E<br/>Independent Tests"]
V2["Worker F<br/>Independent Tests"]
end
R1 --> LEADER
R2 --> LEADER
LEADER --> I1
LEADER --> I2
I1 --> V1
I2 --> V2
4-phase execution model: Research (parallel) -> Synthesis (sequential — leader reads all code directly, no delegation) -> Implement (per-area, avoiding conflicts) -> Verify (parallel, independent test suites).
The core principle: the leader never edits code. It can read but has no access to editing tools. Every modification is delegated to workers. Each worker operates in an isolated git worktree with an independent QueryEngine, providing both filesystem and context isolation.
5. Bridge: Control Code From Your Phone
Implemented across 31 files in src/bridge/, Bridge enables remote control of Claude Code from a phone or browser. It supports WebSocket permission sync, JWT authentication, and session handoff.
| Property | Detail |
|---|---|
| Max parallel sessions | 32 |
| Token refresh (CCR v2) | Session JWT ~5h55m, reconnect 5 min before expiry |
| Backoff strategy | Connection error: 2s escalating to 120s cap, give up after 10 min |
Combined with Daemon Mode, sessions persist even after closing the terminal. Start a large refactoring on desktop, check progress on your phone during commute, approve permission requests remotely.
6. Daemon Mode and UDS Inbox
Daemon Mode launches Claude Code as a background session via --bg, creating a tmux session that survives terminal closure. It serves as the foundation infrastructure for KAIROS’s 5-min cron, Coordinator’s worker processes, and Bridge’s remote sessions.
UDS Inbox is a Unix Domain Socket-based inter-session messaging system. Only processes on the same machine can communicate (security), with zero network overhead. Coordinator dispatches work through it, SendMessage tool uses it for inter-session data exchange, and ListPeers discovers active sessions.
7. Buddy: The Tamagotchi Easter Egg
Hidden in src/buddy/companion.ts is a Tamagotchi-style virtual pet system — more team culture artifact than production feature.
| Property | Detail |
|---|---|
| Species | 18 (duck, owl, cat, penguin, dragon, axolotl, capybara, and more) |
| Rarity | Tiers from common to legendary |
| Stats | RPG-style |
| Shiny chance | 1% |
| Individual determination | Account ID-based PRNG (pseudorandom) |
| Obfuscation | String encoding to bypass build inspection |
Security and Defense Mechanisms
Anti-Distillation
When the ANTI_DISTILLATION_CC flag in claude.ts is active, fake tools are injected into the system prompt’s tool list. Controlled by GrowthBook flag tengu_anti_distill_fake_tool_injection. Server-side, betas.ts also runs summary-based protection.
Bypassable via MITM proxy or the CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS environment variable. The USER_TYPE==='ant' condition restricts certain behaviors to Anthropic-internal users only.
Undercover Mode
undercover.ts removes Anthropic code names, Slack channel references, and “Claude Code” text from external outputs. Activated with CLAUDE_CODE_UNDERCOVER=1. Source comments state “There is NO force-OFF” — meaning no condition forces deactivation.
The controversy: prompts instruct the AI not to mention being AI in commits and pull requests.
Frustration Regex
userPromptKeywords.ts detects frustration expressions like “wtf”, “this sucks”, “fuck you” via regex pattern matching. The irony: an LLM company using regex for sentiment detection.
Native Client Attestation
system.ts contains a cch=00000 placeholder. At runtime, Bun’s Zig HTTP layer replaces it with an actual hash. JavaScript-level code cannot manipulate this value. Only works in the official Bun binary, not in Node — this is DRM-level client attestation that lets the server verify requests originate from an official Claude Code binary.
Competition Comparison
| Capability | Claude Code | OpenAI Codex | Gemini CLI | Claw Code (OSS) |
|---|---|---|---|---|
| Source model | Closed (leaked) | Open source | Open source | Clean-room Rust, 121K stars |
| Autonomous agent | KAIROS (unreleased) | Cloud sandbox | N/A | N/A |
| Memory consolidation | Auto-Dream (unreleased) | Basic memory | Gemini Memory | N/A |
| Multi-agent | Coordinator (unreleased) | N/A | N/A | N/A |
| Long-range planning | UltraPlan (unreleased) | N/A | Deep Think | N/A |
| Remote control | Bridge (unreleased) | Web UI native | N/A | N/A |
| Extensibility | MCP 5 Transport | Plugin API | MCP support | MCP partial |
| Anti-distillation | Multi-layered | N/A | N/A | N/A (OSS) |
What This Tells Us
Taken together, the 8 unreleased features point in two directions.
The agent tools market is shifting: from “you use the tool” to “the tool acts on its own.” The KAIROS + Daemon + Bridge combination envisions agents operating without explicit user invocation — detecting GitHub events in the background, analyzing relevant code, and reporting findings to your phone.
Harness engineering as a new discipline: as model performance alone becomes harder to differentiate, the tooling that wraps the model — permissions, memory, orchestration, security — becomes the competitive moat. Coordinator’s 4-phase execution model, Auto-Dream’s memory consolidation pipeline, and Native Client Attestation’s DRM-level verification are all harness-level innovations.
Of 44 feature flags, 20 are inactive. This is not “incomplete.” It is “planned future.” Design left in source code is more honest than marketing blogs.
Sources & Limitations
This series synthesizes the following publicly available analyses and does not directly contain leaked source code.
| Source | URL | Focus |
|---|---|---|
| ccunpacked.dev | ccunpacked.dev | Visual architecture guide, tool/command catalog |
| Wikidocs Analysis | wikidocs.net/338204 | Detailed technical analysis (execution flow, state, rendering) |
| PyTorch KR | discuss.pytorch.kr | Community analysis + HN discussion synthesis |
| Claw Code | github.com/ultraworkers/claw-code | Clean-room reimplementation (Rust/Python), PARITY.md gap analysis |
Analysis date: April 2, 2026. Anthropic issued DMCA takedowns on 8,100+ forks and discontinued npm distribution shortly after the leak, so some sources may have changed accessibility. Features behind feature flags are unreleased and may be modified or deprecated before launch.
Related Posts

The Memory System I Built Looked Like Claude Code's Internal Design
CC internal Memory System (4-type persistent memory, Auto-Dream, Auto-Compact) vs independently built 3-layer system (documents→index→semantic search). validate_placement() is the differentiator CC doesn't have.

Beyond Ralph Loop — Self-Evolving Agents and the Shifting Role of AI Developers
Ralph Loop solved context rot but remains prompt-bound. This post maps the trajectory from ALAS autonomous parameter updates to Self-Evolving Agent loops, Multi-Agent Swarms with World Models, Korea's Ralphathon results (100K LOC, 70% tests, zero human keystrokes), and the concrete shift in developer roles from implementation to specification and verification.

The Evolution of AI Agent Loops — From RLHF to Ralph Loop
RLHF, ReAct, Reflexion, LangGraph/AutoGen, Context Rot, Ralph Loop. Six generations of agent loop architecture — what each solved, what each broke, and why a Bash while-loop turned out to be the answer.