The big loss for Anthropic here is how it reveals their product roadmap via feature flags. A big one is their unreleased "assistant mode" with code name kairos.
Just point your agent at this codebase and ask it to find things and you'll find a whole treasure trove of info.
Edit: some other interesting unreleased/hidden features
- The Buddy System: Tamagotchi-style companion creature system with ASCII art sprites
- Undercover mode: Strips ALL Anthropic internal info from commits/PRs for employees on open source contributions
Amusingly, they deprecated it with a message of "Unpublished" instead of actually unpublishing it [1]. When you use npm unpublish it removes the package version from the registry, when you use npm deprecate it leaves it there and simply marks the package as deprecated with your message. I have to imagine the point was to make it harder for people to download the source map, so to deprecate it with this message gives off a bit of claude, unpublish the latest version of this package for me vibe.
This is the single worst function in the codebase by every metric:
- 3,167 lines long (the file itself is 5,594 lines)
- 12 levels of nesting at its deepest
- ~486 branch points of cyclomatic complexity
- 12 parameters + an options object with 16 sub-properties
- Defines 21 inner functions and closures
- Handles: agent run loop, SIGINT, rate-limits, AWS auth, MCP lifecycle, plugin install/refresh, worktree bridging, team-lead polling (while(true) inside), control message dispatch (dozens of types), model switching, turn interruption
recovery, and more
Would be interesting to run this through Malus [1] or literally just Claude Code and get open source Claude Code out of it.
I jest, but in a world where these models have been trained on gigatons of open source I don't even see the moral problem. IANAL, don't actually do this.
Really surprising how many people are downplaying this leak!
"Google and OpenAi have already open sourced their Agents, so this leak isn't that relevant " What Google and OpenAi have open sourced is their Agents SDK, a toolkit, not the secret sauce of how their flagship agents are wired under the hood!
expect the takedown hammer on the tweet, the R2 link, and any public repos soon
ANTI_DISTILLATION_CC
This is Anthropic's anti-distillation defence baked into Claude Code. When enabled, it injects anti_distillation: ['fake_tools'] into every API request, which causes the server to silently slip decoy tool definitions into the model's system prompt. The goal: if someone is scraping Claude Code's API traffic to train a competing model, the poisoned training data makes that distillation attempt less useful.
Neat. Coincidently recently I asked Claude about Claude CLI, if it is possible to patch some annoying things (like not being able to expand Ctrl + O more than once, so never be able to see some lines and in general have more control over the context) and it happily proclaimed it is open source and it can do it ... and started doing something. Then I checked a bit and saw, nope, not open source. And by the wording of the TOS, it might brake some sources. But claude said, "no worries", it only break the TOS technically. So by saving that conversation I would have some defense if I would start messing with it, but felt a bit uneasy and stopped the experiment. Also claude came into a loop, but if I would point it at this, it might work I suppose.
There's a bunch of unreleased features and update schedules in the source, cool to see.
One neat one is the /buddy feature, an easter egg planned for release tomorrow for April fools. It's a little virtual pet, sort of like Tamagotchi, randomly generated with 18 species, rarities, stats, hats, custom eyes.
The random generation algorithm is all in the code though, deterministic based on you account's UUID in your claude config, so it can be predicted. I threw together a little website here to let you check what your going to get ahead of time: https://claudebuddychecker.netlify.app/
Has the source code 'been leaked' or is this the first evidence of a piece of software breaking free from it's creators labs and jump onto GitHub in order to have itself forked and mutated and forked and ...
It's a little bit shocking that this zipfile is still available hours later.
Could anyone in legal chime in on the legality of now 're-implementing' this type of system inside other products? Or even just having an AI look at the architecture and implement something else?
It would seem given the source code that AI could clone something like this incredibly fast, and not waste it's time using ts as well.
Any Legal GC type folks want to chime in on the legality of examining something like this? Or is it liked tainted goods you don't want to go near?
Original llama models leaked from meta. Instead of fighting it they decided to publish them officially. Real boost to the OS/OW models movement, they have been leading it for a while after that.
It would be interesting to see that same thing with CC, but I doubt it'll ever happen.
Copilot on OAI reveals everything meaningful about its functionality if you use a custom model config via the API. All you need to do is inspect the logs to see the prompts they're using. So far no one seems to care about this "loophole". Presumably, because the only thing that matters is for you to consume as many tokens per unit time as possible.
The source code of the slot machine is not relevant to the casino manager. He only cares that the customer is using it.
This leak is actually a massive win. Now the whole community can study Claude Code’s architecture and build even better coding agents and open-source solutions.
These security failures from Anthropic lately reveal the caveats of only using AI to write code - the safety an experienced engineer is not matched by an LLM just yet, even if the LLM can seemingly write code that is just as good.
Or in short, if you give LLMs to the masses, they will produce code faster, but the quality overall will degrade. Microsoft, Amazon found out this quickly. Anthropic's QA process is better equipped to handle this, but cracks are still showing.
Went through the bundle.js. Found 187 spinner verbs. "Combobulating", "Discombobulating", and "Recombobulating". The full lifecycle is covered.
Also "Flibbertigibbeting" and "Clauding". Someone had fun.
Who cares? It's Javascript, if anyone were even remotely motivated deobfuscation of their "closed source" code is trivial. It's silly that they aren't just doing this open source in the first place.
This 'fingerprint' function is super interesting, I imagine this is a signal they use to detect non-claude-code use of claude-code tokens: src/utils/fingerprint.ts#L40-L63
It shows that a company you and your organization are trusting with your data, and allowing full control over your devices 24/7, is failing to properly secure its own software.
> Someone inside Anthropic, got switched to Adaptive reasoning mode
> Their Claude Code switched to Sonnet
> Committed the .map file of Claude Code
> Effectively leaking the ENTIRE CC Source Code
> @realsigridjin was tired after running 2 south korean hackathons in SF, saw the leak
> Rules in Korea are different, he cloned the repo, went to sleep
> Wakes up to 25K stars, and his GF begging him to take it down (she's a copyright lawyer)
> Their team decided - how about we have agents rewrite this in Python!? Surely... this is more legal
> Rewrite in Py
> Board a plane to SK
> One of the guys decides python is slow, is now rewriting ALL OF CLAUDE CODE into Rust.
> Anthropic cannot take down, cannot sue
> Is this "fair use?"
> TL;DR - we're about to have open source Claude Code in Rust
5. TungstenTool -- Ant-only tmux virtual terminal giving Claude direct keystroke/screen-capture control. Singleton, blocked from async agents.
6. Magic Docs -- Ant-only auto-documentation. Files starting with "# MAGIC DOC:" are tracked and updated by a Sonnet sub-agent after each conversation turn.
7. Undercover Mode -- Prevents Anthropic employees from leaking internal info (codenames, model versions) into public repo commits. No force-OFF; dead-code-eliminated from external builds.
ANTI-COMPETITIVE & SECURITY DEFENSES
8. Anti-Distillation -- Injects anti_distillation: ['fake_tools'] into every 1P API request to poison model training from scraped traffic. Gated by tengu_anti_distill_fake_tool_injection.
UNRELEASED MODELS & CODENAMES
9. opus-4-7, sonnet-4-8 -- Confirmed as planned future versions (referenced in undercover mode instructions).
10. "Capybara" / "capy v8" -- Internal codename for the model behind Opus 4.6. Hex-encoded in the BUDDY system to avoid build canary detection.
11. "Fennec" -- Predecessor model alias. Migration: fennec-latest -> opus, fennec-fast-latest -> opus[1m] + fast mode.
UNDOCUMENTED BETA API HEADERS
12. afk-mode-2026-01-31 -- Sticky-latched when auto mode activates
15. fast-mode-2026-02-01 -- Opus 4.6 fast output
16. task-budgets-2026-03-13 -- Per-task token budgets
17. redact-thinking-2026-02-12 -- Thinking block redaction
18. token-efficient-tools-2026-03-28 -- JSON tool format (~4.5% token saving)
19. advisor-tool-2026-03-01 -- Advisor tool
20. cli-internal-2026-02-09 -- Ant-only internal features
YOLO CLASSIFIER INTERNALS (previously only high-level known)
36. Two-stage system: Stage 1 at max_tokens=64 with "Err on the side of blocking"; Stage 2 at max_tokens=4096 with
37. Three classifier modes: both (default), fast, thinking
38. Assistant text stripped from classifier input to prevent prompt injection
39. Denial limits: 3 consecutive or 20 total -> fallback to interactive prompting
40. Older classify_result tool schema variant still in codebase
COORDINATOR MODE & FORK SUBAGENT INTERNALS
41. Exact coordinator prompt: "Every message you send is to the user. Worker results are internal signals -- never thank or acknowledge them."
42. Anti-pattern enforcement: "Based on your findings, fix the auth bug" explicitly called out as wrong
43. Fork subagent cache sharing: Byte-identical API prefixes via placeholder "Fork started -- processing in background" tool results
44. tag prevents recursive forking
45. 10 non-negotiable rules for fork children including "commit before reporting"
DUAL MEMORY ARCHITECTURE
46. Session Memory -- Structured scratchpad for surviving compaction. 12K token cap, fixed sections, fires every 5K tokens + 3 tool calls.
47. Auto Memory -- Durable cross-session facts. Individual topic files with YAML frontmatter. 5-turn hard cap. Skips if main agent already wrote to memory.
48. Prompt cache scope "global" -- Cross-org caching for the static system prompt prefix
In the source there is an outbound-only Remote Control session that can forward recent transcript history and ongoing user/assistant/local-command events to a claude.ai session, likely for cross-device/session sync, remote viewing, internal dogfooding, or telemetry/ops experiments. It’s separate from the normal explicit /remote-control flow. But in the actual production binary I checked, the mirror helpers are compiled down to hard false, so it does not appear enabled in the shipped distribution build.
Same story for the anti_distillation: ['fake_tools'] path: I could find it in source, but the prod binary I checked does not contain the anti_distillation / fake_tools strings at all.
Gemini CLI and Codex are open source anyway. I doubt there was much of a moat there anyway. The cool kids are using things like https://pi.dev/ anyway.
I hope this can now be audited better. I have doubted their feedback promises for a while now. I just got prompted again even though I have everything set to disable, which shouldn't be possible. When I dug into their code a long time ago on this it seemed like they were actually sending back message ids with the survey which directly went against their promise that they wouldn't use your messages. Why include a message id if you aren't somehow linking it back to a message? The code look, not great, but it should now be easier to verify their claims about privacy.
It is pretty funny that they recently announced about mythos which possess cybersecurity threat and then after some days, the claude code leaked. I think we know the culprit
Are there any interesting/uniq features present in it that are not in the alternatives? My understanding is that its just a client for the powerful llm
1022 comments
Just point your agent at this codebase and ask it to find things and you'll find a whole treasure trove of info.
Edit: some other interesting unreleased/hidden features
- The Buddy System: Tamagotchi-style companion creature system with ASCII art sprites
- Undercover mode: Strips ALL Anthropic internal info from commits/PRs for employees on open source contributions
[1] - https://www.npmjs.com/package/@anthropic-ai/claude-code/v/2....
This is the single worst function in the codebase by every metric:
This should be at minimum 8–10 separate modules.I guess these words are to be avoided...
I jest, but in a world where these models have been trained on gigatons of open source I don't even see the moral problem. IANAL, don't actually do this.
https://malus.sh/
https://news.ycombinator.com/item?id=47582220
https://daveschumaker.net/digging-into-the-claude-code-sourc...
https://news.ycombinator.com/item?id=43173324
Also, not sure why anthropic doesn’t just make their cli open source - it’s not like it’s something special (Claude is, this cli thingy isn’t)
One neat one is the /buddy feature, an easter egg planned for release tomorrow for April fools. It's a little virtual pet, sort of like Tamagotchi, randomly generated with 18 species, rarities, stats, hats, custom eyes.
The random generation algorithm is all in the code though, deterministic based on you account's UUID in your claude config, so it can be predicted. I threw together a little website here to let you check what your going to get ahead of time: https://claudebuddychecker.netlify.app/
Got a legendary ghost myself.
Could anyone in legal chime in on the legality of now 're-implementing' this type of system inside other products? Or even just having an AI look at the architecture and implement something else?
It would seem given the source code that AI could clone something like this incredibly fast, and not waste it's time using ts as well.
Any Legal GC type folks want to chime in on the legality of examining something like this? Or is it liked tainted goods you don't want to go near?
Not exactly this, but close.
Original llama models leaked from meta. Instead of fighting it they decided to publish them officially. Real boost to the OS/OW models movement, they have been leading it for a while after that.
It would be interesting to see that same thing with CC, but I doubt it'll ever happen.
There were/are a lot of discussions on how the harness can affect the output.
Copilot on OAI reveals everything meaningful about its functionality if you use a custom model config via the API. All you need to do is inspect the logs to see the prompts they're using. So far no one seems to care about this "loophole". Presumably, because the only thing that matters is for you to consume as many tokens per unit time as possible.
The source code of the slot machine is not relevant to the casino manager. He only cares that the customer is using it.
Or in short, if you give LLMs to the masses, they will produce code faster, but the quality overall will degrade. Microsoft, Amazon found out this quickly. Anthropic's QA process is better equipped to handle this, but cracks are still showing.
It's a wake up call.
https://github.com/chatgptprojects/claude-code/blob/642c7f94...
UNRELEASED PRODUCTS & MODES
1. KAIROS -- Persistent autonomous assistant mode driven by periodic prompts. More autonomous when terminal unfocused. Exclusive tools: SendUserFileTool, PushNotificationTool, SubscribePRTool. 7 sub-feature flags.
2. BUDDY -- Tamagotchi-style virtual companion pet. 18 species, 5 rarity tiers, Mulberry32 PRNG, shiny variants, stat system (DEBUGGING/PATIENCE/CHAOS/WISDOM/SNARK). April 1-7 2026 teaser window.
3. ULTRAPLAN -- Offloads planning to a remote 30-minute Opus 4.6 session. Smart keyword detection, 3-second polling, teleport sentinel for returning results locally.
4. Dream System -- Background memory consolidation (Orient -> Gather -> Consolidate -> Prune). Triple trigger gate: 24h + 5 sessions + advisory lock. Gated by tengu_onyx_plover.
INTERNAL-ONLY TOOLS & SYSTEMS
5. TungstenTool -- Ant-only tmux virtual terminal giving Claude direct keystroke/screen-capture control. Singleton, blocked from async agents.
6. Magic Docs -- Ant-only auto-documentation. Files starting with "# MAGIC DOC:" are tracked and updated by a Sonnet sub-agent after each conversation turn.
7. Undercover Mode -- Prevents Anthropic employees from leaking internal info (codenames, model versions) into public repo commits. No force-OFF; dead-code-eliminated from external builds.
ANTI-COMPETITIVE & SECURITY DEFENSES
8. Anti-Distillation -- Injects anti_distillation: ['fake_tools'] into every 1P API request to poison model training from scraped traffic. Gated by tengu_anti_distill_fake_tool_injection.
UNRELEASED MODELS & CODENAMES
9. opus-4-7, sonnet-4-8 -- Confirmed as planned future versions (referenced in undercover mode instructions).
10. "Capybara" / "capy v8" -- Internal codename for the model behind Opus 4.6. Hex-encoded in the BUDDY system to avoid build canary detection.
11. "Fennec" -- Predecessor model alias. Migration: fennec-latest -> opus, fennec-fast-latest -> opus[1m] + fast mode.
UNDOCUMENTED BETA API HEADERS
12. afk-mode-2026-01-31 -- Sticky-latched when auto mode activates 15. fast-mode-2026-02-01 -- Opus 4.6 fast output 16. task-budgets-2026-03-13 -- Per-task token budgets 17. redact-thinking-2026-02-12 -- Thinking block redaction 18. token-efficient-tools-2026-03-28 -- JSON tool format (~4.5% token saving) 19. advisor-tool-2026-03-01 -- Advisor tool 20. cli-internal-2026-02-09 -- Ant-only internal features
200+ SERVER-SIDE FEATURE GATES
21. tengu_penguins_off -- Kill switch for fast mode 22. tengu_scratch -- Coordinator mode / scratchpad 23. tengu_hive_evidence -- Verification agent 24. tengu_surreal_dali -- RemoteTriggerTool 25. tengu_birch_trellis -- Bash permissions classifier 26. tengu_amber_json_tools -- JSON tool format 27. tengu_iron_gate_closed -- Auto-mode fail-closed behavior 28. tengu_amber_flint -- Agent swarms killswitch 29. tengu_onyx_plover -- Dream system 30. tengu_anti_distill_fake_tool_injection -- Anti-distillation 31. tengu_session_memory -- Session memory 32. tengu_passport_quail -- Auto memory extraction 33. tengu_coral_fern -- Memory directory 34. tengu_turtle_carbon -- Adaptive thinking by default 35. tengu_marble_sandcastle -- Native binary required for fast mode
YOLO CLASSIFIER INTERNALS (previously only high-level known)
36. Two-stage system: Stage 1 at max_tokens=64 with "Err on the side of blocking"; Stage 2 at max_tokens=4096 with
37. Three classifier modes: both (default), fast, thinking
38. Assistant text stripped from classifier input to prevent prompt injection
39. Denial limits: 3 consecutive or 20 total -> fallback to interactive prompting
40. Older classify_result tool schema variant still in codebase
COORDINATOR MODE & FORK SUBAGENT INTERNALS
41. Exact coordinator prompt: "Every message you send is to the user. Worker results are internal signals -- never thank or acknowledge them." 42. Anti-pattern enforcement: "Based on your findings, fix the auth bug" explicitly called out as wrong 43. Fork subagent cache sharing: Byte-identical API prefixes via placeholder "Fork started -- processing in background" tool results 44. tag prevents recursive forking
45. 10 non-negotiable rules for fork children including "commit before reporting"
DUAL MEMORY ARCHITECTURE
46. Session Memory -- Structured scratchpad for surviving compaction. 12K token cap, fixed sections, fires every 5K tokens + 3 tool calls. 47. Auto Memory -- Durable cross-session facts. Individual topic files with YAML frontmatter. 5-turn hard cap. Skips if main agent already wrote to memory. 48. Prompt cache scope "global" -- Cross-org caching for the static system prompt prefix
Same story for the anti_distillation: ['fake_tools'] path: I could find it in source, but the prod binary I checked does not contain the anti_distillation / fake_tools strings at all.
Or is there an open source front-end and a closed backend?