Claude Code Source Leak: What the Community Actually Found Inside

Today

You saw the Claude Code leak everywhere this week. Most of the noise was about code quality. Here are the interesting findings in the noise:

  • AutoDream mode. After 5 sessions and 24 hours of silence, Claude spawns a background agent that reviews and consolidates its own memory files. Separately, KAIROS is an assistant mode that turns Claude into a persistent agent. Resumable sessions, team coordination, auto-backgrounding of blocking tasks. It plugs into the Agent SDK. Two unreleased systems. Both gated behind feature flags.

    • src/services/autoDream/autoDream.ts:64-65 — minHours: 24, minSessions: 5
    • src/services/autoDream/config.ts:16-20 — Feature flag tengu_onyx_plover
    • src/services/autoDream/consolidationPrompt.ts:10-65 — 4-phase dream: Orient/Gather/Consolidate/Prune
    • src/bootstrap/state.ts:72 — kairosActive: boolean; src/bridge/bridgeMain.ts:1523 — feature('KAIROS') gate

  • Coordinator Mode. One Claude instance spawns and manages multiple worker agents in parallel. Task distribution, result aggregation, output conflict resolution. The orchestration strategy? A ~258-line system prompt. The model decides what to parallelize at inference time. Code underneath handles spawning, lifecycle, and notifications.

    • src/coordinator/coordinatorMode.ts:111-369 — ~258-line system prompt defining orchestration behavior
    • src/utils/swarm/spawnInProcess.ts:104-216 — spawnInProcessTeammate() worker creation
    • src/tasks/LocalAgentTask/LocalAgentTask.tsx:197-262 — XML <task-notification> result aggregation
    • src/coordinator/coordinatorMode.ts:213-217 — "Parallelism is your superpower. Workers are async."

  • Fake tools as anti-copying poison. Claude Code mixes fake tool definitions into API calls. If a competitor captures traffic to train a copycat, their model learns to call tools that don't exist. A second mechanism compresses assistant output between tool calls server-side. Different technique, same goal.

    • src/services/api/claude.ts:301-313 — anti_distillation: ['fake_tools'] opt-in, gated behind tengu_anti_distill_fake_tool_injection
    • src/utils/betas.ts:279-284 — Separate mechanism: server-side connector-text summarization. Ant-only, requires Capability.ANTHROPIC_INTERNAL_RESEARCH

  • Frustration detection via regex. Not their own LLM. A regex. Pattern-matching for "wtf" and "this sucks" to track analytics. Peak irony.

    • src/utils/userPromptKeywords.ts:4-11 — negativePattern regex matching wtf, this sucks, damn it, profanity variants
    • src/utils/processUserInput/processTextPrompt.ts:59-64 — Logs tengu_input_prompt event with is_negative flag
    • src/screens/REPL.tsx:104-110 — useFrustrationDetection conditionally loaded (Ant-only/dogfooding)

  • A two-tier web. 89 documentation sites are hardcoded for full content extraction. Every other website gets a 125-character paraphrased snippet.

    • src/tools/WebFetchTool/preapproved.ts:14-131 — PREAPPROVED_HOSTS Set with 89 domains
    • src/tools/WebFetchTool/prompt.ts:31 — "Enforce a strict 125-character maximum for quotes from any source document"
    • src/tools/WebFetchTool/prompt.ts:23-46 — makeSecondaryModelPrompt() splits logic on isPreapprovedDomain

  • CLAUDE.md is part of the system prompt. Four hierarchy levels: managed, user, project, local. Every line you write costs tokens on every single message. Every line.

    • src/utils/claudemd.ts:1-26 — Hierarchy: 1. Managed (/etc/claude-code/CLAUDE.md), 2. User (~/.claude/CLAUDE.md), 3. Project (CLAUDE.md, .claude/CLAUDE.md, .claude/rules/*.md), 4. Local (CLAUDE.local.md)
    • src/utils/claudemd.ts:790-920 — getMemoryFiles() loading logic with upward directory walk
    • src/context.ts:155-189 — getUserContext() injects into system prompt via getClaudeMds()

  • 250K wasted API calls per day. A code comment dated March 10, 2026 reveals 1,279 sessions hit 50+ consecutive autocompact failures (one session reached 3,272). The fix? One constant: MAX_CONSECUTIVE_AUTOCOMPACT_FAILURES = 3. Three strikes and compaction disables for the session. That bug sat in production burning a quarter million API calls daily until someone checked the metrics.

    • src/services/compact/autoCompact.ts:67-70 — // BQ 2026-03-10: 1,279 sessions had 50+ consecutive failures (up to 3,272) wasting ~250K API calls/day globally.
    • src/services/compact/autoCompact.ts:70 — const MAX_CONSECUTIVE_AUTOCOMPACT_FAILURES = 3
    • src/services/compact/autoCompact.ts:260-265 — Circuit breaker check; src/services/compact/autoCompact.ts:338-349 — Failure counter increment and trip log

  • Capybara v8 (internal codename, likely Sonnet 4.6) false-claims rate: 29-30%. Their next-gen model has nearly double the false-claims rate of v4 (16.7%). Anthropic's fix? Extra system prompt rules that forbid the model from claiming tests pass when they don't, from suppressing failing checks, from hedging confirmed results. Currently internal-only, being validated via A/B testing before external rollout. The most honest production comment in the entire codebase.

    • src/constants/prompts.ts:237-240 — // @[MODEL LAUNCH]: False-claims mitigation for Capybara v8 (29-30% FC rate vs v4's 16.7%)
    • src/constants/prompts.ts:240 — "Never claim 'all tests pass' when output shows failures, never suppress or simplify failing checks to manufacture a green result"

  • ULTRAPLAN. Claude offloads complex planning to a cloud container that thinks for up to 30 minutes. The plan appears in your browser for approval. You can reject and iterate. Then choose: execute locally or let the cloud session code it and open a PR. This is where AI agents are headed.

    • src/commands/ultraplan.tsx:24 — const ULTRAPLAN_TIMEOUT_MS = 30 * 60 * 1000
    • src/commands/ultraplan.tsx:464 — "~10–30 min · Claude Code on the web drafts an advanced plan you can edit and approve"
    • src/utils/ultraplan/ccrSession.ts:66 — UltraplanPhase = 'running' | 'needs_input' | 'plan_ready'
    • src/commands/ultraplan.tsx:100-101 — executionTarget === 'remote': "User chose execute in CCR in the browser PlanModal"

I use Claude Code every day. This was an accidental product roadmap, it is the most interesting one I have read for a while.