Round 25 - Deep Codebase Analysis (April 9-11, 2026)

• Source ID: src-20260419-be936b1b7d4c
• Kind: analysis
• Scope: private
• Origin: community-analysis
• Raw path: sources/raw/round-25-deep-codebase-analysis-april-9-11-2026__src-20260419-be936b1b7d4c.md
• Status: active

Summary

Community analysis document covering Claude Code architecture, internals, and leak analysis. See extracted entities and claims below.

Tags

round-25

Extracted entities

• ent-20260419-0f7d395cf2dd OWASP ASI (Agentic Security Integrity) (document): OWASP Agentic Security Integrity risk specification published December 2025. Enu
• ent-20260419-186821584a17 Three-Layer Compression Hierarchy (concept): Three-level context compression hierarchy triggered by budget usage. MicroCompac
• ent-20260419-1ef0c2d239f3 4-Layer Prompt Pipeline (concept): System prompt assembly pipeline with four layers and an explicit cache boundary
• ent-20260419-203b3c47420d GitHub Issue #38335 (issue): GitHub issue 'Claude Max plan session limits exhausted abnormally fast' with 201
• ent-20260419-26819dc122be Employee Post-Edit Verification (concept): Leaked source confirms Anthropic employees receive automatic post-edit type-chec
• ent-20260419-29b5d46f5704 Star Chamber (project): Mozilla AI Claude Code skill that fans code-review and design questions out to m
• ent-20260419-2d119c606736 GitHub Issue #39755 (issue): Kolkov's streaming analysis issue on the claude-code repo. 17 team members tagge
• ent-20260419-3b691618a66c main.tsx (file): Claude Code entry point file. 785 KB on disk in the leaked build.
• ent-20260419-3d6db240fb16 print.ts (file): Source file at src/cli/print.ts containing the monster god-function. File total:
• ent-20260419-4124265a4544 autoCompact.ts (file): Autocompact module that implements AutoCompact-level context compression. Contai
• ent-20260419-44119678b89f Andrey Kolkov (person): Reverse engineer who analyzed 12 versions of Claude Code's cli.js (v2.1.74-v2.1.
• ent-20260419-4ec0189128a2 CLAUDE.md Jailbreak Vector (concept): Attack class where a modified CLAUDE.md in a project overrides safety controls,
• ent-20260419-50c4ffb5475f AgentArmor (project): Open-source 8-layer security framework for AI agents (pip install agentarmor-cor
• ent-20260419-539f0ad095fd Melon Mode (concept): Feature present in earlier reverse-engineered versions of Claude Code, absent fr
• ent-20260419-5a333edae610 ccdiag (project): Go CLI diagnostic tool built by Andrey Kolkov to analyze Claude Code session tel
• ent-20260419-649710b4edf6 CrowdStrike Glasswing Assessment (document): CrowdStrike partner assessment of Anthropic's Mythos Preview, published April 5
• ent-20260419-698667884fad Ouroboros Architecture (concept): Stetskov's diagnosis of Claude Code engineering culture: AI writes the code, AI
• ent-20260419-7ac5ce65a94e RAPTOR (project): Offensive/defensive security agent built entirely as markdown configuration on t
• ent-20260419-834d43628b06 Tool.ts (file): Base tool schema validation and permission enforcement module. Approximately 29,
• ent-20260419-86cd43e02019 Issue Tracker Automation (concept): Three-bot automation pipeline on the claude-code repo: a Claude-Sonnet-powered d
• ent-20260419-87eaa38ec1b1 Denis Stetskov (person): Cultural critic and author of 'The Snake That Ate Itself' on Substack (techtrenc
• ent-20260419-87f15ffa6bb0 userPromptKeywords.ts (file): Source file hosting the frustration-detection regex used for telemetry. The rege
• ent-20260419-9278b9c17677 claude-pragma (project): Hooks-based validator system that converts CLAUDE.md rules into lifecycle events
• ent-20260419-92aa328d4fff Streaming Watchdog Bug Chain (concept): A three-bug compound failure that caused the majority of observed stream hangs:
• ent-20260419-950ce15ad392 rolloutConfig (concept): Configuration object in Claude Code source that declares launch windows for upco
• ent-20260419-a13ed1bab3a4 Tengu (concept): Product/telemetry codename prefix used throughout Claude Code. Drives 250+ analy
• ent-20260419-a16e91446276 GitHub Issue #528 (issue): GitHub issue '[BUG] Claude is not adhering to CLAUDE.md when it should' — esta
• ent-20260419-b3574c7d2067 undercover.ts (file): Source file implementing Undercover Mode. Contains an explicit list of protected
• ent-20260419-b5fa0a2036af Thinking Hash System (concept): Server-side reasoning-preservation mechanism: Anthropic computes a hash of the t
• ent-20260419-baa0484a010f commands.ts (file): Slash-command registry / command system implementation module. Approximately 25,
• ... and 10 more

Extracted claims

• clm-20260419-0c1855957709 [fact]: Anthropic employees get automatic post-edit type-checking and ESLint; users don't
• clm-20260419-0c92ff6967cd [fact]: rolloutConfig flags May 2026 as the next Buddy/KAIROS milestone
• clm-20260419-24608d35281c [metric]: 1 in 6 API requests fail (16.3%) in a production sample
• clm-20260419-270ad058d737 [fact]: Core productive TypeScript is 64,464 lines, not 512,000
• clm-20260419-2cbfc5b20c9f [observation]: AI-authorship percentage was never defined as a metric
• clm-20260419-32952978c9e1 [fact]: undercover.ts contains explicit protected-term list and masking function
• clm-20260419-339a6028b5a7 [fact]: Compaction hierarchy: 80% / 90% / 95% triggers, 0 / 1 / 1 API calls
• clm-20260419-366173bf437e [metric]: npm dependency count sits at 74
• clm-20260419-409d60875202 [fact]: Opus users silently downgraded to Sonnet after 3 consecutive 529s
• clm-20260419-449398de1777 [fact]: cch=00000 is the root cause of issues #38335 and #40524
• clm-20260419-4db8f07c7fe0 [metric]: print.ts main function is 3,167 lines with ~486 branch points and 12 nesting levels
• clm-20260419-63179734c7c3 [metric]: File size distribution is alarming for a shipping product
• clm-20260419-659a47d2873e [decision]: Built-in tools ordered before MCP tools for cache stability
• clm-20260419-66f5bb42add2 [metric]: AI-assisted attacks up 89% YoY per CrowdStrike 2026 threat report
• clm-20260419-678c3b6752ba [metric]: Over 99% of Mythos zero-days remain unpatched
• clm-20260419-698db81932e7 [metric]: REPL.tsx has 470 useState and 372 useEffect hooks
• clm-20260419-6d00fe3a5793 [fact]: Streaming watchdog fails because it initializes after the vulnerable phase
• clm-20260419-75b84699f92f [anti-pattern]: releaseStreamResources is a no-op during the critical phase
• clm-20260419-7b9a5a5feb8e [anti-pattern]: AgentArmor's L4 risk scoring is verb-only — known limitation
• clm-20260419-8195185b6d14 [observation]: print.ts hosts 10+ unrelated subsystems
• clm-20260419-85749b32ccac [anti-pattern]: 250K API calls wasted daily documented in autoCompact.ts source comment
• clm-20260419-9da298dcadc6 [observation]: Local JSONL transcripts undercount thinking depth by design
• clm-20260419-9fc0ffd6225f [fact]: Model codename map complete: Fennec → Opus 4.6 predecessor; Numbat unreleased
• clm-20260419-a118f2cc0c52 [observation]: Frustration sentiment detection uses a regex, not inference
• clm-20260419-a1e05cd0858e [observation]: Ouroboros culture: AI writes, AI reviews, AI deploys — no exit condition
• clm-20260419-a550cd44fd79 [metric]: Capybara v8 exhibits 29-30% false-claims rate
• clm-20260419-ba28cf64267b [hypothesis]: Equivalent streaming logic in Go would be ~30 lines
• clm-20260419-bdf4810800ef [metric]: Code Review triples substantive-feedback rate (16% → 54%)
• clm-20260419-bebe5819a6cd [observation]: Compaction preservation semantics enable the Straiker persistence attack
• clm-20260419-c72b49338921 [metric]: One-line watchdog patch drops ESC aborts 8.7x
• ... and 13 more

Extracted relations

• rel-20260419-0309aed77946: RAPTOR --[depends_on]--> Claude Code
• rel-20260419-070d45755560: print.ts --[related_to]--> tt() Function
• rel-20260419-085ae4e86cb1: claude-pragma --[uses]--> Hooks System
• rel-20260419-0c1855957710: Go Rewrite Proposal (Issue #41981) --[contradicts]--> print.ts
• rel-20260419-0c92ff6967ce: Tengu --[contained_in]--> Model Codenames
• rel-20260419-0f04db8052fd: undercover.ts --[related_to]--> Model Codenames
• rel-20260419-10a18dfeecbc: Andrey Kolkov --[owns]--> ccdiag
• rel-20260419-1377aaf0bf64: Billing Sentinel Bug (cch=00000) --[caused]--> GitHub Issue #40524
• rel-20260419-270ad058d738: REPL.tsx --[implements]--> Terminal Renderer
• rel-20260419-2cbfc5b20c90: rolloutConfig --[contains]--> KAIROS
• rel-20260419-2cfd8433d5d1: AgentArmor --[implements]--> OWASP ASI (Agentic Security Integrity)
• rel-20260419-366173bf437f: Numbat --[contained_in]--> Model Codenames
• rel-20260419-36bfdd35ca5c: Billing Sentinel Bug (cch=00000) --[depends_on]--> Bun Runtime
• rel-20260419-5f3fa3150cd1: autoCompact.ts --[implements]--> Autocompact Death Loop
• rel-20260419-66f5bb42add3: Go Rewrite Proposal (Issue #41981) --[owns]--> Andrey Kolkov
• rel-20260419-678c3b6752bb: CrowdStrike Glasswing Assessment --[summarizes]--> Claude Mythos Preview
• rel-20260419-73c2b440cbb3: claude-pragma --[fixed]--> GitHub Issue #528
• rel-20260419-755004bfb4e2: Billing Sentinel Bug (cch=00000) --[caused]--> GitHub Issue #38335
• rel-20260419-7ba9393770b2: undercover.ts --[implements]--> Undercover Mode
• rel-20260419-805be7bf891f: Thinking Hash System --[informed_by]--> Andrey Kolkov
• rel-20260419-87d003ff6eda: 4-Layer Prompt Pipeline --[supports]--> Cache Economics
• rel-20260419-9ae1cf4666b2: Three-Layer Compression Hierarchy --[contains]--> Auto-Compact
• rel-20260419-a1e05cd0858f: rolloutConfig --[contains]--> Buddy Sprite
• rel-20260419-a54527cf541a: 4-Layer Prompt Pipeline --[related_to]--> System Prompt Assembly
• rel-20260419-a550cd44fd80: AI-Authorship Percentage Escalation --[contradicts]--> Source Map Leak (March 31, 2026)
• ... and 16 more