“ Deep Codebase Analysis & Community Feedback Round 22 (April 8, 2026)

• Source ID: src-20260409-35a0e20bf159
• Kind: analysis
• Scope: shared
• Origin: community-analysis
• Raw path: sources/raw/deep-codebase-analysis-community-feedback-round-22-april-8-2026__src-20260409-35a0e20bf159.md
• Status: active

Summary

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

Tags

analysis

Extracted entities

• ent-20260419-a1b2c3d4e5f6 Two-Version Output Efficiency Directive (concept): Conditional branch in the Claude Code system prompt that renders structurally di
• ent-20260419-a3b4c5d6e7f8 Claude Self-Observation Passage (concept): Passage written by Claude itself after analyzing its own session logs, most-quot
• ent-20260419-a7b8c9d0e1f2 Anthropic Bun Fork Native String Replacement (service): Anthropic's custom Bun fork that ships with the standalone Claude Code binary co
• ent-20260419-b2c3d4e5f6a1 Numeric Length Anchors (Anthropic-Only) (concept): Hard numeric word-count limits appended to the system prompt only for Anthropic-
• ent-20260419-b4c5d6e7f8a9 Context-Is-Architecture Thesis (concept): Thesis elevated in the HackerNews code-quality debate (item 47665285) that with
• ent-20260419-b8c9d0e1f2a3 CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING (concept): Environment variable confirmed by Boris Cherny as the interim workaround for the
• ent-20260419-c3d4e5f6a1b2 Verification Agent Gate (concept): Conditional system-prompt section (rendered only when verification_agent feature
• ent-20260419-c9d0e1f2a3b4 CLAUDE_CODE_AUTO_COMPACT_WINDOW (concept): Environment variable controlling the token threshold at which auto-compact trigg
• ent-20260419-d0e1f2a3b4c5 David Breunig System Prompt Interactive Map (document): Interactive analysis by David Breunig (dbreunig.com, April 4 2026) reconstructin
• ent-20260419-d4e5f6a1b2c3 Explore Agent Read-Only System Prompt (concept): Hard-coded system prompt block for the Explore subagent that STRICTLY PROHIBITS
• ent-20260419-e1f2a3b4c5d6 keepCodingInstructions Flag (concept): Configuration flag controlling inclusion of the 'Doing Tasks' coding-philosophy
• ent-20260419-e5f6a1b2c3d4 Three-Model Subagent Execution Matrix (concept): Cross-dimension comparison of the three subagent delegation models confirmed by
• ent-20260419-f2a3b4c5d6e7 Q1 2026 Feature Release Cluster (concept): Dense feature-shipping window in Feb-March 2026 during which 12+ features shippe
• ent-20260419-f6a1b2c3d4e5 Claude Flow V3 vs TeammateTool 92% Overlap (document): Architectural comparison by rUv (GitHub gist 18dc8d060194017b989d1f8993919ee4) q

Extracted claims

• clm-20260419-a1b2c3d4e5f6 [fact]: Anthropic internal users receive person-oriented prompt; external users receive brevity directive
• clm-20260419-a3b4c5d6e7f8 [observation]: Model cannot introspect real-time reasoning budget
• clm-20260419-a7b8c9d0e1f2 [metric]: Teammate mode costs 3-4x single-session token budget; Fork 1.3-1.5x; Worktree 1x per session
• clm-20260419-b2c3d4e5f6a1 [fact]: Adaptive thinking allocated zero reasoning tokens on fabrication turns
• clm-20260419-b4c5d6e7f8a9 [pattern]: MAGIC DOC tool-level permission boundary supersedes prompt-level instructions
• clm-20260419-b8c9d0e1f2a3 [fact]: Teammate mode speedup comes from eliminating parent-waits-for-child bottleneck
• clm-20260419-c3d4e5f6a1b2 [metric]: Shallow thinking drives 80x API request count and 122x billing despite stable prompt count
• clm-20260419-c5d6e7f8a9b0 [observation]: Claude Flow V3 and Claude Code TeammateTool share 92% structural overlap across 7 primitives
• clm-20260419-c9d0e1f2a3b4 [fact]: Bun fork mutates request body between JSON serialize and TLS encrypt, breaking prompt cache
• clm-20260419-d0e1f2a3b4c5 [fact]: Cache Boundary Marker is invisible to the model
• clm-20260419-d4e5f6a1b2c3 [constraint]: Verification Agent gate requires adversarial verifier on 3+ file, backend, or infrastructure changes
• clm-20260419-d6e7f8a9b0c1 [observation]: Q1 2026 feature cluster shipped 12+ features concurrent with unannounced regression
• clm-20260419-e1f2a3b4c5d6 [fact]: Numeric length anchors apply only to Anthropic employees
• clm-20260419-e5f6a1b2c3d4 [constraint]: Explore subagent is the most-constrained agent prompt in the leaked source
• clm-20260419-e7f8a9b0c1d2 [hypothesis]: Context is the architecture: LLM products' real architecture is prompt-side, not code-side
• clm-20260419-f2a3b4c5d6e7 [constraint]: System prompt architecturally outranks CLAUDE.md regardless of content length
• clm-20260419-f6a1b2c3d4e5 [constraint]: Only Fork supports recursive subagent nesting

Extracted relations

• rel-20260419-a1b2c3d4e5f6: Two-Version Output Efficiency Directive --[contains]--> System Prompt Assembly
• rel-20260419-a3b4c5d6e7f8: CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING --[fixed]--> GitHub Issue #42796
• rel-20260419-a7b8c9d0e1f2: Three-Model Subagent Execution Matrix --[contains]--> Forked Agent Pattern
• rel-20260419-a9b0c1d2e3f4: Claude Self-Observation Passage --[derived_from]--> GitHub Issue #42796
• rel-20260419-b0c1d2e3f4a5: Claude Self-Observation Passage --[related_to]--> Claude Mythos Preview
• rel-20260419-b2c3d4e5f6a1: Two-Version Output Efficiency Directive --[caused]--> GitHub Issue #42796
• rel-20260419-b4c5d6e7f8a9: CLAUDE_CODE_AUTO_COMPACT_WINDOW --[uses]--> Auto-Compact
• rel-20260419-b8c9d0e1f2a3: Three-Model Subagent Execution Matrix --[contains]--> Agent Teams
• rel-20260419-c1d2e3f4a5b6: Context-Is-Architecture Thesis --[related_to]--> System Prompt Assembly
• rel-20260419-c3d4e5f6a1b2: Numeric Length Anchors (Anthropic-Only) --[depends_on]--> USER_TYPE === 'ant' Gate
• rel-20260419-c5d6e7f8a9b0: David Breunig System Prompt Interactive Map --[summarizes]--> System Prompt Assembly
• rel-20260419-c9d0e1f2a3b4: Three-Model Subagent Execution Matrix --[contains]--> Worktree Isolation
• rel-20260419-d0e1f2a3b4c5: Claude Flow V3 vs TeammateTool 92% Overlap --[related_to]--> Agent Teams
• rel-20260419-d2e3f4a5b6c7: David Breunig System Prompt Interactive Map --[summarizes]--> 4-Layer Prompt Pipeline
• rel-20260419-d4e5f6a1b2c3: Verification Agent Gate --[depends_on]--> AgentTool
• rel-20260419-d6e7f8a9b0c1: keepCodingInstructions Flag --[supports]--> System Prompt Assembly
• rel-20260419-e1f2a3b4c5d6: Anthropic Bun Fork Native String Replacement --[caused]--> Billing Sentinel Bug (cch=00000)
• rel-20260419-e5f6a1b2c3d4: Explore Agent Read-Only System Prompt --[implements]--> AgentTool
• rel-20260419-e7f8a9b0c1d2: Q1 2026 Feature Release Cluster --[caused]--> GitHub Issue #42796
• rel-20260419-f2a3b4c5d6e7: Anthropic Bun Fork Native String Replacement --[depends_on]--> Bun Runtime
• rel-20260419-f6a1b2c3d4e5: Explore Agent Read-Only System Prompt --[implements]--> Self-Contained Worker Context Principle
• rel-20260419-f8a9b0c1d2e3: Q1 2026 Feature Release Cluster --[related_to]--> Thinking Depth Regression (pre-redaction)