Claude Code Codebase - Deep Technical Understanding Guide

• Source ID: src-20260419-f46b0a2fccd9
• Kind: analysis
• Scope: private
• Origin: community-analysis
• Raw path: sources/raw/claude-code-codebase-deep-technical-understanding-guide__src-20260419-f46b0a2fccd9.md
• Status: active

Summary

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

Tags

architecture internals

Extracted entities

• ent-20260419-026ec2d354e8 bashSecurity.ts Numbered Checks (concept): 23 sequentially numbered security validators inside bashSecurity.ts (~2,000 line
• ent-20260419-0416c1181feb extractMemories.ts (file): ~600-line file implementing a fire-and-forget hidden Opus call per turn used to
• ent-20260419-0847d810164d query.ts 10-Step Loop (concept): Fully dissected per-iteration structure of the query.ts state-machine while-loop
• ent-20260419-1a2d3f40f0cf Streaming Tool Execution Parallelism (concept): Design where tool invocations are executed as their definitions stream in from t
• ent-20260419-1b3359938871 14 Cache-Breaking Vectors (concept): Enumerated list of 14 events that invalidate the prompt cache: CLAUDE.md mid-ses
• ent-20260419-250777fe7786 Harrison Security Deep Dive (Query Loop) (document): Harrison Security blog analysis dissecting the query.ts 1,421-line while-loop wi
• ent-20260419-267b85260b02 Denial Tracking System (concept): 46-line subsystem that, when a user repeatedly denies actions, shifts the agent
• ent-20260419-284b005a6bd9 Session JSONL Log Format (dataset): Structured per-session log file containing tool-call durations, per-operation to
• ent-20260419-3545bbd2eff9 250,000 Wasted API Calls Bug (issue): Defect documented in an internal code comment describing ~250,000 wasted API cal
• ent-20260419-3c940f253835 compaction/ Directory (file): ~4,000-line directory implementing the 5-layer context compression pipeline.
• ent-20260419-46db6f59a6d8 Module Size Map (concept): Reference distribution of Claude Code module sizes used to navigate the codebase
• ent-20260419-522c1fccbf67 main.tsx Entry Point (file): CLI entry point of Claude Code, ~20,000 lines / 785KB bundle; hosts the 3,167-li
• ent-20260419-53d061ec7c78 Memory as Hint, Not Truth (concept): Architectural stance that the agent's stored beliefs are hints to be verified ag
• ent-20260419-5baa864a2734 Velocity-Pressure Code Pattern (concept): Anti-pattern diagnosis: features added to the safest-feeling existing location r
• ent-20260419-5e6a96d8cc9b Deferred Tool Loading Convergence (concept): Pattern of loading tool definitions in batches as context requires them rather t
• ent-20260419-69c4acf29c08 15 Reverse-Engineered Design Intents (concept): Ji_ai's catalog of 15 encoded design decisions extracted from the leaked TypeScr
• ent-20260419-6bd0f130227c Reactive Compact Circuit Breaker (concept): Safety mechanism in the query-loop state that gives up after 3 attempts to recov
• ent-20260419-73e6e77a5f70 Code Authorship Anomaly (concept): Finding from static code authorship analysis that the leaked Claude Code codebas
• ent-20260419-77cbb6b979b1 3,167-Line God Function Metrics (concept): Quantified signature of the single largest function in Claude Code: 3,167 lines,
• ent-20260419-7e876aebce6f Strict Write Discipline (concept): Invariant that the agent only updates its memory index after a confirmed success
• ent-20260419-8fbb653f4f83 kairos/ Directory (file): ~8,000-line directory containing the KAIROS daemon implementation with its 4-pha
• ent-20260419-935506c2d7ac Conditional-Statement Fingerprint (concept): Stylistically unique pattern of if/else/ternary structures produced by Claude Co
• ent-20260419-95541d9e4103 commands.ts (Slash Command Implementations) (file): File (~25,000 lines) holding 100+ slash-command implementations.
• ent-20260419-9bca6bf795bf Fingerprinting AI Coding Agents (MSR '26 Paper) (document): Peer-reviewed paper (MSR '26, January 2026) that analyzed 33,580 PRs across five
• ent-20260419-a18dd883ee7c /insights Command (concept): Pre-leak slash command that reads all session JSONL logs and generates a frictio
• ent-20260419-a42fd6fb7d05 TechTrenches God-Function Analysis (document): TechTrenches article that named the single 3,167-line / 486-branch-point functio
• ent-20260419-a76eebe71f47 Async Generator Query Architecture (concept): QueryEngine.submitMessage() returns an AsyncGenerator rather than a
• ent-20260419-a8b4b5213c3b Cached vs Non-Cached Billing Multiple (concept): At 200K-context Opus rates, the difference between cached and non-cached billing
• ent-20260419-b64d78e2cd15 Codified Behavior Over Model Self-Discipline (concept): Design principle that behaviors such as 'don't refactor, rename, or reorganize b
• ent-20260419-bcbc0edfd509 Ji_ai 15 Design Intents Analysis (document): Ji_ai dev.to article reverse-engineering 15 design intents from ~80K lines of th
• ... and 14 more

Extracted claims

• clm-20260419-037504d29835 [observation]: Design intents are most effectively discovered via /insights
• clm-20260419-081f996bdfd2 [decision]: query.ts is a 10-step state machine, not recursion
• clm-20260419-0851f529de85 [pattern]: Deferred tool loading is converging across agent frameworks
• clm-20260419-2ef4f0e7dc2b [fact]: 14 vectors break the prompt cache
• clm-20260419-2f4ce3d699bc [anti-pattern]: 250,000 wasted API calls were documented but shipped anyway
• clm-20260419-30688600d859 [fact]: /insights reads JSONL logs to produce workflow friction reports
• clm-20260419-3509efd69456 [decision]: submitMessage returns AsyncGenerator, not Promise
• clm-20260419-41c46dbfedc4 [fact]: Leaked Claude Code occupies its own authorship tier
• clm-20260419-51aa6c05d850 [observation]: bashSecurity.ts numbering suggests incident-driven check accretion
• clm-20260419-5f03621d0ace [anti-pattern]: Velocity-pressure pattern is diagnostic, not unusual
• clm-20260419-60161a02c2c5 [constraint]: Memory index only updates after a confirmed successful write
• clm-20260419-60e215fb40dd [decision]: Cache-break latch is intentionally sticky for the session
• clm-20260419-6190c3aa137d [observation]: query.ts is the recommended practical entry point
• clm-20260419-63f5d68049a3 [metric]: Streaming tool execution cuts wall time by about 40%
• clm-20260419-7818a10df16f [pattern]: Denial tracking is a 46-line graceful-degradation subsystem
• clm-20260419-7840c2e5a971 [metric]: Claude Code module-size distribution concentrates complexity in QueryEngine
• clm-20260419-a93940c70218 [fact]: Session JSONL logs carry structured per-operation metrics
• clm-20260419-aaa7342d9931 [metric]: prompts.ts contains 141 system-prompt assembly modules
• clm-20260419-ae9a9732224b [decision]: Minimal-impact behavior is codified in the system prompt
• clm-20260419-aeb5e3f2f7b8 [metric]: A single 3,167-line function has 486 branch points and 12 nesting levels
• clm-20260419-b9b061b980d2 [decision]: Agent treats its own memory as hint, not truth
• clm-20260419-baa89a05ec07 [metric]: KAIROS daemon occupies ~8,000 lines with 4-phase dream loop
• clm-20260419-baae6fb64527 [fact]: extractMemories.ts makes a fire-and-forget Opus call per turn
• clm-20260419-c0331cdbc911 [metric]: Fingerprinting paper reaches 97.2% F1 on agent identification
• clm-20260419-c2d9b9a0940f [metric]: Cached vs non-cached billing differs by ~5x at 200K Opus
• clm-20260419-d27e828089e9 [hypothesis]: Leaked CC style is consistent with execution-feedback training
• clm-20260419-dbace1b4bdc4 [observation]: Undercover Mode does not defeat the structural fingerprint
• clm-20260419-dc35f9649f6a [pattern]: /insights enables a compounding workflow-improvement loop
• clm-20260419-dd70f0d9885d [constraint]: Reactive compact tries up to 3 times before giving up
• clm-20260419-e3f55c90074d [metric]: commands.ts holds 100+ slash commands in 25,000 lines
• ... and 4 more

Extracted relations

• rel-20260419-00000000a001: Conditional-Statement Fingerprint --[contradicts]--> Undercover Mode
• rel-20260419-00000000a002: query.ts 10-Step Loop --[implements]--> query.ts
• rel-20260419-00000000a003: Nine Continue Points --[contains]--> query.ts 10-Step Loop
• rel-20260419-00000000a004: Reactive Compact Circuit Breaker --[implements]--> query.ts 10-Step Loop
• rel-20260419-00000000a005: Streaming Tool Execution Parallelism --[implements]--> StreamingToolExecutor
• rel-20260419-00000000a006: Async Generator Query Architecture --[implements]--> QueryEngine.ts
• rel-20260419-00000000a007: 3,167-Line God Function Metrics --[related_to]--> tt() Function
• rel-20260419-00000000a008: God Function Responsibilities --[contains]--> tt() Function
• rel-20260419-00000000a009: 250,000 Wasted API Calls Bug --[caused]--> Velocity-Pressure Code Pattern
• rel-20260419-00000000a010: 14 Cache-Breaking Vectors --[contains]--> Prompt Cache Section Registry
• rel-20260419-00000000a011: Prompt Cache Section Registry --[supports]--> Cache Economics
• rel-20260419-00000000a012: Cache Sticky Latch --[depends_on]--> Prompt Cache Section Registry
• rel-20260419-00000000a013: Compounding /insights Loop --[depends_on]--> /insights Command
• rel-20260419-00000000a014: /insights Command --[uses]--> Session JSONL Log Format
• rel-20260419-00000000a015: Module Size Map --[contains]--> QueryEngine.ts
• rel-20260419-00000000a016: Module Size Map --[contains]--> Tool.ts (Built-in Tool Registry)
• rel-20260419-00000000a017: main.tsx Entry Point --[contains]--> tt() Function
• rel-20260419-00000000a018: bashSecurity.ts Numbered Checks --[implements]--> Bash Security Pipeline
• rel-20260419-00000000a019: Denial Tracking System --[depends_on]--> Permission Pipeline
• rel-20260419-00000000a020: Deferred Tool Loading Convergence --[related_to]--> ToolSearch System
• rel-20260419-00000000a021: prompts.ts (141-Module Assembly) --[implements]--> System Prompt Assembly
• rel-20260419-00000000a022: extractMemories.ts --[implements]--> Auto-Memory
• rel-20260419-00000000a023: kairos/ Directory --[implements]--> KAIROS
• rel-20260419-00000000a024: compaction/ Directory --[implements]--> Compaction Pipeline
• rel-20260419-00000000a025: buddy/ Directory --[implements]--> Buddy Sprite
• ... and 9 more