6 Pillars of Claude Code's Harness Architecture (from the Source Leak)

Sebastian Raschka analyzed the leaked Claude Code TypeScript codebase (~512K lines, ~1,900 files) after it was accidentally exposed via an npm source map file. The breakdown reveals 6 architectural pillars that make Claude Code effective — and the key insight is that the harness is the secret sauce, not the model.

The 6 Pillars

1. Live Repo Context — Automatically loads git branch, recent commits, and CLAUDE.md at session start. The model always knows where it is.

2. Aggressive Prompt Caching — Boundary markers separate static from dynamic content. Static parts are cached globally across calls, cutting costs and latency.

3. Dedicated Tools — Custom Grep, Glob, and LSP tools (not bash wrappers). The LSP integration gives call hierarchy and cross-file references — much richer than grep alone.

4. Context Bloat Management — File-read deduplication, large outputs written to disk with only a preview in context, plus auto-truncation and compaction when the window fills up.

5. Structured Session Memory — Markdown files per conversation tracking: task spec, touched files, workflow state, errors, learnings, and a worklog.

6. Forks & Subagents — Parallel work via forked agents that reuse the parent’s prompt cache and are aware of mutable state.

The Core Insight

Better harness + mediocre model > bad harness + best model.

This confirms a pattern we keep seeing: drop Claude’s harness onto other models (DeepSeek, MiniMax, Kimi) with some optimization, and you get surprisingly strong coding performance. The model matters less than the scaffolding around it.

Pillar #4 (context bloat management) is particularly interesting — it’s essentially what systems like OpenClaw’s LCM plugin solve: keeping the context window useful without drowning it in stale content.