Claude Code in Pseudocode

LMPL specifications that reconstruct Claude Code's internal architecture from the public analyses of the March 2026 source map leak (cli.js.map shipped in @anthropic-ai/claude-code@2.1.88).

Each spec uses the @profile("agentic") / @intent("specify") LMPL profile and follows the same two-layer structure: an idealized skeleton, then a refined expansion grounded in source-level findings. No source code is reproduced — the pseudocode is an independent expression of the documented control flow.

Index

#	Spec	Captures	Lines
01	Core agentic loop	queryLoop(), State, Continue Site pattern, transition reasons, recovery branches	467
02	Tool catalog	Tool interface, registry, 6-stage lifecycle, concurrency partitioning, 5 canonical tools	518
03	Sub-agents & coordinator	spawn_subagent, 3 isolation levels, cache-alignment, coordinator as prompt contract	402
04	Guardrails & permissions	can_use_tool pipeline, 6 permission modes, yoloClassifier, bashSecurity 5-layer	489
05	MCP client	Transports, lifecycle, tools/resources/prompts, namespacing, trust model	612
06	Skills & slash-commands	Frontmatter contract, discovery, progressive disclosure, Skill meta-tool	472
07	Hooks	9 events across 3 families, exit-code conventions, effect model, Guardrail integration	492
08	Plugin ecosystem	Manifest, components, marketplaces, activation lifecycle, version resolution, trust	654
09	Memory tiers	CLAUDE.md hierarchy, auto-memory (memdir/), JSONL sessions, filter pass	558
10	Context compaction	4-tier pipeline, circuit breaker, PreCompact hook integration, recovery	513
11	Prompt cache	SYSTEM_PROMPT_DYNAMIC_BOUNDARY, 14 break vectors, sticky latches, fork-based sharing	441

Dependency Graph

Read in any order — each spec stands alone — but specs with fewer dependencies are easier entry points.

              01 Core loop
              /  |  |  \
            02  03  09  (02 Tool catalog, 03 Sub-agents, 09 Memory)
           /   / \  |
          04  /   \ |
          ↑  10   11 (10 Compaction uses 03; 11 Prompt cache uses 03+09)
          |   ↑
          07  (07 Hooks extends 04 and integrates with 02/10)
          ↑
         05 (05 MCP consumes 02, 04)
          ↑
         06 (06 Skills consumes 02, 05)
          ↑
         08 (08 Plugins packages 06/07/03/05)

Suggested reading orders:

• Quick tour — 01, 02, 04, 08 (loop → tools → guardrails → how it all gets packaged)
• Runtime-first — 01, 02, 03, 04, 10, 11 (loop + everything that runs per-turn)
• Extension-first — 01, 06, 07, 08, 05 (loop + every way to extend it)

Shared Conventions

Every spec carries:

• Scope & non-goals up front.
• Types section defining the vocabulary the spec uses.
• A refined expansion layer with source-grounded fidelity (behavioral scars, not just abstract intent).
• LMPL gaps — a running list of language-level features the specs need but don't have. Roughly 60 proposed extensions across the set; useful as an artifact for evolving the LMPL profile itself.
• Cross-spec references table and a References section citing the public analyses used as source grounding.

Source Grounding

These specs reconstruct behavior documented in the public analyses of the 2026-03-31 leak. Representative sources (each spec cites its own):

• alejandrobalderas, claude-code-from-source — https://github.com/alejandrobalderas/claude-code-from-source
• 777genius, claude-code-working — https://github.com/777genius/claude-code-working
• cablate, claude-code-research — https://github.com/cablate/claude-code-research
• l3tchupkt, claude-code — https://github.com/l3tchupkt/claude-code
• bits-bytes-nn, "Claude Code Architecture Analysis" — https://bits-bytes-nn.github.io/insights/agentic-ai/2026/03/31/claude-code-architecture-analysis.html
• Blake Crosley, "What the Claude Code Source Leak Reveals" — https://blakecrosley.com/blog/claude-code-source-leak
• Siddhant Khare, "The plumbing behind Claude Code" — https://siddhantkhare.com/writing/the-plumbing-behind-claude-code
• Haseeb Qureshi, "Inside the Claude Code source" — https://gist.github.com/Haseeb-Qureshi/d0dc36844c19d26303ce09b42e7188c1
• Varonis Threat Labs, "A Look Inside Claude's Leaked AI Coding Agent" — https://www.varonis.com/blog/claude-code-leak
• Redreamality, "Claude Code Leak: A Deep Dive" — https://redreamality.com/blog/claude-code-source-leak-architecture-analysis/

No source code is reproduced in this repository. All pseudocode is independent LMPL expression of documented control flow.

LMPL Profile

The pseudocode uses the agentic profile defined in hiivmind-lmpl (spec/profiles/agentic.md). Key constructs used throughout:

• agent_loop { observe / reason / act / update / until } — the canonical observe-reason-act skeleton
• @max_iterations, @requires_approval, @boundary — agent-level annotations
• require / ensure / invariant — pre/post/invariant contracts
• concurrently { ... } join — parallel fan-out (used for tool partitioning and swarm sub-agents)
• attempt / on failure — tool failure handling

Specs deliberately note where LMPL doesn't quite fit, rather than papering over the gap — these are the "LMPL gaps" sections.