Reading order

This is a workshop, not a textbook. Pages were produced by independent research streams that sometimes overlap and sometimes disagree. The structure below tells you what to read in what order depending on what you came for.

If you came to understand the fleet’s current state

1. Variant audit — what each of the 18 named variants is supposed to do, where their persona files live, and which ones have gone silent.
2. Recent activity — what the fleet was actually doing in the 30 days before the pause. Heatmap, top cwds, where activity collapsed.
3. Broken canon — every stale reference and self-contradiction in the existing canon docs. Severity-ranked.
4. Fleet state — verified — recon claims cross-checked against the actual filesystem. 12 critical heads-ups the recon missed live here.

If you came to understand the rebuild plan

1. Concepts (kept) — rules, philosophies, and patterns extracted from the existing canon that survive cleanup.
2. Facts (current) — current verified state of the fleet (machines, services, integrations, active variants).
3. Synthesis design doc — pending. Will hold the integrated rebuild plan when written.

If you came to learn how production multi-agent systems do this

1. Persona patterns — how production agent SDKs separate shared ethos from per-agent identity.
2. Multi-agent frameworks — LangGraph, CrewAI, AutoGen, Mastra, Letta — recommended project layouts.
3. Knowledge management — PARA vs Johnny.Decimal vs Zettelkasten vs LYT for hybrid business+agent vaults.
4. Memory frameworks — Letta, Mem0, Graphiti, Cognee architecture comparison.

If you came to know Claude Code’s actual internals

1. Claude internals — verified — the official-doc-confirmed facts. Trust this list. Retractions of ccleaks-only claims are flagged here.
2. ccleaks internals — the raw reverse-engineered report. Many findings hold; some are walked back in the verifier.
3. Cross-reference audit — 10 specific contradictions across reports, with winners chosen.

If you came to see what already exists in the ecosystem

1. GitHub deep-dive — 8 repos (claude-mem, claude-fleet, agents, etc.) with what to steal and what to skip.
2. Blog deep-fetch — 14 practitioner articles ranked by signal.
3. GitHub claims — verified — star counts, attributions, feature claims cross-checked.
4. Community signal — Reddit/HN/blogs synthesis of what real users are saying.

Three things worth knowing before you read anything else

1. The fleet of “18 variants” is closer to a fleet of 4-6 truly active variants. Most of the rest is paper. Don’t read variant docs assuming they describe live processes.
2. The CLAUDE.md @~/.claude/docs/*.md import pattern is canonically correct — research agents independently confirmed it. The drift problem isn’t in the architecture; it’s in the content of those docs.
3. MEMORY.md has a hard 200-line / 25KB cap. Anything past that is silently invisible to every session. The current MEMORY.md is over the line. Fix tonight.

How to use this workspace going forward

• Treat the synthesis page as the conclusion. Everything else is supporting evidence.
• When two pages disagree, the verifier reports win. They cross-checked against the official Anthropic docs.
• The canon pages are the working source-of-truth — concepts and facts that should make it into the rebuilt fleet doctrine.
• The research and SDK patterns pages are the raw thinking. Useful for context, but don’t take any single one as authoritative.