Apr 2026 research refresh: 4 new parts + Harness Thesis + 5 Anthropic patterns + memory-on-SLM + worktrees + SDD

[OnlyTerp]

Summary

Biggest content pass since the guide started. The week of April 10–17, 2026 produced a wave of convergent signal — nine independent writers on one thesis, an official Anthropic taxonomy, a new research cluster on cheap-model memory, and a handful of new harness patterns. This PR folds all of it into the guide as one coherent release so newcomers see the 2026 state-of-the-art and returning readers see what changed.

Guide goes from 28 → 32 parts.

Type of change

• New part (four of them — 29, 30, 31, 32)
• Correction to an existing part (Parts 2, 5, 15, 22, 24, 25, 26, 28)
• Version bump (still 2026.4.15 stable)
• Benchmark / data addition
• Tooling / CI / meta

What landed

New parts

• Part 29 — The Hook Catalog — 8 copy-paste hooks (shell-guard, secret-redact, cost-tripwire, skill-install-deny, dreaming-gatekeeper, cache-TTL-enforcer, audit-log, tool-name-collision-alarm), exit-code semantics, StopFailure + Amit Kothari's debugging gotcha.
• Part 30 — The Ralph Loop In OpenClaw — 30-line bash wrapper, PRD.json template, budget/exit-condition wiring. Referenced against frankbria/ralph-claude-code (8.7K stars) and the DEV Apr 13 OpenClaw-specific port.
• Part 31 — The LLM Wiki Pattern In OpenClaw — Karpathy's Apr 10, 2026 three-tier pattern mapped 1:1 onto OpenClaw's SOUL / AGENTS / MEMORY / DREAMS / skills.
• Part 32 — Self-Evolving Skills With SkillClaw — AMAP-ML/SkillClaw (created Apr 10, 2026; 687 stars week-one; lists OpenClaw as first-class), scoring + mutation + retirement mechanics, Mem²Evolve as the co-evolution roadmap.

Refactors / expansions

• README — new Harness Thesis hero + mermaid diagram (95%/5% split, 9-source convergence), File Hierarchy At A Glance diagram + table, Navigate-By-Goal expanded with "Run autonomous / overnight work" and "Enforce safety the agent can't ignore" rows, themed TOC updated for Parts 29–32.
• Part 2 — renamed Context Bloat → Context Engineering — The Discipline (Karpathy framing, now picked up by Gartner). New appendix: The 5-Minute Prompt Cache TTL Trap (March 2026) — silent Anthropic default change that is tripling bills across the ecosystem; explicit ttl fix.
• Part 5 — lede rewritten around "Sub-agents are context garbage collection, not a speed hack" with a 3-trigger decision table. New section on Anthropic's 5 multi-agent coordination patterns (claude.com Apr 10, 2026) + decision flowchart.
• Part 15 — new Parallel OpenClaw With Git Worktrees section (three independent April writeups converged on "one worktree per agent, one OpenClaw process per worktree"), 20-line spawner, gotchas.
• Part 22 — new Memory You Can Afford block: vbfs (92.1% Recall@5 with zero LLM calls, Apr 15), LightMem (83 ms retrieval, +2.5 F1 on LoCoMo, Apr 12), Mem²Evolve (+18.53%, Apr 14), AMFS (Apache 2.0 MCP, Apr 13). Reference dreaming.models config that runs Light/REM on local Qwen 3:4b and Deep on Cerebras qwen-3-235b.
• Part 24 — new callout OpenClaw vs. Cloud-Sandbox Delegation (Twill / Amika): 8-column comparison table, when each wins.
• Part 25 — new The Harness Thesis opening section pulling the 9 convergent sources onto the architecture primer.
• Part 26 — new section Pair OpenClaw With A Machine-Readable Spec (Spec-Driven Development) with the AWS Kiro 18-months → 76-days case study and a minimum-viable PRD.json.
• Part 28 Glossary — 15 new entries: Anthropic's 5 patterns, Cloud sandbox delegation, Context Engineering, Exit codes (hooks), Git worktrees, Harness Thesis, LightMem, LLM Wiki pattern, Mem²Evolve, PRD.json, Progressive disclosure, Prompt cache TTL, Ralph loop (rewritten), SDD, SkillClaw, StopFailure, Sub-agents-as-context-GC. Existing Hooks entry expanded.
• AWESOME.md — added SkillClaw, vbfs/agent-memory-store, LightMem, Mem²Evolve, AMFS, Anthropic's 5-pattern blog post, frankbria/ralph-claude-code, DEV Ralph-OpenClaw post, sub-agents-as-GC reframe (heyuan110, Builder.io), Git worktrees for parallel agents (Upsun), Karpathy's LLM Wiki talk, Aaron Fulkerson's Exo writeup, AaronRoeF/claude-code-patterns, The Harness Is Everything (Medium + Trensee + Atlan), Ido Green, SDD Time piece, Progressive Disclosure (MindStudio), Amit Ray file-hierarchy cheat sheet, Claude Code hooks writeups, Amit Kothari hooks debugging, DEV cache-TTL writeup, Twill, Amika, Andon Labs.

Sources (all published Apr 10–17, 2026)

• Anthropic — Multi-Agent Coordination Patterns (Apr 10)
• Karpathy — LLM Wiki talk (Apr 10, 4.3K views)
• AMAP-ML/SkillClaw (created Apr 10)
• Claude Lab — Hooks automation master guide (Apr 10)
• DEV — Claude Code Worktrees (Apr 10)
• HN — Twill YC S25 launch (Apr 11)
• LightMem — arXiv 2604.07798 (Apr 12)
• Atlan — Agent Harness 2026 (Apr 12)
• Trensee — The Harness Is Everything (Apr 12)
• Aaron Fulkerson — Karpathy's LLM Wiki pattern in production (Apr 12)
• DEV — How I Run 20 Claude Code Agents In Parallel (Apr 12)
• Medium — Stop Stuffing Your Custom Instructions (Apr 12)
• heyuan110 — Harness + sub-agent architecture (Apr 13)
• Medium — The Harness Is Everything (Apr 13)
• DEV — Ralph Loops + OpenClaw (Apr 13)
• AMFS — the amnesia fix (Apr 13)
• Amika cloud-sandbox walkthrough (Apr 13)
• Upsun — Git worktrees for parallel AI agents (Apr 14)
• Mem²Evolve — arXiv 2604.10923 (Apr 14)
• Time — Spec-Driven Development (Apr 14)
• Amit Ray — File hierarchy cheat sheet (Apr 14)
• DEV — Claude API Prompt Caching 80% off (Apr 14)
• vbfs/agent-memory-store (Apr 15)
• MindStudio — Progressive Disclosure for AI Agents (Apr 15)
• Builder.io — Claude Code sub-agents (Apr 16)

Review checklist

• Added or updated cross-links from the README TOC and related parts (Navigate-By-Goal, themed TOC, decision-tree callouts on all new parts)
• Sources / release notes linked above, not inline as footnotes
• Ran markdownlint-cli2 (0 errors) and lychee --offline (0 errors, 248 valid, 164 excluded by allowlist) locally before push
• No speculation presented as fact — the few uncertain items are flagged below

Notes / flagged items

1. YouTube view counts, repo stars, and publish dates quoted from the research pass (Apr 17, 2026). A couple were not re-verified at PR time — if any number feels off at merge, ping me and I'll re-stamp.
2. Cerebras qwen-3-235b-a22b-instruct-2507 as the Deep-phase model pick in Part 22 is an opinionated default; swap in whatever fast-cheap-capable model your stack allows. The structure (local small for Light/REM, fast-cheap capable for Deep) is the part worth keeping.
3. Ralph Loop bash script in Part 30 is intentionally minimal — 30 lines, POSIX, no dependencies beyond OpenClaw CLI. If your org needs the full budgeting/observability/retry-backoff version, use it as a starting skeleton.
4. SkillClaw in Part 32 is a week-old project. Its API will churn; the concepts (population-level scoring, mutation, retirement) are what the part teaches. The config block should be read as "the shape of integration," not a pinned contract.
5. Anthropic's 5 patterns in Part 5 were named this week. If Anthropic renames any of them before the next refresh, re-stamp — but the shapes themselves (generator-verifier vs orchestrator-subagent vs teams vs hierarchical vs network) are stable.
6. Harness Thesis percentages (95/5) are rhetorical shorthand from the 9-source convergence, not a measured number. The guide treats them as the framing, not a benchmark.

Link to Devin session: https://app.devin.ai/sessions/df6f8c16f82e448b915735660ed94fb7
Requested by: @OnlyTerp

---

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment and CI monitoring