Ecosystem Update — 2026-04-21

Highlights

• Current setup is saturated against the "easy wins" surface — 167 previously-seen items and a mature 10-agent / 39-skill / 10-hook-event config mean most new signal is already covered or rejected under the anti-overengineering gate
• Zero true Quick Wins today. One candidate ("agent" default-agent field in settings.json) is plausible but unverified against official docs — parked in Build Queue until confirmed rather than auto-applied
• Highest-signal NEW items are two new hook *types* (agent and http) from the 2026 early-features cheatsheet and a meta-review arxiv paper (2604.18071) surveying architectural design decisions across 70 agent harnesses

New Tools, Skills & Patterns

• Skill-scoped safety blocks — /careful and /freeze skill-pattern
  shanraisshan
  On-demand hook blocks embedded inside a skill definition that block destructive commands (/careful) or edits outside a target directory (/freeze). Useful for high-risk skills like drive, orchestrate-local, and audit
• idle-timing hook hook
  rohitg00 toolkit
  Injects hidden timing context so Claude knows how long a conversation has been paused
• rulecheck composite (skill + agent + hook) composite
  coleam00 cheatsheet
  Autonomous rule-violation scanner that edits code, validates the fix, and notifies Slack. Partial overlap with insecure-defaults/sharp-edges plugins but the *autonomous remediation* leg is new
• New hook types: agent and http hook
  coleam00 cheatsheet
  Current settings.json only uses type: command (and type: prompt is in the seen list from prior runs). The agent type fires a subagent on a lifecycle event; http posts to a URL. Verify these are real in current Claude Code spec before wiring
• Observable Loop Prevention Guards (5-layer) pattern
  affaan-m/everything-claude-code
  Re-entrancy detection with throttling, tail sampling, and lazy-start logic to prevent runaway autonomous loops
• Strategic Compaction Hooks (proactive trigger) hook
  affaan-m
  PreCompact, but it only fires at the context-window threshold. Proactive trigger = suggest /compact at logical task breakpoints (slice acceptance, track closure). Lower-priority since it overlaps with existing PreCompact + memory gates
• deep-dive skill skill
  rohitg00
  DAG-based research breaking questions into parallel dependency graphs with sourced reports. Overlap with rlm-scan + explorer subagent pattern but adds explicit dependency DAG management
• memory: project agent frontmatter agent-pattern
  coleam00 cheatsheet
  Persistent project context across runs and worktrees via one frontmatter field. Would be a Quick Win if verified
• "agent" default-agent setting in settings.json config
  Boris's tips
• VibeGuard-style rule pack (top 10, not 88) plugin
  rohitg00
  88 hallucination-prevention rules + 13 hooks is too heavy to adopt wholesale; cherry-pick the 10 most aligned with Python/TypeScript workflows
• additionalDirectories in settings.json config
  Boris's tips
  Relevant only if multi-repo work becomes a pattern. Currently single-repo per session, so deferred

Research Worth Reading

• arXiv:2604.18071 — Architectural Design Decisions in AI Agent Harnesses
  arXiv
  Empirical study of 70 agent projects identifying five recurring design dimensions
• arXiv:2604.17699 — SelfHeal: Multi-Agent ReAct for Bug Repair
  arXiv
  Fix agent + critic agent with internal rules and web search. Partial match for implementation-checker → reviewer → validator chain; compare against their ablation to see which link we're missing
• arXiv:2604.18349 — HiGMem: Hierarchical LLM-Guided Memory
  arXiv
  Two-level event-turn memory allowing inspection of summaries before deep retrieval
• arXiv:2604.18240 — AJ-Bench: Agent-as-a-Judge for Environment-Aware Evaluation
  arXiv
  Judges that actively interact with environments to acquire verifiable evidence outperform passive LLM judging. Direct applicability: acceptance_check shell gate is a weak form of this; the paper likely has sharper patterns for reviewer/validator agents

Considered, Not Adopting

Items reviewed and explicitly declined this cycle, with the reason. Curation discipline matters more than coverage.

• AuraKit (46-mode fullstack skill, 23 sub-agents, 55% token-savings claim) — fails anti-overengineering gate; no one-sentence proof that existing 39 skills + 10 agents cannot satisfy the same use cases
• moyu anti-overengineering skill
• Axiom contract-enforced pipeline
• Bouncer Gemini quality-gate plugin — functional duplicate of second-opinion@trailofbits which is already enabled
• /code-review cloud-based PR analysis — already covered by codex-branch, codex-security, review skill, and second-opinion plugin
• Technology Scout agent — "build-vs-buy evaluator" doesn't match a recurring workflow for CW work
• AutoResearch agent — ML experiment automation via tree search; wrong domain
• Global AGENTS.md cross-harness parity file — would only matter if Codex and Claude Code session pairs became a regular pattern; currently each owns its own home per policy. No pain point justifies the duplication
• Multi-Language Rules Architecture — reorganizing skills into per-language subdirectories is a structural change without a concrete recurring problem; current flat structure works
• Spinner verbs, Output Styles tuning — cosmetic; no behavior impact
• Dispatch (mobile → desktop session handoff) — platform-level feature, no local config needed
• naming skill — not a recurring workflow
• https://github.com/hesreallyhim/awesome-claude-code
• https://howborisusesclaudecode.com/
• https://github.com/shanraisshan/claude-code-best-practice
• https://arxiv.org/search/?searchtype=all&query=LLM+agent+coding&order=-announced_date_first
• https://github.com/rohitg00/awesome-claude-code-toolkit
• https://code.claude.com/docs/
• https://github.com/coleam00/claude-code-new-features-early-2026/blob/main/CHEATSHEET.md
• https://github.com/affaan-m/everything-claude-code

Sources Reviewed

• github.com
• howborisusesclaudecode.com
• github.com
• arxiv.org
• github.com
• code.claude.com
• github.com
• github.com