Ecosystem Update — 2026-04-19

Highlights

• Very quiet cycle: all 4 "latest additions" from awesome-claude-code and all major Boris/shanraisshan items already in seen_items (148 prior items)
• Net-new discoverables are mostly built-in CLI flags and commands (--bare, --add-dir, --agent, /sandbox, /voice) — knowledge-level, not config-level. No Quick Wins today
• Tier 2 (arxiv) surfaced 8 new April-2026 papers — notable: AgentForge (execution-grounded multi-agent with mandatory Docker verification), ORBIT (dependency-aware translation graph), and Three Roles One Model (role scaffolding at inference time)

New Tools, Skills & Patterns

• Prompt-hook guardrail hook
  github.com/coleam00/claude-code-new-features-early-2026
  A PreToolUse hook that calls a cheap LLM to validate a dangerous tool call before it executes. Cost: pennies per check. Would complement pre_tool_guard.py (which is a deterministic rule checker) with a semantic layer. Impact 3 / Effort 3 / Priority 1.0 / Align Y.
• /careful and /freeze skill blocks skill-pattern
  github.com/shanraisshan/claude-code-best-practice
  On-demand permission gates embeddable in SKILL.md to restrict destructive commands and scope edits to named directories during that skill's execution. Useful for the refactor, audit, and security-audit skills which touch many files. Impact 2 / Effort 2 / Priority 1.0 / Align Y.
• /code-review multi-agent PR analysis command skill
  github.com/rohitg00/awesome-claude-code-toolkit
  Multi-agent PR analyzer (bugs + security + regressions via parallel agents). Partially redundant with existing reviewer + security-review + codex-branch trio, but packaged as a single /code-review <PR> entry point. Impact 2 / Effort 2 / Priority 1.0 / Align Y — marginal, consider only if current reviewer workflow feels under-integrated

Research Worth Reading

• AgentForge: Execution-Grounded Multi-Agent LLM Framework
  arXiv
  Planner/Coder/Tester/Debugger coordination with mandatory Docker sandbox verification between turns. Relevant: current validator agent is stdlib-only; a container-verified stage could strengthen R3/R4 acceptance gates
• ORBIT: Guided Agentic Orchestration for C-to-Rust Transpilation
  arXiv
  Dependency-aware translation graph with 91.7% success on >1000 LoC. Relevant: general pattern for dependency-ordered slice execution in auto-runtime
• Three Roles, One Model: Role Orchestration at Inference Time
  arXiv
  Role scaffolding at inference closes the small/large-model gap without training
• λ_A: Typed Lambda Calculus for LLM Agent Composition
  arXiv
  Formalizes type safety and termination for agent composition, identifies structural incompleteness in real configs. Relevant: possible framework for auditing route_manifest.json composition rules
• FM-Agent: Hoare-Style Reasoning at 143K LoC
  arXiv
  Top-down compositional function-spec derivation from caller expectations. Relevant to large-codebase R4 work
• COEVO: Joint Functional Correctness + PPA Optimization
  arXiv
  Single evolutionary loop with fine-grained testbench scoring. Relevant to verifier convergence design
• WebXSkill: Executable Skill Learning for Web Agents
  arXiv
  Parameterized programs paired with NL guidance, dual-mode (direct exec + agent-adapted). Relevant to the Skill tool design philosophy
• HWE-Bench: LLM Hardware Bug Repair
  arXiv
  Benchmark showing fault localization and cross-artifact coordination as dominant failure modes — applies to any multi-file agent task, not just hardware

Considered, Not Adopting

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

• claude --name / auto-naming — session naming; already covered by rename-session in seen_items and not a config change
• /rewind / /compact vs /clear — built-in context controls, no install needed
• BigQuery integration / Slack MCP — workflow-specific, not general-purpose; alignment fail (no concrete recurring need in )
• Chrome Extension for frontend verification — browser-tier feature, parallel to existing Playwright MCP (already installed)
• Comprehensive toolkit drops — (rohitg00, travisjneuman, vinicius91carvalho) — bulk catalogs; extract individual items on demand rather than installing wholesale. Alignment fail at the catalog level
• PostCompact hook as Quick Win — in seen_items but still not wired; hard-limit forbids registering a hook whose script does not yet exist. Move to Build Queue next cycle if the script is authored first