Ecosystem Update — 2026-04-05
TL;DR
- Auto-mode (background safety classifier) eliminates permission prompts — platform feature, just enable
- Preflight MCP catches vague/underspecified prompts before they waste cycles — 24-tool server, one install
- SkillDeck gives a local dashboard to manage all Claude Code config in one place — visual skills/hooks/agents/settings management
Quick Wins
| Item | Source | Type | Impact | Effort | Action |
|---|---|---|---|---|---|
| auto-mode | claude-code-best-practice | hook | 3 | 1 | Enable background safety classifier in settings — replaces manual permission prompts with auto-classification |
| preflight | awesome-claude-code-toolkit | mcp | 3 | 1 | Install 24-tool MCP server that scores prompt clarity, estimates cost, catches vague requests before execution |
| skilldeck | skilldeck | mcp | 2 | 1 | Install local dashboard to manage skills, agents, teams, commands, hooks, settings, and CLAUDE.md visually |
| once-true-hooks | claude-code-new-features | hook | 2 | 1 | Add once: true to one-shot hooks (session-start, preflight) to prevent duplicate execution |
Build Queue
- harness (skill) — revfactory/harness — Meta-skill that designs domain-specific agent teams, defines specialized agents, and generates the skills they use. Could auto-scaffold new project agent configurations.
- codetape (hook) — 888wing/codetape — Flight recorder for AI coding sessions — auto-records semantic traces, syncs README/CHANGELOG. Useful for audit trail.
- bouncer (hook) — buildingopen/bouncer — Independent quality gate using Gemini to audit Claude Code output. Cross-model verification adds a second opinion.
- channels (hook) — claude-code-best-practice — Push events from Telegram, Discord, or webhooks into running sessions. Enables async notification beyond desktop.
- ccmanager (mcp) — kbwo/ccmanager — Multi-agent session manager supporting Claude Code, Gemini, Codex, Cursor, Copilot. Unified orchestration across models.
- claude-devtools (mcp) — matt1398/claude-devtools — Desktop observability app: turn-based context data, compaction visualization, subagent execution trees. Good for debugging complex agent runs.
- ultraplan (agent-pattern) — claude-code-best-practice — Draft plans in cloud with browser review, inline comments, flexible execution. Planning-gate alternative with visual review.
Research
- Agentic Verification of Software Systems — AutoRocq: first LLM agent for program verification via iterative refinement with theorem prover. Applicable to generate-and-validate loops in multi-agent coding.
- AgentSpec: Customizable Runtime Enforcement — Framework for safe LLM agent operation through customizable runtime enforcement policies. Directly relevant to governance/routing.
- DeepCode: Open Agentic Coding — Autonomous document-to-codebase synthesis with structured indexing and closed-loop error correction. Parallels to our build skill's spec→code pipeline.
- PRDJudge: Automated Code Validation — Lightweight evaluation agent that auto-validates code agent output using executable tests. Could enhance our evaluate skill.
- From LLM Reasoning to Autonomous AI Agents — Comprehensive review of autonomous agent architectures — planning, tool use, memory, multi-agent collaboration. Good survey of the landscape.
Already Have
context-fork-pattern, isolation-worktree (worker/reviewer/planner/python-reviewer/typescript-reviewer), session-teleportation, batch-command, instinct-scoring-agentshield, dynamic-shell-injection, compound-engineering-plugin, tdd-guard-hook, cc-devops-skills, type-prompt-hook, worktree-hooks, skill-scoped-hooks, postcompact-hook, precompact-hook, permissionrequest-hook, subagentstart-hook, explorer-read-only, parry-prompt-injection-scanner, claude-rules-doctor, container-use, sandbox-isolation, cc-tools, fullstack-dev-skills, claude-mem (similar to claude-recap/supermemory)
Rejected
- claude-recap — We already have claude-mem with semantic search + type tagging; claude-recap is shell-hook markdown only. Covered by existing primitive.
- fractal (recursive project management) — Predicate decomposition is interesting but our planning-gate + build skill already handle task breakdown. Adding a new decomposition engine = overengineering.
- vibe-kanban — Kanban board + worktree orchestration. Our govern skill already routes and tracks work packets. Redundant orchestration layer.
- oh-my-claudecode — Teams orchestration with 19 agents. We already have chad-agent + governed routing. Adding another 19-agent framework = complexity without proof of gap.
- agent-teams-exercises — Educational repo, not a tool. No action needed.
- reporecall — AST-indexed codebase memory. We have Grep/Glob + claude-mem. No proof existing primitives can't cover this.
Sources checked: awesome-claude-code, claude-code-best-practice, awesome-claude-code-toolkit, claude-code-new-features-early-2026, arxiv LLM agent search, GitHub search 2026 Tier 2 fetched: yes Tier 3 fetched: yes Run at: 2026-04-05T21:55:00Z Mode: dry-run (no changes applied)