Ecosystem Update — 2026-05-01 (dry-run)

TL;DR

• Two genuinely new hook events surfaced since the last run — ConfigChange and InstructionsLoaded — that target a real gap in the current setup: drift detection when settings.json, CLAUDE.md, or agents/ change mid-session.
• Hook types agent and http (parked in last run's Build Queue) are now confirmed in multiple sources; agent-type hooks are the most useful new primitive — they let a verifier subagent read files before allowing a tool, which is sharper than the current shell-script command hooks.
• Surface remains saturated. chad-twin already carries memory: project and maxTurns, the worker/reviewer pair already use isolation: worktree, and 13 of the spec's 21 hook events are wired. New signal lands mostly in Build Queue (verification or scripting required) rather than as zero-effort frontmatter wins.

Quick Wins

Rationale: one frontmatter line, mirrors a pattern already validated on chad-twin, gives chad-agent persistent project-scoped memory across runs and worktrees. Skipped from auto-apply per the --dry-run flag.

Everything else that scored ≥2.0 priority either requires a new script (disqualifies it from the Quick Win lane), needs verification against official Anthropic docs that aren't yet detailed enough, or duplicates an existing primitive.

Build Queue

• ConfigChange hook (hook) — Fires when settings.json / project-level settings change mid-session. Genuine gap: nothing in the current config notices when an external editor or a sibling agent rewrites ~/.claude/settings.json during a long autonomous run. Build: a ~/.claude/bin/configchange_audit.py that diffs against a stored hash and writes to state/config-drift.jsonl.
• InstructionsLoaded hook (hook) — Fires when CLAUDE.md, rules, or agent definitions reload. Same gap as above for the constitutional surface. Build: tiny shell hook that journals the reload event with file SHA so post-run forensics can see whether policy drifted during execution.
• agent-type hook (hook-type) — Spawns a subagent for verification instead of running a shell script. Highest-leverage new primitive: replaces the current "shell script greps for dangerous patterns" approach with a real LLM check on PreToolUse for things like ambiguous rm -rf, opaque git push --force, or arbitrary curl-pipe commands. Build: rewrite the highest-risk Bash matcher in settings.json from type: command to type: agent with a 1-line verifier prompt.
• http-type hook (hook-type) — POSTs the hook payload to a URL endpoint. Useful for the notify_done.sh path — instead of shelling out, post to a local notification webhook directly from the hook. Lower priority than agent-type.
• /fewer-permission-prompts skill (skill) — Boris's permission-pruner. Scans recent session history, identifies safe commands that keep prompting, suggests allowlist entries. Direct fit for our permissions block in settings.json which has accumulated noise. Build: install or replicate the skill, run it weekly.
• Routines (Anthropic-managed cron) (feature) — Schedules runs on Anthropic infra rather than local laptop. Replaces /loop 24h /ecosystem-update for the daily ecosystem run so it survives laptop reboots. Build: migrate the daily ecosystem-update from local cron/loop to a Routine.
• idle-timing hook by clankercode (hook) — Carried forward from 2026-04-21. Injects elapsed-since-last-reply into context. Still relevant for /drive and autonomous loops where the model has no signal that a 30-minute gap occurred between turns.
• Loop-operator agent + 5-layer re-entrancy guard (pattern) — Carried forward. Tail sampling, throttling, and lazy-start to prevent runaway autonomous loops. Direct applicability to /loop, /drive, /orchestrate-local, /evolve. Build: Python helper in ~/.claude/bin/ that inspects recent event log for re-entry signatures and aborts.
• ECC_HOOK_PROFILE / ECC_DISABLED_HOOKS env vars (config) — Runtime hook strictness tuning without editing settings.json. Useful for /drive where you want stricter gates than during normal interactive work. Build: read the env-var pattern from affaan-m and add an analogous switch to our hook scripts.
• exa-search skill (skill) — Neural web/code/company search via Exa MCP. Cleaner than ad-hoc WebSearch for evidence-gathering during research-first development. Verify whether Exa is something we want to depend on first (vendor lock-in question).
• leapfrog-mcp (15 parallel Chromium) (mcp) — Multi-session browser MCP. Only relevant if browser-automation work becomes a recurring pattern; deferred otherwise.
• PUIUX Pilot auto-config (tool) — Auto-configures hooks/MCPs across 95+ project types with dry-run/atomic-write/rollback. Possibly useful as a reference for safer Quick Win application logic in this very skill, not as a runtime dependency.

Research

• arXiv:2604.27283 — Risk-Sensitive Contextual Bandits for Abstention-Aware Memory Retrieval — Coding agents that learn when not to retrieve memory, preventing false-positive contamination. Direct check against our omni-mem retrieval pattern in auto_runtime.py — we always pull memory; abstention might cut noise.
• arXiv:2604.25737 — SAFEdit: Multi-Agent Decomposition for Code Editing — Planner/Editor/Verifier roles with a "Failure Abstraction Layer" for structured diagnostic feedback. Comparable to our planner → worker → reviewer chain; the Failure Abstraction Layer is the part we don't have explicitly.
• arXiv:2604.24550 — Mono2Sls: Multi-Stage Pipeline with Explicit Intermediate Artifacts — Four sequential agents that communicate through explicit intermediate artifacts with static-analysis grounding; 100% deployment success on benchmark. Validates our existing artifact-passing pattern in auto_runtime.py event log; worth reading for the static-analysis grounding step.
• arXiv:2604.25849 — ADEMA: Knowledge-State Orchestration with Epistemic Bookkeeping — Checkpoint-resumable persistence and segment-level memory condensation across long-horizon work. Closest analogue to our omni-mem + Stop-hook persistence model; the explicit epistemic bookkeeping (what the agent knows vs. doesn't) is novel.
• arXiv:2604.26963 — MARS: Co-Scheduling for Heterogeneous Agentic Systems — GPU-CPU resource coordination across multi-turn LLM-tool loops. Relevant if/when local-model orchestration via Codex or LM Studio becomes a recurring pattern; not urgent today.
• arXiv:2604.26805 — Bian Que: Flexible Skill Arrangement — Skills auto-generate from LLM or refine via natural language; unified self-evolving mechanism driven by correction signals. Compare against our evolve and skill-creator skills — Bian Que's correction-driven loop is sharper than ours.

Already Have

memory: project on chad-twin, maxTurns on chad-twin/chad-agent/worker/reviewer, isolation: worktree on worker/reviewer/planner/typescript-reviewer/python-reviewer/explorer (read-only), model: per-agent overrides, tools: whitelist on every agent, hooks: SubagentStop / SessionStart / SessionEnd / PreCompact / PostCompact / WorktreeCreate / WorktreeRemove / PostToolUseFailure / TaskCompleted / Stop / PreToolUse / PostToolUse / UserPromptSubmit (13 of the spec's 21 events), auto_runtime.py track lifecycle, auto-mode setting, agent: chad-twin default-agent setting, xhigh effort level, CLAUDE_CODE_AUTO_COMPACT_WINDOW=400000, omni-mem MCP at port 8765, skill-installer, skills-janitor, planning-gate, memory-adaptation, govern, drive, evolve, caveman, what-would-chad-do, notify_done.sh, /loop, /batch, /schedule, instinct-scoring (in caveman), context-fork (in explorer), claude-code-hooks-mastery, claude-code-hooks-battle-tested, type: prompt hook, /go composite skill, ReviewGrounder pattern (in reviewer), agent-teams-env flag, teammate-mode-in-process, save-task-list, route-classifier in UserPromptSubmit hook.

Rejected

• Claude Forge (11 agents + 33 commands + 24 skills + 15 hooks) — fails anti-overengineering gate; the current setup already has 10 agents / ~40 skills / 13 hook events with a justified mapping to actual workflow problems. Wholesale-installing a parallel 6-layer security harness would duplicate trailofbits/insecure-defaults, second-opinion, and the existing hook pipeline.
• Multi-Agent Coordinator agent (rohitg00) — already covered by /govern + auto_runtime.py packet DAG.
• AutoResearch agent (rohitg00) — ML experiment automation via tree search; wrong domain.
• PUIUX Pilot as runtime dep — useful as a reference pattern (see Build Queue), not as something to install.
• Claudify "OS-layer with 1,727 skills" — fails the simplicity budget and the "prove an existing primitive cannot satisfy this" gate. Hard pass.
• /rewind (double-tap Escape) — UX feature, not a config item; already available, no action needed.
• /focus mode — UI-level visual filter; no behavior impact on autonomous runs.
• Recaps — ambient feature when returning to long-running sessions; nothing to wire.
• microsoft/agents-v2-py — container-based agents; wrong abstraction for our local-skill model.
• alirezarezvani/claude-skills (232+ skill catalog) — bulk import would violate the simplicity budget; cherry-pick if a specific skill matches a recurring need (none identified today).

Sources checked: - https://github.com/hesreallyhim/awesome-claude-code (README in transition; no useful diff) - 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/sangrokjung/claude-forge - https://github.com/VoltAgent/awesome-agent-skills - https://github.com/affaan-m/everything-claude-code - https://www.developersdigest.tech/guides/config-change-hook

Tier 2 fetched: yes (last 2026-04-16, >24h) Tier 3 fetched: yes (last 2026-04-13, >7d) State file: NOT updated (dry-run) Auto-implement: NOT executed (dry-run) Run at: 2026-05-01T20:50:00Z