Ecosystem Update — 2026-04-12

TL;DR

• Dry-run rerun only: no Quick Wins were applied.
• The earlier same-day Quick Win is already present: SessionStart now matches startup|resume|clear, so the current release-level clear-session hook gap is closed.
• Today's fresh Tier 1 signal is backlog-shaped: official skill catalog gaps, bounded heartbeat/scheduled-work patterns, and lightweight config/rules linting are worth tracking, but none pass the safe auto-apply gate today.
• GitHub Trending is strongly aligned with our direction today: agent coworkers, repeat-until-done loops, session memory, and deterministic coding harnesses are all trending.

Quick Wins

Build Queue

• Official skill catalog triage (skill) — https://github.com/openai/skills/tree/main/skills/.curated — the curated catalog includes potentially relevant skills not currently installed locally, including gh-fix-ci, gh-address-comments, security-threat-model, security-ownership-map, sentry, and deployment skills. Run codex-skill-audit --strict before installing anything; do not bulk-install.
• Bounded heartbeat playbook (workflow) — https://howborisusesclaudecode.com/ — Boris's schedule/loop examples reinforce the direction we already chose: invoked heartbeat for recurring safe work, with durable state and authority stops. This maps to manager-heartbeat, not a permanent daemon.
• Config/rules lint check evaluation (workflow) — https://github.com/hesreallyhim/awesome-claude-code — tools like agnix and rules/config doctors are relevant as a concept: add a small local validation/eval only if we see repeated AGENTS/SKILL/hook drift. Outside tools need audit first.
• Plugin enablement decision (mcp) — https://github.com/shanraisshan/codex-cli-best-practice and https://developers.openai.com/codex/ — Codex plugins are now a documented surface, and plugin-creator exists locally, but [features].plugins = false. Keep disabled until a specific plugin use case beats direct MCP/skill wiring.
• GitHub Trending daily scan source (workflow) — https://github.com/trending?since=daily — add Trending as an ecosystem-update Tier 1 source, filtered for agent runtimes, memory systems, verification harnesses, MCP servers, and codebase-intake tools. Today's signal would have caught hermes-agent, multica, Archon, claude-mem, ralph, and markitdown.

GitHub Trending

• NousResearch/hermes-agent — https://github.com/NousResearch/hermes-agent — self-improving agent with learning loop, skills, memory/search, scheduled automations, subagents, and messaging gateways. Classification: PARTIAL. We already have pieces through /auto, omni-mem, skills, and manager-run-task, but Hermes is a good reference for gateway/channel UX and skill evolution.
• multica-ai/multica — https://github.com/multica-ai/multica — managed agents platform for assigning tasks, tracking progress, and compounding skills. Classification: PARTIAL. This maps to our coworker control-plane/dashboard/inbox direction, not a reason to add an external platform.
• coleam00/Archon — https://github.com/coleam00/Archon — deterministic/repeatable AI coding harness. Classification: PARTIAL. Strongly aligned with manager-task-eval and final-review gates; worth reviewing for harness UX, not importing wholesale.
• thedotmack/claude-mem — https://github.com/thedotmack/claude-mem — session capture, compression, and future-context injection plugin. Classification: PARTIAL/HAVE. We already have omni-mem and session recall; useful as a product-shape comparison for memory UX.
• snarktank/ralph — https://github.com/snarktank/ralph — repeat-loop agent that runs fresh coding instances until PRD items are complete, persisting through git history/progress files/PRD JSON. Classification: HAVE/PARTIAL. This is very close to the philosophy behind manager-run-task; the remaining delta is mostly our invoked heartbeat and dashboard layer.
• forrestchang/andrej-karpathy-skills — https://github.com/forrestchang/andrej-karpathy-skills — single-file behavior guidance focused on assumptions, confusion management, tradeoffs, and overengineering. Classification: HAVE. Our AGENTS rules already cover these principles.
• microsoft/markitdown — https://github.com/microsoft/markitdown — document-to-Markdown converter with MCP support. Classification: MISSING. Useful for intake/document parsing, but not a Quick Win unless we hit repeated PDF/Office intake failures.

Research

• No new Tier 2 paper scan in this rerun. The state file shows Tier 2 was already fetched on 2026-04-12T05:56:25Z, which is within the skill's 24-hour skip window.

Already Have

SessionStart, Stop, and PreCompact hooks; command hooks with status messages and timeouts; SessionStart clear-source coverage; $CODEX_HOME/config.toml; $CODEX_HOME/hooks.json; TOML custom agents; agent model/reasoning overrides; read-only reviewer/planner/validator agents; worker agent; [agents] thread/depth/runtime limits; omni-mem MCP; stitch MCP; reusable skills under $AGENTS_HOME/skills; system skills under $CODEX_HOME/skills/.system; plugin-creator system skill; plugin feature flag present but disabled; Playwright/browser verification skill; invoked manager-run-task; bounded manager heartbeat roadmap; final-review hard gate; task-run reports; manager dashboard/inbox/outbox roadmap; anti-overengineering gate; governed/autonomous runtime rules.

Rejected

• TaskCreated/StopFailure/PermissionDenied hook sounds from codex-cli-best-practice recent commits — notification sounds are low signal and would add noise without improving autonomy correctness.
• Bulk-install official or third-party skills — outside skills require codex-skill-audit --strict; installing multiple skills is outside dry-run scope and outside Quick Win limits.
• Enable plugins globally — no concrete current plugin requirement; direct skills/MCP already cover current needs.
• Adopt an external orchestration framework from the Claude ecosystem — the useful Ralph/loop/control-plane concepts are already represented in /auto manager-run-task and the bounded heartbeat plan; importing external framework machinery would violate the anti-overengineering gate.
• Adopt trending agent platforms wholesale — Hermes, Multica, Archon, and Ralph mostly validate our architecture; the right move is targeted comparison/backlog extraction, not replacement.
• Fetch Tier 2/Tier 3 again — both were fetched earlier today and remain within their skip windows.

Sources checked: https://github.com/hesreallyhim/awesome-claude-code, https://howborisusesclaudecode.com/, https://github.com/shanraisshan/codex-cli-best-practice, https://github.com/openai/skills/tree/main/skills/.curated, https://developers.openai.com/codex/, https://github.com/trending?since=daily, https://github.com/trending/typescript?since=daily, https://github.com/trending/python?since=daily Tier 2 fetched: no — skipped because tier2_last_run is within 24 hours Tier 3 fetched: no — skipped because tier3_last_run is within 7 days Run at: 2026-04-12T17:38:07Z