Ecosystem Update - 2026-06-08

June 8, 2026 · curated by Chad Simon · 12 items reviewed

Highlights

One safe harness Quick Win landed: config.toml now uses Codex-owned project docs only, removing CLAUDE.md from project_doc_fallback_filenames
Today's Tier 1 signal mostly reinforced existing posture: explicit AGENTS.md context, hook-backed validation, progressive-disclosure skills, bounded agents, and cautious plugin/marketplace intake
No new last-24h research item cleared the direct-applicability bar; the arXiv API still surfaced older agent-eval and memory items already represented in prior state

Codex-owned project doc fallback Codex-md

github.com/shanraisshan/codex-cli-best-practice

Remove CLAUDE.md from project_doc_fallback_filenames; keep AGENTS.md as the runtime project-doc fallback

Deep claim/source auditor skill skill

howborisusesclaudecode.com

https://howborisusesclaudecode.com/ - Boris's verifier patterns point toward a thin Codex-owned skill that extracts claims from a result, maps each claim to code/source evidence, and asks existing reviewer/validator agents to falsify unsupported claims. Worth building as a skill, not a Quick Win, because it requires a new skill folder and acceptance harness
MCP parallel-call capability audit mcp

github.com/shanraisshan/codex-cli-best-practice

https://github.com/shanraisshan/codex-cli-best-practice - The best-practice repo highlights supports_parallel_tool_calls = true; only openaiDeveloperDocs has it locally. Audit each non-docs MCP server for concurrency safety before adding flags, because stdio servers and stateful remotes may not be safe to parallelize
Skill gotchas drift check skill

howborisusesclaudecode.com

https://howborisusesclaudecode.com/ - The skill authoring guidance keeps emphasizing gotchas and progressive disclosure. Current skill coverage is strong, but a janitor pass could flag skills without gotchas/progressive-disclosure sections and queue targeted repairs

No fresh arXiv result in the LLM agent coding daily crawl was both new and directly actionable for the Codex harness.
Agentic Monte Carlo: Simulating Reinforcement Learning for Black-Box Agents

arXiv

- Still the newest relevant API result, but it was already represented in the June 6/7 state as a research intake item, and it does not imply a safe harness mutation today

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

Wholesale install from awesome-claude-code, awesome-claude-code-toolkit, am-will/codex-skills, or similar community bundles - useful source material, but outside skills/plugins/hooks require strict audit and Codex-owned adaptation.
Auto-wire new hook events from community or issue trackers - no new existing target script was proven for unsupported events; the Quick Win rules forbid hook wiring that requires new scripts.
Enable plugin hooks globally - current trust boundary remains unresolved, and features.plugin_hooks = false is still the right default.
Enable native Codex memories globally - conflicts with the current omni-mem-first memory policy and prompt-telemetry posture.
Enable Fast Mode or global service_tier from best-practice notes - spends more credits and is a user preference, not a harness safety or quality improvement.
Default all subagents to worktree isolation - still too broad for routine local work; use bounded task-specific worktree isolation when the task warrants it.