Ecosystem Update - 2026-05-09

TL;DR

• The only safe automatic Quick Win was narrowing the startup RLM hook so /clear no longer pays the cached-context preflight cost.
• Today's Codex ecosystem signal mostly validates the current setup: gpt-5.5 default, live web search, hooks, role-specific agents, plugins, omni-mem, planning-gate, and explicit verification loops are already in place.
• Local CLI is still codex-cli 0.128.0 while OpenAI published 0.130.0 on 2026-05-08; keep that as a deliberate upgrade-and-smoke task, not an automatic ecosystem-update mutation.

Quick Wins

Auto-Implemented

• Backed up ~/.codex/config.toml, ~/.codex/hooks.json, and ~/.codex/agents/*.toml under ~/.codex/backups/2026-05-09/.
• Changed the existing SessionStart RLM preflight hook matcher to startup|resume, preserving the same script and avoiding new hook wiring.
• Verified ~/.codex/hooks.json with python3 -m json.tool.

Build Queue

• Codex 0.130.0 stable upgrade and smoke (Codex-md) - https://github.com/openai/codex/releases/tag/rust-v0.130.0 - current binary is codex-cli 0.128.0; upgrade should be deliberate because it changes the installed runtime. Smoke after upgrade: hooks, plugins, subagents, /review, MCP, browser plugin, and /goal if still enabled.
• PreToolUse / PostToolUse contract review (hook) - https://developers.openai.com/codex/hooks - official docs now describe PreToolUse, PermissionRequest, PostToolUse, UserPromptSubmit, and Stop wire formats. Do not add new hook entries until existing scripts are reviewed against current schemas.
• Config rules and permissions profile audit (config) - https://developers.openai.com/codex/config-reference#configtoml - rules, named permissions, granular approvals, and app tool policies are now documented. Add only as optional constrained profiles; keep the power-user default intact.
• Skill Gotchas maintenance pass (skill) - https://github.com/shanraisshan/codex-cli-best-practice - the repo emphasizes trigger-focused skill descriptions and high-signal Gotchas sections. Current skill inventory is broad; audit high-use skills before adding new ones.
• Native Codex memories pilot decision (config) - https://developers.openai.com/codex/config-basic#supported-features - native memories are stable but disabled by default. Current omni-mem remains the primary memory substrate; a pilot needs explicit scope to avoid duplicate recall paths.
• Worktree isolation wrapper for parallel local runs (workflow) - https://howborisusesclaudecode.com/ - Boris's strongest recurring workflow pattern is parallel worktrees. Codex already has subagents; a local wrapper is only worth building if it replaces repeated manual worktree setup.

Research

• TriEx: A Game-based Tri-View Framework for Explaining Internal Reasoning in Multi-Agent LLMs - relevant to future verifier design because it compares action-bound self-reasoning, belief state, and oracle audit traces over time.
• Kill-Chain Canaries: Stage-Level Tracking of Prompt Injection Across Attack Surfaces and Model Safety Tiers - directly relevant to memory/write-path governance; reinforces keeping untrusted ingestion away from durable memory writes.
• No new last-24-hour arXiv result appeared for the exact LLM agent coding query; Tier 2 was still fetched because the user asked for today's crawl.

Already Have

gpt-5.5 default, approval_policy = "never", danger-full-access, prompt telemetry off, config schema directive, top-level web_search = "live", review_model = "gpt-5.4", OpenAI developer docs MCP, omni-mem MCP, Stop and PreCompact memory hooks, RLM session preflight hook, role-specific custom agents, read-only planner/explorer/reviewer/validator agents, Python and TypeScript reviewers, worker and chad-twin agents, [agents] runtime caps, plugins enabled, browser-use, computer-use, Gmail, documents, spreadsheets, presentations, live marketplace entries, planning-gate, /auto, build/backlog/evaluate/govern skills, security-audit and codex-security skills, session recall via codex-session-search, skill audit before external installs, and ecosystem-update state tracking.

Rejected

• Adopt Claude ecosystem wholesale - awesome-claude-code is currently an update-in-progress/TODO README, and Claude-specific files or ~/.claude runtime dependencies conflict with Codex ownership.
• Add new hook scripts today - the safe Quick Win only narrowed an existing matcher. New PreToolUse, PermissionRequest, PostToolUse, or UserPromptSubmit entries require existing scripts first.
• Switch to shell-only edits for hook governance - third-party hook guidance suggests routing edits through shell to improve interception, but local Codex editing policy requires apply_patch, and official docs now document apply_patch hook aliases with caveats.
• Enable native Codex memories immediately - stable does not mean automatically correct for this harness; omni-mem is already wired into Stop/PreCompact and avoids a second recall/write path.
• Enable experimental app/tool policies broadly - current plugins already expose the needed Browser/Gmail/Documents/Spreadsheets/Presentations capabilities. Extra app policy churn needs a concrete failure mode.
• Auto-upgrade Codex CLI - upgrade is high-leverage but changes the runtime binary; keep it in Build Queue for an explicit upgrade-and-smoke slice.

Sources checked: https://github.com/hesreallyhim/awesome-claude-code, https://howborisusesclaudecode.com/, https://github.com/shanraisshan/codex-cli-best-practice, https://arxiv.org/search/?searchtype=all&query=LLM+agent+coding&order=-announced_date_first, https://export.arxiv.org/api/query?search_query=all:%22LLM%20agent%20coding%22&start=0&max_results=10&sortBy=submittedDate&sortOrder=descending, https://developers.openai.com/codex/hooks, https://developers.openai.com/codex/config-basic#supported-features, https://developers.openai.com/codex/config-reference#configtoml, https://developers.openai.com/codex/subagents#custom-agent-file-schema, https://github.com/openai/codex/releases/tag/rust-v0.130.0 Tier 2 fetched: yes Tier 3 fetched: partial - official Codex docs/releases checked because runtime facts are volatile; weekly toolkit source skipped because tier3_last_run is within 7 days Run at: 2026-05-09T10:31:57Z