Ecosystem Update - 2026-05-23

TL;DR

• One safe harness delta was found and implemented: SessionStart hooks now cover the official compact source in addition to startup|resume|clear.
• Today's community signal is mostly orchestration packaging, context packets, run ledgers, and scoped worker handoffs; the local setup already has the core primitives through auto, planning-gate, AgentOps contract checks, and role-specific agents.
• No website deployment was performed; this run only touched Codex harness files plus the ecosystem report/state.

Quick Wins

Auto-Implemented

• Backups written to ~/.codex/backups/2026-05-23/ for config.toml, hooks.json, and all current agent TOMLs.
• Updated ~/.codex/hooks.json so both existing SessionStart hooks match startup|resume|clear|compact.
• Verified ~/.codex/hooks.json with python3 -m json.tool.
• Verified active feature posture with codex features list; hooks, goals, multi-agent, plugins, browser, computer-use, and OpenAI developer surfaces remain enabled, while plugin_hooks and native memories remain disabled.
• Ran codex doctor; config loaded successfully, with the only failing check caused by this non-interactive runner's TERM=dumb terminal setting. Doctor also noted unrestricted filesystem/network posture and large rollout state, both pre-existing.

Build Queue

• Context-packet/run-ledger compatibility intake (agent-pattern) - https://github.com/douglasmonsky/codex-orchestrate and https://github.com/ZypherHQ/agent-orchestration-skill - Both packages emphasize root-only routing metadata, compact worker packets, terminal lifecycle states, and handoff validation. Do not install wholesale; compare their packet/ledger schemas against the existing AgentOps task envelope and planning-gate implementation.
• Feature-specific subagent review pass (agent-pattern) - https://github.com/shanraisshan/codex-cli-best-practice - Local agents already cover planner, explorer, validator, generic reviewer, Python, TypeScript, worker, and chad-twin. A narrow follow-up could identify one or two genuinely recurring domains that deserve dedicated TOML agents instead of expanding the roster generically.
• Skill trigger and Gotchas audit (skill) - https://github.com/shanraisshan/codex-cli-best-practice - Current skills are numerous and mostly mature; a bounded audit should check whether high-use skills have sharp trigger descriptions and concise failure-mode notes without rewriting bodies.

Research

• Tier 2 was skipped because the previous run fetched research at 2026-05-22T10:37:39Z, and this run started at 2026-05-23T10:30:39Z, inside the skill's 24-hour skip window.
• No new research papers were fetched in this run.

Already Have

Hooks enabled, PreToolUse Bash guard, PostToolUse verification/failure hooks, UserPromptSubmit route classifier, Stop omni-mem save hook, PreCompact omni-mem capture hook, SessionStart startup/resume/clear/compact coverage, custom read-only explorer/planner/reviewer/validator agents, Python and TypeScript reviewer agents, worker and chad-twin agents, max_threads = 3, OpenAI developer docs MCP with parallel calls, omni-mem MCP, browser/chrome/computer-use/openai-developers plugins, web_search = "live", conservative/review profiles, goals = true, prevent_idle_sleep = true, skill-audit, rlm-scan, planning-gate, auto, orchestrate-local, ecosystem-update, codex-security, and security-audit.

Rejected

• Edit|Write matcher-only hook wiring - Official docs and community hook catalogs confirm apply_patch/Edit/Write matcher support, but the existing local edit_verify_async.py and tool_failure_context.py scripts intentionally return unless tool_name == "Bash". Changing only the matcher would be a no-op; any file-edit hook coverage needs a real script design and belongs outside Quick Wins.
• Global AGENTS.override.md adoption - Official docs support overrides, but a global override could silently shadow the byte-identical ~/.codex/AGENTS.md and ~/AGENTS.md contract. Treat this as an explicit policy decision, not an automatic harness tweak.
• Wholesale install of codex-orchestrate, agent-orchestration-skill, or am-will/codex-skills - Conflicts with the Codex-owned runtime rule and duplicates existing auto/planning-gate/orchestrate-local primitives without proving those primitives are insufficient.
• llm-council or multi-provider planning UI - Requires extra providers/API keys and a web UI; overengineered for the current local harness.
• read-github, markdown-url, Context7, or alternate browser skills - Already covered well enough by live web, OpenAI docs MCP, Browser/Chrome plugins, and local rg/GitHub reads.

Sources checked: https://github.com/hesreallyhim/awesome-claude-code, https://howborisusesclaudecode.com/, https://github.com/shanraisshan/codex-cli-best-practice, https://developers.openai.com/codex/hooks#matcher-patterns, https://developers.openai.com/codex/hooks#posttooluse, https://developers.openai.com/codex/guides/agents-md#customize-fallback-filenames, https://github.com/douglasmonsky/codex-orchestrate, https://github.com/ZypherHQ/agent-orchestration-skill, https://github.com/am-will/codex-skills, supplemental web search for Codex hooks/agents/skills. Tier 2 fetched: no; previous fetch was less than 24 hours old. Tier 3 fetched: no; previous fetch was less than 7 days old. omni-mem note: durable memory saved with id c30fadea-71d0-4863-a3d4-6059bd8d7445; state file remains source of truth. Run at: 2026-05-23T10:33:50Z.