Ecosystem Update - 2026-06-17
Highlights
- Implemented one safe harness Quick Win:
CODEX_DISABLED_HOOKSnow lets hook profiles suppress named hooks at runtime without editinghooks.json - Today's Tier 1 signal reinforces context isolation, workflow triggers, and hook runtime controls; this setup already has most of the useful primitives
- No new hook wiring, policy-doc edits, plugin-hook enablement, native memories, or deeper nested agents were applied because they either require trust review, violate current posture, or exceed Quick Win scope
Quick Wins (implemented today)
-
Runtime disabled-hook list hookAdd
CODEX_DISABLED_HOOKShandling to existinghook_profile.pyprofile gate
New Tools, Skills & Patterns
-
Nested subagent depth eval agent-patternhttps://howborisusesclaudecode.com/ - Boris's latest nested-agent guidance argues for deeper delegated context isolation. Current Codex config pins
agents.max_depth = 1; evaluate a temporarymax_depth = 2profile only under explicit autonomous-lab rollback notes before considering any default change -
Forked skill context compatibility audit skillhttps://howborisusesclaudecode.com/ - Claude's
fork: trueskill isolation maps partly to Codex/forkand side-thread primitives, but Codex skills do not expose that frontmatter. Build a compatibility note or eval instead of adding unsupported metadata -
Portable skill sanity-test harness skillhttps://github.com/grahama1970/agent-skills - The shared skill catalog includes
skills/run-all-sanity.shas a pattern. This setup hasskill-auditand many skills, but no single local sanity pass that verifies installed skill metadata plus referenced files and scripts across all roots -
Portability fidelity rubric skillhttps://github.com/camoa/claude-skills/blob/main/PORTABILITY.md - The portability guide separates portable
SKILL.mdknowledge from non-portable commands/agents/hooks. Turn that into a local audit mode for outside skill intake so Codex can classify "knowledge only" vs. "workflow fidelity lost" before installing -
Adaptive reasoning-effort watch item agent-patternhttps://github.com/openai/codex/issues/26961 - Community request asks for hooks/profile managers to adjust reasoning effort dynamically. Current setup has a prompt route classifier, but Codex does not expose a stable hook output to mutate effort mid-session, so this remains a watch item
Research Worth Reading
-
No arXiv Tier 2 papers fetched. The prior Tier 2 run was
2026-06-16T06:32:36-04:00, and today's crawl began at2026-06-17T06:30:59-04:00, still inside the 24-hour gate.
Considered, Not Adopting
Items reviewed and explicitly declined this cycle, with the reason. Curation discipline matters more than coverage.
-
Raise
agents.max_depthfor nested subagents — - Rejected as a Quick Win because it increases autonomous fan-out and cost. Evaluate under a temporary profile first -
Add
fork: trueto local skills — - Rejected because Codex does not document that skill frontmatter; unsupported metadata would be noise -
Enable plugin hooks globally — - Rejected because current posture keeps
plugin_hooks = false, and official docs require hook trust review for plugin-bundled hooks - Wire new Stop validation hooks immediately — - Rejected because prior state already tracked stop-gate wiring, existing hooks require trust hashes, and new/changed non-managed hooks are skipped until trusted
-
Bootstrap AGENTS.md from an external repo each session — - Rejected because this machine owns
~/.codex/AGENTS.mdand~/AGENTS.mdbyte identity; self-updating policy docs conflict with runtime ownership - Wholesale install shared skill/hook catalogs — - Rejected because outside skills require strict audit first and many catalog hooks target Claude paths or unknown trust semantics
- Switch to Claude/Fable model guidance — - Rejected because it is not a Codex runtime target and does not map to the configured OpenAI model posture