Ecosystem Update - 2026-04-26

April 26, 2026 · curated by Chad Simon · 14 items reviewed

Highlights

Dry run only: no config/control-plane files, hooks, skills, agents, or trusted project entries were changed
Best new candidate is light observability in ~/.codex/config.toml: explicit TUI status line / terminal title and explicit live web search settings from the official Codex config surface
Main caution from current research: AGENTS.md can become counterproductive when it accumulates broad requirements. The global contract should be periodically compressed into smaller mandatory rules plus linked references

TUI status/title observability config

OpenAI Codex config reference

Add an explicit [tui] status line and terminal title so long-running sessions show model/project/context state at a glance
Explicit web search posture config

OpenAI Codex config reference

Set top-level web_search = "live" and optional [tools].web_search context metadata if current-data lookup should remain intentional under danger-full-access
Skill gotchas maintenance pass skills

Boris skill guidance / codex-cli-best-practice

Audit high-use skills for missing Gotchas sections and keep trigger descriptions focused on when the skill should fire

AGENTS.md contract compression/eval policy/runtime

arXiv 2602.11988 warns that overbroad context files can reduce task success and raise cost. Current ~/.codex/AGENTS.md is intentionally comprehensive, but it should be split into a smaller hot-path contract plus references for autonomous runtime internals
Codex TUI observability config config

official config exposes tui.status_line and tui.terminal_title; local config only has notifications = false
Explicit web search config config

official config supports top-level web_search and [tools].web_search; local config relies on full-access defaults
Skill trigger/gotchas audit skills

current skill inventory is large. The most useful ecosystem guidance is not adding packs, but tightening descriptions and failure-specific gotchas in the skills already used
Repository intelligence graph watch item

arXiv 2601.10112 supports deterministic repo/build/test maps. Local RLM preflight partly covers this; future work should compare RLM outputs against a repo graph contract before adding another extractor

Evaluating AGENTS.md

arXiv

- argues that repository context files can lower success rates and increase cost when they include unnecessary requirements. This is a direct caution for the growing global runtime contract
Repository Intelligence Graph

arXiv

- reports that deterministic build/test architecture maps improve coding-agent accuracy and efficiency. Current RLM preflight is the closest local analogue

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

Import broad Claude skill/plugin collections — - not appropriate for dry run, and the local runtime already has a large skill set plus mandatory external-skill audit
Enable native Codex memories
Rebuild the disabled keep-working LaunchAgent — - recent cleanup intentionally disabled the dead watcher; ecosystem sources do not justify reviving it
Add more autonomous orchestration layers — - current AGENTS.md already warns against this, and recent research favors leaner context and deterministic repo maps over more workflow text