Ecosystem Update - 2026-05-26

May 26, 2026 · curated by Chad Simon · 12 items reviewed

Highlights

One safe Quick Win was implemented: the existing SessionStart cached-repo-context hook now has a bounded timeout, matching the rest of the hook harness
Today's strongest new research signal is CODESKILL: trajectory-derived, self-evolving skill banks for coding agents. It belongs in eval/design intake, not as an automatic runtime change
The local Codex binary reports codex-cli 0.133.0-alpha.1; given the runtime posture against global under-development features, channel/version posture should be audited explicitly before any upgrade or rollback

SessionStart cached-context timeout hook

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

Add a timeout to the existing rlm_session_preflight.py hook so cached context cannot hold session startup indefinitely

CODESKILL trajectory-to-skill eval intake

arXiv

https://arxiv.org/abs/2605.25430v1 - Distill coding-agent trajectories into compact procedural skills with a stable skill bank. Current setup has evaluate, analyze, memory-consolidation, and many domain skills, but no explicit admission test for converting repeated failures into durable skill changes
Session-scoped guardrail commands hook

howborisusesclaudecode.com

https://howborisusesclaudecode.com/ - Boris/Thariq patterns mention on-demand guardrails like careful/freeze modes. Current setup has global Bash safety via pre_tool_guard.py; session-scoped modes would need a small design for how a user toggles them and how hooks read that state
Codex version channel posture audit Codex-md

github.com/openai/codex

https://github.com/openai/codex/releases - Local codex --version reports an alpha build while AGENTS.md says not to enable under-development Codex features globally. This needs explicit user direction because rollback/upgrade touches the runtime install, not just harness config

CODESKILL: Learning Self-Evolving Skills for Coding Agents

arXiv

- Directly relevant to turning verified Codex task trajectories into compact, reviewable skill updates without unbounded skill growth
IterInject: Indirect Prompt Injection Against LLM Agents via Feedback-Guided Iterative Optimization

arXiv

- Supports keeping outside content, memory promotion, and tool permissions under strict admission control
FrontierOR: Benchmarking LLMs' Capacity for Efficient Algorithm Design in Large-Scale Optimization

arXiv

- Useful as a reminder that executable code is not enough; optimization/design tasks need hidden or objective performance checks, not only smoke tests

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

Wholesale import Boris, Awesome Claude Code, or Codex community skill packs — - rejected: outside skills require strict audit, and the local skill set already covers the recurring workflows
Auto-format hooks from community tips — - rejected as a Quick Win: formatter hooks need repo-specific detection and an existing script; adding broad format commands globally would violate the no-new-script/no-arbitrary-hook limit
Enable native Codex memories because community docs recommend it
Add or rely on AGENTS.override.md as a Quick Win — - rejected: AGENTS policy docs are explicitly out of Quick Win scope, and current global policy requires the two Codex-owned AGENTS files to stay byte-identical
Clone Agent View or other Claude-specific control planes — - rejected: the already has local task/goal/reporting primitives; a new session-control service would be orchestration without proof an existing primitive cannot satisfy the need