Ecosystem Update - 2026-05-30

TL;DR

• Safe Quick Win implemented: config posture cleanup now catches missing /private/tmp/auto-task-eval-* trusted roots and pruned 3 stale entries from config.toml.
• Official Codex stable is still 0.135.0; this machine is on codex-cli 0.133.0, so upgrade + smoke remains a queued runtime action, not an automatic Quick Win.
• Today's research signal is strong around risk-stratified review, semantic checkpoints, and conflicting-memory abstention; those map to the current AgentOps/omni-mem harness as eval and gate work, not wholesale new orchestration.

Quick Wins

Build Queue

• Codex 0.135.0 stable upgrade and smoke (runtime) - https://github.com/openai/codex/releases - Current CLI is 0.133.0; latest stable 0.135.0 adds richer codex doctor, named permission profile display, packaged zsh helper discovery, Python SDK sandbox presets, resume/TUI fixes, and thread idle lifecycle work. Queue because upgrading the installed CLI crosses the runtime authority boundary.
• Named permission profile migration evaluation (config) - https://developers.openai.com/codex/config-reference#configtoml - Current config uses top-level sandbox_mode plus [profiles.*]; docs expose default_permissions and [permissions.<name>], but warn not to combine default_permissions with sandbox_mode. Needs a deliberate migration plan.
• Thread idle lifecycle hook intake (hook) - https://github.com/openai/codex/releases - Release notes include thread idle lifecycle work. Current hooks.json covers PreToolUse, PostToolUse, SessionStart, UserPromptSubmit, Stop, and PreCompact; no idle-specific existing script is available, so this cannot be wired as a Quick Win.
• Missing non-temp trust-root cleanup policy (config hygiene) - local codex-runtime-doctor - Doctor now reports 1 remaining stale trusted project, /Users/chadsimon/Documents/New project 7. It is missing on disk, but it is not an auto-task temp root, so cleanup should be explicit.
• RADAR-style review admission policy (review gate) - https://arxiv.org/abs/2605.30208 - Meta's RADAR paper supports layered low-risk diff eligibility, static heuristics, LLM review, and deterministic validation before automation. Current review posture is strong, but codex_review_gate.py could add an explicit low-risk/needs-human-review classifier for autonomous closures.
• Semantic checkpoint evidence contract (AgentOps/autonomy) - https://arxiv.org/abs/2605.30042 - The paper's "semantic checkpoints" pattern matches the existing evidence-ref contract; the gap is a deterministic action-intent to outcome check in autonomous slice ledgers.
• Conflicting memory abstention eval (omni-mem) - https://arxiv.org/abs/2605.30087 - Omni-mem is the active memory system, but today's paper suggests adding an eval where context builders must abstain or mark uncertainty when sources conflict instead of flattening memory into one answer.

Research

• Automating Low-Risk Code Review at Meta: RADAR, Risk Calibration, and Review Efficiency - Directly relevant to review-gated autonomous closure and risk-tiered approval.
• Learning to Choose: An Empowerment-Guided Multi-Agent System with semantic communication for Adaptive Method Selection - Useful framing for preventing semantic drift between selected strategy, delegated implementation, and verified result.
• Selective QA over Conflicting Multi-Source Personal Memory - Good candidate for an omni-mem eval around conflicts, source weighting, and abstention.
• Discovering Cooperative Pipelines: Autoresearch for Sequential Social Dilemmas - Relevant to /evolve, but should stay behind eval-gated self-modification rather than runtime default behavior.
• Code as Agent Harness - Reinforces the current code-first harness direction: executable checks, typed state, and deterministic verification matter more than model-only claims.

Already Have

PreToolUse Bash guard, PostToolUse verification ledger, PostToolUse failure context, SessionStart startup/resume/clear/compact coverage, UserPromptSubmit route classifier, Stop omni-mem save hook, PreCompact omni-mem hook, OpenAI developer docs MCP, omni-mem MCP, browser/chrome/computer-use plugins, read-only planner/reviewer/validator agents, workspace-write worker agent, custom skills, planning-gate, auto runtime, what-would-chad-do reflection, codex-runtime-doctor official doctor summary, execpolicy rules for destructive git and rm -rf, prompt telemetry disabled, plugin hooks disabled, native Codex memories disabled, conservative profiles, project doc byte cap.

Rejected

• Auto-upgrade Codex to 0.135.0 as a Quick Win - rejected because upgrading the installed CLI is a user-authority/runtime-change boundary and prior state rejected auto-upgrades.
• Clone dynamic workflows wholesale - rejected because Boris's May 28 dynamic workflow pattern is a research-preview, hundred-agent orchestration shape; current max_threads = 3, /auto, planning-gate, and explicit slices already cover the safe subset.
• Enable native Codex memories from community advice - rejected because omni-mem is the active memory system and prompt/memory telemetry posture remains intentionally conservative.
• Wholesale import am-will/codex-skills, oh-my-codex, or Claude community skill catalogs - rejected because local skills already cover the useful primitives; bulk imports add supply-chain and trigger-surface risk. oh-my-codex is also archived.
• Global auto-format hooks from community tips - rejected because safe formatter hooks need repo-specific existing formatter commands; no universal formatter hook script should be wired globally.
• Remove non-temp missing trusted roots automatically - rejected because /Users/chadsimon/Documents/New project 7 is missing but not an auto-task temp root; leave it for an explicit cleanup pass.

Auto-Implemented

• Backups written under ~/.codex/backups/2026-05-30/ for config, hooks, agents, and codex_config_posture.py.
• Patched ~/.codex/bin/codex_config_posture.py so stale temp trust-root detection includes /private/tmp/auto-task-eval-* paths.
• Ran python3 /Users/chadsimon/.codex/bin/codex_config_posture.py --fix-stale-temp-roots; removed 3 missing temp trusted project entries.
• Verification passed: python3 -m py_compile /Users/chadsimon/.codex/bin/codex_config_posture.py.
• Verification passed: python3 -m json.tool /Users/chadsimon/.codex/hooks.json.
• Verification passed: tomllib parse of ~/.codex/config.toml and all ~/.codex/agents/*.toml.
• Verification passed: python3 /Users/chadsimon/.codex/bin/codex_config_posture.py --mode check returned ok: true.
• Verification passed with known residual warnings: python3 /Users/chadsimon/.codex/bin/codex-runtime-doctor completed with errors=0 warnings=3; stale trusted project warning dropped from 4 entries to 1, and official codex doctor still reports TERM=dumb in the non-interactive shell.
• Omni-mem search was attempted but failed with Missing observation ... for memory embedding; state file remains the source of truth.

Sources checked: https://github.com/hesreallyhim/awesome-claude-code, https://howborisusesclaudecode.com/, https://github.com/shanraisshan/codex-cli-best-practice, https://github.com/openai/codex/releases, https://developers.openai.com/codex/, https://developers.openai.com/codex/config-reference, https://arxiv.org/search/?searchtype=all&query=LLM+agent+coding&order=-announced_date_first, https://github.com/am-will/codex-skills, https://github.com/scalarian/oh-my-codex, web search supplement for Codex hooks/agents/skills. Tier 2 fetched: yes. Tier 3 fetched: yes, targeted official Codex docs/releases; weekly toolkit crawl skipped because the state file shows Tier 3 ran on 2026-05-29. Run at: 2026-05-30T10:34:13Z.