Ecosystem Update - 2026-06-11

Highlights

• One safe harness Quick Win landed: default.rules now blocks short-form git push -f, matching the existing pre_tool_guard.py force-push protection
• Today's community sources mostly reinforce the existing setup: progressive-disclosure skills, rules-backed permissions, verification loops, read-only reviewers, planning gates, and conservative rejection of broad auto-upgrades

Quick Wins (implemented today)

• Force-push short flag execpolicy parity guard rule
  github.com/shanraisshan/codex-cli-best-practice
  Extend the existing Starlark command policy from git push --force to also forbid git push -f

New Tools, Skills & Patterns

• Codex 0.137+ upgrade and smoke Codex-md
  github.com/shanraisshan/codex-cli-best-practice
  https://github.com/shanraisshan/codex-cli-best-practice - Local codex --version is still codex-cli 0.133.0, while the best-practice repo references 0.137.0 session archive/history behavior. Verify against official OpenAI release notes before upgrading; do not auto-upgrade from this daily skill
• Self-gated clarification adapter for R5 routing agent-pattern
  arXiv
  https://arxiv.org/abs/2606.11349 - ACTION-RATING's "ask vs act" framing maps to Codex's R5 ambiguity boundary
• Disaggregated evidence audit for accepted slices agent-pattern
  arXiv
  https://arxiv.org/abs/2606.11522 - Search Discipline shows aggregate scores can hide local failure. Adapt this into auto closure checks that require per-slice evidence, not just a terminal pass/fail summary
• SkillJuror-style skill organization eval skill
  arXiv
  https://arxiv.org/abs/2606.11543 - The local skill library already uses progressive disclosure, but there is no regression eval proving which skills are too flat, too deep, or poorly routed
• HORMA-style hierarchical memory navigation spike mcp
  arXiv
  a bounded spike could test hierarchy-first retrieval against today's hybrid search packs
• TreeSeeker branch-and-return research mode skill
  arXiv
  https://arxiv.org/abs/2606.11662 - Useful for deep research or ecosystem scans where branches need explicit evidence, uncertainty, conflicts, and pruning state. Build as a scoped research skill, not a global agent loop
• Infrastructure-aware orchestration budget policy agent-pattern
  arXiv
  https://arxiv.org/abs/2606.11440 - INFRAMIND supports simpler topologies under load and richer ones when capacity is available. This belongs in route budgeting or local-worker dispatch after telemetry exists
• Claim-level verification harness for numerical/tabular tasks skill
  arXiv
  https://arxiv.org/abs/2606.11537 - MOCA-Agent's atomic claim verification maps well to spreadsheet, finance, and proof-heavy tasks. Build only as a domain-specific checker

Research Worth Reading

• Knowing When to Ask: Self-Gated Clarification for Hierarchical Language Agents
  arXiv
  - Strong fit for R5 detection: make clarification an explicit action competing with execution
• INFRAMIND: Infrastructure-Aware Multi-Agent Orchestration
  arXiv
  - Relevant to future local/remote worker routing and budget-aware agent topology
• Search Discipline for Long-Horizon Research Agents
  arXiv
  - Reinforces evidence-by-slice closure and external audits over single aggregate success scores
• MoCA-Agent: A Market-of-Claims Code Agent for Financial and Numerical Reasoning
  arXiv
  - Useful pattern for claim-level verification before executable synthesis in high-stakes numerical work
• SkillJuror: Measuring How Agent Skill Organization Changes Runtime Behavior
  arXiv
  - Directly applicable to local skill quality: progressive disclosure changes runtime behavior and should be evaluated, not assumed
• TreeSeeker: Tree-Structured Trial, Error, and Return in Deep Search
  arXiv
  - Candidate control pattern for disciplined branch-and-return research sessions
• Organize then Retrieve: Hierarchical Memory Navigation for Efficient Agents
  arXiv
• Defeater Cards: Characterizing and Managing Safety Assurance Case Defeaters
  arXiv
  - Good artifact pattern for surfacing hidden assumptions and reasoning gaps in planning-gate reviews

Considered, Not Adopting

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

• Auto-upgrade Codex CLI from the ecosystem run - version drift remains real, but upgrades require official release verification and smoke testing outside the safe Quick Win lane.
• Enable nested agents beyond current max_depth = 1 because Boris references depth 5 - this is an autonomy-risk and cost-control decision, not a one-line daily harness fix.
• Add fork: true skill frontmatter across local skills - experimental cross-agent pattern; skill bodies must not be rewritten by this Quick Win workflow.
• Trigger dynamic workflows from the phrase "use a workflow" - useful product pattern, but current Codex policy makes /auto the canonical autonomous runtime and avoids broad keyword routing without an eval.
• Retire planning/alignment in favor of context minimalism or always-on auto mode - conflicts with this machine's non-trivial-work alignment and AgentOps contract.
• Enable plugin hooks globally - still a trust-boundary decision; current features.plugin_hooks = false remains the safer default.
• Enable native Codex memories globally - conflicts with the omni-mem-first memory policy and current prompt-telemetry posture.
• Wholesale install from community skill/plugin catalogs - outside skills/plugins require strict audit and Codex-owned adaptation.
• Add global auto-format, browser-QA, or visual-QA hooks - useful per project, but too broad as a machine-wide hook because repo stacks vary.
• Edit AGENTS.md as a Quick Win - constitutional policy docs require explicit user direction.
• Adopt Fable 5 / Claude-specific model guidance - informative for cross-agent awareness, but not applicable to the local Codex model configuration.

Sources Reviewed

• github.com
• howborisusesclaudecode.com
• github.com
• arxiv.org
• arxiv.org
• arxiv.org
• arxiv.org