Ecosystem Update - 2026-07-02

Highlights

• No safe Quick Wins were auto-applied today; the new signal is useful but needs scoped design work rather than blind harness edits
• Today's strongest research signal is skill supply-chain risk: public skills increasingly need dependency/provenance manifests, lockfile-like records, and audit commands
• Runtime-grounded verification keeps showing up: compatibility rescue, bug reproduction diagnosis, and benchmark reliability all point toward better evidence capture in the local harness

Quick Wins (implemented today)

• None admitted -
  Current scoring pass
  No automatic harness mutation passed the anti-overengineering and safety gates

New Tools, Skills & Patterns

• Skill dependency manifest and audit spike skill
  arXiv
  Skills Are Not Islands - Add a local design for typed skill dependency/provenance metadata and a codex-skill-audit extension that can warn about hidden package, service, MCP, or recursive skill dependencies before outside skills are trusted
• Source-only compatibility rescue mode agent-pattern
  arXiv
  RepoRescue - Add an opt-in harness/eval mode for compatibility-rescue tasks that records source-only repair evidence and can block test-file edits when the task contract requires it
• Runtime-diagnosis bug repair adapter skill
  arXiv
  SWE-Doctor - Extend bug-miner or fix-issue with multi-faceted bug reproduction records that separate observed failures, diagnosis, localization, and patch guidance
• Benchmark reliability scoring for agent evals
  arXiv
  Are Performance-Optimization Benchmarks Reliably Measuring Coding Agents? - Add an eval-harness note or check that scores task instability and reference-patch replay reliability before using benchmark results as promotion evidence
• Instruction-conflict and progress-report taxonomy
  arXiv
  Adversarial Pragmatics for AI Safety Evaluation - Map instruction conflict, embedded commands, policy ambiguity, scaffold failures, and false progress claims into existing completion/advisor evidence
• Memory sycophancy and scope tests
  arXiv

Research Worth Reading

• AutoMem: Automated Learning of Memory as a Cognitive Skill
  arXiv
  - Treats memory management as an independently improvable agent skill
• RepoRescue: An Empirical Study of LLM Agents on Whole-Repository Compatibility Rescue
  arXiv
  - Useful for designing source-only repair contracts and practical validation beyond passing legacy suites
• Are Performance-Optimization Benchmarks Reliably Measuring Coding Agents?
  arXiv
  - Warns that aggregate benchmark scores can hide replay instability and scoring-rule artifacts
• Skills Are Not Islands: Measuring Dependency and Risk in Agent Skill Supply Chains
  arXiv
  - Directly applicable to the local outside-skill audit and trust workflow
• Cheap Code, Costly Judgment: A Case Study on Governable Agentic Software Engineering
  arXiv
  - Reinforces the local AgentOps emphasis on inspectable evidence, review pressure, and maintainable generated changes
• SWE-Doctor: Guiding Software Engineering Agents with Runtime Diagnosis from Multi-Faceted Bug Reproduction Tests
  arXiv
  - Strong candidate pattern for bug-miner/fix-issue once adapted to local logs and test runners

Considered, Not Adopting

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

• Auto-enable native Codex memories — needs an explicit migration/pilot, not a daily Quick Win
• Switch back to deprecated features.codex_hooks — - community best-practice text still mentions codex_hooks, but official docs mark features.hooks as canonical and the local config already uses it
• Auto-import public skill catalogs — - fails the skill supply-chain risk gate; outside skills require codex-skill-audit --strict and explicit selection
• Add dynamic workflow or nested-agent runtime layers from Claude patterns — - conflicts with anti-overengineering unless tied to a recurring local failure and implemented in Codex-owned surfaces
• Default all agents to deeper nesting or worktree isolation — - prior runs already rejected this; current max_depth = 1 and bounded roles are intentional
• Skip all SessionStart hooks on /clear — - the current heavy context hook already excludes clear; remaining clear hooks are lightweight readiness/posture checks, so this is low-impact and not worth a mutation
• Global test-edit blocking — - useful for compatibility-rescue evals, but harmful as a blanket rule because normal TDD requires adding or updating tests
• Automatic Codex CLI upgrade — - update checks are already enabled; changing binaries remains outside the safe Quick Win boundary

Sources Reviewed

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