Ecosystem Update - 2026-07-11

July 11, 2026 · curated by Chad Simon · 16 items reviewed

Highlights

The daily community sources were quiet: awesome-claude-code had only automated ticker refreshes, while Boris's site and codex-cli-best-practice exposed no post-run workflow changes
The strongest runtime delta is on Codex main: materialized listed-workspace plugins can have hook hashes trusted after refresh, and terminal/plugin events are gaining richer completion and scheduled-task metadata. The local runtime deliberately has plugin_hooks = false, so these are upgrade-and-audit items, not safe automatic changes
New research reinforces the current code-owned contract and verifier posture. The useful remaining gaps are contamination-resistant long-horizon task evals and a bounded way to identify repeated SOP steps that deserve validated tools

None passed the automatic-change gate n/a

Local diff + today's sources

No harness files changed

Contamination-resistant long-horizon task eval slice research, Impact 3 / Effort 2 = 1.5

arXiv

https://arxiv.org/abs/2607.07946 - Add a small set of original tasks with hand-written behavioral verifiers to the existing task-eval path so closure quality is measured independently of public issue history and inherited fix tests
Stable plugin-hook trust upgrade smoke hook, Impact 3 / Effort 2 = 1.5

github.com/openai/codex

https://github.com/openai/codex/commit/076a110eb83445fd33fcd40f9a70f3ad84120f76 - When the behavior reaches a stable release, audit install-policy boundaries and verify that only listed workspace plugins for the active account receive current hook hashes before considering plugin_hooks enablement
Terminal completion metadata adapter agent-pattern, Impact 2 / Effort 2 = 1.0

github.com/openai/codex

https://github.com/openai/codex/commit/c4318c386de365bd0dd9595a08d55a30bb142d11 - After a stable upgrade exposes terminal errors in completion events, test whether the existing failure-context/completion-evidence path can consume them without another logging layer
Repeated-SOP tool candidate profiler skill, Impact 2 / Effort 2 = 1.0

arXiv

https://arxiv.org/abs/2607.08010 - Extend existing evaluation or analysis surfaces to identify recurring deterministic steps and propose versioned tools; keep synthesis, validation, and promotion human-gated

TTHE: Test-Time Harness Evolution

arXiv

- Relevant to evolve and autoconfig, but its unsupervised population/judge loop should remain research until execution-derived proxy reliability is independently verified
From Prompts to Contracts: Harness Engineering for Auditable Enterprise LLM Agents

arXiv

- Supports the current pattern of moving guarantees into code, manifests, schemas, traces, and fault-detecting validators instead of prompt-only policy
DeepSWE: Measuring Frontier Coding Agents on Original, Long-Horizon Engineering Tasks

arXiv

- Its original tasks and hand-written verifiers offer a concrete design for stronger local task-eval evidence
Tool-Making and Self-Evolving LLM Agents in Low-Latency Systems

arXiv

- Suggests compiling repeated, validated procedures into versioned tools rather than regenerating identical code on every run
Aleena: Alignment Agent for Research Software Engineering Collaborations

arXiv

- Useful as a comparison point for decision continuity across chat, meetings, issues, and pull requests

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

Automatically enable plugin hooks — - Rejected because the feature is intentionally disabled and today's hook-trust behavior is on main, not established in the installed stable runtime; enabling executable plugin hooks requires a security review and stable smoke test
Automatically upgrade Codex from main — - Rejected because source commits are not a stable release boundary, and binary upgrades are outside the Quick Win mutation limits
Test-time population evolution in the live harness — - Rejected as overengineered and insufficiently verified; it introduces persistent self-modification driven by proxy signals where the local loop contract requires deterministic verifiers, budgets, stop conditions, and human gates
Plugin scheduled-task metadata as a harness edit — - Rejected as low-impact UI/observability metadata with no safe local config action; revisit only after it ships and a concrete visibility gap appears
Ticker-only catalog refreshes or wholesale community imports — - Rejected because the only awesome-claude-code changes were generated popularity data, and importing external bundles would bypass selective skill audit
Policy-document edits — - Rejected by the ecosystem-update hard limit; no source finding justified explicit constitutional changes