Ecosystem Update — 2026-07-13

TL;DR

• No candidate cleared the automatic-change gate, so config.toml, hooks.json, all agent TOMLs, and both AGENTS.md policy copies remain unchanged.
• Upstream Codex main added explicit per-spawn model/reasoning overrides with backend filtering, warning-gated Max/Ultra selection with resume persistence, and Guardian/session-config parity; all three require a stable owned-runtime smoke rather than a harness Quick Win. (curl)
• The ChatGPT app now bundles Codex 0.144.2, which contains the Guardian regression rollback, while PATH remains 0.144.1; the already-seen split-reconciliation item covers that drift, and 0.144.3 is documented as version-only. (curl)

Quick Wins

Auto-Implemented

• None. The missing or partial candidates require a stable runtime upgrade smoke, touch approval/reasoning behavior, or add a duplicate client/control plane. Dated backups, this report, dedup state, memory, planning evidence, and notification are workflow artifacts rather than harness Quick Wins.

Build Queue

• Multi-agent v2 per-spawn model override compatibility smoke (agent-pattern, Impact 2 / Effort 2 = 1.0) — override exposure, backend filtering — Upstream main exposes optional model and reasoning_effort only for explicit partial/context-free forks, then restricts choices to the active multi-agent backend. Local role TOMLs already pin Sol/Terra/Luna, but the current spawn_agent schema lacks these per-call fields and the app reports multi_agent_v2 as under development/disabled. Test profile precedence, fork restrictions, and backend filtering after a stable app-bundled release. (curl; local)
• Advanced reasoning explicit-selection and resume smoke (agent-pattern, Impact 3 / Effort 2 = 1.5) — https://github.com/openai/codex/commit/769a5de257e8b38204329263e3781c043f3f618d — Upstream main moves Max/Ultra behind a warning path, prevents shortcuts from silently escalating into them, and persists applied thread settings across resume. This matches the local policy that Ultra needs separate authorization; validate the picker, thread metadata, resume behavior, and explicit overrides once stable. (curl)
• Guardian session-configuration parity smoke (agent-pattern, Impact 3 / Effort 2 = 1.5) — https://github.com/openai/codex/commit/ea0fd84d948097761d0e8f14ae1e098ae0ed8eed — Upstream main now includes session permission instructions and the configured tool plan in Guardian reviews. Because this affects an approval boundary and the active default remains approvals_reviewer = "user", wait for a stable app release and test the inactive conservative auto-review profile rather than editing config now. (curl; OpenAI docs MCP)

Research

• Tier 2 was not fetched: its 2026-07-12T06:43:20-04:00 state timestamp was still within 24 hours at the 2026-07-13T06:31:05-04:00 crawl decision. No arXiv freshness claim or paper recommendation is made.

Already Have

Byte-identical global/workspace AGENTS.md, concise Codex-owned policy, power-user config.toml, prompt telemetry off, live web search, stable hooks, guarded command execution, startup/posture checks, PreCompact/Stop omni-mem capture, planner/reviewer/validator/explorer/worker roles with model and sandbox metadata, fixed role-level Sol/Terra/Luna routing, AgentOps/planning-gate/grounded implementation, artifact-backed verification, skill trigger reliability gates, skill audit and installer controls, OpenAI docs MCP, omni-mem MCP, plugin support with plugin hooks disabled, native memories disabled, the prior PATH/app split-reconciliation queue item, app-bundled Codex 0.144.2 Guardian rollback, prior ticker-only-refresh rejection, and mergeability-first engineering behavior already covered by the grounded workflow and reviewer posture. (local; curl; OpenAI docs MCP)

Rejected

• Claude Overlay — The new Awesome Claude Code entry is a Windows-only, always-on-top client that captures every monitor. It does not fit this macOS harness and duplicates installed Computer Use/browser surfaces. (curl)
• Codex 0.144.3 automatic upgrade — Official notes say it is a version-only release with no merged pull-request changes after 0.144.2; no harness benefit justifies an automatic binary change. (curl)
• Megapowers marketplace/orchestration suite — Its trigger evals, artifact-backed done checks, irreversible-action gate, role routing, and multi-model orchestration substantially overlap memory-adaptation, planning-gate, AgentOps, and route_manifest.json; wholesale installation would duplicate the owned control plane. (curl)
• Awesome Claude Code ticker churn as a new candidate — Eight of nine post-run commits again changed only generated ticker/SVG data. This is a semantic duplicate of the July 12 rejection and was not requeued. (curl)
• Wholesale adoption of the remaining GitHub harness/catalog hits — The search returned metadata-level discovery leads, but no additional fetched source established a discrete, non-duplicative Quick Win. Discovery snippets were not promoted to facts. [snippet]
• Website render or deployment — Explicitly excluded by the user; the wrapper owns both after this run.

Sources checked: https://github.com/hesreallyhim/awesome-claude-code, https://howborisusesclaudecode.com/, https://github.com/shanraisshan/codex-cli-best-practice, GitHub repository search for Codex hooks/agents/skills, https://github.com/lawzava/megapowers, https://github.com/hudrazine/velkross, https://github.com/openai/codex/commits, https://github.com/openai/codex/releases/tag/rust-v0.144.2, https://github.com/openai/codex/releases/tag/rust-v0.144.3, and official Codex subagent/config documentation through the OpenAI developer docs MCP. Tier 2 fetched: no — skipped at the crawl decision because tier2_last_run was still within 24 hours. Tier 3 fetched: no — full weekly crawl skipped because tier3_last_run is 2026-07-08T06:30:35-04:00; official releases/main commits were a narrow compatibility supplement. Run at: 2026-07-13T06:42:26-04:00