openai/codex merged commit 81b000421d on June 19, adding config switches for orchestrator-owned skills and Codex Apps MCP so hosts can disable those surfaces without turning off local skills or regular MCP servers.
Facts
- Codex added
[orchestrator.skills].enabledand[orchestrator.mcp].enabled, plus tests that hidden orchestrator skills and Codex Apps MCP tools no longer reach model turns. NousResearch/hermes-agentadded a largehtml-artifactskill bundle in commit 9362ce257, then reverted it the same day in 92451151c.earendil-works/picommit 0d89a333 added selectivepi-aiandpi-agentbase entrypoints so builders can import narrower package surfaces.
Evidence
The receipts are Codex's config schema and MCP exposure tests, Hermes Agent's add-and-revert skill commits, and Pi's package export plus new base.ts entrypoint files.
Context
For agent builders and workspace operators, the practical question is shifting from "can the agent use tools?" to "which capability surfaces are loaded, visible, and paid for before the task starts?"
Limits
These are independent patches, not a shared standard. Hermes's revert is also negative evidence: capability expansion can move too fast for a repo's own release and documentation surface.