6 commits landed across Codex in the last 3 hours, all from jif, and they’re squarely about making the agent runtime more observable and more controllable when it fans out.
What Changed
MCP clients no longer hang on startup (#32781). Apply MCP startup timeouts during client creation puts a timeout on MCP client creation so a slow or dead MCP server can’t wedge the whole spawn.
Skill search shadow selection is on by default (#32780). Enable skill search shadow selection by default flips the shadow path live, so the newer skill-search ranking runs alongside the old one for comparison.
Shadow selection matches observable sources (#32768). Align shadow skill selection with observable sources makes the shadow ranking pull from the same telemetry the real path uses, so the A/B is actually apples-to-apples.
Lexical skill selection gets metrics (#32761). Add shadow metrics for lexical skill selection instruments the lexical (text-matching) skill picker so the team can see hit rates and regressions instead of guessing.
Spawned agents stay on the active backend (#32751). Restrict spawned-agent models to the active backend keeps sub-agents from drifting to a different model backend than the one you chose — no more surprise routing under multi-agent runs.
Multi-agent v2 exposes model overrides (#32749). Expose model overrides for multi-agent v2 spawns lets you pin models on the v2 spawn path, giving finer control over what each spawned agent runs.
Why It Matters
This is the kind of batch that makes Codex safer to run at scale: MCP timeouts kill a whole class of hang-on-boot failures, and the skill-search shadow work is disciplined experimentation — measure before you switch. The backend restriction and override exposure close the gap between “what you asked for” and “what the sub-agents actually used.” For the broader Codex-versus-CLI story these runtime changes feed into, our Codex terminal battle coverage compares where the agent stands against the field.
Smart developers shop around. Find the best AI tool deals at aiFiesta.