[charlie-ps]

A launch with effort "Auto" passes no --effort or model_reasoning_effort, so the agent picks its own level. The agents frame did not say which level that is, so an extension's model picker could not place Auto anywhere.

• Each model in the agents frame, and in the client api.agents(), has an optional autoEffort: the level the agent runs that model at when no effort is passed. It is left out when unknown.
• Claude: resolved in Claude Code's documented order using static reads only.
  • CLAUDE_CODE_EFFORT_LEVEL from a settings env block. Launches strip it from the shell environment, so a shell export does not count.
  • Managed settings: server-managed remote-settings.json, the MDM plist, then managed-settings.json with its managed-settings.d/ drop-ins. The first source with a policy key wins, unless it opts into merge.
  • User settings. A modelSettings entry beats the top-level effortLevel. The top-level key in user settings only reaches models older than Opus 5.5.
  • The model's documented default: medium for Opus, Sonnet and Haiku 5.5, xhigh for Opus 4.7, high otherwise.
  • The lowest maxEffortLevel cap from any source then applies.
• Codex: the top-level model_reasoning_effort in config.toml, then the model's default_reasoning_level from codex debug models. That level is now kept in the catalog and the bundled snapshot.
• The frame is re-sent when a source file changes (a stat poll every 15s; files are re-read only when their mtime or size moves) and on a catalog refresh.
• Host API 1.49.0 documents autoEffort.

Validation:

• New server/agents/auto-effort.test.js and a codex-catalog.test.js case. Full npm test: 4,659 passed. npm run docs:check passed.
• On an isolated dev board, the agents frame showed opus medium, fable high, and each Codex model at config.toml's level. That matches the effort recorded in 478 earlier Auto-effort Opus transcripts.

Made with Agent Wrangler and the PortSwigger/wrangler-squared extension