Trace Ultra inheritance through child tasks and failover

[iam-brain]

Question

What state does OpenCode and this plugin propagate into child tasks, retries, resumed sessions, and account failover, and what inheritance policy is required for Ultra?

Establish with concrete traces or focused experiments:

• selected model and model alias;
• logical Ultra versus wire max reasoning state;
• catalog-derived instructions and proactive delegation policy;
• agent/profile identity and x-openai-subagent classification;
• session IDs, prompt-cache keys, sticky/hybrid account affinity, and catalog scope;
• parent/child visibility in logs and snapshots;
• whether child tasks should inherit maximum reasoning but switch to explicit-request-only delegation to prevent recursive fan-out;
• behavior when a retry changes account or catalog scope mid-turn;
• behavior across compaction, resume, and fork-like flows available in OpenCode.

The resolution should document observed behavior and the minimum inheritance policy needed by the architecture decision; it should not implement that policy.