Support per-agent context-window overrides in agent roles

What variant of Codex are you using?

Codex multi-agent/subagent configuration, observed with current Codex behavior and checked against the current open-source implementation.

What feature would you like to see?

Please allow an agent role to set its own model_context_window and model_auto_compact_token_limit instead of always inheriting those values from the parent session.

Example use case:

• Primary Astra orchestrator: native/default context window
• Luna worker: 1M requested context window

A role file can currently contain the Luna model, reasoning effort, and context settings. The model and reasoning effort are applied to the spawned child, but the context-window settings remain parent-derived.

Current implementation observation

In codex-rs/core/src/agent/role.rs, the typed role override layer applies fields such as model and reasoning effort, but does not include the context-window or auto-compaction settings. As a result, those values from the role config do not become effective child configuration.

Expected behavior

Agent roles should be able to request their own context window and auto-compaction threshold, subject to the selected model's normal catalog/runtime maximums.

This would allow context allocation to follow task shape and model cost. For example, an orchestrator can stay on its normal context window while a lower-cost implementation or debugging worker uses a larger window.

Related issues

• #40429 requests per-thread context profiles; this request specifically concerns spawned agent-role configuration.
• #43229 discusses context behavior in mixed-model worker trees; this request asks for a direct role-level setting.

Start in codex-rs/core/src/agent/role.rs, where typed role overrides currently apply model and reasoning effort but omit context-window and auto-compaction settings. Trace how role configuration becomes child configuration and how selected-model catalog/runtime maximums are enforced. Done means a role's model_context_window and model_auto_compact_token_limit affect its spawned child without exceeding those limits.