Codex model + reasoning policy (Dyads)

All dyads should run the latest Codex model:

Reasoning level is controlled via:

These values are applied in two places:

Default Codex config: si-codex-init writes ~/.codex/config.toml with model and model_reasoning_effort.
Critic-driven execution: when a critic runs interactive codex/codex resume inside an actor, it explicitly passes -m $CODEX_MODEL and -c model_reasoning_effort=$CODEX_REASONING_EFFORT.

Complexity-based overrides

When a dyad task includes complexity=low|medium|high, critics map complexity to model/effort:

Model: CODEX_MODEL_LOW|MEDIUM|HIGH if set, otherwise CODEX_MODEL.
Reasoning: CODEX_REASONING_EFFORT_LOW|MEDIUM|HIGH if set, otherwise the complexity level itself (low|medium|high), with low clamped to medium.

If complexity is empty, critics fall back to priority and then to the base env defaults.

Recommended defaults (can be overridden per container via env):

infra dyad
- actor: xhigh
- critic (driver): xhigh
web dyad
- actor: medium
- critic (driver): high
research dyad
- actor: high
- critic (driver): high
pm dyad
- actor: high (planning)
- critic (driver): xhigh (program-level reasoning)
default dyad
- actor: medium
- critic (driver): medium

si dyad spawn sets:

CODEX_MODEL (defaults to gpt-5.2-codex)
CODEX_REASONING_EFFORT for actor/critic based on ROLE
CODEX_PER_DYAD=1 by default so Codex state (~/.codex) is not shared across dyads. Override to 0 only if you explicitly want a shared store.

Overrides:

si dyad spawn sets CODEX_MODEL and CODEX_REASONING_EFFORT per container (actor/critic).