Add AGENTD_ZARVIS_REASONING_EFFORT knob for codex-oauth

[edwin-zvs]

What

Adds an optional AGENTD_ZARVIS_REASONING_EFFORT env var (low | medium | high) that sets reasoning.effort on the codex-oauth responses body, mirroring Codex CLI's model_reasoning_effort.

let mut reasoning = json!({ "summary": "auto" });
if let Ok(effort) = std::env::var("AGENTD_ZARVIS_REASONING_EFFORT") {
    let effort = effort.trim();
    if !effort.is_empty() {
        reasoning["effort"] = json!(effort);
    }
}

Why

The codex-oauth provider previously sent reasoning: { summary: "auto" } with no effort field, so every request ran at the backend's default reasoning effort with no way to change it. The Codex CLI exposes model_reasoning_effort; this gives zarvis the same lever.

Motivation was benchmark parity: on a DeepSWE task (etree-xml-diff-patch, gpt-5.5 via the codex-oauth backend), zarvis at the default effort scored 56/57; with AGENTD_ZARVIS_REASONING_EFFORT=high it scored 57/57 — matching the Codex CLI harness (which defaults to high).

Compatibility

Unset (the default) preserves the exact previous body (reasoning: { summary: "auto" }), so existing behavior is unchanged.

Test plan

• cargo build -p agentd-adapter-zarvis (clean).
• Existing build_responses_body unit tests pass unchanged (env unset → identical body).
• Ran end-to-end via the DeepSWE/Pier harness with AGENTD_ZARVIS_REASONING_EFFORT=high; observed deeper reasoning and 57/57 on the task above.

🤖 Generated with Claude Code