From be04c8beea9604a421d0cf8e16b281fe0cd4308c Mon Sep 17 00:00:00 2001
From: Fini <fini.yang@gmail.com>
Date: Wed, 29 Apr 2026 09:49:59 +0800
Subject: [PATCH] fix(ab-corpus): pin codex reasoning effort to medium
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

gpt-5.5 在 codex CLI 默认走 xhigh，单调用 1-2 分钟，把 ab-v5 全
量 sweep 从 ~15 分钟拉到预测 13 小时。pin 到 medium 跟 ab-v4
gpt-5.4 历史值同档，保留 AB_CORPUS_CODEX_REASONING env override
留给以后想测全火力上限。
---
 scripts/ab-corpus/clients/codex-cli.ts | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/scripts/ab-corpus/clients/codex-cli.ts b/scripts/ab-corpus/clients/codex-cli.ts
index 1812eec32..48c5a5c92 100644
--- a/scripts/ab-corpus/clients/codex-cli.ts
+++ b/scripts/ab-corpus/clients/codex-cli.ts
@@ -61,6 +61,13 @@ export async function callCodex(args: CallCodexArgs): Promise<ChatCallResult> {
   const prompt = `SYSTEM CONTEXT (treat as authoritative system prompt):\n<<<\n${args.system}\n>>>\n\nUSER REQUEST:\n${args.user}`;
   try {
     const content = await new Promise<string>((resolve, reject) => {
+      // Pin reasoning effort to `medium` so ab-corpus stays comparable
+      // across model bumps. Codex CLI's per-model default isn't fixed:
+      // gpt-5.4 historically ran ~medium; gpt-5.5 defaults to `xhigh`,
+      // which sent ab-v5 wall time from ~15 min (ab-v4) to a projected
+      // ~13 h. Override via `AB_CORPUS_CODEX_REASONING` if you ever
+      // need to measure full-effort ceiling.
+      const reasoningEffort = process.env.AB_CORPUS_CODEX_REASONING || 'medium';
       const proc = spawn(
         'codex',
         [
@@ -69,6 +76,8 @@ export async function callCodex(args: CallCodexArgs): Promise<ChatCallResult> {
           '--ephemeral',
           '--sandbox',
           'read-only',
+          '-c',
+          `model_reasoning_effort=${reasoningEffort}`,
           '-m',
           args.model,
           '--output-last-message',