fix(ci): switch AI review to sensenova-6.8-flash-lite and treat curl timeout as non-fatal

[watney1024]

概述

修复 ai-code-review.yml AI Code Review workflow 在大 PR 上因模型超时导致 job 失败（exit code 28）的问题。

背景

PR #35 的 CI 在 "AI Code Review" step 失败：

Reviewing: .github/workflows/ci.yml
  [ERROR] HTTP 429 for .github/workflows/ci.yml  (insufficient_quota)
Reviewing: benchmarks/bench_const_merge.py
Error: Process completed with exit code 28.

根因（本地用真实 PR 数据完整复现确认）：

1. deepseek-v4-flash 在此 key 下处理大文件 diff（如 ci.yml ~14KB payload）时持续挂起，超过 curl 的 60s 超时。
2. 随后 --retry-all-errors 又把 HTTP 429（配额）当成可重试错误，重试 3 次，把一次超时拖成 180s+ 的挂起，最终 curl 返回 exit 28。
3. GitHub Actions 默认 bash -e -o pipefail 使任何非零退出立即中止整个 step → job 标红。

对照实验（同 key、同 PR #35 真实 diff、workflow 原样 curl 参数，仅换模型）：

模型	结果
deepseek-v4-flash	第一个文件即超时 exit 28，job 崩 ❌
glm-5.2	全 429 Workspace allocated quota exceeded（key 无该模型额度）❌
sensenova-6.8-flash-lite	10 文件全 200，稳定通过 ✅

改动内容

1. 换模型：deepseek-v4-flash → sensenova-6.8-flash-lite（唯一实测稳定通过的模型）。
2. 去掉 --retry-all-errors：HTTP 429（配额）不再被重试成长时间挂起。
3. 超时中性化：curl 失败用 || { echo "[WARN] ..."; continue; } 包裹——单文件超时被当作"本次无结果"跳过，job 仍正常成功，不再因 set -e 崩溃。
4. 收紧超时预算：--max-time 60 --retry 3 → --max-time 30 --retry 1，单个慢文件最多占用 ~60s（原 ~240s），确保 job 落在 timeout-minutes: 15 内。

验证

本地端到端：以 sensenova-6.8-flash-lite 跑完整 21 文件 PR #35 模拟，全部文件成功返回 review，无 exit 28 / 429 崩溃。bash -n 语法检查通过，YAML 校验通过。

备注

• 超时语义：单文件超时会打印 [WARN] curl timeout ... skipped (no result) 并跳过该文件，job 保持绿色成功。
• glm-5.2 已验证此 key 无额度（Workspace allocated quota exceeded），故未采用。

[github-actions]

🤖 AI Code Review

共审查 1 个变更文件

📁 .github/workflows/ai-code-review.yml

🔴 No blockers found.

🟡 HTTP error codes are silently ignored — curl exits 0 for 4xx/5xx responses; the || { … continue; } only catches transport-level failures (timeout, DNS, connection reset). A 401/429/500 will still produce a non-empty RESPONSE, and the script will proceed to parse it, likely failing later with invalid JSON or an empty BODY.
Suggestion: after parsing HTTP_CODE, add:

if [ "$HTTP_CODE" -lt 200 ] || [ "$HTTP_CODE" -ge 300 ]; then
  echo "  [WARN] API returned $HTTP_CODE for $FILE — skipped"
  continue
fi

🟡 Dropping retries increases flakiness — --retry 3 --retry-all-errors was replaced with --retry 0. Transient network blips now skip files entirely instead of retrying, which can silently reduce review coverage. The new || fallback is good, but it should be the safety net, not the primary mechanism.
Suggestion: keep --retry 3 --retry-all-errors alongside the || { continue; } guard. With --max-time 90, worst-case time becomes 270s, but successful reviews won't be lost to a single dropped packet.

💭 --retry 0 is redundant — it's equivalent to removing the flag entirely. Either delete it or set an actual retry count.

💭 Verify the new model name — sensenova-6.8-flash-lite must match the exact string exposed by the Sensenova API. A typo here will surface as a 400 error (which the missing HTTP-code check above would swallow). Double-check against the provider's model list.

---