Auto review fails closed on always-thinking models (GLM 1210): the review request carries no reasoningEffort, so the catalog default dispatches thinking=disabled

[DMIAOCHEN]

环境

• dsh 0.2.0-rc.2（npx -y @deepseek-ai/dsh web；registry 的 latest 与 next 均为此版本，无可升级修复）
• @deepseek-ai/dsh-experimental-auto-review 0.2.0-rc.2（web profile，已启用，/permission auto）
• LLM 路由：pi-ai 目录 provider zai，model glm-5.3-flash（智谱"始终思考"模型，仅接受 low/high/max 思考级别）
• 主对话完全正常（会话请求携带显式思考级别）；仅 Auto review 的审查请求失败

现象

开启 Auto review 后，每次工具调用（原生与 PTC 内层）都失败：

auto-review: reviewer ended with error INVALID_REQUEST: 400
{"code":"1210","message":"This model always engages in thinking and cannot be disabled; please use low, high, or max"}

失败即 fail-closed：工具体不执行。Auto 模式在该模型路由下完全不可用。

根因链

1. packages/experimental/auto-review/src/index.ts 的 classifyRisk() 构造审查请求时，GenerateOptions 只包含 provider / model / system / messages / temperature: 0 / signal —— 没有携带任何 reasoning 级别（也没有沿用当前会话请求的思考级别）。
2. packages/llm/llm-pi-ai/src/adapter.ts 的解析回退是 options.reasoningEffort ?? profile.reasoning；provider profile 未配置 reasoning 字段时两者均为 undefined，reasoning 选项被整体省略。
3. pi-ai（@earendil-works/pi-ai@0.87.1）对带 reasoning 元数据的目录模型在无显式级别时套用自己的缺省级别，在线上载荷中变成了显式的 thinking 关闭参数。
4. glm-5.3-flash 这类"始终思考"的模型拒绝该参数（API 1210）→ reviewer 失败 → 工具调用失败。

注意这与 adapter 现有注释自述的语义相悖（profileOptions 对 off 的处理）："Omitting reasoning entirely ... a provider whose own default is to think would keep thinking" —— 省略 reasoning 本应让 provider 端用自己的默认（对 GLM 即保持思考），但 pi-ai 对目录 reasoning 模型的缺省分发实际发送了 disable。

主对话不受影响的原因：会话请求携带用户在模型选择中显式指定的思考级别，只有审查请求落到"缺省"路径。

期望行为 / 建议修复（按优先级）

1. reviewer 沿用会话当前的 reasoning effort（与主对话请求保持一致），而不是落到 provider profile 缺省或 pi-ai 内部缺省。审查请求本就是"同一模型多跑一次"，思考行为理应一致。
2. 或者：当请求未显式指定 reasoning 且模型目录声明了思考能力时，不要主动分发 disable —— 省略参数、由 provider 端默认行为接管（与上面引用的注释语义对齐）。
3. 至少：在 auto-review 的已知限制中标注该场景，并在选择 Auto 时对"始终思考"类模型给出能力提示。

临时绕过（供其他用户参考）

• 给该 provider 配置 reasoning: low（provider 级字段；审查请求的回退值会变成 low，API 明确接受 low/high/max；主对话的显式级别优先，不受影响）；
• 或改用手工声明的 provider（模型条目无 reasoning 元数据 → 审查请求完全省略思考参数 → provider 端"始终思考"默认生效）。

感谢！Auto review 本身的设计（审查请求不干预思考配置）是对的，问题出在缺省分发这一层，期待修复。

[DMIAOCHEN]

Workaround available for this one. I hit it while running Auto review against an always-thinking model (the review was dispatched with the catalog default thinking=disabled) and shipped a small plugin that covers this issue and #8764.

What it changes for this report: the review request now carries the session's reasoning level, plus a configurable fallback level that is capability-checked against the review route before it is sent. If the review route is pinned to a model that cannot take the session's level, the review retries once without the level instead of failing every call of that session (fail-closed is preserved: a review that still fails denies the call).

Also in the same plugin (the #8764 half): the request carries GenerateOptions.sessionId, so providers that require a per-session routing header (e.g. opencode-go) work.

Extra: a per-session reviewer-model picker — review with a different model than the one under review, stored per session, resettable to "follow the session model".

Install (dsh 0.2.0-rc.2, Node >= 22):

npx -y @deepseek-ai/dsh plugin --profile web add github:DMIAOCHEN/dsh-auto-review-plus#v0.1.1

then restart dsh web.

Disclosure — please read before installing:

• It re-registers the auto preset, and a shipped patch disables the official auto-review row while it is installed. Uninstalling restores the official one (live Auto sessions are migrated to Full access on uninstall).
• Verified end-to-end in a real session, including a pinned opencode-go review route. It is still one person's stopgap: it decides your approvals, so read the source (MIT, small) before trusting it.
• Derived from packages/experimental/auto-review and packages/client/ui-permission-presets; the upstream copyright line and the derived-file list are in NOTICE.
• It should be dropped as soon as this is fixed upstream — that is written into its README.

Repo + design notes: https://github.com/DMIAOCHEN/dsh-auto-review-plus