Bug reports: Claude Code `system`-role message in `messages[]` causes HTTP 422 on GLM

[chgwan]

System Info / 系統信息

Notes

• Affected models observed: glm-4.6, glm-5.1.
• Client: Claude Code 2.1.156.
• Current client-side workaround: a local proxy that moves system-role
  messages into the top-level system field before the request reaches GLM.

Your Anthropic-compatible endpoint (/api/anthropic/v1/messages) returns
HTTP 422 when the messages array contains an entry with role: "system":

422 {"detail":[{"type":"literal_error","loc":["body","messages",1,"role"],
 "msg":"Input should be 'user' or 'assistant'","input":"system",
 "ctx":{"expected":"'user' or 'assistant'"}}]}

The official Anthropic Messages API accepts such requests (it effectively folds
system-role message content into system handling rather than rejecting the
request). Tools built against the Anthropic API — notably Claude Code, which
injects hook/session context as a system-role message in messages[] — send
this shape and therefore fail against GLM while succeeding against Anthropic.

Who can help? / 谁可以帮助到您？

No response

Information / 问题信息

• The official example scripts / 官方的示例脚本
• My own modified scripts / 我自己修改的脚本和任务

Reproduction / 复现过程

A) Accepted today → HTTP 200

curl -s -w "\nHTTP %{http_code}\n" \
  https://open.bigmodel.cn/api/anthropic/v1/messages \
  -H "Authorization: Bearer <TOKEN>" \
  -H "anthropic-version: 2023-06-01" -H "content-type: application/json" \
  -d '{"model":"glm-4.6","max_tokens":20,"system":"You are helpful.",
       "messages":[{"role":"user","content":"Hello"}]}'

B) Rejected today → HTTP 422 (only difference: a system entry in messages)

curl -s -w "\nHTTP %{http_code}\n" \
  https://open.bigmodel.cn/api/anthropic/v1/messages \
  -H "Authorization: Bearer <TOKEN>" \
  -H "anthropic-version: 2023-06-01" -H "content-type: application/json" \
  -d '{"model":"glm-4.6","max_tokens":20,
       "messages":[{"role":"user","content":"ctx"},
                   {"role":"system","content":"injected"},
                   {"role":"user","content":"Hello"}]}'

C plugin test in Claude Code

1. Enable any plugin with a SessionStart hook that emits additionalContext
   (e.g. the superpowers plugin, which injects its using-superpowers skill text).
2. Point Claude Code at an Anthropic-compatible endpoint that strictly validates
   message roles (GLM/Zhipu):
   • ANTHROPIC_BASE_URL=https://open.bigmodel.cn/api/anthropic
   • ANTHROPIC_AUTH_TOKEN=<token>, model glm-4.6 / glm-5.1
3. Start an interactive session and send any message ("Hello").
4. → 422 on the first turn, every turn.

Expected behavior / 期待表现

Requested fix (either is sufficient)

1. Tolerate system-role entries in messages[]: fold their content into
   the system prompt instead of returning 422 (match Anthropic's behavior); or
2. Document the strict behavior and, ideally, normalize on the server side so
   Anthropic-targeted clients work unmodified.

[JaredforReal]

This is expected behavior to mitigate prompt injection attacks.
So we currently don't plan to change this validation behavior, but we can improve the documentation/error message to make the design clearer.
@chgwan Thanks for your report