Repository navigation
z.ai/zhipuai finish_reason: "model_context_window_exceeded" is treated as unknown, causing loop instead of stopping #18813
Description
Description
Some providers appear to return a non-standard finish reason:
finish_reason: "model_context_window_exceeded"
Based on internet search, this was introduced on Anthropic side, and z.ai seems to follow the same behavior when context is too long.
In OpenCode, this finish reason is currently treated as unknown (instead of length), which can cause the session loop to continue rather than stop/compact, resulting in repeated/infinite loops sending request.
This is one of the repeated assistant message exported from session:
{
"message": {
"role": "assistant",
"time": {
"created": 1774204420869,
"completed": 1774204423742
},
"parentID": "msg_d16d2a7f7001GyKBIDcFE3dGOY",
"modelID": "glm-5-turbo",
"providerID": "zhipuai-coding-plan",
"mode": "normal",
"agent": "normal",
"path": {
"cwd": "/home/hmsy/workspace",
"root": "/home/hmsy/workspace"
},
"cost": 0,
"tokens": {
"input": 0,
"output": 0,
"reasoning": 0,
"cache": {
"read": 0,
"write": 0
}
},
"finish": "unknown",
"id": "msg_d16d30305001Nyw0BNeLsTjxCe",
"sessionID": "ses_2e9440126ffeOGeX4DKrphB4tT"
},
"parts": [
{
"type": "step-start",
"snapshot": "51df3df3638f2d50ef5530301f1f3889bcf12c29",
"id": "prt_d16d30e2c001TYpfLIMd5SSeof",
"sessionID": "ses_2e9440126ffeOGeX4DKrphB4tT",
"messageID": "msg_d16d30305001Nyw0BNeLsTjxCe"
},
{
"type": "step-finish",
"reason": "unknown",
"snapshot": "51df3df3638f2d50ef5530301f1f3889bcf12c29",
"cost": 0,
"tokens": {
"input": 0,
"output": 0,
"reasoning": 0,
"cache": {
"read": 0,
"write": 0
}
},
"id": "prt_d16d30e2d001V047uR4bXcaluN",
"sessionID": "ses_2e9440126ffeOGeX4DKrphB4tT",
"messageID": "msg_d16d30305001Nyw0BNeLsTjxCe"
}
]
}This is the request and response of z.ai
References:
OpenClaw already handled this case: openclaw/openclaw#35934
Link OpenCode issue: #15778
Expected behavior:
Treat model_context_window_exceeded as length (or equivalent overflow handling).
Stop current generation and trigger normal overflow/compaction flow, instead of continuing with unknown.
Plugins
no
OpenCode version
v1.3.0
Steps to reproduce
No response
Screenshot and/or share link
z.ai model context limit reached
Operating System
Linux
Terminal
Alacritty
Activity
github-actions commentedon Mar 23, 2026
This issue might be a duplicate of existing issues. Please check:
heimoshuiyu commentedon Mar 23, 2026
I think this is a problem of
vercel ai sdkor z.ai not follow the standard value.[-]Anthropic/z.ai finish_reason: "model_context_window_exceeded" is treated as unknown, causing loop instead of stopping[/-][+]z.ai/zhipuai finish_reason: "model_context_window_exceeded" is treated as unknown, causing loop instead of stopping[/+]heimoshuiyu commentedon Mar 24, 2026
The z.ai development team stated that
model_context_window_exceededmeans that the validation exceeded the allowed length before the model start processing, resulting in an error.lengthmeans that the output was incomplete. They believe their approach is correct.fix(provider): map model_context_window_exceeded to length finish reason
okuyam2y commentedon Mar 26, 2026
Fix PR: #19342 — removes
"unknown"from the non-terminal finish reason list so the prompt loop exits correctly.github-actions commentedon Jul 3, 2026
To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.