bug: intermittent Empty completion with BigModel glm-5-turbo on long Sage runs

[chindris-mihai-alexandru]

Summary

I hit repeated retries with:

Retryable(Empty completion received - no content, tool calls, or valid finish reason)

when using Forge against BigModel (open.bigmodel.cn) with glm-5-turbo during a long Sage session.

This seems to happen after a successful upstream connection/response event, but before Forge can build a non-empty final completion.

Environment

• Forge: local build (0.1.0-dev)
• OS: macOS
• Provider: big_model
• Model: glm-5-turbo
• Endpoint: https://open.bigmodel.cn/api/paas/v4/chat/completions

What happened

During a deep analysis run, previous turns worked, then at message_count: 82 Forge started retrying the same turn and failed 8 times with the empty completion error.

Relevant log excerpt (sanitized):

{"timestamp":" 297.233908375s","level":"INFO","fields":{"message":"Connecting Upstream","url":"https://open.bigmodel.cn/api/paas/v4/chat/completions","model":"glm-5-turbo","message_count":"82","message_cache_count":"81"}}
{"timestamp":" 300.078292042s","level":"DEBUG","fields":{"message":"Received completion from Upstream"}}
{"timestamp":" 300.078443209s","level":"ERROR","fields":{"message":"Retry attempt due to error","error":"Retryable(Empty completion received - no content, tool calls, or valid finish reason)","model":"glm-5-turbo"}}

Then the same pattern repeated multiple times for the same turn.

Raw SSE capture / replay

I captured/replayed the exact request payload shape (82 messages) via direct curl to BigModel and got valid SSE endings in repeated runs:

• some runs ended with finish_reason: "stop" + [DONE]
• some runs ended with finish_reason: "tool_calls" + [DONE]
• no raw replay run reproduced an empty terminal completion

Example tail (stop):

data: {"choices":[{"index":0,"finish_reason":"stop","delta":{"role":"assistant","content":""}}],"usage":{...}}
data: [DONE]

Example tail (tool_calls):

data: {"choices":[{"index":0,"finish_reason":"tool_calls","delta":{"role":"assistant","content":""}}],"usage":{...}}
data: [DONE]

Why I think this may be Forge-side (or at least needs better diagnostics)

The error is thrown when the aggregated stream result has:

• empty content
• no tool calls
• no finish reason
• no thought signature

Given the upstream was connected and a completion event was observed, this is hard to debug without raw response chunk logging at failure time.

Web/docs checks

• No obvious duplicate issue found in this repo for the exact error text.
• Z.ai docs note that abnormal SSE termination should indicate reason via finish_reason.

Request

Could maintainers please:

1. Investigate this edge case in OpenAI-compatible stream aggregation for BigModel/GLM-5, and
2. Add optional debug logging/dump of terminal parsed chunk state when EmptyCompletion is raised (to distinguish provider empty stream vs local parse/aggregation gap).

Artifacts

Sanitized artifacts are in this gist:

https://gist.github.com/chindris-mihai-alexandru/43f711cbb45e214d3307025d25a28425

(contains core Forge log excerpt, SSE tails, replay summary, and duplicate-search notes)

[tusharmath]

HI @chindris-mihai-alexandru
Thanks for the detailed report - Can you try this on your end to help us debug this further:

1. In .forge.toml add a

debug_requests = "request.jsonl"

2. Give forge a task and wait for this error to occur.
3. All the requests will be logged in the request.jsonl file.
4. Pick the last request and try with a tool like postman or equivalent.
5. If the response is still valid and content is non-empty, then its definitely a ForgeCode issue.