feishu-gateway-debug

Trace the gateway log timeline in order: Received raw message, Flushing text batch, inbound message, response ready, and Sending response. The first missing entry marks the breakpoint, which you then correlate with agent.log and errors.log entries for the same session id.

A 404 in the agent log usually comes from a vLLM inference node, not the Feishu API. Check the base_url in the log line; if it points to an internal IP on port 8000, run curl against /v1/models on that node to confirm the model name exists, then fix the node or model name.

Feishu MEDIA attachments larger than 10MB fail silently with no error shown. Verify the file exists with ls -la, check its size, and compress oversized PDFs with Ghostscript using the /screen setting, which typically reduces a 33MB file to about 2.4MB before resending.

No. CLI sessions (prefixed cli:) and Feishu sessions (prefixed feishu:dm:) are fully isolated in the gateway. A CLI execute_code timeout, for example, will not block or interfere with any Feishu conversation, so cross-platform interference can be ruled out during diagnosis.

This error indicates the vLLM node's default 180 second timeout fired, closing the stream before the model finished responding. It is typically caused by high node load or overly long responses; check node load and concurrency, then resend or adjust timeout configuration.