#13600 feat(providers): Empty assistant response after tool_calls completion (104×, Gemini flash variants)

OmniRoute Version: 3.8.50 Installation Method: npm (global) Operating System: Linux OS Version: Ubuntu / Linux 6.17.0-1020-oracle (arm64) Node.js Version: v24.16.0 Provider(s) Involved: OpenRouter, TokenRouter, BAI Model(s) Involved: google/gemini-3.5-flash-lite (70×), google/gemini-3.6-flash (22×), google/gemini-3-flash-preview (8×), qwen3.8-27b (3×), z-ai/glm-5.3-free (1×) Client Tool: Any client routed through OmniRoute

Description

After a model returns tool_calls and the client submits tool results, the streamed assistant response comes back empty: [STREAM] Empty assistant response after tool_calls completion — 104 warnings in ~41 h.

Concentration on Gemini-flash variants via OpenRouter/TokenRouter suggests either (a) OmniRoute mishandles the post-tool stream continuation for these models, or (b) the upstream legitimately returns empty content and OmniRoute should auto-retry/failover rather than passing an empty completion to the client.

Steps to Reproduce

Run an agentic client through OmniRoute to openrouter:google/gemini-3.5-flash-lite.
Let the model emit tool_calls, then submit results.
Observe [STREAM] Empty assistant response after tool_calls completion — the client receives an empty assistant turn.

Expected Behavior

Post-tool-call continuations return actual assistant content; if the upstream legitimately returns empty, OmniRoute should retry/failover to an alternate rather than forwarding an empty response.

Actual Behavior

Empty assistant responses reach the client (~104×/41h), degrading agentic flows — tools execute but the model never narrates/continues.

Test Impact

Needs a new integration test — tool_calls → tool_result → streamed continuation returns non-empty assistant content for the affected Gemini variants.

Error Logs / Output

[STREAM] Empty assistant response after tool_calls completion (openrouter:google/gemini-3.5-flash-lite) — sessionId=...   # ×70
[STREAM] Empty assistant response after tool_calls completion (openrouter:google/gemini-3.6-flash)   # ×22
[STREAM] Empty assistant response after tool_calls completion (tokenrouter:google/gemini-3-flash-preview)   # ×8

Possible Solutions

Detect empty post-tool continuations and retry once (or failover to the next route) before returning to the client.
Verify the tool-result → Gemini functionResponse conversion — a malformed functionResponse can make Gemini return empty turns.
Track empty-response rate per model; auto-demote models above a threshold.

Validation Plan

Replay a captured tool_calls→tool_result sequence against openrouter:google/gemini-3.5-flash-lite → non-empty continuation.

Additional Context

Related conversion issue filed separately for Bedrock toolResult pairing — likely the same code area.