gemini3-tools

Gemini 3 Flash/Pro Preview — function calling broken via nexos API

Problem

Gemini 3 models (Flash Preview, Pro Preview) require thought_signature on every functionCall content block in multi-turn conversations. This is a Vertex AI requirement for all Gemini 3 models, even with thinking disabled.

The official way to pass thought_signature in the OpenAI-compatible API is via extra_content.google.thought_signature on tool_calls entries. However, nexos API strips extra_content (and any unknown fields) from tool_calls before forwarding to Vertex AI.

This means function calling with Gemini 3 models is broken — the first request works, but the follow-up request (containing the tool result) always fails with:

error: Unable to submit request because function call `<name>` in the 2. content block
is missing a `thought_signature`.

Gemini 2.5 Pro is not affected (no thought_signature requirement).

Reproduce

Step 1 — Get a tool call response (works)

curl -s -X POST "https://api.nexos.ai/v1/chat/completions" \
  -H "Authorization: Bearer $NEXOS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Gemini 3 Flash Preview",
    "messages": [
      {"role": "user", "content": "Run ls -la in the current directory"}
    ],
    "tools": [{"type":"function","function":{"name":"bash","description":"Run a shell command","parameters":{"type":"object","properties":{"command":{"type":"string"}},"required":["command"]}}}]
  }'

Returns a tool call with id — note it down for step 2.

Step 2 — Send tool result back (fails)

curl -s -X POST "https://api.nexos.ai/v1/chat/completions" \
  -H "Authorization: Bearer $NEXOS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Gemini 3 Flash Preview",
    "messages": [
      {"role": "user", "content": "Run ls -la in the current directory"},
      {"role": "assistant", "content": null, "tool_calls": [{"id":"TOOL_CALL_ID_FROM_STEP1","type":"function","function":{"name":"bash","arguments":"{\"command\":\"ls -la\"}"}}]},
      {"role": "tool", "tool_call_id": "TOOL_CALL_ID_FROM_STEP1", "content": "total 40\ndrwxrwxr-x 3 user user 4096 Feb 8 12:16 .\n-rw-rw-r-- 1 user user 3666 Feb 8 12:14 index.mjs\n-rw-rw-r-- 1 user user 163 Feb 8 11:35 package.json"}
    ],
    "tools": [{"type":"function","function":{"name":"bash","description":"Run a shell command","parameters":{"type":"object","properties":{"command":{"type":"string"}},"required":["command"]}}}]
  }'

Returns 400:

{"error":{"message":"error: Unable to submit request because function call `bash` in the 2. content block is missing a `thought_signature`.","type":"INVALID_ARGUMENT","param":null,"code":"400"},"provider":"vertex-ai"}

Step 3 — Adding extra_content does not help (nexos strips it)

curl -s -X POST "https://api.nexos.ai/v1/chat/completions" \
  -H "Authorization: Bearer $NEXOS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Gemini 3 Flash Preview",
    "messages": [
      {"role": "user", "content": "Run ls -la in the current directory"},
      {"role": "assistant", "content": null, "tool_calls": [{"id":"TOOL_CALL_ID_FROM_STEP1","type":"function","function":{"name":"bash","arguments":"{\"command\":\"ls -la\"}"},"extra_content":{"google":{"thought_signature":"skip_thought_signature_validator"}}}]},
      {"role": "tool", "tool_call_id": "TOOL_CALL_ID_FROM_STEP1", "content": "total 40\ndrwxrwxr-x 3 user user 4096 Feb 8 12:16 .\n-rw-rw-r-- 1 user user 3666 Feb 8 12:14 index.mjs\n-rw-rw-r-- 1 user user 163 Feb 8 11:35 package.json"}
    ],
    "tools": [{"type":"function","function":{"name":"bash","description":"Run a shell command","parameters":{"type":"object","properties":{"command":{"type":"string"}},"required":["command"]}}}]
  }'

Same 400 error — extra_content is stripped by nexos before reaching Vertex AI.

Tested placements for thought_signature (all fail)

Nexos strips all non-standard fields from tool_calls entries (only id, type, function survive). The following placements were all tested and failed:

Placement	Result
tool_calls[].thought_signature	Stripped by nexos
tool_calls[].extra_content.google.thought_signature	Stripped by nexos
tool_calls[].extra_content.thought_signature	Stripped by nexos
tool_calls[].google.thought_signature	Stripped by nexos
tool_calls[].metadata.thought_signature	Stripped by nexos
message.thought_signature	Stripped by nexos
tool_calls[].function.thought_signature	Passed to Vertex but rejected — thought_signature is not a valid field inside function_call
tool_calls[].function.google.thought_signature	Stripped by nexos
tool_calls[].function.extra_content.google.thought_signature	Stripped by nexos
tool_calls[].function.metadata.thought_signature	Stripped by nexos
tool_calls[].function.parts[].thought_signature	Stripped by nexos
tool_calls[].function.extensions.thought_signature	Stripped by nexos
tool_calls[].function.vendor_extensions.thought_signature	Stripped by nexos
Inside arguments JSON string	Stripped by nexos

In the native Vertex AI format, thought_signature is a sibling of functionCall at parts[] level. Nexos converts OpenAI format to native format but does not map any client field to parts[].thought_signature.

Provider workaround

The provider now rewrites tool call history for Gemini 3/3.1 models (rewriteToolCallHistory in fix-gemini.mjs). Instead of sending the standard assistant(tool_calls) → tool(result) format, it converts:

• assistant messages with tool_calls → assistant with text description of the calls
• tool result messages → user messages with the result text

This avoids the thought_signature requirement entirely. First tool call works natively (no history to rewrite). Multi-turn tool use works via the rewritten format.

Upstream fix (still needed)

Nexos API should propagate extra_content.google.thought_signature on tool_calls entries to Vertex AI, as documented in: https://docs.cloud.google.com/vertex-ai/generative-ai/docs/thought-signatures

Alternatively, nexos could inject skip_thought_signature_validator server-side for Gemini 3 models when the client does not provide a thought_signature.

Test script

Run test-tool-call.sh to test multi-turn tool calling across Gemini models:

./test-tool-call.sh

Output shows which models pass/fail each step:

=== Multi-turn tool calling test ===

   MODEL                          STEP1      STEP2 NOTES
   ------------------------- ---------- ---------- ------------------------------
✅ Gemini 2.5 Pro                    OK         OK 
✅ Gemini 2.5 Flash                  OK         OK 
⚠️  Gemini 3 Flash Preview            OK      ERROR error: Unable to submit request because function call `bash`
⚠️  Gemini 3 Pro Preview              OK      ERROR error: Unable to submit request because function call `bash`

Reference

• https://docs.cloud.google.com/vertex-ai/generative-ai/docs/thought-signatures
• Affected models: Gemini 3 Flash Preview, Gemini 3 Pro Preview
• Not affected: Gemini 2.5 Pro, Gemini 2.5 Flash