[Bug] Gemini API fails with 400: Requests ending with a model turn are not supported

[herliansyah]

Affected area

Coding agent and CLI

What happened?

Summary

When using Gemini-based models (e.g., antigravity/gemini-3.7-flash-medium or Google Generative AI endpoints), requests fail repeatedly with
HTTP 400 INVALID_ARGUMENT: Requests ending with a model turn are not supported..

This error causes consecutive retry loops until maximum retry attempts are exhausted, blocking the agent session.

---

Environment & Version Info

• Prime Agent Version: 0.9.1
• Node.js Version: v22.22.3
• npm Version: 10.9.8
• Python Version: 3.11.15
• OS / Platform: Linux x86_64 (Linux-6.8.0-138-generic-x86_64-with-glibc2.39)
• Model in use: antigravity/gemini-3.7-flash-medium

---

Error Log / Output

Error: 400 [antigravity/gemini-3.7-flash-medium] [400]: {
   "error": {
     "code": 400,
     "message": "Requests ending with a model turn are not supported.",
     "status": "INVALID_ARGUMENT"
   }
}
(reset after 30s)

Error: 400 [antigravity/gemini-3.7-flash-medium] [400]: {
   "error": {
     "code": 400,
     "message": "Requests ending with a model turn are not sup (reset after 28s)

Error: 400 [antigravity/gemini-3.7-flash-medium] [400]: {
   "error": {
     "code": 400,
     "message": "Requests ending with a model turn are not sup (reset after 24s)

Error: 400 [antigravity/gemini-3.7-flash-medium] [400]: {
   "error": {
     "code": 400,
     "message": "Requests ending with a model turn are not sup (reset after 16s)

Error: Retry failed after 3 attempts: 400 [antigravity/gemini-3.7-flash-medium] [400]: {
   "error": {
     "code": 400,
     "message": "Requests ending with a model turn are not sup (reset after 16s)

### Steps to reproduce

1. Run prime-agent configured with a Gemini provider/model (antigravity/gemini-3.7-flash-medium).
2. Execute tasks that trigger intermediate agent turns or prompt/context updates without a preceding user message.
3. Observe the outgoing request ending with a role: "model" turn.
4. The Gemini API rejects the request with HTTP 400.

### Expected behavior

- The Gemini message adapter should validate and sanitize the contents array before dispatching:
  - If the last turn has role: "model", either strip the trailing turn or insert an appropriate continuation/user turn.
- Non-retryable client schema errors (INVALID_ARGUMENT / HTTP 400) should fail fast rather than triggering exponential retry timeouts.

### Prime Agent version

0.9.1

### Environment

Prime Agent Version: 0.9.1 Node.js Version: v22.22.3 npm Version: 10.9.8 Python Version: 3.11.15 OS / Platform: Linux x86_64 (Linux-6.8.0-138-generic-x86_64-with-glibc2.39) Model in use: antigravity/gemini-3.7-flash-medium

### Additional context

### Root Cause Analysis

1. Gemini Turn Structure Requirement:
Google's Generative Language API strictly enforces alternating turn sequences (user -> model -> user -> model). The API specification explicitly
requires that the last message in the contents array must be a user turn (or a function_response / tool output acting as user feedback), not a
model (assistant) turn.
2. Trigger in prime-agent:
During multi-step tool orchestration, prompt notes updates, or context formatting:
  - An assistant/model turn was appended to the conversation history.
  - A subsequent payload was dispatched where the trailing element in contents remained role: "model".
  - Google Gemini rejected this payload immediately with HTTP 400.
3. Ineffective Retries:
Because HTTP 400 INVALID_ARGUMENT is a deterministic request schema validation failure, retrying with exponential backoff (e.g. waiting 30s) will
always fail and stalls execution.