AI SDK 6 warnings on every multi-turn Agent CLI run

[konard]

Summary

@link-assistant/agent 0.25.0 emits repeated AI SDK warnings during a normal multi-turn run against an OpenAI-compatible provider:

AI SDK Warning: System messages in the prompt or messages fields can be a security risk ... Set allowSystemInMessages to true ...
AI SDK Warning (opencode.chat / big-pickle): The feature "specificationVersion" is used in a compatibility mode. Using v2 specification compatibility mode.

This was found while auditing link-assistant/formal-ai#717. The successful Hive Mind self-coding E2E emits the system-message warning eight times and the compatibility warning four times.

Reproduction

bun add -g @link-assistant/agent@0.25.0
# Configure an @ai-sdk/openai-compatible provider and run a multi-turn tool loop.
agent --model formalai/formal-ai --permission-mode auto \
  --output-format stream-json --compact-json --disable-stdin \
  --prompt 'Create file self-coding-result.txt containing self-coding=passed'

Formal AI's exact harness is examples/self-coding/run.sh in link-assistant/formal-ai.

Root cause

js/src/session/prompt.ts calls AI SDK 6 streamText with system-role messages inside messages:

messages: [
  ...system.map((x): ModelMessage => ({ role: 'system', content: x })),
  ...safeModelMessages,
],

AI SDK 6.0.225's standardizePrompt warns unless the caller either uses the top-level system option or explicitly sets allowSystemInMessages. Since Agent intentionally builds these messages itself, it should use system where feasible, or set allowSystemInMessages: true after reviewing the injection boundary. The provider compatibility warning should likewise be resolved at the model/provider adapter boundary rather than globally silenced.

Expected

A successful standard Agent CLI run should not emit warnings. Please add a regression that captures stderr for a multi-turn OpenAI-compatible-provider run.