Fix AI SDK multi-turn warnings

[konard]

Summary

Fixes #279.

• route Agent-owned system prompts through AI SDK 6's dedicated system option
  in normal generation, compaction, and title generation
• upgrade @ai-sdk/openai-compatible to the Language Model v3 adapter and pin
  AI SDK to the release that reports unsafe system-role placement
• add a credential-free, two-turn OpenAI-compatible regression that captures
  AI SDK's console.warn/stderr warning path
• add a patch changeset

Reproduction and root cause

The regression performs two streamText turns against a local mocked
OpenAI-compatible SSE response.

Before the adapter upgrade, it captured the compatibility warning twice:

The feature "specificationVersion" is used in a compatibility mode. Using v2 specification compatibility mode.

With the v3 adapter but system-role entries still inside messages, it captured
the system-message warning twice:

System messages in the prompt or messages fields can be a security risk...

The two root causes were therefore independent:

1. the direct OpenAI-compatible dependency was still v1 / Language Model v2;
2. trusted prompts constructed by Agent were passed as ordinary message history.

The fix uses AI SDK's dedicated system input at every Agent generation
boundary and updates the adapter to its AI SDK 6-compatible v3 implementation.
Warnings remain enabled globally, so future SDK warnings are still visible.

Verification

• bun test ./tests/ai-sdk-warnings.ts — 1 pass
• npm test — 610 pass, 4 todo, 0 fail
• npm run check — ESLint, Prettier, and file-size checks pass
• cargo test — all Rust unit, integration, and doc tests pass
• git diff --check — pass

[konard]

Working session summary

Implemented and finalized PR #284.

Key changes:

• Moved trusted system prompts to AI SDK’s dedicated system option.
• Upgraded the OpenAI-compatible adapter to Language Model v3.
• Added a credential-free two-turn warning regression test.
• Added a patch changeset.
• Merged the latest main and resolved its .gitkeep deletion conflict.

Verification:

• JavaScript: 610 passed, 0 failed.
• Rust tests: all passed.
• Lint, formatting, file-size, and package-install checks passed.
• Fresh CI run passed on all platforms.
• Working tree is clean, PR is mergeable, and it is ready for review.

---

This summary was automatically extracted from the AI working session output.

[konard]

🤖 Solution Draft Log

This log file contains the complete execution trace of the AI solution draft process.

💰 Cost estimation:

• Model: GPT-5.6 Sol
• Provider: OpenAI
• Public pricing estimate: $6.981880

📊 Context and tokens usage:

• 150.2K / 200K (75%) input tokens, 12.8K / 128K (10%) output tokens

Total: (150.2K + 4.9M cached) input tokens, 12.8K output tokens, $6.981880 cost

🤖 Models used:

• Tool: OpenAI Codex
• Requested: gpt-5.6-sol
• Thinking level: low (~7999 tokens)
• Model: GPT 5.6 Sol (gpt-5.6-sol)

📎 Log file uploaded as Gist (1944KB)

• View complete solution draft log

---

Now working session is ended, feel free to review and add any feedback on the solution draft.

[konard]

🎉 Auto-merged

This pull request has been automatically merged by hive-mind.

• All CI checks have passed

---

Auto-merged by hive-mind with --auto-merge flag