fix(opencode): Fixes image reading with OpenAI-compatible providers like Kimi K2.5.

[zhming0]

What does this PR do?

Fixes #11304.
Fixes #11306.
Fixes #11550.

Problem: When using the Read tool to read an image file, Claude works fine but OpenAI-compatible providers (Kimi, etc.) fail to see the actual image content. The model either dumps base64 data as text or hallucinates the image content.

Root cause: OpenAI-compatible Chat Completions API only supports string content in tool results. When the Read tool returns an image as an attachment, it gets JSON.stringify()'d and the model can't interpret it as an image.

Solution: For providers using @ai-sdk/openai-compatible SDK that support image input:

• Extract media (images, PDFs) from tool results
• Inject them as a synthetic user message after the assistant message
• The model then sees the media as user-provided content, which is supported

How did you verify your code works?

1. Tested with Kimi K2.5 - asked it to read /tmp/test.png (a dashboard screenshot)
2. Before fix: Model dumped base64 or hallucinated "Hello World" text
3. After fix: Model correctly described the dashboard UI with "Explore", "Dashboards", "Monitors" etc.

[github-actions]

Hey! Your PR title Fixes image reading with OpenAI-compatible providers like Kimi K2.5. doesn't follow conventional commit format.

Please update it to start with one of:

• feat: or feat(scope): new feature
• fix: or fix(scope): bug fix
• docs: or docs(scope): documentation changes
• chore: or chore(scope): maintenance tasks
• refactor: or refactor(scope): code refactoring
• test: or test(scope): adding or updating tests

Where scope is the package name (e.g., app, desktop, opencode).

See CONTRIBUTING.md for details.

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

No duplicate PRs found

[rekram1-node]

Will look tmr, there is some granularity with other providers and packages, need to verify that this properly fixes all the providers

Some have special cases like good for example, u can send attachments as tool results for the gemini-3 series models but no others

[robinr2]

I have the same issue but I am using Opus 4.5 via Google/Antigravity and Copilot providers. It works with some other models of these two providers. Claude models can't read images using the Read tool.

Also tested it with this branch. Doesn't fix the issue for me.

[rekram1-node]

The copilot one makes sense why it'd have an issue, Ik how to fix w/ this branch, for antigravity im not really sure

[rekram1-node]

Okay I tested various providers, it should work fine now. Also added a text part too (some providers are more strict about that than others)

[rekram1-node]

/review