bedrock: GPT-6 Sol rejects images returned by read tool in v2.0.20

[kianpu-lila]

Summary

On OpenCode v2.0.20, reading a PNG with amazon-bedrock/global.openai.gpt-6-sol succeeds, but the following model request fails with an image-field error. The same PNG attached directly to a user message does not produce that error.

Environment

• opencode version: v2.0.20
• OS: macOS (Darwin 25.6.0, arm64)
• Terminal: TERM=xterm-256color, COLORTERM=truecolor; TERM_PROGRAM unset
• Shell: /bin/zsh
• Install/channel: v2.0.20 release
• Active plugins: None found in global configuration or the isolated reproduction directory

Reproduction

1. Create a valid PNG at /path/to/test.png.
2. Run:

   opencode run --standalone \
     --model amazon-bedrock/global.openai.gpt-6-sol \
     'Use the read tool to open /path/to/test.png and say what color it is. Do not use other tools.'

3. Observe that Read completes, then the model request fails.

Expected Behavior

Sol processes the image returned by read and answers the question.

Actual Behavior

Sanitized output (local PNG filename replaced with test.png):

> build · global.openai.gpt-6-sol
→ Read test.png
Error: This model doesn't support the image field for user messages. Remove image and try again.

The command exits with code 1. In a separate control run, supplying the same PNG with opencode run --file /path/to/test.png completed without this image-field error.

Additional Context

• Reproduced on v2.0.20 with a newly generated, non-confidential 16×16 PNG. No screenshot or image is attached.
• Related V1 issue: Bedrock GPT-6 Astra fails after the read tool returns an image #48069
• The V1 fix was merged in fix(opencode): hoist Bedrock tool images except Claude, Nova, and Llama 4 #50272.
• The v2.0.20 Bedrock Converse implementation still lowers file content from a tool result into toolResult.content: https://github.com/anomalyco/opencode/blob/v2.0.20/packages/ai/src/protocols/bedrock-converse.ts#L288-L319

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• #48069: The same Bedrock failure when the read tool returns an image, reported for GPT-6 Astra.
• #49443: A broader Bedrock Converse issue where tool-result images are rejected for non-Anthropic/Nova models.

[rekram1-node]

the api is a lil inconsistent in terms of which models support what, have a decent approximation fix for now but will make better abstractions soon.

Will release later today.

[alohaninja]

Thanks for the V2 fix. For the model-capability follow-up, my earlier Bedrock Converse tests are here: #49197 (comment). Pixtral, Mistral Large 3, and Gemma accepted user-level images but rejected tool-result images; Claude, Nova, and Llama 4 Maverick accepted tool-result images. The Llama 4 result was empirical even though the AWS documentation listed only Anthropic/Nova. Those tests predate this V2 fix; I have not retested the newly released implementation.