Read tool cannot pass image data to vision-capable models

Bug Description

The Read tool cannot pass image data to vision-capable models (such as qwen3.5-plus). When attempting to analyze an image file, the tool chain fails to convert the image into a proper visual input format.

Steps to Reproduce

Use opencode v1.2.15
Reference an image file using @path/to/image.png
Ask the model to analyze the image

Expected Behavior

The image should be passed to the vision-capable model for analysis.

Actual Behavior

Read tool returns Image read successfully
System error: Cannot read image (this model does not support image input)
Model never receives the image data

Environment

opencode version: 1.2.15
Model: qwen3.5-plus (vision-capable)
OS: Linux
Provider: 阿里云百炼 (Bailian)

Additional Context

The issue appears to be in the tool chain - the Read tool can read image files but cannot convert them to the proper attachment format for vision models.

Even though qwen3.5-plus natively supports vision, the Read tool's implementation only returns a success status without actually passing the image data in a format the model can process.

Suggested Fix

The Read tool should detect image files and either:

Convert them to base64 and pass as an attachment
Use the MCP attachment protocol to pass image data directly
Return an error message indicating image files cannot be processed via Read tool

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

Z.AI GLM-4.7 cannot read images from Read tool (model supports vision) #13310: Z.AI GLM-4.7 cannot read images from Read tool (model supports vision) - same error message and same root cause around supportsMediaInToolResults
Read tool fails to process images with direct Anthropic provider (Claude Opus 4) on v1.2.15 #15566: Read tool fails to process images with direct Anthropic provider (Claude Opus 4) on v1.2.15 - same symptom on v1.2.15
Read tool accesses JPEG file but vision model hallucinates incorrect description (no actual image content processed) #11437: Read tool accesses JPEG file but vision model hallucinates incorrect description - related Read tool vision handling issue
Unable to read images for some models #5359: Unable to read images for some models - broader version of the same issue

If your issue is specifically about the @ai-sdk/openai-compatible provider not being included in the supportsMediaInToolResults check (which is what affects Alibaba Cloud Bailian / qwen models), #13310 covers the same root cause.

I'm not sure if this is a tooling issue or if this is the agent believing that the back-end LLM model is not vision-capable. Because, when supplying already base64 encoded image data, the agent refuses to send it to the model with the same reason.

I, unfortunately, been stuck with this particular issue too.

Got it working by adding

          "modalities": {
            "input": ["text", "image"],
            "output": ["text"]
          }

to the model inside opencode.json

fix(session): use image part type for images injected from tool results

@Clusters

@Clusters thank you, it works! For all JSON newbies (me included), please see if you need to add a comma in the line before modalities section. This is my full /home/username/.config/opencode/opencode.json (on Linux Mint):

{
  "$schema": "https://opencode.ai/config.json",
  "provider": {
    "lmstudio": {
      "npm": "@ai-sdk/openai-compatible",
      "name": "LM Studio (local)",
      "options": {
        "baseURL": "http://127.0.0.1:1234/v1"
      },
      "models": {
        "qwen_qwen3.5-35b-a3b@q8_0": {
          "name": "Qwen3.5-35B-A3B-Q8 (local)",
          "modalities": {
            "input": ["text", "image"],
            "output": ["text"]
          }
        },
        "qwen_qwen3.5-27b@q8_0": {
          "name": "Qwen3.5-27B-Q8 (local)",
          "modalities": {
            "input": ["text", "image"],
            "output": ["text"]
          }
        },
        "zai-org_glm-4.7-flash": {
          "name": "GLM-4.7-Flash-Q8 (local)"
        }
      }
    }
  }
}

Got similar issue. Somehow the model does not see the image via read tool, and will attempt to read them via playwright. This is Claude Opus 4.6.

After more testing, it seems like Claude is still able to read images, but other models tested (MiniMax, Kimi) cannot.

This is actually a severe problem for people whose workflow depends on visual understanding. The agent still attempts to read the image and there are returns coming back from the tool calls, but they don't really read it. If the agent happened to have seen the source code, they will start hallucinate the content of the images. This is disastrous for front end workflows.

roadmap: #220 filed — Image/vision input is structurally impossible a…

roadmap: ultraworkers#220 filed — Image/vision input is structurally …

please see if you need to add a comma in the line before modalities, add to this inpu and output should be in small case

roadmap: ultraworkers#220 filed — Image/vision input is structurally …

roadmap: #220 filed — Image/vision input is structurally impossible a…

You tried to replicate all the changes but still with the @ and even pasting the image didn't work! unfortunately I hope they leave this natively

roadmap: ultraworkers#220 filed — Image/vision input is structurally …

A work-around for this that I used was to setup a vision mcp. I then explicitly told the agent to use that MCP for any image reads. This has worked pretty well and I was able to handle my immediate needs.

A work-around for this that I used was to setup a vision mcp. I then explicitly told the agent to use that MCP for any image reads. This has worked pretty well and I was able to handle my immediate needs.

Which MCP?

A work-around for this that I used was to setup a vision mcp. I then explicitly told the agent to use that MCP for any image reads. This has worked pretty well and I was able to handle my immediate needs.

Which MCP?

I used this one and modified it to my model preference https://github.com/MikeyBeez/mcp-vision

To stay organized issues are automatically closed after 60 days of no activity. If the issue is still relevant please open a new one.

Navigation Menu

Repository navigation

Read tool cannot pass image data to vision-capable models #15728

Bug Description

Steps to Reproduce

Expected Behavior

Actual Behavior

Environment

Additional Context

Suggested Fix

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Read tool cannot pass image data to vision-capable models #15728

Description

Bug Description

Steps to Reproduce

Expected Behavior

Actual Behavior

Environment

Additional Context

Suggested Fix

Activity

github-actions commented on Mar 2, 2026

Clusters commented on Mar 2, 2026

Clusters commented on Mar 2, 2026

mtomas7 commented on Mar 10, 2026

sghng commented on Apr 1, 2026

sghng commented on Apr 1, 2026

subbu124 commented on Apr 28, 2026

srsaviorangell commented on Apr 28, 2026

kconnor1991 commented on May 16, 2026

Maxiviper117 commented on May 17, 2026

kconnor1991 commented on May 17, 2026

github-actions commented on Jul 17, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions