[BUG] - Multiple Image Reads causing Api errors.

[wala0003]

Problem (one or two sentences)

Tool Results in wrong order, and image read content not nested in the content array key,

Date/time: 2026-02-23T18:47:53.595Z
Extension version: 3.50.3
Provider: anthropic
Model: claude-sonnet-4-6

400 {"type":"error","error":{"type":"invalid_request_error","message":"messages.20: tool_use ids were found without tool_result blocks immediately after: toolu_012LvAhvmtDL3fCaukBiqKG2. Each tool_use block must have a corresponding tool_result block in the next message."},"request_id":"req_0"}

I think anthropic expects tool_result array entries grouped at the beginning of the user -> content array.

The fix should be in how pushToolResult handles array results. When result is a content block array (text + image), the entire array needs to go into the tool_result's content field:
json{
"type": "tool_result",
"tool_use_id": "toolu_xxx",
"content": [
{ "type": "text", "text": "File: tests/screenshots/seed-smoke-dashboard.png\nNote: Image file (15 KB)" },
{ "type": "image", "source": { "type": "base64", ... } }
]
}

log:
"role": "assistant",
"content": [
{
"type": "reasoning",
"text": "Good, I can see the login screen screenshot. Let me now view the dashboard and emails screenshots.",
"summary": []
},
{
"type": "tool_use",
"id": "toolu_01KVMKVoYinYuaZBShijKw2h",
"name": "read_file",
"input": {
"path": "tests/screenshots/seed-smoke-dashboard.png"
}
},
{
"type": "tool_use",
"id": "toolu_012LvAhvmtDL3fCaukBiqKG2",
"name": "read_file",
"input": {
"path": "tests/screenshots/seed-smoke-emails.png"
}
}
],
"ts": 1771713465053
},
{
"role": "user",
"content": [
{
"type": "tool_result",
"tool_use_id": "toolu_01KVMKVoYinYuaZBShijKw2h",
"content": "File: tests/screenshots/seed-smoke-dashboard.png\nNote: Image file (15 KB)"
},
{
"type": "image",
"source": {
"type": "base64",
"media_type": "image/png",
"data": "iVBORw0"
}
},
{
"type": "tool_result",
"tool_use_id": "toolu_012LvAhvmtDL3fCaukBiqKG2",
"content": "File: tests/screenshots/seed-smoke-emails.png\nNote: Image file (359 KB)"
},
{
"type": "image",
"source": {
"type": "base64",
"media_type": "image/png",
"data": "iVBORw0KGgoAAAA",
}
},

Context (who is affected and when)

multiple read file tool calls of images in parallel.

Reproduction steps

did a new session with the prompt: "im testing something - was getting api errors when in a previous task i asked you to read multiple images at once. can you retry that - find the screenshots and attempt to read 3 images in one read file tool call"

This immediately failed with the same error

Expected result

reading of two or more image files at once works.

Actual result

400 error from API

Variations tried (optional)

No response

App Version

3.50.3

API Provider (optional)

Anthropic

Model Used (optional)

Sonnet 4.6

Roo Code Task Links (optional)

No response

Relevant logs or errors (optional)

[
  {
    "role": "user",
    "content": [
      {
        "type": "text",
        "text": "<user_message>\nim testing something - was getting api errors when in a previous task i asked you to read multiple images at once. can you retry that - find the screenshots and attemtp to read 3 images in one read file tool call\n</user_message>"
      },
      {
        "type": "text",
        "text": "<environment_details>\n# VSCode Visible Files"
      }
    ],
    "ts": 1771873240271
  },
  {
    "role": "assistant",
    "content": [
      {
        "type": "reasoning",
        "text": "The user wants me to find screenshots in the workspace and attempt to read 3 images in one read_file tool call (as parallel calls). Let me first find where the screenshots are stored.",
        "summary": []
      },
      {
        "type": "text",
        "text": "\nLet me find the screenshots first."
      },
      {
        "type": "tool_use",
        "id": "toolu_01G6SxhtwAPtbwDLVfzhc9ze",
        "name": "execute_command",
        "input": {
          "command": "find /home/administrator/pulse/Pulse-DSL -name \"*.png\" -o -name \"*.jpg\" -o -name \"*.jpeg\" 2>/dev/null | head -20",
          "cwd": "/home/administrator/pulse/Pulse-DSL",
          "timeout": 10
        }
      }
    ],
    "ts": 1771873244680
  },
  {
    "role": "user",
    "content": [
      {
        "type": "tool_result",
        "tool_use_id": "toolu_01G6SxhtwAPtbwDLVfzhc9ze",
        "content": "Command executed in terminal within working directory '/home/administrator/pulse/Pulse-DSL'. Exit code: 0\nOutput:\n/home/administrator/pulse/Pulse-DSL/tests/screenshots/seed-smoke-dashboard.png\r\n/home/administrator/pulse/Pulse-DSL/tests/screenshots/seed-smoke-home.png\r\n/home/administrator/pulse/Pulse-DSL/tests/screenshots/seed-smoke-emails.png\r\n/home/administrator/pulse/Pulse-DSL/apps/example/tests/screenshots/deals.png"
      },
      {
        "type": "text",
        "text": "<environment_details>\n# VSCode Visible Files"
      }
    ],
    "ts": 1771873245431
  },
  {
    "role": "assistant",
    "content": [
      {
        "type": "reasoning",
        "text": "The user wants me to read 3 images at once in a single read_file tool call. However, the read_file tool only reads exactly one file per call. The instructions say \"IMPORTANT: This tool reads exactly one file per call. If you need multiple files, issue multiple parallel read_file calls.\"\n\nSo I need to issue 3 parallel read_file calls simultaneously. Let me pick 3 screenshots and read them all at once.",
        "summary": []
      },
      {
        "type": "text",
        "text": "\nThe [`read_file`](packages/parser/src/lexer.ts) tool reads **one file per call**, but the instructions explicitly say to issue **multiple parallel calls** for multiple files. I'll fire all 3 simultaneously now:"
      },
      {
        "type": "tool_use",
        "id": "toolu_0133TC2SNR1wkGpjqQh7z4av",
        "name": "read_file",
        "input": {
          "path": "apps/example/tests/screenshots/dashboard.png"
        }
      },
      {
        "type": "tool_use",
        "id": "toolu_01M8GVA46BwaGNeDKRQ3sT3e",
        "name": "read_file",
        "input": {
          "path": "apps/example/tests/screenshots/contacts.png"
        }
      },
      {
        "type": "tool_use",
        "id": "toolu_01ALGnpNKo7xtWptMYEGrRr7",
        "name": "read_file",
        "input": {
          "path": "apps/example/tests/screenshots/deals.png"
        }
      }
    ],
    "ts": 1771873251098
  },
  {
    "role": "user",
    "content": [
      {
        "type": "tool_result",
        "tool_use_id": "toolu_0133TC2SNR1wkGpjqQh7z4av",
        "content": "File: apps/example/tests/screenshots/dashboard.png\nNote: Image file (17 KB)"
      },
      {
        "type": "image",
        "source": {
          "type": "base64",
          "media_type": "image/png",
          "data": "iVBORw0KGgoAAAA"
        }
      },
      {
        "type": "tool_result",
        "tool_use_id": "toolu_01M8GVA46BwaGNeDKRQ3sT3e",
        "content": "File: apps/example/tests/screenshots/contacts.png\nNote: Image file (17 KB)"
      },
      {
        "type": "image",
        "source": {
          "type": "base64",
          "media_type": "image/png",
          "data": "iVBORw0KGgoAAAANS"
        }
      },
      {
        "type": "tool_result",
        "tool_use_id": "toolu_01ALGnpNKo7xtWptMYEGrRr7",
        "content": "File: apps/example/tests/screenshots/deals.png\nNote: Image file (17 KB)"
      },
      {
        "type": "image",
        "source": {
          "type": "base64",
          "media_type": "image/png",
          "data": "iVBORw0KGgoAAAANS"
        }
      },
      {
        "type": "text",
        "text": "<environment_details>\n# VSCode Visible Files"
      }
    ],
    "ts": 1771873251364
  }
]

[roomote]

I've analyzed this issue and here's my proposed implementation plan:

Root Cause: In src/core/assistant-message/presentAssistantMessage.ts, both pushToolResult implementations (for mcp_tool_use at ~line 136 and for regular tool_use at ~line 449) handle image content incorrectly. When a tool result contains image blocks (e.g., from read_file on an image), the code:

1. Creates a tool_result with only the text as a string content
2. Pushes image blocks as separate sibling entries in userMessageContent

This means when multiple image-returning tools run in parallel, the result looks like:

[
  {"type": "tool_result", "tool_use_id": "xxx", "content": "File: img1.png"},
  {"type": "image", "source": {...}},
  {"type": "tool_result", "tool_use_id": "yyy", "content": "File: img2.png"},
  {"type": "image", "source": {...}}
]

Anthropic expects each tool_result to be immediately followed by the next tool_result. The orphaned image blocks between them break the pairing.

Fix: Nest image blocks inside each tool_result's content field as an array, per the Anthropic API spec:

[
  {"type": "tool_result", "tool_use_id": "xxx", "content": [
    {"type": "text", "text": "File: img1.png"},
    {"type": "image", "source": {...}}
  ]},
  {"type": "tool_result", "tool_use_id": "yyy", "content": [
    {"type": "text", "text": "File: img2.png"},
    {"type": "image", "source": {...}}
  ]}
]

Changes:

1. Modify pushToolResult in the mcp_tool_use case (~line 169-178) to include image blocks in the tool_result.content array instead of pushing them as siblings
2. Modify pushToolResult in the regular tool_use case (~line 481-489) with the same fix
3. Add tests to verify image blocks are nested correctly inside tool_result content

Please mention @roomote in a comment saying that you approve this plan, and I'll proceed with the implementation.