Dispatch: Unrecoverable API 400 error when pasting multiple screenshots into conversation

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

What happened

During a multi-turn Dispatch session (Cowork mode, Claude Desktop App), I pasted 5 screenshots into the conversation and received a hard API 400 error that permanently broke the session. No recovery possible — further messages do not resolve the error.

Error:

API Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"Could not process image"},"request_id":"req_011CZSSbTsmvAUsai1mqyF8d"}

Steps to reproduce

1. Start a long-running agentic Dispatch session with multiple tool calls active (Gmail, Google Drive, file creation)
2. Claude generates an interactive HTML artifact with form inputs (e.g., a scorecard with number fields and text areas)
3. User fills in the HTML form fields directly in the browser
4. User realizes input is trapped in browser memory — the HTML artifact has no save/export function
5. Claude suggests screenshotting the filled form as a workaround
6. User takes 5 screenshots of the completed form and pastes them into the conversation
7. API returns HTTP 400: "Could not process image"
8. Session is permanently stuck — cannot send further messages or recover

Two issues in one

Issue 1: API failure on multi-image paste (the bug)

Pasting 5 screenshots simultaneously into a Dispatch conversation triggers an unrecoverable 400 error. Possible factors: total image payload size, number of concurrent images, image format from clipboard paste, or interaction with an already-long agentic session context.

Issue 2: HTML artifact design gap (contributing factor)

Claude generated an interactive HTML scorecard but did not include any mechanism to persist or export user input. The artifact is static once rendered — form data exists only in the browser DOM. Claude cannot read it back. This forced the user to resort to screenshots as a workaround, which then triggered the API bug.

Impact

• Lost work: Extensive multi-turn agentic session (email scanning, knowledge base updates, artifact generation) terminated mid-workflow
• Lost user input: Scorecard scores and qualitative feedback entered by the user were never captured
• Broken trust: User followed Claude's own suggestion ("take a screenshot and share it") and was punished with a session-ending error
• No recovery path: Session cannot be resumed; user must start over in a new session

Recommendations

1. Graceful degradation: If image processing fails, strip the problematic image(s) and continue the conversation with a user-facing message, rather than bricking the session
2. Session recovery: Allow users to continue after a transient API error (auto-retry, or ability to edit/remove the failed message)
3. Artifact data path: Interactive HTML artifacts that collect user input should always include a save/export mechanism (e.g., a button that outputs JSON or copies to clipboard)

What Should Happen?

1. Claude accepts and processes 5 pasted screenshots, reads the scorecard data, and continues the session normally.
2. If image processing fails for any reason, the error is graceful: the user is prompted to retry, share fewer images, or share data in a different format. The session remains functional.
3. Interactive HTML artifacts that collect user input should include a built-in export mechanism so users are never forced into screenshot workarounds.

Environment: Claude Desktop App (Dispatch/Cowork mode), March 2026, long-running agentic session with multiple MCP tools active.

Error Messages/Logs

API Error: 400 {"type":"error","error":{"type":"invalid_request_error","message":"Could not process image"},"request_id":"req_011CZSSbTsmvAUsai1mqyF8d"}

Steps to Reproduce

1. Start a long-running agentic Dispatch session (Cowork mode) with multiple tool calls (Gmail, Google Drive, file creation)
2. Have Claude generate an interactive HTML artifact with form inputs (e.g., a scorecard with number fields 1-10 and text areas for feedback)
3. Fill in the HTML form fields directly in the browser
4. Attempt to share the filled-in form by pasting 5 screenshots into the conversation simultaneously
5. Observe: API returns HTTP 400 "Could not process image"
6. Observe: Session is permanently stuck — further messages cannot be sent

Note: The session had been running for an extended period with substantial context (email thread analysis, knowledge base updates, HTML artifact generation) before the screenshots were pasted. The long context may be a contributing factor.

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

Claude Desktop App (Dispatch/Cowork mode) — March 2026

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

Terminal.app (macOS)

Additional Information

This bug occurred in Claude Desktop App's Dispatch/Cowork mode, not the Claude Code CLI. Filing here as the closest public bug tracker for Anthropic's developer-facing tools.

Key context:

• The session was long-running with substantial agentic context (multiple Gmail threads analyzed, Google Drive scanned, HTML artifact generated, knowledge base updates proposed)
• The 5 screenshots were pasted from clipboard (macOS screenshot → paste), not uploaded as files
• The error was immediate and permanent — no partial processing, no retry option
• Claude itself had suggested the screenshot workaround moments before the error occurred

The secondary issue (HTML artifacts lacking data export) is a design pattern problem worth addressing separately — interactive form artifacts should never be dead-ends for user input.