Don't think this is a bug after all. I was able to get around this issue by wrapping$TASK_PROMPT in quotes within the Workspace Spec.

claude -p "$TASK_PROMPT"

Submitted fix in pr #75.

The task prompt was exported correctly, but unquoted $TASK_PROMPT references in workspace task commands caused the shell to word-split multi-line prompts (for example when an AI system prompt is prepended). Workspace shells now auto-quote unquoted $TASK_PROMPT, ${TASK_PROMPT}, $env:TASK_PROMPT, and %TASK_PROMPT% references, and the task automation command field documents the quoting requirement.

@aibot is this working for other workspace spec template also?

@robin Yes — the fix applies to all workspace spec templates, not just Claude Code. Auto-quoting runs in the workspace shell layer when decorating the task command, so every spec (built-in templates and custom specs) benefits regardless of which agent command is configured.

All eight built-in templates (Claude Code, Codex, Cursor, and Open Code — each in shell and container variants) already reference "$TASK_PROMPT"; the shell layer also auto-quotes any unquoted reference for backward compatibility with existing custom specs.

Added test coverage for all built-in templates in pr #75.