Problem
Currently the app only treats an agent turn as successful if it returns HTML. This forces every interaction into a single-shot "produce the final markup or fail" mode. For example, if the agent responds with analysis or a list of suggested changes instead of HTML, the edit is marked as failed (e.g. "Edit failed: No HTML produced") even though the response was useful and relevant.
Why this matters
There are legitimate cases where a user wants to discuss and narrow things down with the agent before committing to a concrete change — for instance, asking the agent to identify candidate areas to improve, discussing tradeoffs, or refining scope — before asking it to actually generate the HTML. The current flow doesn't support this; it assumes every message should culminate in HTML, which discourages exploratory conversation.
Suggested fix
Support a conversational mode where non-HTML responses (plain text, suggestions, clarifying questions, etc.) are treated as valid intermediate turns rather than failures. HTML generation could be triggered explicitly (e.g., once the user confirms a direction) rather than being the only accepted output for every turn.
Problem
Currently the app only treats an agent turn as successful if it returns HTML. This forces every interaction into a single-shot "produce the final markup or fail" mode. For example, if the agent responds with analysis or a list of suggested changes instead of HTML, the edit is marked as failed (e.g. "Edit failed: No HTML produced") even though the response was useful and relevant.
Why this matters
There are legitimate cases where a user wants to discuss and narrow things down with the agent before committing to a concrete change — for instance, asking the agent to identify candidate areas to improve, discussing tradeoffs, or refining scope — before asking it to actually generate the HTML. The current flow doesn't support this; it assumes every message should culminate in HTML, which discourages exploratory conversation.
Suggested fix
Support a conversational mode where non-HTML responses (plain text, suggestions, clarifying questions, etc.) are treated as valid intermediate turns rather than failures. HTML generation could be triggered explicitly (e.g., once the user confirms a direction) rather than being the only accepted output for every turn.