fix(web): suppress Task tool prompt text from leaking into chat

[hqhq1025]

Summary

• When Claude calls the Task/Agent tool, it writes the prompt text as a regular text content block before the tool_use block in the same assistant message
• This causes the prompt to appear twice: as a standalone rendered message AND inside the tool card
• Fix: skip text blocks in agent messages that also contain a Task tool_use, since the prompt is already visible in the tool card

Before

The Agent tool's full prompt text (description + detailed instructions) renders as a standalone markdown message below the tool card, duplicating the content already shown inside the card.

After

Only the Agent tool card is shown. The prompt text is accessible by clicking the card to open the detail dialog.

Test plan

• All 76 existing tests pass (no regression)
• Manual: trigger an Agent/Task tool call and verify the prompt text only appears inside the tool card, not as a separate message

[github-actions]

Findings

• [Major] Task tool-call suppresses all text blocks, not only the duplicated prompt — any non-prompt assistant text in the same message is dropped, losing user-visible content. Evidence web/src/chat/reducerTimeline.ts:76
  Suggested fix:

  const taskTool = msg.content.find(
      (content) => content.type === 'tool-call' && content.name === 'Task'
  )
  const promptText = taskTool
      && typeof taskTool.input === 'object'
      && taskTool.input !== null
      && 'prompt' in taskTool.input
      && typeof (taskTool.input as { prompt: unknown }).prompt === 'string'
      ? (taskTool.input as { prompt: string }).prompt
      : null
  
  if (promptText && c.text.trim() === promptText.trim()) continue

Summary
Review mode: initial. 1 issue. Potential loss of non-prompt text in Task tool-call messages.

Testing
Not run (automation). Suggested: manual Task tool-call that includes extra assistant text; verify only prompt suppressed.

[github-actions]

[MAJOR] Any agent message that contains a Task tool-call will now drop all text blocks, not just the prompt text. That can hide non-prompt assistant content in the same message. Evidence web/src/chat/reducerTimeline.ts:76.

Suggested fix:

const taskTool = msg.content.find(
    (content) => content.type === 'tool-call' && content.name === 'Task'
)
const promptText = taskTool
    && typeof taskTool.input === 'object'
    && taskTool.input !== null
    && 'prompt' in taskTool.input
    && typeof (taskTool.input as { prompt: unknown }).prompt === 'string'
    ? (taskTool.input as { prompt: string }).prompt
    : null

if (promptText && c.text.trim() === promptText.trim()) continue

[github-actions]

Review mode: initial

• 1 major issue; see inline comment.
• Testing not run (automation).

[github-actions]

[MAJOR] CLI output cards suppressed

normalizeUserOutput now drops messages starting with <command-name> / <local-command-caveat>, but those tags are what the CLI output renderer parses for terminal command cards. This removes command output from chat.

Suggested fix:

if (
    trimmed.startsWith('<task-notification>') ||
    trimmed.startsWith('<system-reminder>')
) {
    return null
}