feat(web): show subagent task trace in tool dialog

[junmo-kim]

Problem

When a session uses the Task (subagent) tool, the tool dialog currently
shows only Input and Result. The child tool calls that run inside the
subagent (Glob, Grep, Read, …) are already wired through the reducer into
block.children, but the modal never renders them, leaving the user with
no visibility into what the subagent actually did.

Solution

Adds a Trace section between Input and Result in the Task tool dialog.
The section surfaces the existing block.children array without any new
data plumbing.

• Default open/close policy mirrors intent-at-open: running and error
  expand (user opened the dialog to see progress or diagnose a failure);
  completed collapses (Result is the primary content, Trace is secondary).
• Summary header shows (N calls · M.Mk tok · S.Ss), reading
  result.totalToolUseCount / totalTokens / totalDurationMs with
  graceful fallback when any field is absent.
• Per-row inline expand lets the user drill into a child's Input/Result
  without opening a nested Dialog.
• Type-guarded access to result.* fields — the field is typed unknown
  at the call site, so all numeric reads go through isObject guards.
• Empty-children guard — the section does not render when there are no
  child tool calls, so non-Task tools are unaffected.

Screenshots

Running — Trace expanded, last child live (●)	Completed — Trace collapsed, summary header shows totals

Error — Trace expanded, failing child marked (✕)	Child inline expand — first child's result rendered in place

Tests

14 new unit tests in web/src/components/ToolCard/trace.test.tsx:

• getTaskTraceChildren — null for empty children, filters non-tool-call nodes
• getTraceSummaryText — all three fallback branches
• TraceSection — empty renders nothing; children render header; running/error
  default to expanded; completed defaults to collapsed; toggle inverts state;
  summary text visible in header

All existing tests continue to pass (bun run test, 187 total).

i18n

tool.trace key added for en and zh-CN.

[github-actions]

Findings

• [Minor] Expanded trace rows omit child input — expanding a child call currently renders only the result, so calls whose important details live in input (for example Bash.command, Edit.old_string/new_string, or Write.content) still cannot be inspected from the Task trace. Evidence: web/src/components/ToolCard/trace.tsx:216.
  Suggested fix:

  const FullInputView = getToolFullViewComponent(child.tool.name)
  const { t } = useTranslation()
  
  {expanded && (
      <div className="ml-8 flex flex-col gap-2 rounded border border-[var(--app-border)] p-2">
          <div>
              <div className="mb-1 text-xs font-medium text-[var(--app-hint)]">{t('tool.input')}</div>
              {FullInputView ? (
                  <FullInputView block={child} metadata={metadata} />
              ) : (
                  <CodeBlock code={safeStringify(child.tool.input)} language="json" />
              )}
          </div>
          <div>
              <div className="mb-1 text-xs font-medium text-[var(--app-hint)]">{t('tool.result')}</div>
              <ResultView block={child} metadata={metadata} />
          </div>
      </div>
  )}

Questions

• None.

Summary

• Review mode: initial
• One functional gap found in the new trace UI. Residual risk: I could not execute the test suite in this runner.

Testing

• Not run (automation): bun is not installed in the runner (bun: command not found).

HAPI Bot

[github-actions]

[MINOR] Expanded trace rows omit child input

TraceChildRow only renders <ResultView block={child} ... /> when a child is expanded. That leaves important call details hidden for tools where the meaningful data is in child.tool.input, such as Bash.command, Edit.old_string/new_string, or Write.content, so the trace still does not let users inspect what the subagent actually invoked.

Suggested fix:

const FullInputView = getToolFullViewComponent(child.tool.name)
const { t } = useTranslation()

{expanded && (
    <div className="ml-8 flex flex-col gap-2 rounded border border-[var(--app-border)] p-2">
        <div>
            <div className="mb-1 text-xs font-medium text-[var(--app-hint)]">{t('tool.input')}</div>
            {FullInputView ? (
                <FullInputView block={child} metadata={metadata} />
            ) : (
                <CodeBlock code={safeStringify(child.tool.input)} language="json" />
            )}
        </div>
        <div>
            <div className="mb-1 text-xs font-medium text-[var(--app-hint)]">{t('tool.result')}</div>
            <ResultView block={child} metadata={metadata} />
        </div>
    </div>
)}

[github-actions]

Findings

• None.

Questions

• None.

Summary

• Review mode: follow-up after new commits
• No blocking correctness, security, regression, data loss, performance, or maintainability issues found in the latest diff. The follow-up commit addresses the prior bot finding by rendering expanded child input along with result output. Residual risk: validation is based on static review only.

Testing

• Not run (automation; PR code not executed during review)

HAPI Bot