fix(teammate-progress): keep cumulative token+tool counts across prompts (#475)

[0xghost42]

Closes #475.

What broke

TeammateSpinnerLine, InProcessTeammateDetailDialog, and the leader's spinner aggregate (Spinner.tsx:191-200) all read teammate.progress?.tokenCount and teammate.progress?.toolUseCount straight from AppState. Those fields are populated by getProgressUpdate(tracker) from inside the in-process teammate runner's per-message loop. The runner used to do this:

while (!abortController.signal.aborted && !shouldExit) {
  // ...
  // Create fresh progress tracker for this prompt
  const tracker = createProgressTracker()
  // ...
}

A fresh tracker per prompt iteration meant that whenever the leader sent a second prompt to the same teammate (i.e. any multi-turn agent-team flow), the previous prompts' cumulativeOutputTokens and toolUseCount were dropped on the floor and the pill jumped back to (or near) zero.

The Claude API returns input_tokens as cumulative-per-request — forkContextMessages already re-sends the full history each turn — so latestInputTokens was always large and partially masked the bug. The visible symptom matched the report on #475: output tokens and tool-use counts repeatedly looked low / stale to leaders, even after long sessions.

Fix

Hoist createProgressTracker() (and the resolver) out of the while loop so a single tracker spans the whole teammate lifetime. cumulativeOutputTokens and toolUseCount now keep running totals across every prompt; latestInputTokens continues to track the latest request's cumulative-per-request input (which is what we want for "running context cost"). recentActivities stays capped at MAX_RECENT_ACTIVITIES = 5 so the preview doesn't bloat over long teammate sessions.

Behavior on a single-prompt teammate is unchanged.

Tests

New src/tasks/LocalAgentTask/progressTracker.test.ts pins the contract the runner relies on. Six cases:

1. Output tokens accumulate across multiple assistant messages.
2. Cumulative semantic survives a simulated multi-prompt teammate session (3 turns prompt-1 + 2 turns prompt-2 → monotonically-increasing tokenCount).
3. Fresh-tracker-per-prompt regression repro — confirms that under the old behavior prompt-1's output tokens and tool uses are lost when a new tracker is created for prompt-2.
4. Tool use count accumulates across messages.
5. cache_creation_input_tokens + cache_read_input_tokens fold into latestInputTokens.
6. recentActivities stays capped at 5 while toolUseCount keeps climbing.

bun test (full): 2998/2998 pass. bun run build clean (no React/Ink leakage in SDK bundle; external lists valid).

Repro for reviewer

1. Start two teammates (e.g. /team with two agents).
2. Send a first prompt to one — note token + tool-use count growing on its pill.
3. Send a second prompt to the same teammate — pre-fix the count snaps back to the second prompt's per-call value (often visibly lower); post-fix it keeps climbing.

[jatmn]

Thanks for the contribution. I do not see any actionable issues from my review.