fix: use current model's context window for usage_update size

[timvisher-dd]

The usage_update notification reports size: 200000 even when the active model has a 1M context window (e.g. opus[1m]), causing clients to display incorrect context utilization (e.g. 689k/200k (344.3%) instead of 689k/1000k (68.9%)).

Four bugs fixed:

• Min across all models: The original code used Math.min across all modelUsage entries, so subagent models (Sonnet/Haiku with 200k windows) dragged down the reported size for the main Opus 1M model. Now tracks the top-level assistant model and looks up its context window specifically.
• Model name mismatch: The SDK's streaming path keys modelUsage by the requested model alias (e.g. claude-opus-4-6) while BetaMessage.model on assistant messages has the resolved API response model (e.g. claude-opus-4-6-20250514). The exact-match lookup always missed, falling back to the hardcoded 200k default. Now falls back to prefix matching, preferring the longest/most-specific match.
• Synthetic messages corrupt model tracking: /compact and similar commands emit assistant messages with model: "<synthetic>". These were updating lastAssistantModel, causing the next usage_update to miss the modelUsage lookup and fall back to the 200k default. Now filters out <synthetic> models.
• Stale usage after compaction: No usage_update was sent on compact_boundary, so clients kept showing the pre-compaction context size (e.g. 944k/1m) right after "Compacting completed" until the next full turn. Now sends used: 0 immediately on compaction. This is a deliberate approximation — the exact post-compaction size isn't known until the SDK's next API call, which replaces it within seconds. The alternative (no update) is worse UX: showing a full context bar right after compaction.

Eight new tests cover: token sum correctness, current-model context window lookup, model switching, subagent isolation, prefix matching in both directions, and synthetic message filtering.

Note: bin/test (local CI validation script) is cherry-picked from #353.

Would fix agent-shell's usage indicator which currently has to defend against this broken math: xenodium/agent-shell#364

Testing

• Manual riding with it for a bit.

[SteffenDE]

In my testing, the result modelUsage contained a claude-opus-4-6[1m] and the assistant message claude-opus-4-6 (no suffix). But assuming there can be a suffix - if the assistant message contains claude-opus-4-6-20250514 or similar, this would not match, because it checks for startsWith.

Maybe do something like this and actually calculate the common prefix:

function commonPrefixLength(a: string, b: string) {
  let i = 0;
  while (i < a.length && i < b.length && a[i] === b[i]) {
    i++;
  }
  return i;
}

function getMatchingModelUsage(modelUsage: Record<string, ModelUsage>, currentModel: string) {
  let bestKey: string | null = null;
  let bestLen = 0;

  for (const key of Object.keys(modelUsage)) {
    const len = commonPrefixLength(key, currentModel);
    if (len > bestLen) {
      bestLen = len;
      bestKey = key;
    }
  }

  if (bestKey) {
    return modelUsage[bestKey];
  }
}

and then use it as

const matchingModelUsage = lastAssistantModel ? getMatchingModelUsage(message.modelUsage, lastAssistantModel) : null;
const contextWindowSize = matchingModelUsage?.contextWindow ?? 200000;
lastContextWindowSize = contextWindowSize;

[SteffenDE]

I'd do one change, other than that looks good to me. (I'd drop the bin/test file from the PR).