#595 release: subagent live progress + transcript (#594)

Summary

• Converts taskTool.execute to async function* yielding open-agents' 4-stage progress sequence (initial → tool-call pending → finish-step usage → final transcript)
• toModelOutput extracts last assistant text from output.final for parent agent context
• New sumLanguageModelUsage helper (SRP)
• Drive-by: fixes askUserQuestionTool.toModelOutput signature for the new SDK (was a latent bug from the older beta SDK)

Test plan

• 25 new/updated unit tests; full suite 3114/3114 pass
• E2E on preview: 6 tool-output-available chunks emitted in expected order
• Cross-validated yield sequence against open-agents production — byte-identical (same chunk count, same field sets on every yield)
• toModelOutput confirmed end-to-end: parent agent correctly relayed subagent's summary

🤖 Generated with Claude Code

---

Summary by cubic

Adds live progress streaming for the task subagent by converting taskTool.execute to an async generator, so the UI can show pending tool-calls, running usage totals, and the final transcript. Also extracts the last assistant text for parent context and fixes askUserQuestionTool.toModelOutput to read the SDK’s { output } shape.

• New Features

  • taskTool.execute now yields: initial { startedAt, modelId, toolCallCount: 0 }, a pending { name, input } on each tool-call, accumulated usage on each finish-step, and a final { final: ModelMessage[] } transcript.
  • taskTool.toModelOutput returns the last assistant text from output.final.
  • Added sumLanguageModelUsage to safely accumulate per-step usage.

• Bug Fixes

  • askUserQuestionTool.toModelOutput now uses ({ output }) so user answers are correctly surfaced.