feat: add streaming support for LLM responses #423
Description
Summary
Add streaming support to clawbolt's LLM calls so users see responses as they're generated rather than waiting for the full completion. This would improve perceived latency in the Telegram UX.
Context
All 5 current acompletion call sites use non-streaming (stream=None). As part of #422 we're migrating to the Messages API (amessages), which supports streaming natively. Streaming should be implemented as a separate effort after that migration lands.
Scope
- Stream agent loop responses to Telegram as they arrive (primary value)
- Evaluate whether compaction, heartbeat, and vision calls benefit from streaming (likely not, since they're short internal calls)
- Handle partial tool call chunks correctly in the agent loop
- Consider Telegram API rate limits for message editing (progressive updates)
Related
- investigate: switch off of chatCompletion API and over to Messages API #422 (Messages API migration, prerequisite)