Clawbolt Files

Stream the three LLM calls whose output budget exceeds what a 120 second proxy read timeout can carry unstreamed: compaction (16,000 tokens), heartbeat and vision (12,000 each).

An unstreamed request sends nothing until generation finishes, so a proxy measuring its timeout against silence cuts a healthy connection mid-answer. A production compaction took 150.4s behind a 120s timeout: the origin served 13,610 completion tokens and billed for them, the caller got a 524, and because the trim watermark advances before the call runs, the conversation had already left the context.

amessages_streamed accumulates the event stream back into the same MessageResponse the blocking call returned, so usage logging, payload capture, and response parsing are unchanged. A stream that ends without the provider signalling completion now raises rather than returning blocks that look finished.

The agent loop and eval runner stay unstreamed at 8,192 tokens, which is inside the window.

[#1548]