Assistants: no answer when a proposal sends the model too much issue #248 closed: completed andymccall opened this · 1 comment
andymccall set labels bug ·
andymccall referenced by change c08cec372c81 on main ·
andymccall
Andy, 2026-10-17: it was Claude (API). Configured: Claude on repo 1 (claude-opus-5-5) and on org 1 (claude-opus-5), budget 120 000 characters (~30k tokens, far inside Claude's window). Likely cause: the 90 s limit on interactive asks for a non-streamed answer of up to 16k tokens (it now says 'no answer within 90 seconds' instead of nothing). Decided: stream the answers, so the limit applies to silence, not to the whole answer.
andymccall referenced by change 5436f0dd9429 on main ·
closed by change 119fd17d213f ·
Andy, 2026-10-16: the AI assistant doesn't respond when too much data is sent to it. The hub cuts a proposal's changes and diff to the assistant's budget (characters, assistants.rs), but a large proposal can still overflow the model's context, or take longer than the request allows, and the person sees nothing useful. Wanted: (1) say plainly what happened (too large for the model, timed out, the provider's error) instead of no answer; (2) a budget that fits the model: by tokens (estimated), not characters, defaulting from the model's context where the provider says it (model lists, issue #109); (3) send less, better: the files' list and stats always, then whole hunks in order of size until the budget, saying what was left out; for very large proposals, summarise per file first and then the whole (map-reduce); (4) a test with a stand-in provider that refuses over a limit.