Spaces:
Paused
Paused
Commit History
-
Retry a 502/504 from the model gateway instead of failing the turn
Reproduced the reported "504 Gateway Time-out" against router.huggingface.co directly: a full multi-tool briefing (several aerodromes + METAR/TAF + NOTAM + route/fuel + alternates all answered in one turn) builds a large enough context that the follow-up completion took 107s in a live test — well inside LLM_TIMEOUT (180s), but in the range where an intermediary gateway in front of the actual inference provider gives up first and hands back a bare HTML 504 instead of a JSON error. chat_completion() now retries a 502/504 up to twice with a short backoff before surfacing it, and agent.py reports each retry to the UI as a status line so a slow step reads as "retrying" instead of a silent stall. A plain 4xx (bad token, bad model name, etc.) is unaffected — those are not transient and retrying them would just waste time.aaaf5af
-
Make tool-call datetime args schema-aware, not just tz-stripped
plan_route was still rejecting a clean, offset-free ISO datetime ("Invalid isoformat string: '2026-08-18T11:00:00'") after the earlier tz-suffix fix, because that error is what date.fromisoformat() raises on a full datetime — the tool likely has a date-only argument that was being fed a combined date+time. Since plan_route's real parameter names live behind the private vfrplan-mcp Space, _sanitize_arguments() now reads each tool's own declared JSON Schema format (date/time/other) per argument and reshapes the value to match, instead of guessing field names. Also documents the earlier stated-figures-are-never-re-asked PROMPT.md rule that hadn't been committed yet.d42109e -
Canvis finals avui
3d90a9a -
Cambis de logica
ce95a69 -
changed into docker
4d97285