Models & Routing
Platform key vs BYOK, tolerant key matching, wire-spelling rewrites, reasoning budgets.
Two key sources
- Platform key — Flowmanner's own provider keys back your calls within plan limits.
- BYOK — your stored provider keys are preferred for your missions; keys are encrypted at rest and relayed server-side (see Connect BYOK).
Tolerant matching, strict gateways
Routing matches a requested model id against each key's declared model list
suffix-tolerantly (llm_router.py:858-870). When the match was
suffix-only, Flowmanner sends the key's declared spelling on the wire,
because strict gateways reject spellings they did not declare — e.g. a bare
deepseek-v4-flash sent to a gateway declaring
deepseek-chat/deepseek-v4-flash fails with
400 invalid model format. Expected format: modelType/model
(llm_router.py:512-524, defect fixed 2026-08-24).
Your results and logs keep showing your original id; only the outbound payload spelling changes.
Reasoning-model quirk
Reasoning-style models spend part of the context window on internal reasoning before answering. When you cap output tokens, leave headroom or answers can come back truncated.
Last updated 2026-08-25 (git-derived)