Skip to content

Router: fail over on upstream_unavailable instead of 503ing clients (DeepSeek/Kimi outage took down tax chat) #9

Description

@drewstone

2026-07-14: the router returns 503 upstream_unavailable for deepseek/deepseek-chat, deepseek-v4-flash, and kimi-k2.7-code (probed with a valid sk-tan key; glm-5.2, anthropic/claude-haiku-4.5, gemini-2.5-flash-lite, openai/gpt-5-mini all 200).

Impact: taxes.tangle.tools defaulted to deepseek/deepseek-chat, so every chat turn on the product died with "chat backend returned 503" until we changed the app default (tax-agent#373). Any consumer pinning an affected model gets a hard outage for the duration of the upstream incident. Also affected as defaults: physim host-supervisor (deepseek-chat), tuner visible-reasoning (deepseek-v4-flash).

Ask: when an upstream is unavailable, the router should fail over to an equivalent healthy model (same capability class: tool-calling, context) — or at minimum expose health in /v1/models (routeability.status) so product catalogs drop dead models automatically. A per-provider outage should degrade quality, not availability.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions