v0.1.92

Changelog

Features

• cfaf75f feat(anthropic): support response_format instead of rejecting it (#963)
• b1de0a9 feat(dashboard): hide user rows with only inactive keys (#975)
• f70bb91 feat(models): add GET /v1/models/{model} retrieve endpoint (#959)
• 1d4ae81 feat(observability): flag and count empty 200 provider responses (#951)
• 6c3d39e feat(plugins): add tag_replace plugin for request value tags (#1005)

Bug Fixes

• 2cace9a fix(audio): forward unknown audio parameters to the provider (#968)
• 28e6447 fix(cache): route /v1/embeddings through the response cache (#962)
• 0d34b58 fix(cache): scope exact response-cache entries to the guardrail chain (#971)
• d1c152d fix(ci): retry the MongoDB image pull when the registry rate-limits it (#1007)
• 62373ae fix(deepseek): apply DeepSeek compatibility to OpenCode Go routes (#985)
• e540765 fix(gateway): name provider instances, keep rate-limit subject spelling, hide upstream URLs (#967)
• f0a7bf4 fix(guardrails): keep llm_judge enforcing on a reasoning judge model (#969)
• 5f4d077 fix(mcp): explain 404 upstream dials and bump go-sdk to v1.8.0 (#987)
• fa24442 fix(mcp): point 405 handshake failures at the transport (#994)
• 6d042df fix(messages): always render a signature on thinking blocks (#964)
• 2b5fc18 fix(providers): adapt provider-hostile parameters and normalize Groq reasoning (#973)
• 4b91505 fix(providers): drop unsupported xAI reasoning effort, parse Groq thinking, strip x_groq (#954)
• ac672ad fix(responses): complete background responses and drop the model check on lifecycle routes (#961)
• eb2aedd fix(responses): drop replayed output item ids on chat translation (#952)
• 957bd86 fix(responses): name the predecessor on streamed chained responses (#950)
• fbe3911 fix(responses): renumber stream sequence numbers when a guardrail edits the stream (#953)
• cd0599f fix(responses): replay chained history for native providers when needed (#956)
• 071fb8e fix(responses): report truncation and normalize the translated output shape (#955)
• 4471cb9 fix(usage): bill media calls by one unit, not two or none (#970)
• 6ca6eb7 fix(usage): keep explicit zero audio usage and document the models dialect (#974)
• 63cc886 fix(zai): forward reasoning effort as reasoning_effort (#999)

Performance

• 0d0b061 perf(live): skip live-only audit encoding without subscribers and ignore null usage chunks (#1006)