mnfst/llm-gateway manifest@6.29.0

manifest v6.29.0

on GitHub

✨ Minor Changes

• 8dec00d: Manifest Cloud is Pro only: a workspace without a subscription routes nothing and gets M204 "Subscription required". Self-hosted is unchanged.

🐛 Patch Changes

• ac7c89e: Bedrock discovery now lists the Claude Fable 5 and 5.1 inference profiles. AWS only serves them when the Region's data retention mode is aws_review (or legacy provider_data_share), which retains prompts and completions for up to 30 days and lets AWS review them; otherwise Bedrock answers 400.

• 9c87bbe: Serve Chat Completions requests with function tools and reasoning for Amazon Bedrock GPT cross-Region inference profiles through the Bedrock Runtime Responses API, which Runtime Chat Completions rejects for these models. Set BEDROCK_RUNTIME_CHAT_VIA_RESPONSES=false to turn it off. Sampling parameters (temperature, top_p, stop, seed, ...) are not forwarded on this route. Also read Responses stream events that carry their name only in data.type, and number streamed tool calls from 0 when a reasoning item comes first.

  This numbering also applies to the other Responses-backed routes (OpenAI, Copilot, xAI, ChatGPT subscription and Bedrock Mantle): their streamed tool calls now start at index 0 when a reasoning item precedes them. GPT-6 Astra on Bedrock Runtime now serves Chat Completions requests with function tools instead of failing; without a reasoning_effort it runs at the model's default effort.

• ac0d8bf: Bump modelparams to 0.0.72: GPT-6 (luna, astra, sol), GPT-6.1 sol and Claude Fable 5.1 parameter specs are now available in the bundled catalog, including offline and cold-boot installs. Also fixes the catalog loader, which could not read modelparams 0.0.58 or newer.

• 394d9a9: Report Gemini replies blocked for BLOCKLIST, PROHIBITED_CONTENT, SPII or IMAGE_SAFETY as content_filter instead of stop.