Forge

A bug-fix release for the llamafile backend. When llama.cpp's tool-call parser rejects malformed model output with a 500, the raw error JSON no longer leaks into the conversation as assistant text — complete tool calls are rescued out of the error body and executed, and unrecoverable ones trigger a clean re-sample nudge.

Added

• Tool-call rescue from malformed-500 bodies. llama.cpp's Failed to parse input message embeds the rejected generation; forge now leniently re-parses <tool_call> blocks out of it (Qwen-coder XML format) and returns each block that names a tool from the request's tools array as a real ToolCall — deduped, with parameters coerced to their declared schema types. Skeleton/preview blocks are passed through too: dispatch rejects them with a [ToolError] on the tool channel, the canonical corrective signal, while complete calls simply execute. Unknown tool names are never fabricated. Rescues are counted on LlamafileClient.rescued_tool_calls and logged, so rescued runs stay auditable.

Fixed

• Malformed tool-call 500s no longer leak error JSON into the conversation. When llama.cpp rejects a malformed or incomplete tool call and nothing can be rescued from the body, forge returns a targeted retry nudge (naming the stutter pattern when a <tool_call> block was visible, generic otherwise) so the retry loop re-samples a clean call. Previously the raw 500 body was echoed back as if the model had said it.

Changed

• Arbitrary llamafile 500s now fail loud. A 500 that is not a tool-call parse rejection raises BackendError instead of being returned as a TextResponse carrying the error body — genuine backend failures now cascade, matching the other clients, rather than entering the conversation as model text.