LiteLLM’s LTS patch targets lost tool history and misleading AI bills
LiteLLM’s LTS patch targets lost tool history and misleading AI bills
LiteLLM’s November 12 LTS release, version 1.108.7, addresses a problem that can look like model confusion: streamed MCP follow-ups retaining only the latest tool round, causing repeated calls. The patch also fixes missing or duplicated sequence numbers in bridged Responses streams and a failed-turn chain error on Bedrock.
The same release corrects Gemini/Vertex cost calculations and batch calls recorded as zero-cost with an unknown model. There is an explicit migration requirement: Responses-only parameters such as max_tool_calls now produce a 400 response on bridged providers unless parameter dropping is enabled. The project advises either removing unsupported fields or setting drop_params: true. Official release
Guth analysis: a compatibility option can change a control
These are infrastructure changes, not evidence of a smarter underlying model. An agent that repeats a completed action may have received an incomplete history rather than decided to ignore it. A dashboard that says zero dollars may have failed attribution rather than delivered free inference.
The parameter warning deserves particular care. If a system relies on a field to constrain behavior, silently dropping it is not equivalent to enforcing that constraint elsewhere. Operators should identify which controls are implemented by the model provider and which must remain in their own runtime. Tests should cover multiple tool rounds, failed turns, usage attribution and the chosen unsupported-parameter behavior before deployment.
Evidence limits
This account verifies the upstream release notes, not a running installation. No performance or safety guarantee is inferred from the patch. No software was installed or tested for this report. The extracted page shows November 12 and a time without an explicit timezone; no precise UTC release time is asserted.
Sources and citations
Revision history
Published .