Open WebUI Proposes Tool Search Fix to Prevent Invalid Calls from Stopping Subsequent Tool Use

On November 7, the Open WebUI community submitted PR #33872 to adjust Tool Search error handling: when a model sends a malformed tool call, it should receive an error while retaining the ability to retry. As of this check, the proposal remains open and should not be treated as a fix in a released version. Proposed fix

The issue report says that after Tool Search is enabled, if call_tool receives an empty or missing calls list, a batch item without params, or parameters that are not a valid JSON object, the system returns an error but clears the tool list in the next request to the model provider and sets tool choice to none. As a result, even if the model recognizes that its parameters were invalid, it cannot correct the call in the same response. The report locates the problem in the development branch after batch calls were added. Issue and reproduction steps

This path is relevant to using large tool collections. According to the official documentation, Tool Search defers loading larger tool definitions: the model first uses search_tools to find the full specifications, then executes them through call_tool. This reduces the tool descriptions included with each request and keeps the prompt prefix stable. The feature is still experimental, is disabled by default, and applies only to Native Mode. Tool mechanism documentation

The proposal author says malformed calls and genuinely incomplete model responses share an error field, and the stop condition ends tool use whenever that field has a value. The fix changes this so that only incomplete responses trigger a stop, while other errors are returned to the model for handling. Technically, this preserves the agent's error recovery loop: a parameter mistake can still be corrected instead of forcing an answer without the requested results. Author's explanation

Engineering teams should next monitor the review and merge outcome, and whether regression tests cover single-call errors, batch errors, and genuinely truncated responses. The public information currently provides reproduction steps and the rationale for the fix, but is not enough to confirm end-to-end behavior across model backends or to conclude that all released versions are affected.