chore(ui): regenerate dashboard API types

ui/litellm-dashboard/src/lib/http/schema.d.ts

@@ -42512,7 +42512,7 @@ export interface components {
             classifier_context_include_assistant_turns: boolean;
             /**
              * Classifier Context Include Tool Turns
-             * @description Send the tool calls and tool results that followed the current ask to the classifier as the work done on it so far, so an agent loop's continuation turn is classified on what its tools turned up rather than on the ask alone: 'What is in TASK.md?' stays SIMPLE until the file it reads describes a complex task, and the turns after that classify as that task. With classification_mode='every_request' every turn of the loop is classified this way. With classification_mode='user_turn' the tier pinned for the ask is a floor for its tool turns: each one is classified, and only a higher tier moves the session up and replaces the pin, so a loop never drops back mid-task, at the cost of a classifier call per tool turn, which the turn waits on, up to the classifier timeout, before it routes. That call is skipped when nothing could outrank the pin: the top configured tier, or an ask a keyword_tier_rules match decides. A tool turn whose classifier call fails keeps the held tier. With the circuit breaker on, tool turns in either mode trip a breaker of their own, so a timeout on one opens it for the worker's tool turns while new asks keep classifying. Under heuristic_first and hybrid a turn with tool traffic always goes to the classifier model, since the local scorer reads only the ask. The block is bounded on its own by classifier_context_budget_chars and classifier_context_per_turn_chars, newest first, and once a loop overflows that budget each entry is cut to a quarter of it, so one large file read cannot push out the rest, and the first tool result after the ask, usually the read that described the task, is kept ahead of the newest ones. It is not bounded by classifier_context_window_size, since one ask can run dozens of tool calls; setting the window size to 0 still turns all context off. Tool output such as file contents is sent to the classifier model, which may be a different deployment or provider than the routed completion model. Off by default because enabling it shifts tier decisions, and therefore spend, for an already-deployed router. Requires classifier_type llm, oss_classifier, heuristic_first, or hybrid, the ones that send the conversation to a classifier model, and cannot be combined with session_affinity, which replays its pin without classifying.
+             * @description Send the tool calls and tool results that followed the current ask to the classifier as the work done on it so far, so an agent loop's continuation turn is classified on what its tools turned up rather than on the ask alone: 'What is in TASK.md?' stays SIMPLE until the file it reads describes a complex task, and the turns after that classify as that task. With classification_mode='every_request' every turn of the loop is classified this way. With classification_mode='user_turn' the tier pinned for the ask is a floor for its tool turns: each one is classified, and only a higher tier moves the session up and replaces the pin, so a loop never drops back mid-task, at the cost of a classifier call per tool turn, which the turn waits on, up to the classifier timeout, before it routes. That call is skipped when nothing could outrank the pin: the top configured tier, or an ask a keyword_tier_rules match decides. A tool turn whose classifier call fails keeps the held tier. With the circuit breaker on, tool turns in either mode trip a breaker of their own, so a timeout on one opens it for the worker's tool turns while new asks keep classifying. Under heuristic_first and hybrid a turn with tool traffic always goes to the classifier model, since the local scorer reads only the ask. The block is bounded on its own by classifier_context_budget_chars and classifier_context_per_turn_chars, newest first, and once a loop overflows that budget each entry is cut to a quarter of it, so one large file read cannot push out the rest, and the oldest entries after the ask, usually the reads that described the task, fill half of it ahead of the newest ones. Tool traffic before the newest ask is left out, since it served an earlier ask, so a follow-up such as 'yes, go ahead' that approves that work needs classifier_context_include_assistant_turns to be classified on it. It is not bounded by classifier_context_window_size, since one ask can run dozens of tool calls; setting the window size to 0 still turns all context off. Tool output such as file contents is sent to the classifier model, which may be a different deployment or provider than the routed completion model. Off by default because enabling it shifts tier decisions, and therefore spend, for an already-deployed router. Requires classifier_type llm, oss_classifier, heuristic_first, or hybrid, the ones that send the conversation to a classifier model, and cannot be combined with session_affinity, which replays its pin without classifying.
              * @default false
              */
             classifier_context_include_tool_turns: boolean;