Enforce deterministic gateway task policies

Make the server authoritative for thinking, model, search, tools, retry, and output budgets while preserving legacy client behavior.

docs/openapi.yaml

@@ -356,6 +356,11 @@ paths:
   /v1/gateway/llm/{capability}:
     post:
       summary: Run a metered polish, AI, or agent request
+      description: |
+        The server deterministically selects model, thinking, search, tools, retry,
+        and output-budget policy from `capability` plus optional `taskKind`. It
+        never infers task type from `input` or `context`, and clients cannot
+        supply provider parameters. Search and tools are currently disabled.
       parameters:
         - $ref: "#/components/parameters/RequestId"
         - name: capability
@@ -1374,9 +1379,35 @@ components:
       properties:
         input: { type: string, minLength: 1, maxLength: 32000 }
         context: { type: ["string", "null"], maxLength: 32000 }
-        maxOutputTokens: { type: integer, minimum: 1, maximum: 4096, default: 512 }
+        maxOutputTokens:
+          type: integer
+          minimum: 1
+          maximum: 4096
+          default: 512
+          description: |
+            Requested output budget. The server clamps dictation polish and
+            edit-last-input to 512 tokens; translation, clipboard transform,
+            and custom skill to 2,048; and reasoning tasks to 4,096.
         temperature: { type: number, minimum: 0, maximum: 1, default: 0.2 }
         stream: { type: boolean, default: false }
+        taskKind:
+          type: ["string", "null"]
+          enum:
+            - dictation_polish
+            - translation
+            - edit_last_input
+            - ai_question
+            - clipboard_transform
+            - custom_skill
+            - agent_planning
+            - null
+          description: |
+            Optional deterministic task selector. Allowed combinations are:
+            `polish` with `dictation_polish`, `translation`, or `edit_last_input`;
+            `ai` with `ai_question`, `clipboard_transform`, or `custom_skill`;
+            and `agent` with `agent_planning`. Omission defaults respectively to
+            `dictation_polish`, `ai_question`, and `agent_planning`. A mismatch
+            returns `400 invalid_request`.
         requestSource:
           type: ["string", "null"]
           enum: [hotword, null]