Add per-model generation settings to the model selector

Each text and vision harness setting now carries the generation settings
its model takes (temperature, sampling limits, penalties, seed, reasoning
effort), edited in a modal under the model pick so the harness block stays
short. Nothing is set by default, so models keep their provider's own
values until a super admin or admin changes them.

A setting is only offered and only sent when the catalog says the model
takes it: OpenRouter's supported_parameters is now stored on
harness_models and refreshed by the hourly sync. Other providers list no
parameters, so their models get them all.

src/harness/media/read-image.ts

@@ -3,6 +3,7 @@ import HarnessError from "../harness-error";
 import grabHarnessAdapter from "../providers/grab-harness-adapter";
 import resolveHarnessCandidates from "../resolve/resolve-harness-candidates";
 import withCandidates from "../resolve/with-candidates";
+import grabChatParams from "../text/grab-chat-params";
 
 type Params = {
     /**
@@ -35,6 +36,12 @@ export default async function readImage({ url, focus, signal, bot_id, workspace_
         workspace_id,
     });
 
+    const saved_params = await grabChatParams({
+        harness: "vision",
+        bot_id,
+        workspace_id,
+    });
+
     const { result } = await withCandidates({
         harness: "vision",
         candidates,
@@ -57,7 +64,14 @@ export default async function readImage({ url, focus, signal, bot_id, workspace_
                         ],
                     },
                 ],
-                temperature: 0,
+                /**
+                 * Reading is a lookup, so it stays literal unless a
+                 * temperature was set for the vision harness.
+                 */
+                params: {
+                    temperature: 0,
+                    ...saved_params(candidate.model),
+                },
                 signal: signal || new AbortController().signal,
             });
 

src/harness/providers/openai-compatible/chat.ts

@@ -12,7 +12,7 @@ const chat: HarnessAdapter["chat"] = async ({
     messages,
     tools,
     max_tokens,
-    temperature,
+    params,
     signal,
 }) => {
     const { provider, model } = candidate;
@@ -32,7 +32,7 @@ const chat: HarnessAdapter["chat"] = async ({
                   }
                 : {}),
             ...(max_tokens ? { max_tokens } : {}),
-            ...(temperature !== undefined ? { temperature } : {}),
+            ...(params || {}),
         },
         signal,
         candidate_key: candidate.key,

src/harness/providers/openrouter/openrouter-model-to-catalog.test.ts

@@ -35,6 +35,7 @@ describe("openRouterModelToCatalog", () => {
             context_length: 1000000,
             max_output_tokens: 65536,
             voices: undefined,
+            supported_parameters: ["tools", "temperature"],
             released_at: 1792000000000,
             expires_at: Date.parse("2026-12-01"),
             intelligence_score: 26.5,

src/harness/providers/openrouter/openrouter-model-to-catalog.ts

@@ -56,6 +56,9 @@ export default function openRouterModelToCatalog({
         voices: model.supported_voices?.length
             ? model.supported_voices
             : undefined,
+        supported_parameters: model.supported_parameters?.length
+            ? model.supported_parameters
+            : undefined,
         released_at: model.created ? model.created * 1000 : undefined,
         expires_at: Number.isNaN(expires_at) ? undefined : expires_at,
         intelligence_score: scores?.intelligence_index ?? undefined,

src/harness/text/generate-text.ts

@@ -8,6 +8,7 @@ import type {
     HarnessMessage,
     HarnessToolDefinition,
 } from "../types";
+import grabChatParams from "./grab-chat-params";
 import prepareMessagesForModel from "./prepare-messages-for-model";
 import stripThinkTags from "./strip-think-tags";
 
@@ -23,7 +24,10 @@ export type GenerateTextParams = {
      */
     prefer_key?: string;
     max_tokens?: number;
-    temperature?: number;
+    /**
+     * Overrides the saved generation settings (media reading wants 0).
+     */
+    params?: Record<string, number | string | { effort: string }>;
     signal?: AbortSignal;
     /**
      * Media readings shared across calls of one run.
@@ -54,16 +58,22 @@ export default async function generateText({
     candidates,
     prefer_key,
     max_tokens,
-    temperature,
+    params,
     signal = new AbortController().signal,
     media_cache = new Map(),
     bot_id,
     workspace_id,
 }: GenerateTextParams): Promise<GenerateTextResult> {
+    const resolved =
+        candidates ||
+        (await resolveHarnessCandidates({ harness: "text", bot_id, workspace_id }));
+    const saved_params = params
+        ? undefined
+        : await grabChatParams({ harness: "text", bot_id, workspace_id });
+
     const { result, candidate } = await withCandidates({
         harness: "text",
-        candidates:
-            candidates || (await resolveHarnessCandidates({ harness: "text", bot_id, workspace_id })),
+        candidates: resolved,
         prefer_key,
         run: async (candidate) => {
             const answer = await grabHarnessAdapter({
@@ -80,7 +90,7 @@ export default async function generateText({
                 }),
                 tools,
                 max_tokens,
-                temperature,
+                params: params || saved_params?.(candidate.model),
                 signal,
             });
 

src/harness/text/grab-chat-params.ts

@@ -0,0 +1,56 @@
+import {
+    HarnessParameters,
+    modelTakesParameter,
+} from "@/src/dict/harness-parameter-dict";
+import grabHarnessSetting from "@/src/functions/backend/db/harness/grab-harness-setting";
+import type { HarnessModelInfo } from "../types";
+
+type Params = {
+    harness: "text" | "vision";
+    bot_id?: number | string;
+    workspace_id?: number | string | null;
+};
+
+/**
+ * A model's generation settings for one call: what is saved in the
+ * harness's setting (the workspace's, the bot's, else the default),
+ * keyed by the field sent on the wire, keeping only what the model says it
+ * takes. Empty when none are set, so the model gets its provider's
+ * defaults.
+ */
+export default async function grabChatParams({
+    harness,
+    bot_id,
+    workspace_id,
+}: Params): Promise<
+    (model: HarnessModelInfo) => Record<string, number | string | { effort: string }>
+> {
+    const setting = await grabHarnessSetting({
+        harness,
+        bot_id,
+        workspace_id,
+    });
+
+    return (model) =>
+        Object.fromEntries(
+            HarnessParameters.flatMap((parameter) => {
+                const value = setting.options[parameter.key];
+
+                if (typeof value != "number" && typeof value != "string") return [];
+                if (!value && value !== 0) return [];
+                if (
+                    !modelTakesParameter({
+                        parameter,
+                        supported: model.supported_parameters,
+                    })
+                )
+                    return [];
+
+                const key = "wire" in parameter ? parameter.wire : parameter.key;
+
+                return [
+                    [key, key == "reasoning" ? { effort: String(value) } : value],
+                ];
+            }),
+        );
+}

src/harness/types.ts

@@ -82,6 +82,12 @@ export type HarnessModelInfo = {
     context_length?: number;
     max_output_tokens?: number;
     voices?: string[];
+    /**
+     * The request fields the provider says it takes (OpenRouter's
+     * `supported_parameters`). Undefined when the provider doesn't say:
+     * then every generation setting is sent.
+     */
+    supported_parameters?: string[];
 };
 
 /**
@@ -137,7 +143,11 @@ export type HarnessChatParams = {
     messages: HarnessMessage[];
     tools?: HarnessToolDefinition[];
     max_tokens?: number;
-    temperature?: number;
+    /**
+     * Sampling and reasoning fields, keyed by their wire name. The caller
+     * only includes what the model advertises.
+     */
+    params?: Record<string, number | string | { effort: string }>;
     signal: AbortSignal;
 };