omp-llm-role 1.1.0 → 1.2.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "omp-llm-role",
-  "version": "1.1.0",
+  "version": "1.2.0",
   "type": "module",
   "description": "omp plugin: daily model-role updater driven by llm-stats/OpenRouter rankings",
   "license": "MIT",
@@ -90,7 +90,7 @@
       },
       "features.providerPinning": {
         "type": "boolean",
-        "description": "Price a model on its own lab's route where one exists (per-model policy; never drops a model)",
+        "description": "Bind each request to the model's best provider route (per-model policy; never drops a model) and prefer routes at or above 50 tok/s",
         "default": false
       },
       "features.costCap": {
@@ -147,7 +147,7 @@
       },
       "roles.default.features.providerPinning": {
         "type": "boolean",
-        "description": "Enable the Prefer the model's own lab endpoint capability for the default role",
+        "description": "Enable the Best-route provider pinning capability for the default role",
         "default": false
       },
       "roles.default.features.costCap": {
@@ -157,7 +157,7 @@
       },
       "roles.default.preferOwnProvider": {
         "type": "boolean",
-        "description": "Price default models on their own lab's route where one exists (soft preference; never drops a model)",
+        "description": "Bind default requests to each model's best provider route (soft preference; never drops a model)",
         "default": false
       },
       "roles.default.required": {
@@ -193,6 +193,12 @@
         "default": 0,
         "min": 0
       },
+      "roles.default.filters.preferredMinThroughput": {
+        "type": "number",
+        "description": "Soft throughput floor for default routes (tok/s; 0 = off) — a below-floor route sorts last but is never dropped",
+        "default": 0,
+        "min": 0
+      },
       "roles.default.lambda": {
         "type": "number",
         "description": "Explicit λ ($ per quality point) override for default",
@@ -295,7 +301,7 @@
       },
       "roles.smol.features.providerPinning": {
         "type": "boolean",
-        "description": "Enable the Prefer the model's own lab endpoint capability for the smol role",
+        "description": "Enable the Best-route provider pinning capability for the smol role",
         "default": false
       },
       "roles.smol.features.costCap": {
@@ -305,7 +311,7 @@
       },
       "roles.smol.preferOwnProvider": {
         "type": "boolean",
-        "description": "Price smol models on their own lab's route where one exists (soft preference; never drops a model)",
+        "description": "Bind smol requests to each model's best provider route (soft preference; never drops a model)",
         "default": false
       },
       "roles.smol.required": {
@@ -341,6 +347,12 @@
         "default": 0,
         "min": 0
       },
+      "roles.smol.filters.preferredMinThroughput": {
+        "type": "number",
+        "description": "Soft throughput floor for smol routes (tok/s; 0 = off) — a below-floor route sorts last but is never dropped",
+        "default": 0,
+        "min": 0
+      },
       "roles.smol.lambda": {
         "type": "number",
         "description": "Explicit λ ($ per quality point) override for smol",
@@ -435,7 +447,7 @@
       },
       "roles.slow.features.providerPinning": {
         "type": "boolean",
-        "description": "Enable the Prefer the model's own lab endpoint capability for the slow role",
+        "description": "Enable the Best-route provider pinning capability for the slow role",
         "default": false
       },
       "roles.slow.features.costCap": {
@@ -445,7 +457,7 @@
       },
       "roles.slow.preferOwnProvider": {
         "type": "boolean",
-        "description": "Price slow models on their own lab's route where one exists (soft preference; never drops a model)",
+        "description": "Bind slow requests to each model's best provider route (soft preference; never drops a model)",
         "default": false
       },
       "roles.slow.required": {
@@ -481,6 +493,12 @@
         "default": 0,
         "min": 0
       },
+      "roles.slow.filters.preferredMinThroughput": {
+        "type": "number",
+        "description": "Soft throughput floor for slow routes (tok/s; 0 = off) — a below-floor route sorts last but is never dropped",
+        "default": 0,
+        "min": 0
+      },
       "roles.slow.lambda": {
         "type": "number",
         "description": "Explicit λ ($ per quality point) override for slow",
@@ -575,7 +593,7 @@
       },
       "roles.vision.features.providerPinning": {
         "type": "boolean",
-        "description": "Enable the Prefer the model's own lab endpoint capability for the vision role",
+        "description": "Enable the Best-route provider pinning capability for the vision role",
         "default": false
       },
       "roles.vision.features.costCap": {
@@ -585,7 +603,7 @@
       },
       "roles.vision.preferOwnProvider": {
         "type": "boolean",
-        "description": "Price vision models on their own lab's route where one exists (soft preference; never drops a model)",
+        "description": "Bind vision requests to each model's best provider route (soft preference; never drops a model)",
         "default": false
       },
       "roles.vision.required": {
@@ -621,6 +639,12 @@
         "default": 0,
         "min": 0
       },
+      "roles.vision.filters.preferredMinThroughput": {
+        "type": "number",
+        "description": "Soft throughput floor for vision routes (tok/s; 0 = off) — a below-floor route sorts last but is never dropped",
+        "default": 0,
+        "min": 0
+      },
       "roles.vision.lambda": {
         "type": "number",
         "description": "Explicit λ ($ per quality point) override for vision",
@@ -723,7 +747,7 @@
       },
       "roles.plan.features.providerPinning": {
         "type": "boolean",
-        "description": "Enable the Prefer the model's own lab endpoint capability for the plan role",
+        "description": "Enable the Best-route provider pinning capability for the plan role",
         "default": false
       },
       "roles.plan.features.costCap": {
@@ -733,7 +757,7 @@
       },
       "roles.plan.preferOwnProvider": {
         "type": "boolean",
-        "description": "Price plan models on their own lab's route where one exists (soft preference; never drops a model)",
+        "description": "Bind plan requests to each model's best provider route (soft preference; never drops a model)",
         "default": false
       },
       "roles.plan.required": {
@@ -769,6 +793,12 @@
         "default": 0,
         "min": 0
       },
+      "roles.plan.filters.preferredMinThroughput": {
+        "type": "number",
+        "description": "Soft throughput floor for plan routes (tok/s; 0 = off) — a below-floor route sorts last but is never dropped",
+        "default": 0,
+        "min": 0
+      },
       "roles.plan.lambda": {
         "type": "number",
         "description": "Explicit λ ($ per quality point) override for plan",
@@ -855,7 +885,7 @@
       },
       "roles.commit.features.providerPinning": {
         "type": "boolean",
-        "description": "Enable the Prefer the model's own lab endpoint capability for the commit role",
+        "description": "Enable the Best-route provider pinning capability for the commit role",
         "default": false
       },
       "roles.commit.features.costCap": {
@@ -865,7 +895,7 @@
       },
       "roles.commit.preferOwnProvider": {
         "type": "boolean",
-        "description": "Price commit models on their own lab's route where one exists (soft preference; never drops a model)",
+        "description": "Bind commit requests to each model's best provider route (soft preference; never drops a model)",
         "default": false
       },
       "roles.commit.required": {
@@ -901,6 +931,12 @@
         "default": 0,
         "min": 0
       },
+      "roles.commit.filters.preferredMinThroughput": {
+        "type": "number",
+        "description": "Soft throughput floor for commit routes (tok/s; 0 = off) — a below-floor route sorts last but is never dropped",
+        "default": 0,
+        "min": 0
+      },
       "roles.commit.lambda": {
         "type": "number",
         "description": "Explicit λ ($ per quality point) override for commit",
@@ -987,7 +1023,7 @@
       },
       "roles.tiny.features.providerPinning": {
         "type": "boolean",
-        "description": "Enable the Prefer the model's own lab endpoint capability for the tiny role",
+        "description": "Enable the Best-route provider pinning capability for the tiny role",
         "default": false
       },
       "roles.tiny.features.costCap": {
@@ -997,7 +1033,7 @@
       },
       "roles.tiny.preferOwnProvider": {
         "type": "boolean",
-        "description": "Price tiny models on their own lab's route where one exists (soft preference; never drops a model)",
+        "description": "Bind tiny requests to each model's best provider route (soft preference; never drops a model)",
         "default": false
       },
       "roles.tiny.required": {
@@ -1033,6 +1069,12 @@
         "default": 0,
         "min": 0
       },
+      "roles.tiny.filters.preferredMinThroughput": {
+        "type": "number",
+        "description": "Soft throughput floor for tiny routes (tok/s; 0 = off) — a below-floor route sorts last but is never dropped",
+        "default": 0,
+        "min": 0
+      },
       "roles.tiny.lambda": {
         "type": "number",
         "description": "Explicit λ ($ per quality point) override for tiny",
@@ -1111,7 +1153,7 @@
       },
       "roles.task.features.providerPinning": {
         "type": "boolean",
-        "description": "Enable the Prefer the model's own lab endpoint capability for the task role",
+        "description": "Enable the Best-route provider pinning capability for the task role",
         "default": false
       },
       "roles.task.features.costCap": {
@@ -1121,7 +1163,7 @@
       },
       "roles.task.preferOwnProvider": {
         "type": "boolean",
-        "description": "Price task models on their own lab's route where one exists (soft preference; never drops a model)",
+        "description": "Bind task requests to each model's best provider route (soft preference; never drops a model)",
         "default": false
       },
       "roles.task.required": {
@@ -1157,6 +1199,12 @@
         "default": 0,
         "min": 0
       },
+      "roles.task.filters.preferredMinThroughput": {
+        "type": "number",
+        "description": "Soft throughput floor for task routes (tok/s; 0 = off) — a below-floor route sorts last but is never dropped",
+        "default": 0,
+        "min": 0
+      },
       "roles.task.lambda": {
         "type": "number",
         "description": "Explicit λ ($ per quality point) override for task",
@@ -1259,7 +1307,7 @@
       },
       "roles.advisor.features.providerPinning": {
         "type": "boolean",
-        "description": "Enable the Prefer the model's own lab endpoint capability for the advisor role",
+        "description": "Enable the Best-route provider pinning capability for the advisor role",
         "default": false
       },
       "roles.advisor.features.costCap": {
@@ -1269,7 +1317,7 @@
       },
       "roles.advisor.preferOwnProvider": {
         "type": "boolean",
-        "description": "Price advisor models on their own lab's route where one exists (soft preference; never drops a model)",
+        "description": "Bind advisor requests to each model's best provider route (soft preference; never drops a model)",
         "default": false
       },
       "roles.advisor.required": {
@@ -1305,6 +1353,12 @@
         "default": 0,
         "min": 0
       },
+      "roles.advisor.filters.preferredMinThroughput": {
+        "type": "number",
+        "description": "Soft throughput floor for advisor routes (tok/s; 0 = off) — a below-floor route sorts last but is never dropped",
+        "default": 0,
+        "min": 0
+      },
       "roles.advisor.lambda": {
         "type": "number",
         "description": "Explicit λ ($ per quality point) override for advisor",
@@ -1391,7 +1445,7 @@
       },
       "roles.designer.features.providerPinning": {
         "type": "boolean",
-        "description": "Enable the Prefer the model's own lab endpoint capability for the designer role",
+        "description": "Enable the Best-route provider pinning capability for the designer role",
         "default": false
       },
       "roles.designer.features.costCap": {
@@ -1401,7 +1455,7 @@
       },
       "roles.designer.preferOwnProvider": {
         "type": "boolean",
-        "description": "Price designer models on their own lab's route where one exists (soft preference; never drops a model)",
+        "description": "Bind designer requests to each model's best provider route (soft preference; never drops a model)",
         "default": false
       },
       "roles.designer.required": {
@@ -1437,6 +1491,12 @@
         "default": 0,
         "min": 0
       },
+      "roles.designer.filters.preferredMinThroughput": {
+        "type": "number",
+        "description": "Soft throughput floor for designer routes (tok/s; 0 = off) — a below-floor route sorts last but is never dropped",
+        "default": 0,
+        "min": 0
+      },
       "roles.designer.lambda": {
         "type": "number",
         "description": "Explicit λ ($ per quality point) override for designer",

package/src/availability.ts

@@ -88,6 +88,24 @@ export type ProbeVerdict = "ok" | "blocked" | "unknown";
 
 const OPENROUTER_COMPLETIONS_URL = "https://openrouter.ai/api/v1/chat/completions";
 
+/** POST a one-token completion to OpenRouter with the shared request skeleton
+ * (Bearer auth, JSON body, 20s timeout). Shared by `probeModel` and
+ * `fetchAllowedProviders`. */
+function completionRequest(token: string, body: unknown, fetchImpl: typeof fetch): Promise<Response> {
+  return fetchImpl(OPENROUTER_COMPLETIONS_URL, {
+    method: "POST",
+    headers: { authorization: `Bearer ${token}`, "content-type": "application/json" },
+    body: JSON.stringify(body),
+    signal: AbortSignal.timeout(20_000),
+  });
+}
+
+/** Extract `error.message` from a parsed OpenRouter error body, or "" when the
+ * body does not carry one. */
+function errorMessage(body: unknown): string {
+  return isRecord(body) && isRecord(body.error) && typeof body.error.message === "string" ? body.error.message : "";
+}
+
 /** 404 body OpenRouter returns when the account's allowed-providers privacy
  * whitelist (or a data-policy toggle) leaves a model without a runnable endpoint. */
 const BLOCKED_RE = /no allowed providers/i;
@@ -102,22 +120,71 @@ const BLOCKED_RE = /no allowed providers/i;
  */
 export async function probeModel(token: string, catalogId: string, fetchImpl: typeof fetch = fetch): Promise<ProbeVerdict> {
   try {
-    const res = await fetchImpl(OPENROUTER_COMPLETIONS_URL, {
-      method: "POST",
-      headers: { authorization: `Bearer ${token}`, "content-type": "application/json" },
-      body: JSON.stringify({ model: catalogId, messages: [{ role: "user", content: "ok" }], max_tokens: 1 }),
-      signal: AbortSignal.timeout(20_000),
-    });
+    const res = await completionRequest(token, { model: catalogId, messages: [{ role: "user", content: "ok" }], max_tokens: 1 }, fetchImpl);
     if (res.ok) return "ok";
     if (res.status !== 404) return "unknown";
     const body: unknown = await res.json().catch(() => null);
-    const message = isRecord(body) && isRecord(body.error) && typeof body.error.message === "string" ? body.error.message : "";
+    const message = errorMessage(body);
     return BLOCKED_RE.test(message) ? "blocked" : "unknown";
   } catch {
     return "unknown";
   }
 }
 
+/** 404 body OpenRouter returns when the account's allowed-providers privacy
+ * whitelist leaves the canary provider without a runnable endpoint; the message
+ * names the whitelist after "permits only:". */
+const ALLOWED_PROVIDERS_RE = /permits only:\s*([^.]*)\./i;
+
+/**
+ * Harvest the account's allowed-providers privacy whitelist. The whitelist is
+ * invisible to `/api/v1/key` and the catalog endpoints; the only place it
+ * appears is the 404 body of a request that fails because of it. A canary
+ * completion on a valid model with a nonexistent provider slug (`provider.only`)
+ * 404s and names the whitelist. Returns null when the body does not name a
+ * whitelist (no whitelist set), the response is not a 404, or the request
+ * throws — never throws.
+ */
+export async function fetchAllowedProviders(token: string, modelId: string, fetchImpl: typeof fetch = fetch): Promise<Set<string> | null> {
+  try {
+    const res = await completionRequest(
+      token,
+      {
+        model: modelId,
+        messages: [{ role: "user", content: "ok" }],
+        max_tokens: 1,
+        provider: { only: ["__omp-llm-role-canary__"] },
+      },
+      fetchImpl,
+    );
+    if (res.status !== 404) return null;
+    const body: unknown = await res.json().catch(() => null);
+    const message = errorMessage(body);
+    const match = ALLOWED_PROVIDERS_RE.exec(message);
+    if (!match) return null;
+    const providers = match[1]
+      .split(",")
+      .map((s) => s.trim())
+      .filter((s) => s.length > 0);
+    return providers.length > 0 ? new Set(providers) : null;
+  } catch {
+    return null;
+  }
+}
+
+/**
+ * Whether a route's `providerSlug` is permitted by the account's
+ * allowed-providers whitelist. Whitelist entries are bare lab names
+ * (`coreweave`) while route slugs are tiered (`coreweave/fp8`), so a match is
+ * the entry itself or an `entry/` prefix.
+ */
+export function providerAllowed(allowed: ReadonlySet<string>, slug: string): boolean {
+  for (const entry of allowed) {
+    if (slug === entry || slug.startsWith(`${entry}/`)) return true;
+  }
+  return false;
+}
+
 // ---------------------------------------------------------------------------
 // Keyed-catalog availability
 // ---------------------------------------------------------------------------
@@ -228,12 +295,23 @@ export function rankingIdOf(catalogId: string): string {
   return noFree.endsWith("-latest") ? noFree.slice(0, -"-latest".length) : noFree;
 }
 
+/** Split a selector's trailing thinking suffix (`provider/model[:level]`):
+ * `base` without the suffix, `level` the suffix or null. The single home for
+ * the `:level` strip shared by `currentRankingId`, `providerPinOfSelector`, and
+ * the extension's selector handling. */
+export function splitThinkingSuffix(selector: string): { base: string; level: string | null } {
+  const colon = selector.lastIndexOf(":");
+  if (colon !== -1 && selector.slice(colon + 1) in THINKING_LEVELS) {
+    return { base: selector.slice(0, colon), level: selector.slice(colon + 1) };
+  }
+  return { base: selector, level: null };
+}
+
 /** Identity of a configured selector for hysteresis/chain keys: drop `:level`,
  * `openrouter/`, and a trailing `@<slug>` provider pin, then rankingIdOf. A
  * tiered pin (`@deepinfra/fp8`) is dropped whole — everything from the last `@`. */
 export function currentRankingId(selector: string): string {
-  const colon = selector.lastIndexOf(":");
-  const base = colon !== -1 && selector.slice(colon + 1) in THINKING_LEVELS ? selector.slice(0, colon) : selector;
+  const { base } = splitThinkingSuffix(selector);
   const unprefixed = base.startsWith("openrouter/") ? base.slice("openrouter/".length) : base;
   const at = unprefixed.lastIndexOf("@");
   const unpinned = at !== -1 ? unprefixed.slice(0, at) : unprefixed;

package/src/cli/create-role.ts

@@ -50,7 +50,7 @@ const USAGE = [
   "  --lambda <n>           explicit λ override ($ per quality point)",
   "  --feature <id,...>     capability flag(s) to author on the role (see --list-features);",
   "                         repeatable, comma-separated, merged",
-  "  --prefer-own-provider  price the role on each model's own lab route where one exists",
+  "  --prefer-own-provider  bind the role to each model's best provider route where one exists",
   "  --list-features        print the capability flags and the settings each applies, then exit",
   "  --lock <path>          settings lock file (default ~/.omp/plugins/omp-plugins.lock.json)",
   "  --dry-run              validate and print without writing",