Merge remote-tracking branch 'origin/main' into chore/autofix-bot-issues-2026-07-31

erin2722 · erin2722 · commit 400beced7822 · 2026-07-31T14:27:45.000Z
diff --git a/.github/workflows/sync-models.yaml b/.github/workflows/sync-models.yaml
@@ -99,6 +99,23 @@ jobs:
 
           pnpm exec tsx packages/proxy/scripts/sync_models.ts sync-baseten --write
 
+      - name: Sync Cohere models from /v1/models
+        env:
+          COHERE_API_KEY: ${{ secrets.COHERE_API_KEY }}
+        run: |
+          set -euo pipefail
+
+          # Cohere's /v1/models is authoritative for the Cohere-hosted chat models
+          # (LiteLLM lags it and even carries models Cohere has removed). Pricing
+          # is overlaid from LiteLLM where available. Skip cleanly if the key is
+          # not configured yet so the rest of the sync still runs.
+          if [ -z "${COHERE_API_KEY:-}" ]; then
+            echo "COHERE_API_KEY not set; skipping Cohere model sync."
+            exit 0
+          fi
+
+          pnpm exec tsx packages/proxy/scripts/sync_models.ts sync-cohere --write
+
       - name: Sync OpenRouter alternates from /api/v1/models
         run: |
           set -euo pipefail
diff --git a/packages/proxy/schema/index.ts b/packages/proxy/schema/index.ts
@@ -1317,6 +1317,18 @@ export const AvailableEndpointTypes: { [name: string]: ModelEndpointType[] } = {
   "command-r-plus-08-2024": ["cohere", "openrouter"],
   "command-r-08-2024": ["cohere", "openrouter"],
   "command-r7b-12-2024": ["cohere", "openrouter"],
+  "c4ai-aya-expanse-32b": ["cohere"],
+  "c4ai-aya-vision-32b": ["cohere"],
+  "command-a-plus-05-2026": ["cohere"],
+  "command-a-reasoning-08-2025": ["cohere"],
+  "command-a-translate-08-2025": ["cohere"],
+  "command-a-vision-07-2025": ["cohere"],
+  "command-r7b-arabic-02-2025": ["cohere"],
+  "north-mini-code-1-0": ["cohere"],
+  "tiny-aya-earth": ["cohere"],
+  "tiny-aya-fire": ["cohere"],
+  "tiny-aya-global": ["cohere"],
+  "tiny-aya-water": ["cohere"],
   "databricks-gpt-5-5-pro": ["databricks"],
   "databricks-gemini-3-5-flash": ["databricks"],
   "databricks-gpt-5-3-codex": ["databricks"],
diff --git a/packages/proxy/schema/model_list.json b/packages/proxy/schema/model_list.json
@@ -10022,6 +10022,102 @@
       "bedrock"
     ]
   },
+  "tiny-aya-water": {
+    "format": "openai",
+    "flavor": "chat",
+    "max_input_tokens": 8192,
+    "available_providers": [
+      "cohere"
+    ]
+  },
+  "tiny-aya-global": {
+    "format": "openai",
+    "flavor": "chat",
+    "max_input_tokens": 8192,
+    "available_providers": [
+      "cohere"
+    ]
+  },
+  "tiny-aya-fire": {
+    "format": "openai",
+    "flavor": "chat",
+    "max_input_tokens": 8192,
+    "available_providers": [
+      "cohere"
+    ]
+  },
+  "tiny-aya-earth": {
+    "format": "openai",
+    "flavor": "chat",
+    "max_input_tokens": 8192,
+    "available_providers": [
+      "cohere"
+    ]
+  },
+  "north-mini-code-1-0": {
+    "format": "openai",
+    "flavor": "chat",
+    "max_input_tokens": 436000,
+    "available_providers": [
+      "cohere"
+    ]
+  },
+  "command-r7b-arabic-02-2025": {
+    "format": "openai",
+    "flavor": "chat",
+    "max_input_tokens": 128000,
+    "available_providers": [
+      "cohere"
+    ]
+  },
+  "command-a-vision-07-2025": {
+    "format": "openai",
+    "flavor": "chat",
+    "max_input_tokens": 128000,
+    "available_providers": [
+      "cohere"
+    ]
+  },
+  "command-a-translate-08-2025": {
+    "format": "openai",
+    "flavor": "chat",
+    "max_input_tokens": 8992,
+    "available_providers": [
+      "cohere"
+    ]
+  },
+  "command-a-reasoning-08-2025": {
+    "format": "openai",
+    "flavor": "chat",
+    "max_input_tokens": 288768,
+    "available_providers": [
+      "cohere"
+    ]
+  },
+  "command-a-plus-05-2026": {
+    "format": "openai",
+    "flavor": "chat",
+    "max_input_tokens": 436000,
+    "available_providers": [
+      "cohere"
+    ]
+  },
+  "c4ai-aya-vision-32b": {
+    "format": "openai",
+    "flavor": "chat",
+    "max_input_tokens": 16384,
+    "available_providers": [
+      "cohere"
+    ]
+  },
+  "c4ai-aya-expanse-32b": {
+    "format": "openai",
+    "flavor": "chat",
+    "max_input_tokens": 128000,
+    "available_providers": [
+      "cohere"
+    ]
+  },
   "command-a-03-2025": {
     "format": "openai",
     "flavor": "chat",
diff --git a/packages/proxy/scripts/model_deprecation.test.ts b/packages/proxy/scripts/model_deprecation.test.ts
@@ -33,6 +33,21 @@ describe("openrouter deprecation adapter", () => {
   });
 });
 
+describe("cohere deprecation adapter", () => {
+  it("is registered as an authoritative, list-only, non-report-only provider", () => {
+    const api = PROVIDER_APIS.cohere;
+    expect(api).toBeDefined();
+    // Absence from Cohere's /v1/models alone marks a model deprecated (Cohere
+    // hosts only its own models), so the list is authoritative and there is no
+    // probe. This is what prunes models Cohere removes but LiteLLM still lists.
+    expect(api.listIsAuthoritative).toBe(true);
+    expect(api.probeModel).toBeNull();
+    expect(api.listModels).not.toBeNull();
+    // Must NOT be report-only, or deprecations would never be applied.
+    expect(REPORT_ONLY_PROVIDERS.has("cohere")).toBe(false);
+  });
+});
+
 describe("classifyProbe", () => {
   it("treats 2xx as active", () => {
     expect(classifyProbe(200, "{}")).toBe("active");
diff --git a/packages/proxy/scripts/model_probe.ts b/packages/proxy/scripts/model_probe.ts
@@ -97,6 +97,14 @@ const listSchema = z
   })
   .passthrough();
 
+// Cohere's /v1/models is paginated: { models: [{ name }], next_page_token }.
+const cohereListSchema = z
+  .object({
+    models: z.array(z.object({ name: z.string() }).passthrough()),
+    next_page_token: z.string().nullish(),
+  })
+  .passthrough();
+
 // OpenAI-compatible /models responses come back either as { data: [{id}] } or,
 // for some providers (e.g. Together), as a bare array of { id }.
 function extractOpenAiModelIds(body: string): Set<string> {
@@ -300,6 +308,46 @@ export const PROVIDER_APIS: Record<string, ProviderApi> = {
       return new Set((parsed.endpoints ?? []).map((e) => e.name));
     },
   },
+  // Cohere hosts only its own models and its paginated /v1/models is the
+  // authoritative directory of what it currently serves. A catalog model with
+  // the cohere provider that is no longer listed is deprecated (list-only, no
+  // probe) — this is what catches models Cohere removes (e.g. `command-r`,
+  // removed 2025-09-15) that LiteLLM still lists. We list ALL models (no endpoint
+  // filter) so a model still served on any endpoint is not falsely deprecated;
+  // Cohere ids are bare names, matching catalog keys directly.
+  cohere: {
+    listIsAuthoritative: true,
+    probeModel: null,
+    listModels: async (secret) => {
+      const names = new Set<string>();
+      let pageToken: string | undefined;
+      // Bound the pagination so a malformed next_page_token cannot loop forever.
+      for (let page = 0; page < 100; page++) {
+        const url = new URL("https://api.cohere.com/v1/models");
+        url.searchParams.set("page_size", "1000");
+        if (pageToken) {
+          url.searchParams.set("page_token", pageToken);
+        }
+        const { status, body } = await request("GET", url.toString(), {
+          authorization: `Bearer ${secret.secret}`,
+        });
+        if (status >= 400) {
+          throw new Error(
+            `list cohere /v1/models -> HTTP ${status}: ${body.slice(0, 160)}`,
+          );
+        }
+        const parsed = cohereListSchema.parse(JSON.parse(body));
+        for (const model of parsed.models) {
+          names.add(model.name);
+        }
+        if (!parsed.next_page_token) {
+          return names;
+        }
+        pageToken = parsed.next_page_token;
+      }
+      return names;
+    },
+  },
   // OpenRouter's /api/v1/models is a public, authoritative directory of what it
   // serves. The audit uses it to proactively prune openrouter: a catalog model
   // with the openrouter provider that is no longer listed loses openrouter (and
diff --git a/packages/proxy/scripts/sync_models.test.ts b/packages/proxy/scripts/sync_models.test.ts
@@ -6,6 +6,9 @@ import {
   applyEquivalentModels,
   canonicalizeLocalModelsContent,
   convertBasetenToLocalModel,
+  convertCohereToLocalModel,
+  applyCohereLiteLLMPricing,
+  isSupportedCohereChatModel,
   convertRemoteToLocalModel,
   applyBasetenPricing,
   applyOpenRouterPricing,
@@ -953,3 +956,96 @@ describe("applyOpenRouterPricing", () => {
     expect(applyOpenRouterPricing("x-ai/grok-4.5", model, orModel)).toBeNull();
   });
 });
+
+describe("convertCohereToLocalModel", () => {
+  it("maps a Cohere chat model to an openai-format cohere entry (no LiteLLM = no pricing)", () => {
+    const spec = convertCohereToLocalModel({
+      name: "command-a-reasoning-08-2025",
+      endpoints: ["chat"],
+      context_length: 288768,
+    });
+    expect(spec.format).toBe("openai");
+    expect(spec.flavor).toBe("chat");
+    expect(spec.max_input_tokens).toBe(288768);
+    expect(spec.available_providers).toEqual(["cohere"]);
+    // No LiteLLM overlay -> no fabricated pricing.
+    expect(spec.input_cost_per_mil_tokens).toBeUndefined();
+    expect(spec.output_cost_per_mil_tokens).toBeUndefined();
+  });
+
+  it("overlays LiteLLM pricing (per-token -> per-mil) when available", () => {
+    const spec = convertCohereToLocalModel(
+      {
+        name: "command-a-03-2025",
+        endpoints: ["chat"],
+        context_length: 256000,
+      },
+      {
+        input_cost_per_token: 2.5e-6,
+        output_cost_per_token: 1e-5,
+        cache_read_input_token_cost: 1.25e-6,
+      },
+    );
+    expect(spec.input_cost_per_mil_tokens).toBe(2.5);
+    expect(spec.output_cost_per_mil_tokens).toBe(10);
+    expect(spec.input_cache_read_cost_per_mil_tokens).toBe(1.25);
+  });
+});
+
+describe("isSupportedCohereChatModel", () => {
+  it("accepts a live chat model", () => {
+    expect(
+      isSupportedCohereChatModel({
+        name: "command-a-03-2025",
+        endpoints: ["chat"],
+      }),
+    ).toBe(true);
+  });
+
+  it("rejects non-chat, fine-tuned, and deprecated models", () => {
+    expect(
+      isSupportedCohereChatModel({ name: "embed-v4.0", endpoints: ["embed"] }),
+    ).toBe(false);
+    expect(
+      isSupportedCohereChatModel({
+        name: "my-finetune",
+        endpoints: ["chat"],
+        finetuned: true,
+      }),
+    ).toBe(false);
+    expect(
+      isSupportedCohereChatModel({
+        name: "command-r",
+        endpoints: ["chat"],
+        is_deprecated: true,
+      }),
+    ).toBe(false);
+  });
+});
+
+describe("applyCohereLiteLLMPricing", () => {
+  const priceless: ModelSpec = {
+    format: "openai",
+    flavor: "chat",
+    available_providers: ["cohere"],
+  };
+
+  it("fills pricing on an existing price-less entry when LiteLLM has it", () => {
+    const priced = applyCohereLiteLLMPricing("command-a-03-2025", priceless, {
+      input_cost_per_token: 2.5e-6,
+      output_cost_per_token: 1e-5,
+      cache_read_input_token_cost: 1.25e-6,
+    });
+    expect(priced?.input_cost_per_mil_tokens).toBe(2.5);
+    expect(priced?.output_cost_per_mil_tokens).toBe(10);
+    expect(priced?.input_cache_read_cost_per_mil_tokens).toBe(1.25);
+  });
+
+  it("returns null when LiteLLM has no entry or pricing is unchanged", () => {
+    expect(applyCohereLiteLLMPricing("x", priceless, undefined)).toBeNull();
+    const already: ModelSpec = { ...priceless, input_cost_per_mil_tokens: 2.5 };
+    expect(
+      applyCohereLiteLLMPricing("x", already, { input_cost_per_token: 2.5e-6 }),
+    ).toBeNull();
+  });
+});
diff --git a/packages/proxy/scripts/sync_models.ts b/packages/proxy/scripts/sync_models.ts
