feat(snowflake-cortex): close model-coverage gap vs Snowflake docs (#851)

docs/docs/configure/providers.md

@@ -280,15 +280,39 @@ Billing flows through your Snowflake credits — no per-token costs.
 
 | Model | Tool Calling |
 |-------|-------------|
-| `claude-sonnet-4-6`, `claude-opus-4-6`, `claude-sonnet-4-5`, `claude-opus-4-5`, `claude-haiku-4-5`, `claude-4-sonnet`, `claude-3-7-sonnet`, `claude-3-5-sonnet` | Yes |
-| `openai-gpt-4.1`, `openai-gpt-5`, `openai-gpt-5-mini`, `openai-gpt-5-nano`, `openai-gpt-5-chat` | Yes |
-| `llama4-maverick`, `snowflake-llama-3.3-70b`, `llama3.1-70b`, `llama3.1-405b`, `llama3.1-8b` | No |
-| `mistral-large`, `mistral-large2`, `mistral-7b` | No |
-| `deepseek-r1` | No |
+| `claude-opus-4-7`, `claude-sonnet-4-6`, `claude-opus-4-6`, `claude-sonnet-4-5`, `claude-opus-4-5`, `claude-haiku-4-5`, `claude-4-sonnet`, `claude-3-7-sonnet`, `claude-3-5-sonnet` | Yes |
+| `openai-gpt-4.1`, `openai-gpt-5`, `openai-gpt-5.1`, `openai-gpt-5.2`, `openai-gpt-5-mini`, `openai-gpt-5-nano`, `openai-gpt-5-chat` | Yes |
+| `llama4-maverick`, `llama4-scout`, `llama3.3-70b`, `snowflake-llama-3.3-70b`, `snowflake-llama-3.1-405b`, `llama3.1-70b`, `llama3.1-405b`, `llama3.1-8b` | No |
+| `mistral-large`, `mistral-large2`, `mistral-7b`, `mixtral-8x7b` | No |
+| `deepseek-r1`, `gemini-3.1-pro` | No |
 
 !!! note
     Model availability depends on your Snowflake region. Enable cross-region inference with `ALTER ACCOUNT SET CORTEX_ENABLED_CROSS_REGION = 'ANY_REGION'` for full model access.
 
+### Adding a model not in the list
+
+Snowflake Cortex adds models faster than this list can be updated. If a model is available on your Cortex account but not yet listed above, you can register it locally without forking the CLI — add it under `provider["snowflake-cortex"].models` in your `altimate-code.json` (or `.altimate-code/altimate-code.json`):
+
+```json
+{
+  "provider": {
+    "snowflake-cortex": {
+      "models": {
+        "your-new-model-id": {
+          "name": "Your New Model",
+          "limit": { "context": 200000, "output": 32000 },
+          "tool_call": true
+        }
+      }
+    }
+  }
+}
+```
+
+The entry merges with the built-in list, so the model appears in the picker and can be selected as `snowflake-cortex/your-new-model-id`. Set `"tool_call": false` for models that don't support tools on Cortex (Llama, Mistral, DeepSeek, Gemini today) — otherwise requests with tools will fail.
+
+The `tool_call` field uses snake_case (matching the rest of the `altimate-code.json` schema) and maps to the picker's `capabilities.toolcall`. The request transform reads the same value, so a user-added model marked `tool_call: true` keeps `tools` and `tool_choice` in outgoing requests — and one marked `tool_call: false` has them stripped, the same as the built-in non-tool entries.
+
 ## Databricks AI Gateway
 
 Connect to Databricks serving endpoints (Foundation Model APIs) via your workspace PAT. Use Databricks-hosted Llama, Claude, GPT, Gemini, DBRX, or Mixtral for agent reasoning — billing flows through your Databricks account.

packages/opencode/src/altimate/plugin/snowflake.ts

@@ -1,16 +1,41 @@
 import type { Hooks, PluginInput } from "@opencode-ai/plugin"
 import { Auth, OAUTH_DUMMY_KEY } from "@/auth"
 
-// Only OpenAI and Claude models support tool calling on Snowflake Cortex.
-// All other models reject tools with "tool calling is not supported".
-const TOOLCALL_MODELS = new Set([
-  // Claude
-  "claude-sonnet-4-6", "claude-opus-4-6", "claude-sonnet-4-5", "claude-opus-4-5",
-  "claude-haiku-4-5", "claude-4-sonnet", "claude-4-opus", "claude-3-7-sonnet", "claude-3-5-sonnet",
-  // OpenAI
-  "openai-gpt-4.1", "openai-gpt-5", "openai-gpt-5-mini", "openai-gpt-5-nano",
-  "openai-gpt-5-chat", "openai-gpt-oss-120b", "openai-o4-mini",
-])
+/**
+ * Build the set of Snowflake Cortex model IDs that support tool calling.
+ * Derived from `provider.models[*].capabilities.toolcall` so the picker's
+ * advertised capability and the request-transform's behavior cannot drift —
+ * adding a tool-capable model to the provider definition (or registering one
+ * via `altimate-code.json`) is the single source of truth.
+ *
+ * Indexes both the picker key and the model's `api.id` (and `id`) so that
+ * aliased models — where a user registers e.g. `"my-alias": { "id":
+ * "claude-opus-4-7" }` in their config — match correctly. The transform
+ * compares against `parsed.model` which is the API id sent in the request
+ * body, not the picker map key.
+ *
+ * Snowflake's documented constraint: only Claude and OpenAI models accept
+ * tools on Cortex; everything else rejects with "tool calling is not supported."
+ */
+export function buildToolCapableSet(
+  models: Record<
+    string,
+    {
+      id?: string
+      api?: { id?: string }
+      capabilities: { toolcall: boolean }
+    }
+  >,
+): ReadonlySet<string> {
+  const set = new Set<string>()
+  for (const [key, m] of Object.entries(models)) {
+    if (!m.capabilities.toolcall) continue
+    set.add(key)
+    if (m.id) set.add(m.id)
+    if (m.api?.id) set.add(m.api.id)
+  }
+  return set
+}
 
 /** Snowflake account identifiers contain only alphanumeric, hyphen, underscore, and dot characters. */
 export const VALID_ACCOUNT_RE = /^[a-zA-Z0-9._-]+$/
@@ -29,8 +54,15 @@ export function parseSnowflakePAT(code: string): { account: string; token: strin
 /**
  * Transform a Snowflake Cortex request body string.
  * Returns a Response to short-circuit the fetch (synthetic stop), or undefined to continue normally.
+ *
+ * @param toolCapable Model IDs that should retain `tools` / `tool_choice` / tool messages.
+ *                    Build via `buildToolCapableSet(provider.models)` at loader time so
+ *                    user-added models with `tool_call: true` in `altimate-code.json` are honored.
  */
-export function transformSnowflakeBody(bodyText: string): { body: string; syntheticStop?: Response } {
+export function transformSnowflakeBody(
+  bodyText: string,
+  toolCapable: ReadonlySet<string>,
+): { body: string; syntheticStop?: Response } {
   const parsed = JSON.parse(bodyText)
 
   // Snowflake uses max_completion_tokens instead of max_tokens
@@ -41,7 +73,7 @@ export function transformSnowflakeBody(bodyText: string): { body: string; synthe
 
   // Strip tools for models that don't support tool calling on Snowflake Cortex.
   // Also remove orphaned tool_calls from messages to avoid Snowflake API errors.
-  if (!TOOLCALL_MODELS.has(parsed.model)) {
+  if (!toolCapable.has(parsed.model)) {
     delete parsed.tools
     delete parsed.tool_choice
     if (Array.isArray(parsed.messages)) {
@@ -97,6 +129,14 @@ export async function SnowflakeCortexAuthPlugin(_input: PluginInput): Promise<Ho
           model.cost = { input: 0, output: 0, cache: { read: 0, write: 0 } }
         }
 
+        // Build the tool-capable allowlist from the live provider definition.
+        // This includes both hardcoded entries in provider.ts AND any models the
+        // user registered via `altimate-code.json` with `"tool_call": true`.
+        // Without this, the documented escape hatch is silently broken — picker
+        // shows the model as tool-capable, but the transform strips tools at
+        // request time because a static hardcoded set never sees user additions.
+        const toolCapable = buildToolCapableSet(provider.models)
+
         return {
           apiKey: OAUTH_DUMMY_KEY,
           async fetch(requestInput: RequestInfo | URL, init?: RequestInit) {
@@ -134,7 +174,7 @@ export async function SnowflakeCortexAuthPlugin(_input: PluginInput): Promise<Ho
                   text = ""
                 }
                 if (text) {
-                  const result = transformSnowflakeBody(text)
+                  const result = transformSnowflakeBody(text, toolCapable)
                   if (result.syntheticStop) return result.syntheticStop
                   body = result.body
                   headers.delete("content-length")

packages/opencode/src/provider/provider.ts

@@ -1015,6 +1015,14 @@ export namespace Provider {
       return m
     }
 
+    // Snowflake Cortex models are hardcoded rather than auto-discovered.
+    // Cortex's models endpoint (`GET /api/v2/cortex/v1/models`) has
+    // non-standard semantics — it requires a JSON body on GET, which breaks
+    // standard OpenAI-compatible clients. Until that gets a probe-able shim,
+    // the source of truth for "what's selectable" is this map. New models
+    // surfaced on Snowflake's regional availability docs should be added here
+    // (or registered via `altimate-code.json` per the
+    // `docs/docs/configure/providers.md` escape-hatch section).
     database["snowflake-cortex"] = {
       id: ProviderID.snowflakeCortex,
       source: "custom",
@@ -1023,6 +1031,7 @@ export namespace Provider {
       options: {},
       models: {
         // Claude models — tool calling supported
+        "claude-opus-4-7": makeSnowflakeModel("claude-opus-4-7", "Claude Opus 4.7", { context: 1000000, output: 128000 }),
         "claude-sonnet-4-6": makeSnowflakeModel("claude-sonnet-4-6", "Claude Sonnet 4.6", {
           context: 200000,
           output: 64000,
@@ -1049,6 +1058,16 @@ export namespace Provider {
         }),
         // OpenAI models — tool calling supported
         "openai-gpt-4.1": makeSnowflakeModel("openai-gpt-4.1", "OpenAI GPT-4.1", { context: 1047576, output: 32768 }),
+        // openai-gpt-5.2: not in Snowflake's per-model restrictions table; using
+        // gpt-5 family defaults as best-effort until docs publish exact limits.
+        "openai-gpt-5.2": makeSnowflakeModel("openai-gpt-5.2", "OpenAI GPT-5.2", {
+          context: 272000,
+          output: 8192,
+        }),
+        "openai-gpt-5.1": makeSnowflakeModel("openai-gpt-5.1", "OpenAI GPT-5.1", {
+          context: 272000,
+          output: 8192,
+        }),
         "openai-gpt-5": makeSnowflakeModel("openai-gpt-5", "OpenAI GPT-5", { context: 1047576, output: 32768 }),
         "openai-gpt-5-mini": makeSnowflakeModel("openai-gpt-5-mini", "OpenAI GPT-5 Mini", {
           context: 1047576,
@@ -1070,10 +1089,35 @@ export namespace Provider {
           { context: 1048576, output: 4096 },
           { toolcall: false },
         ),
+        "llama4-scout": makeSnowflakeModel(
+          "llama4-scout",
+          "Llama 4 Scout",
+          { context: 128000, output: 8192 },
+          { toolcall: false },
+        ),
+        // llama3.3-70b: upstream Meta-hosted variant.
+        "llama3.3-70b": makeSnowflakeModel(
+          "llama3.3-70b",
+          "Llama 3.3 70B",
+          { context: 128000, output: 8192 },
+          { toolcall: false },
+        ),
+        // snowflake-llama-3.3-70b: Snowflake-hosted variant (different routing /
+        // region pinning vs the upstream `llama3.3-70b` above).
         "snowflake-llama-3.3-70b": makeSnowflakeModel(
           "snowflake-llama-3.3-70b",
           "Snowflake Llama 3.3 70B",
-          { context: 128000, output: 4096 },
+          { context: 128000, output: 8192 },
+          { toolcall: false },
+        ),
+        // snowflake-llama-3.1-405b: 8k context per Snowflake docs (much smaller
+        // than the upstream Meta model's window). Snowflake's table lists
+        // output=8192, but output cannot exceed the total token budget — cap
+        // at 4096 (the original sibling default) so prompt+output always fit.
+        "snowflake-llama-3.1-405b": makeSnowflakeModel(
+          "snowflake-llama-3.1-405b",
+          "Snowflake Llama 3.1 405B",
+          { context: 8000, output: 4096 },
           { toolcall: false },
         ),
         "llama3.1-70b": makeSnowflakeModel(
@@ -1113,13 +1157,26 @@ export namespace Provider {
           { context: 32000, output: 4096 },
           { toolcall: false },
         ),
+        "mixtral-8x7b": makeSnowflakeModel(
+          "mixtral-8x7b",
+          "Mixtral 8x7B",
+          { context: 32000, output: 8192 },
+          { toolcall: false },
+        ),
         // DeepSeek — no tool calling
         "deepseek-r1": makeSnowflakeModel(
           "deepseek-r1",
           "DeepSeek R1",
           { context: 64000, output: 32000 },
           { reasoning: true, toolcall: false },
         ),
+        // Gemini — tool calling not verified on Cortex; default to off until confirmed
+        "gemini-3.1-pro": makeSnowflakeModel(
+          "gemini-3.1-pro",
+          "Gemini 3.1 Pro",
+          { context: 1000000, output: 64000 },
+          { toolcall: false },
+        ),
       },
     }
     // altimate_change end

packages/opencode/test/altimate/cortex-snowflake-e2e.test.ts

@@ -30,6 +30,13 @@ import {
   VALID_ACCOUNT_RE,
 } from "../../src/altimate/plugin/snowflake"
 
+// Production builds this set from `provider.models` at loader time; mirror
+// the same shape here for the transform-only e2e checks.
+const TOOLCAPABLE_E2E_FIXTURE: ReadonlySet<string> = new Set([
+  "claude-3-5-sonnet", "claude-sonnet-4-6", "claude-opus-4-7",
+  "openai-gpt-4.1", "openai-gpt-5",
+])
+
 // ---------------------------------------------------------------------------
 // Auth configuration
 // ---------------------------------------------------------------------------
@@ -375,7 +382,7 @@ describe.skipIf(!HAS_CORTEX)("Snowflake Cortex E2E", () => {
         max_tokens: 100,
         stream: true,
       })
-      const { body } = transformSnowflakeBody(input)
+      const { body } = transformSnowflakeBody(input, TOOLCAPABLE_E2E_FIXTURE)
       const parsed = JSON.parse(body)
       expect(parsed.max_completion_tokens).toBe(100)
       expect(parsed.max_tokens).toBeUndefined()
@@ -388,7 +395,7 @@ describe.skipIf(!HAS_CORTEX)("Snowflake Cortex E2E", () => {
         tools: [{ type: "function", function: { name: "read_file" } }],
         tool_choice: "auto",
       })
-      const { body } = transformSnowflakeBody(input)
+      const { body } = transformSnowflakeBody(input, TOOLCAPABLE_E2E_FIXTURE)
       const parsed = JSON.parse(body)
       expect(parsed.tools).toBeUndefined()
       expect(parsed.tool_choice).toBeUndefined()
@@ -403,7 +410,7 @@ describe.skipIf(!HAS_CORTEX)("Snowflake Cortex E2E", () => {
           { role: "assistant", content: "response" },
         ],
       })
-      const { syntheticStop } = transformSnowflakeBody(input)
+      const { syntheticStop } = transformSnowflakeBody(input, TOOLCAPABLE_E2E_FIXTURE)
       expect(syntheticStop).toBeUndefined()
     })
 
@@ -416,7 +423,7 @@ describe.skipIf(!HAS_CORTEX)("Snowflake Cortex E2E", () => {
           { role: "assistant", content: "response" },
         ],
       })
-      const { syntheticStop } = transformSnowflakeBody(input)
+      const { syntheticStop } = transformSnowflakeBody(input, TOOLCAPABLE_E2E_FIXTURE)
       expect(syntheticStop).toBeDefined()
       expect(syntheticStop!.status).toBe(200)
     })