Update Poolside Laguna model metadata

.changeset/update-laguna-limits.md

@@ -0,0 +1,5 @@
+---
+"opencode-provider-poolside": patch
+---
+
+Update Laguna model limits to the verified Poolside metadata and preserve those limits when live discovery reports stale values.

README.md

@@ -18,7 +18,7 @@
 - **Live model discovery** — fetches the current model catalog from the Poolside API at startup, with sensible Laguna fallbacks when the API is unreachable.
 - **OpenAI-compatible** — uses [`@ai-sdk/openai-compatible`](https://ai-sdk.dev/providers/open-source-providers/openai-compatible) under the hood, so streaming, tool calls, and JSON mode work out of the box.
 - **API key management** — built-in auth hook for storing and loading your Poolside API key via OpenCode's `/connect` command.
-- **Reasoning support** — Laguna models expose `none`, `minimal`, `low`, `medium`, `high`, and `xhigh` reasoning effort levels through OpenCode's variant picker.
+- **Reasoning support** — Laguna M.1 and XS 2.1 expose `none`, `minimal`, `low`, `medium`, `high`, and `xhigh`; Laguna S 2.1 exposes its supported `none` and `xhigh` modes through OpenCode's variant picker.
 - **Zero config** — add the plugin to your `opencode.json` and you're ready to go.
 
 ## Quick start
@@ -73,11 +73,13 @@ Or set a default model in your config:
 
 ## Available models
 
-| Model | Description | Reasoning |
-|---|---|---|
-| `poolside/laguna-m.1` | Poolside's most capable coding agent model | ✅ |
-| `poolside/laguna-xs-2.1` | Second-generation coding agent model (XS) | ✅ |
-| `poolside/laguna-s-2.1` | Second-generation coding agent model (S) | ✅ |
+| Model | Context | Max output | Reasoning |
+|---|---:|---:|---|
+| `poolside/laguna-m.1` | 262,144 | 32,768 | `none`–`xhigh` |
+| `poolside/laguna-xs-2.1` | 262,144 | 32,768 | `none`–`xhigh` |
+| `poolside/laguna-s-2.1` | 1,048,576 | 131,072 | `none`, `xhigh` |
+
+The static catalog uses these verified Poolside limits. Authenticated model discovery enriches the catalog with live capabilities, while preserving these limits when the Poolside `/models` response is incomplete or stale.
 
 ## Configuration
 
@@ -125,7 +127,7 @@ export POOLSIDE_API_KEY="your-api-key"
 ## How it works
 
 1. **Config hook** — On startup, the plugin registers the `poolside` provider with `@ai-sdk/openai-compatible`, sets the base URL and environment variable, and populates the model catalog.
-2. **Model discovery** — If `POOLSIDE_API_KEY` is available, the plugin fetches the live model list from `https://inference.poolside.ai/v1/models`. If the API is unreachable, it falls back to a static catalog of known Laguna models.
+2. **Model discovery** — If `POOLSIDE_API_KEY` is available, the plugin fetches the live model list from `https://inference.poolside.ai/v1/models`. If the API is unreachable, it falls back to a static catalog of known Laguna models; verified limits remain in place if the endpoint reports stale values.
 3. **Model aliases** — OpenCode exposes models as `poolside/laguna-*` while sending Poolside's required `poolside/laguna-*` upstream IDs without duplicating the provider prefix.
 4. **Auth hook** — The plugin provides an API key auth method so you can manage your key with OpenCode's `/connect poolside` command.
 

src/constants.ts

@@ -93,10 +93,12 @@ export const MAX_ONLY_VARIANTS = {
 const ZERO_COST = { input: 0, output: 0, cache_read: 0, cache_write: 0 } as const;
 
 /**
- * Default context window and output token limits for Poolside models.
+ * Default context window and output token limits for Laguna M.1 and XS 2.1.
  */
 const DEFAULT_CONTEXT_WINDOW = 262_144;
 const DEFAULT_MAX_TOKENS = 32_768;
+const LAGUNA_S_CONTEXT_WINDOW = 1_048_576;
+const LAGUNA_S_MAX_TOKENS = 131_072;
 
 /**
  * Current Poolside Platform models, available before authenticated refresh.
@@ -130,7 +132,7 @@ export const FALLBACK_MODELS = [
     reasoning: true,
     tool_call: true,
     cost: ZERO_COST,
-    limit: { context: DEFAULT_CONTEXT_WINDOW, output: DEFAULT_MAX_TOKENS },
+    limit: { context: LAGUNA_S_CONTEXT_WINDOW, output: LAGUNA_S_MAX_TOKENS },
     variants: MAX_ONLY_VARIANTS,
   },
 ];

src/models.ts

@@ -95,15 +95,16 @@ export function parseModelsResponse(
             fallback?.cost.cache_write ?? 0
           ),
         },
+        // The verified Laguna catalog is authoritative for known models.
+        // Poolside's /models endpoint can expose stale limits for Laguna S
+        // 2.1, so live metadata must not shrink its advertised capacity.
         limit: {
-          context: positiveNumber(
-            item.context_length,
-            fallback?.limit.context ?? 262_144
-          ),
-          output: positiveNumber(
-            item.max_completion_tokens,
-            fallback?.limit.output ?? 32_768
-          ),
+          context:
+            fallback?.limit.context ??
+            positiveNumber(item.context_length, 262_144),
+          output:
+            fallback?.limit.output ??
+            positiveNumber(item.max_completion_tokens, 32_768),
         },
         variants,
         modalities: {

test/models.test.ts

@@ -192,6 +192,24 @@ describe("parseModelsResponse", () => {
     assert.equal(result[1]!.id, "poolside/laguna-xs-2.1");
     assert.equal(result[2]!.id, "poolside/laguna-s-2.1");
   });
+
+  test("preserves verified Laguna limits when the API reports stale values", () => {
+    const payload: PoolsideModelsResponse = {
+      data: [
+        {
+          id: "poolside/laguna-s-2.1",
+          context_length: 262_144,
+          max_completion_tokens: 32_768,
+        },
+      ],
+    };
+    const result = parseModelsResponse(payload);
+
+    assert.deepEqual(result[0]!.limit, {
+      context: 1_048_576,
+      output: 131_072,
+    });
+  });
 });
 
 describe("modelsToConfigMap", () => {
@@ -242,4 +260,15 @@ describe("FALLBACK_MODELS", () => {
       assert.ok(Object.keys(model.variants).length > 0);
     }
   });
+
+  test("uses the verified context and output limits", () => {
+    const limits = Object.fromEntries(
+      FALLBACK_MODELS.map((model) => [model.id, model.limit])
+    );
+    assert.deepEqual(limits, {
+      "poolside/laguna-m.1": { context: 262_144, output: 32_768 },
+      "poolside/laguna-xs-2.1": { context: 262_144, output: 32_768 },
+      "poolside/laguna-s-2.1": { context: 1_048_576, output: 131_072 },
+    });
+  });
 });