yuhp · Oct 27, 2026 · Oct 26, 2026 · Oct 26, 2026 · Oct 26, 2026 · Oct 27, 2026
diff --git a/README.md b/README.md
@@ -34,7 +34,7 @@ Add the plugin package to your OpenCode v2 `opencode.json` (or `opencode.jsonc`)
 {
   "$schema": "https://opencode.ai/config.json",
   "plugins": [
-    "opencode-models-discovery@2.0.0"
+    "opencode-models-discovery@2.1.0"
   ],
   "providers": {
     "gateway-custom": {
@@ -80,7 +80,7 @@ events without requiring a restart.
 See [Connect and Auth Documentation](docs/connect-and-auth.md) for credential
 resolution, lifecycle handling, and security details.
 
-`modelInfoFormat` and `smartModelName` are optional: without them, discovered names remain the model IDs and no external metadata is fetched. Supported metadata formats are `aiproxy`, `models.dev`, `bifrost`, `litellm`, `vllm`, `lmstudio`, `llama-swap`, `omniroute`, and `openrouter`. `models.dev` fetches `https://models.dev/models.json` by default; `modelInfoEndpoint` can override the catalog URL. The `aiproxy` format composes that catalog with inline metadata from AIProxy's model list. The `openrouter` format reads OpenRouter's inline `/v1/models` metadata without another request. `filterNonChat` defaults to `true`; it can filter non-chat models when the selected metadata source supplies that information. `timeoutMs` defaults to 5000 ms. Models returned by discovery are added alongside explicitly configured models.
+`modelInfoFormat` and `smartModelName` are optional: without them, discovered names remain the model IDs and no external metadata is fetched. Supported metadata formats are `aiproxy`, `models.dev`, `bifrost`, `litellm`, `vllm`, `lmstudio`, `llama-swap`, `omniroute`, `openrouter`, and `ollama`. `models.dev` fetches `https://models.dev/models.json` by default; `modelInfoEndpoint` can override the catalog URL. The `aiproxy` format composes that catalog with inline metadata from AIProxy's model list. The `openrouter` format reads OpenRouter's inline `/v1/models` metadata without another request. The `ollama` format reads per-model context lengths and capabilities from Ollama's native `/api/show` endpoint. `filterNonChat` defaults to `true`; it can filter non-chat models when the selected metadata source supplies that information. `timeoutMs` defaults to 5000 ms. Models returned by discovery are added alongside explicitly configured models.
 
 For a local build, run `npm run compile` and replace the package value above with `"file:///absolute/path/to/opencode-models-discovery/dist"`. OpenCode v2 expects a **directory**, not a path to `index.js` or `server.js`. After rebuilding plugin code, run `opencode service restart` so the background service loads the new bundle. Discovery runs when the plugin initializes. The native TUI commands are `/models-discovery-refresh` and `/models-discovery-console`; the console combines Provider → model → model details browsing with cache inspection and clearing. Use `/models-discovery-console status --provider <provider-id>` to filter the browser. The OpenCode v2 agent tools `models_discovery_refresh` and `models_discovery_status` remain available for automation. Pass `rawCache: true` to `models_discovery_status` to include cached raw models and enrichment results.
 
@@ -103,7 +103,7 @@ For OpenCode v1 (1.18.29+), add the plugin to your `opencode.json`:
 {
   "$schema": "https://opencode.ai/config.json",
   "plugin": [
-    "opencode-models-discovery@latest"
+    "opencode-models-discovery@2.1.0"
   ],
   "provider": {
     "lmstudio": {
@@ -336,6 +336,17 @@ For LM Studio 0.4.0+'s native v1 REST inventory endpoint:
 }
 ```
 
+For Ollama's native `/api/show` per-model metadata:
+
+```json
+{
+  "modelsDiscovery": {
+    "enabled": true,
+    "modelInfoFormat": "ollama"
+  }
+}
+```
+
 For vLLM's inline `max_model_len` metadata, without another metadata request:
 
 ```json

diff --git a/docs/config_example/a2agent.md b/docs/config_example/a2agent.md
@@ -12,7 +12,7 @@ This project is not affiliated with, endorsed by, or sponsored by [A2Agent](http
 {
   "$schema": "https://opencode.ai/config.json",
   "plugins": [
-    "opencode-models-discovery@2.0.0"
+    "opencode-models-discovery@2.1.0"
   ],
   "providers": {
     "a2agent": {

diff --git a/docs/config_example/ai-router.md b/docs/config_example/ai-router.md
@@ -12,7 +12,7 @@ This project is not affiliated with, endorsed by, or sponsored by [AI-ROUTER](ht
 {
   "$schema": "https://opencode.ai/config.json",
   "plugins": [
-    "opencode-models-discovery@2.0.0"
+    "opencode-models-discovery@2.1.0"
   ],
   "providers": {
     "ai-router": {

diff --git a/docs/config_example/deepseek.md b/docs/config_example/deepseek.md
@@ -12,7 +12,7 @@ This project is not affiliated with, endorsed by, or sponsored by DeepSeek.
 {
   "$schema": "https://opencode.ai/config.json",
   "plugins": [
-    "opencode-models-discovery@2.0.0"
+    "opencode-models-discovery@2.1.0"
   ],
   "providers": {
     "deepseek": {

diff --git a/docs/configuration.md b/docs/configuration.md
@@ -14,7 +14,7 @@ OpenCode v2 uses `plugins` and `providers`. The plugin can be declared directly
 {
   "$schema": "https://opencode.ai/config.json",
   "plugins": [
-    "opencode-models-discovery@2.0.0"
+    "opencode-models-discovery@2.1.0"
   ],
   "providers": {
     "my-provider": {
@@ -116,8 +116,8 @@ Each provider can configure discovery behavior through `provider.<name>.options.
 | `provider.<name>.options.modelsDiscovery.enabled` | `boolean` | Force enable or disable discovery for a single provider |
 | `provider.<name>.options.modelsDiscovery.endpoint` | `string` | Provider-specific models endpoint as an origin-relative path beginning with `/`. Defaults to `/v1/models` |
 | `provider.<name>.options.modelsDiscovery.timeoutMs` | positive finite `number` | Per-request timeout for the provider's models and provider-specific metadata endpoints. Defaults to `3000` |
-| `provider.<name>.options.modelsDiscovery.modelInfoEndpoint` | `string` | Override a format-specific metadata endpoint as an origin-relative path or complete URL. Defaults to `/v1/model/info` for `"litellm"` and `/api/v1/models` for `"lmstudio"`; for `"aiproxy"`, it overrides the models.dev fallback URL |
-| `provider.<name>.options.modelsDiscovery.modelInfoFormat` | `string` | Model info response format. Currently supports `"aiproxy"`, `"bifrost"`, `"litellm"`, `"models.dev"`, `"vllm"`, `"lmstudio"`, `"llama-swap"`, `"omniroute"`, and `"openrouter"` |
+| `provider.<name>.options.modelsDiscovery.modelInfoEndpoint` | `string` | Override a format-specific metadata endpoint as an origin-relative path or complete URL. Defaults to `/v1/model/info` for `"litellm"`, `/api/v1/models` for `"lmstudio"`, and `/api/show` for `"ollama"`; for `"aiproxy"`, it overrides the models.dev fallback URL |
+| `provider.<name>.options.modelsDiscovery.modelInfoFormat` | `string` | Model info response format. Currently supports `"aiproxy"`, `"bifrost"`, `"litellm"`, `"models.dev"`, `"vllm"`, `"lmstudio"`, `"llama-swap"`, `"omniroute"`, `"openrouter"`, and `"ollama"` |
 | `provider.<name>.options.modelsDiscovery.filterNonChat` | `boolean` | When model info is available, skip models whose `model_info.mode` is not `chat`. Defaults to `true` |
 | `provider.<name>.options.modelsDiscovery.models.includeRegex` | `string[]` | Shortcut regex allow-list for discovered model ids only |
 | `provider.<name>.options.modelsDiscovery.models.excludeRegex` | `string[]` | Shortcut regex deny-list for discovered model ids only |
@@ -355,7 +355,7 @@ Community provider examples live in [`docs/config_example/`](config_example/).
 
 The generic OpenAI-compatible `/v1/models` endpoint only guarantees a small model list shape. Extra metadata such as context limits, tool calling, reasoning, image input, or structured output is provider-specific, so metadata enrichment is opt-in.
 
-The plugin currently supports nine model info formats:
+The plugin currently supports ten model info formats:
 
 | Format | Source | Requires `modelInfoEndpoint` | Notes |
 |--------|--------|------------------------------|-------|
@@ -368,6 +368,7 @@ The plugin currently supports nine model info formats:
 | `"llama-swap"` | Fields in llama-swap's `/v1/models` response | No | Reads inline context, modalities, and function-calling metadata when present |
 | `"omniroute"` | Fields in OmniRoute's `/v1/models` response | No | Reads OmniRoute inline limits, modalities, and capabilities when present |
 | `"openrouter"` | Fields in OpenRouter's `/v1/models` response | No | Reads OpenRouter inline limits, modalities, pricing, capabilities, and reasoning metadata when present |
+| `"ollama"` | Ollama's `/api/show` per-model endpoint | No | Uses `/api/show` by default; makes one POST per discovered model id; set `modelInfoEndpoint` for another path or a proxy URL |
 
 > **Configuration syntax**: All examples below use the primary **OpenCode v2** configuration shape (`providers.<id>.settings.modelsDiscovery`). The discovery options (`enabled`, `modelInfoFormat`, `models`, `smartModelName`, etc.) are identical between versions; for legacy OpenCode v1, place the same options under `provider.<id>.options.modelsDiscovery` (with `plugin: ["opencode-models-discovery"]` and `npm: "@ai-sdk/openai-compatible"`).
 
@@ -615,6 +616,40 @@ Only models returned by `/v1/models` are injected. A model is enriched only when
 
 When available, the plugin sets `limit.context` from the largest loaded instance `config.context_length`, otherwise it uses `max_context_length`. LM Studio does not report a distinct output limit, so the plugin resolves a safe default output limit bounded by the context window and capped at OpenCode's default 32,000 output tokens. The plugin maps `capabilities.vision` to image input and `capabilities.trained_for_tool_use` to `tool_call`. Its reported reasoning options are the source of truth for variants: `off`, `low`, `medium`, `high`, and `xhigh` become variants, with `off` sent as `reasoningEffort: "none"`; `on` and unknown options are omitted because they are not concrete OpenAI-compatible efforts. Missing or malformed metadata is left unset without preventing discovery.
 
+### Ollama Model Info
+
+Use `modelInfoFormat: "ollama"` for Ollama providers. Ollama's OpenAI-compatible `/v1/models` response only contains ids, so the plugin discovers models normally and then calls Ollama's native `POST /api/show` once per discovered model to read context length and capabilities.
+
+```jsonc
+{
+  "providers": {
+    "ollama": {
+      "package": "@opencode/ai/providers/openai-compatible",
+      "settings": {
+        "baseURL": "http://127.0.0.1:11434/v1",
+        "modelsDiscovery": {
+          "enabled": true,
+          "modelInfoFormat": "ollama"
+        }
+      }
+    }
+  }
+}
+```
+
+The endpoint defaults to `/api/show` on the provider base URL origin, so `http://127.0.0.1:11434/v1` becomes `http://127.0.0.1:11434/api/show`. Set `modelInfoEndpoint` only for a non-standard path or a proxy URL. Requests use the configured `timeoutMs` and are limited to four concurrent calls. A failed, timed-out, or malformed per-model response only disables enrichment for that model; discovery continues and other models are still enriched.
+
+Context length is resolved in this order:
+
+1. `num_ctx` from the model's `parameters` string, when positive;
+2. `model_info.<general.architecture>.context_length`, then `model_info.<details.family>.context_length`;
+3. any other positive `*.context_length` entry in `model_info`;
+4. a top-level `context_length`.
+
+`num_ctx` is preferred because it is the only intentional effective-context signal the model exposes. The `model_info` values are architecture-derived maxima and are used when `num_ctx` is absent or invalid. The plugin does not clamp `num_ctx` to the architecture maximum, and it cannot see server-wide runtime overrides such as `OLLAMA_CONTEXT_LENGTH`. When no value is available, the model keeps the host default context limit.
+
+Capabilities are mapped as follows: `vision` or `images` adds image input, `audio` adds audio input, `thinking` sets `reasoning`, and `tools`, `function_calling`, or `tool_calling` sets `tool_call`. When `filterNonChat` is enabled, a model whose capabilities contain `embedding` but not `completion` is skipped.
+
 ### models.dev Metadata
 
 Use `modelInfoFormat: "models.dev"` to enrich discovered models from the public [models.dev](https://models.dev) metadata index.
@@ -692,7 +727,7 @@ For providers with custom metadata paths or non-standard behavior:
 {
   "$schema": "https://opencode.ai/config.json",
   "plugins": [
-    "opencode-models-discovery@2.0.0"
+    "opencode-models-discovery@2.1.0"
   ],
   "providers": {
     "lmstudio": {
@@ -741,7 +776,7 @@ In this example:
 {
   "$schema": "https://opencode.ai/config.json",
   "plugins": [
-    "opencode-models-discovery@2.0.0"
+    "opencode-models-discovery@2.1.0"
   ],
   "providers": {
     "ollama": {

diff --git a/docs/connect-and-auth.md b/docs/connect-and-auth.md
@@ -18,7 +18,7 @@ OpenCode v2 features a first-class Integration subsystem (`ctx.integration`) and
 {
   "$schema": "https://opencode.ai/config.json",
   "plugins": [
-    "opencode-models-discovery@latest"
+    "opencode-models-discovery@2.1.0"
   ],
   "providers": {
     "deepseek-custom": {
@@ -83,7 +83,7 @@ OpenCode v1 does not have the V2 plugin integration API. Instead, the V1 adapter
 {
   "$schema": "https://opencode.ai/config.json",
   "plugin": [
-    "opencode-models-discovery@latest"
+    "opencode-models-discovery@2.1.0"
   ],
   "provider": {
     "test_provider": {

diff --git a/docs/migrate-v1-to-v2.md b/docs/migrate-v1-to-v2.md
@@ -9,7 +9,7 @@ OpenCode v2 is the stable, primary integration. The package contains separate ad
 | OpenCode v1 | OpenCode v2 |
 |---|---|
 | `plugin` | `plugins` |
-| `plugin: ["opencode-models-discovery"]` | `plugins: ["opencode-models-discovery@2.0.0"]` |
+| `plugin: ["opencode-models-discovery"]` | `plugins: ["opencode-models-discovery@2.1.0"]` |
 | `provider` | `providers` |
 | `provider.<id>.npm` | `providers.<id>.package` |
 | `provider.<id>.options` | `providers.<id>.settings` |
@@ -64,7 +64,7 @@ The V1 package commonly uses:
 {
   "$schema": "https://opencode.ai/config.json",
   "plugins": [
-    "opencode-models-discovery@2.0.0"
+    "opencode-models-discovery@2.1.0"
   ],
   "providers": {
     "gateway": {

diff --git a/docs/providers.md b/docs/providers.md
@@ -118,6 +118,31 @@ The examples below use the OpenCode v1 keys `plugin`, `provider`, `npm`, and `op
 }
 ```
 
+### Ollama Metadata Enrichment
+
+Ollama's OpenAI-compatible model list only exposes ids. Enable `modelInfoFormat: "ollama"` to read context lengths and capabilities from the native `POST /api/show` endpoint:
+
+```json
+{
+  "plugin": ["opencode-models-discovery"],
+  "provider": {
+    "ollama": {
+      "npm": "@ai-sdk/openai-compatible",
+      "name": "Ollama",
+      "options": {
+        "baseURL": "http://127.0.0.1:11434/v1",
+        "modelsDiscovery": {
+          "enabled": true,
+          "modelInfoFormat": "ollama"
+        }
+      }
+    }
+  }
+}
+```
+
+The plugin appends `/api/show` to the provider origin, so the example above calls `http://127.0.0.1:11434/api/show`. It requests each discovered model with bounded concurrency and degrades to no enrichment for models whose request fails. See [Ollama model info](configuration.md#ollama-model-info) for context-length precedence and capability mapping.
+
 ### Non-`/v1` Discovery Endpoint
 
 ```json

diff --git a/package-lock.json b/package-lock.json
diff --git a/package.json b/package.json
@@ -1,7 +1,7 @@
 {
   "$schema": "https://json.schemastore.org/package.json",
   "name": "opencode-models-discovery",
-  "version": "2.0.1",
+  "version": "2.1.0",
   "description": "OpenCode plugin for auto-discovery of OpenAI-compatible models with dynamic provider configuration",
   "type": "module",
   "main": "./dist/index.js",

diff --git a/src/core/enricher-resolver.ts b/src/core/enricher-resolver.ts
@@ -1,6 +1,9 @@
 import { ModelInfoFormat } from '../types/plugin-config'
-import { createModelEnricher, isSupportedModelInfoFormat, type ModelEnricher } from '../utils/model-info/index'
+import { createDeferredModelEnricher, createModelEnricher, isSupportedModelInfoFormat, type ModelEnricher } from '../utils/model-info/index'
 import { fetchModelsDevData, DEFAULT_MODELS_DEV_URL } from '../utils/models-dev-fetcher'
+import { buildModelInfoURL } from './model-info-url'
+
+export { buildModelInfoURL } from './model-info-url'
 
 export const DEFAULT_LITELLM_ENDPOINT = '/v1/model/info'
 export const DEFAULT_LMSTUDIO_ENDPOINT = '/api/v1/models'
@@ -26,15 +29,6 @@ export interface ModelInfoResolverContext {
   readonly logger?: ModelInfoResolverLogger
 }
 
-export function buildModelInfoURL(baseURL: string, endpoint: string): string {
-  if (/^https?:\/\//i.test(endpoint)) {
-    return endpoint
-  }
-  return endpoint.startsWith('/')
-    ? new URL(endpoint, new URL(baseURL).origin).toString()
-    : new URL(endpoint, baseURL.endsWith('/') ? baseURL : `${baseURL}/`).toString()
-}
-
 /**
  * Shared model-info enricher resolution across V1 and V2 adapters.
  * Dispatches to static indexes, inline/local formats, or remote provider endpoints.
@@ -79,6 +73,16 @@ export async function resolveModelInfoEnricher(
     return createModelEnricher(format, null)
   }
 
+  const deferred = createDeferredModelEnricher(format, {
+    baseURL: context.baseURL,
+    apiKey: context.apiKey,
+    endpoint: options.endpoint,
+    timeoutMs: options.timeoutMs ?? DEFAULT_ENRICHER_TIMEOUT_MS,
+    fetcher: context.fetcher,
+    logger: context.logger,
+  })
+  if (deferred) return deferred
+
   if (format === ModelInfoFormat.LiteLLM || format === ModelInfoFormat.LMStudio) {
     const defaultEndpoint = format === ModelInfoFormat.LiteLLM ? DEFAULT_LITELLM_ENDPOINT : DEFAULT_LMSTUDIO_ENDPOINT
     const targetEndpoint = options.endpoint ?? defaultEndpoint

diff --git a/src/core/model-enrichment.ts b/src/core/model-enrichment.ts
@@ -4,6 +4,25 @@ export interface ModelEnrichmentContext {
   readonly filterNonChat: boolean
 }
 
+export interface ModelEnricherLogger {
+  warn(message: string, context?: Record<string, unknown>): void
+  info?(message: string, context?: Record<string, unknown>): void
+}
+
+/**
+ * Request bindings passed to deferred enricher factories at resolution time.
+ * Format modules use these to construct an enricher that performs per-model
+ * metadata requests later in prepare().
+ */
+export interface ModelEnricherDeferredContext {
+  readonly baseURL: string
+  readonly apiKey?: string
+  readonly endpoint?: string
+  readonly timeoutMs?: number
+  readonly fetcher?: typeof fetch
+  readonly logger?: ModelEnricherLogger
+}
+
 export interface ModelEnrichmentResult {
   readonly skip?: boolean
   readonly metadataName?: string
@@ -24,9 +43,39 @@ export interface ModelEnrichmentResult {
 }
 
 export interface ModelEnricher {
+  /**
+   * Optional deferred hydration step. Called at most once per discovery run,
+   * after the provider model list is known and before any enrich() call.
+   *
+   * Implementations may perform network requests to fetch per-model metadata.
+   * Individual failures must degrade to "no enrichment" for that model and
+   * must not reject; enrichment is always best-effort.
+   */
+  prepare?(modelIds: readonly string[]): Promise<void>
+
   enrich(model: DiscoveredRawModel, context: ModelEnrichmentContext): ModelEnrichmentResult
 }
 
+/**
+ * Runs the optional deferred hydration step of an enricher.
+ *
+ * Adapters call this after the provider model list is available and before
+ * building discovery drafts, so enrichers that need per-model detail requests
+ * can keep enrich() synchronous. Preparation is best-effort: a failure must
+ * never abort discovery.
+ */
+export async function prepareModelEnricher(
+  enricher: ModelEnricher | undefined,
+  modelIds: readonly string[],
+): Promise<void> {
+  if (!enricher?.prepare) return
+  try {
+    await enricher.prepare(modelIds)
+  } catch {
+    // Deferred enrichment is best-effort; discovery must continue.
+  }
+}
+
 export interface EnrichedModelDraft {
   readonly draft: DiscoveredModelDraft
   readonly metadataName?: string