feat: add Ollama metadata enrichment (context length + capabilities via /api/show)

README.md

@@ -82,7 +82,7 @@ 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`, and `omniroute`. `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. `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`, 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. `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.
 

docs/configuration.md

@@ -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"`, and `"omniroute"` |
+| `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"`, 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 eight model info formats:
+The plugin currently supports nine model info formats:
 
 | Format | Source | Requires `modelInfoEndpoint` | Notes |
 |--------|--------|------------------------------|-------|
@@ -367,6 +367,7 @@ The plugin currently supports eight model info formats:
 | `"lmstudio"` | LM Studio 0.4.0+ `/api/v1/models` inventory | No | Uses `/api/v1/models` by default; set `modelInfoEndpoint` for another path |
 | `"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 |
+| `"ollama"` | Ollama native `POST /api/show` | No | Uses `/api/show` by default; set `modelInfoEndpoint` to override |
 
 ### AIProxy Metadata
 
@@ -604,6 +605,34 @@ 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 Show
+
+Use `modelInfoFormat: "ollama"` with Ollama's OpenAI-compatible endpoint. Ollama's `GET /v1/models` response carries only the minimal model list shape, so the plugin enriches each discovered model with a native `POST /api/show` request (`{"model": "<id>"}`) after discovery. Set `modelInfoEndpoint` only when Ollama uses another show path.
+
+```json
+{
+  "plugin": ["opencode-models-discovery"],
+  "provider": {
+    "ollama": {
+      "npm": "@ai-sdk/openai-compatible",
+      "name": "Ollama",
+      "options": {
+        "baseURL": "http://localhost:11434/v1",
+        "modelsDiscovery": {
+          "enabled": true,
+          "modelInfoFormat": "ollama"
+        }
+      },
+      "models": {}
+    }
+  }
+}
+```
+
+Only models returned by `/v1/models` are injected, and each one receives its own `/api/show` request in parallel. Each model is enriched from the `/api/show` response stored under the exact id the plugin requested; a failed or malformed per-model response degrades to no enrichment for that model without preventing discovery.
+
+The context limit is resolved in strict order: a valid `num_ctx` value parsed from the `parameters` string first — a `num_ctx` set in the model's Modelfile is Ollama's only intentional effective-context signal, so it wins over the architecture-derived maxima — then `model_info.<details.family>.context_length`, then any other `model_info` key ending in `.context_length`, then a top-level `context_length`. A missing or non-positive `num_ctx` is ignored and the chain applies; the pinned value is not clamped to the architecture maximum. The output limit is bounded by the context window and capped at OpenCode's default 32,000 output tokens. Ollama `capabilities` map as follows: `vision` or `images` add image input, `audio` adds audio input, `thinking` sets `reasoning`, and `tools`, `function_calling`, or `tool_calling` set `tool_call`. Unknown capabilities and missing fields are ignored without inventing metadata. `filterNonChat` has no effect for this format; Ollama's show response exposes no chat-mode field.
+
 ### models.dev Metadata
 
 Use `modelInfoFormat: "models.dev"` to enrich discovered models from the public [models.dev](https://models.dev) metadata index.

docs/providers.md

@@ -193,3 +193,31 @@ This project is not affiliated with, endorsed by, or sponsored by [models.dev](h
 When a safe model id segment match exists, models.dev metadata can add model limits and capabilities such as `tool_call`, `reasoning`, `attachment`, `structured_output`, `temperature`, and `modalities`.
 
 This means providers using `@ai-sdk/anthropic` with OpenAI-compatible backends are supported when the `baseURL` contains `/v1/`, when a provider-specific discovery endpoint is configured, or when provider-level discovery is explicitly forced on. It also means providers like DeepSeek can be discovered from a non-`/v1` baseURL as long as the models endpoint is configured explicitly.
+
+### Ollama Metadata Enrichment
+
+Ollama exposes an OpenAI-compatible `GET /v1/models` endpoint that returns only the minimal model list shape — no context limits and no capabilities. Ollama's native API provides both through `POST /api/show` with a `{"model": "<id>"}` body, so the plugin supports an explicit Ollama metadata format:
+
+```json
+{
+  "plugin": ["opencode-models-discovery"],
+  "provider": {
+    "ollama": {
+      "npm": "@ai-sdk/openai-compatible",
+      "name": "Ollama",
+      "options": {
+        "baseURL": "http://localhost:11434/v1",
+        "modelsDiscovery": {
+          "enabled": true,
+          "modelInfoFormat": "ollama"
+        }
+      },
+      "models": {}
+    }
+  }
+}
+```
+
+This setting is explicit because the plugin issues one additional `POST /api/show` request per discovered model after the `/v1/models` fetch. Without `modelInfoFormat: "ollama"`, the plugin does not contact `/api/show`.
+
+For each discovered model, the plugin resolves the context limit from a valid `num_ctx` in the show body's `parameters` string first (the Modelfile's explicit context pin; missing or non-positive values are ignored), falling back to `model_info.<family>.context_length`, then any other `*.context_length` key, then a top-level `context_length`; and maps Ollama `capabilities` to modalities, `reasoning`, and `tool_call`. A per-model failure degrades to no enrichment for that model only; discovery continues for the rest.

src/core/enricher-resolver.ts

@@ -4,13 +4,15 @@ import { fetchModelsDevData, DEFAULT_MODELS_DEV_URL } from '../utils/models-dev-
 
 export const DEFAULT_LITELLM_ENDPOINT = '/v1/model/info'
 export const DEFAULT_LMSTUDIO_ENDPOINT = '/api/v1/models'
+export const DEFAULT_OLLAMA_ENDPOINT = '/api/show'
 export const DEFAULT_ENRICHER_TIMEOUT_MS = 5000
 
 export interface ModelInfoResolutionOptions {
   readonly format?: ModelInfoFormat | string
   readonly endpoint?: string
   readonly filterNonChat?: boolean
   readonly timeoutMs?: number
+  readonly modelIds?: readonly string[]
 }
 
 export interface ModelInfoResolverLogger {
@@ -78,6 +80,65 @@ export async function resolveModelInfoEnricher(
     return createModelEnricher(format, null)
   }
 
+  if (format === ModelInfoFormat.Ollama) {
+    const modelIds = options.modelIds ?? []
+    if (modelIds.length === 0) {
+      context.logger?.warn('Ollama model info format requires model IDs', {
+        provider: context.providerName,
+      })
+      return undefined
+    }
+
+    const targetEndpoint = options.endpoint ?? DEFAULT_OLLAMA_ENDPOINT
+    const showUrl = buildModelInfoURL(context.baseURL, targetEndpoint)
+    const fetcher = context.fetcher ?? globalThis.fetch
+    const timeoutMs = options.timeoutMs ?? DEFAULT_ENRICHER_TIMEOUT_MS
+
+    const headers = new Headers({ accept: 'application/json', 'content-type': 'application/json' })
+    if (context.apiKey) {
+      headers.set('authorization', `Bearer ${context.apiKey}`)
+    }
+
+    const results = await Promise.all(modelIds.map(async (modelId): Promise<readonly [string, unknown] | undefined> => {
+      try {
+        const res = await fetcher(showUrl, {
+          method: 'POST',
+          headers,
+          body: JSON.stringify({ model: modelId }),
+          signal: AbortSignal.timeout(timeoutMs),
+        })
+        if (!res.ok) {
+          context.logger?.warn('Provider model info discovery failed', {
+            provider: context.providerName,
+            endpoint: targetEndpoint,
+            format,
+            modelId,
+            status: res.status,
+          })
+          return undefined
+        }
+        return [modelId, await res.json()] as const
+      } catch (error) {
+        context.logger?.warn('Provider model info discovery failed', {
+          provider: context.providerName,
+          endpoint: targetEndpoint,
+          format,
+          modelId,
+          error: error instanceof Error ? error.message : String(error),
+        })
+        return undefined
+      }
+    }))
+
+    const shows = new Map<string, unknown>()
+    for (const entry of results) {
+      if (entry) shows.set(entry[0], entry[1])
+    }
+    if (shows.size === 0) return undefined
+
+    return createModelEnricher(format, shows, { filterNonChat })
+  }
+
   if (format === ModelInfoFormat.LiteLLM || format === ModelInfoFormat.LMStudio) {
     const defaultEndpoint = format === ModelInfoFormat.LiteLLM ? DEFAULT_LITELLM_ENDPOINT : DEFAULT_LMSTUDIO_ENDPOINT
     const targetEndpoint = options.endpoint ?? defaultEndpoint

src/plugin/commands.ts

@@ -126,6 +126,7 @@ Supported plugin options under provider.<id>.options.modelsDiscovery:
 - modelInfoFormat="lmstudio": discover callable models through the normal /v1/models endpoint, then enrich exact model-id matches from LM Studio's /api/v1/models inventory
 - modelInfoFormat="llama-swap": read llama-swap's inline /v1/models context, modalities, and function-calling metadata without another request
 - modelInfoFormat="omniroute": read OmniRoute's documented inline /v1/models limits, modalities, and capabilities without another request
+- modelInfoFormat="ollama": enrich exact model-id matches from Ollama's POST /api/show endpoint; modelInfoEndpoint optionally overrides the path
 - filterNonChat: when LiteLLM model info is available, skip non-chat models by default
 - cache.enabled: opt in to a provider-scoped persisted discovery cache of filtered and enriched model configurations; defaults to false
 - cache.ttlSeconds: non-negative finite cache lifetime in seconds; defaults to 86400
@@ -153,6 +154,7 @@ Recommended defaults:
 - use modelInfoFormat="lmstudio" for LM Studio REST metadata enrichment; set modelInfoEndpoint only when LM Studio uses another inventory path
 - use modelInfoFormat="llama-swap" only for llama-swap /v1/models responses; it is an explicit inline-metadata format and does not make another request
 - use modelInfoFormat="omniroute" only for OmniRoute's documented /v1/models response; it is an explicit inline-metadata format and does not make another request
+- use modelInfoFormat="ollama" for Ollama-native metadata enrichment; set modelInfoEndpoint only when Ollama uses another path
 - configure caching only when the user requests it; it is disabled by default
 - the persisted discovery cache belongs to this plugin and is not a replacement for opencode.json
 

src/plugin/enhance-config.ts

@@ -339,6 +339,7 @@ export async function enhanceConfig(
           endpoint: providerDiscoveryConfig.modelInfoEndpoint,
           filterNonChat,
           timeoutMs,
+          modelIds: models.filter(isDiscoveredRawModel).map((model) => model.id),
         })
       }
       if (reprocessPersistedModels) {

src/types/plugin-config.ts

@@ -38,6 +38,7 @@
   LMStudio = 'lmstudio',
   LlamaSwap = 'llama-swap',
   OmniRoute = 'omniroute',
+  Ollama = 'ollama',
 }
 
 export const DEFAULT_CACHE_TTL_SECONDS = 86400
@@ -104,7 +105,7 @@
       fallback: DEFAULT_DISCOVERY_CONFIG.enabled,
     })
   } else {
     console.warn(`[opencode-models-discovery] Ignoring invalid OPENCODE_MODELS_DISCOVERY_DEFAULT_ENABLED value: ${rawValue}`)
   }
 
   return DEFAULT_DISCOVERY_CONFIG
@@ -141,7 +142,7 @@
     if (logger) {
       logger.warn('Ignoring invalid model regex', { category: 'filtering', pattern })
     } else {
       console.warn(`[opencode-models-discovery] Ignoring invalid model regex: ${pattern}`)
     }
     return null
   }
@@ -165,7 +166,7 @@
       if (logger) {
         logger.warn('Ignoring invalid model field regex', { category: 'filtering', field: filter.field, pattern: filter.match })
       } else {
         console.warn(`[opencode-models-discovery] Ignoring invalid model field regex for ${filter.field}: ${filter.match}`)
       }
       return null
     }

src/utils/model-info/index.ts

@@ -1,6 +1,7 @@
 import { createBifrostEnricher } from './bifrost'
 import { createLiteLLMEnricher } from './litellm'
 import { createLMStudioEnricher } from './lmstudio'
+import { createOllamaEnricher } from './ollama'
 import { createLlamaSwapEnricher } from './llamaswap'
 import { createModelsDevEnricher } from './models-dev'
 import { createOmniRouteEnricher } from './omniroute'
@@ -20,6 +21,7 @@ const MODEL_ENRICHERS: Partial<Record<ModelInfoFormat, ModelEnricherFactory>> =
   [ModelInfoFormat.LMStudio]: (data) => createLMStudioEnricher(data),
   [ModelInfoFormat.LiteLLM]: (data) => createLiteLLMEnricher(data),
   [ModelInfoFormat.AIProxy]: (data) => createAIProxyEnricher(data),
+  [ModelInfoFormat.Ollama]: (data) => createOllamaEnricher(data),
 }
 
 /** Creates the neutral contract used by the shared discovery pipeline. */