pi-ollama-cloud 0.11.0 → 0.12.0

package/CHANGELOG.md

@@ -2,7 +2,14 @@
 
 All notable changes to this project will be documented in this file.
 
-## [Unreleased]
+## [0.12.0] - 2026-09-11
+
+- Source per-model thinking levels from models.dev instead of hardcoded maps. `scripts/generate-reasoning.ts` fetches the `ollama-cloud` provider's `reasoning_options` into `reasoning.generated.ts`, and `thinking-levels.ts` maps each model's effort values onto Pi's levels (toggle-only models become a binary on/off map; models with no models.dev entry fall back to `DEFAULT`). The `off` switch is handled by a small override table for models verified not to honor `reasoning_effort:"none"` (`gpt-oss:20b`, `gpt-oss:120b`, `minimax-m2.7`). Removed the now-stale per-family maps and the `docs/think-experiment.md` doc.
+- `generate-models` now also refreshes `reasoning.generated.ts` (runs `generate-pricing`, `generate-reasoning`, then `generate-models`).
+- Fix `generate-pricing` mis-dropping models whose pricing-page cached-input cell is `-` (no cache rate): those rows now match and their `cacheRead` equals `input`. This restored pricing for `mistral-large-3:675b`, `nemotron-3-nano:30b`, and `qwen3.5:397b`, which the earlier regex had left at zero cost.
+- Refresh the model catalog: added `deepseek-v4.1-flash` (probed max output 393216).
+
+## [0.11.0] - 2026-09-07
 
 - Fix `/ollama-cloud-usage` and the usage status bar failing with "unexpected response shape" after the undocumented `/api/usage` endpoint flipped between a single `limits.monthly` bucket and `limits.session` plus `limits.weekly` (the shape has flip-flopped repeatedly as of 2026-09). Any bucket present (`monthly`, `session`, `weekly`) is accepted alone or in combination, and whichever are present are displayed as `5h`/`7d`/`30d` segments. Thanks @johanngyger (#56).
 - Cache `ollama_web_search` results (24h) and `ollama_web_fetch` pages (24h success / 15 min failure) on disk under the pi agent home, so repeated queries and page reads cost 0 API calls. Expired entries are pruned on write, the cache is capped at 500 entries per kind (oldest evicted beyond the cap; `PI_OLLAMA_SEARCH_MAX_ENTRIES`), a partially corrupted cache file is validated per entry and degrades to "no cache" instead of crashing tool calls, and the file is written with `0600` permissions since it stores page content and URLs that can embed credentials. Tune with `PI_OLLAMA_SEARCH_TTL_HOURS`, `PI_OLLAMA_SEARCH_FAIL_TTL_MINUTES`, `PI_OLLAMA_SEARCH_MAX_ENTRIES`, and `PI_OLLAMA_SEARCH_CACHE_PATH`.

package/README.md

@@ -7,7 +7,7 @@ Registers Ollama Cloud as a model provider with dynamically fetched models, and
 ## Features
 
 - **Dynamic model discovery** - Fetches the full model list from `ollama.com/v1/models`, then fetches per-model details via `/api/show` to determine capabilities, context length, and tool support.
-- **Curated thinking levels** - Maps Pi's thinking levels to Ollama Cloud's OpenAI-compatible `reasoning_effort` values via `thinking-levels.ts`, with per-model exceptions based on API testing.
+- **Data-driven thinking levels** - Maps Pi's thinking levels to Ollama Cloud's OpenAI-compatible `reasoning_effort` values via `thinking-levels.ts`, sourced from models.dev per-model reasoning options with a small override table for the models where `none` doesn't disable thinking.
 - **Baked-in model list** - A generated fallback list (`models.generated.ts`) ships with the extension so models are available on first launch without any network calls. It is only a fallback: pi refreshes the live catalog at runtime, so shipping a new release for catalog freshness is no longer needed.
 - **Automatic model refresh** - On startup, `/model` open, and `pi update --models`, pi calls the extension's `refreshModels` callback to fetch the latest models from the API and persists them through pi's own model store. No manual refresh command.
 - **`ollama_web_search` tool** - Search the web for real-time information using Ollama Cloud's `/api/web_search` endpoint. Returns titles, URLs, and content snippets.
@@ -132,7 +132,7 @@ Model metadata is derived from the `/api/show` response:
 | Field | Source |
 |---|---|
 | `reasoning` | `capabilities` includes `"thinking"` |
-| `thinkingLevelMap` | [`thinking-levels.ts`](thinking-levels.ts) with 5 maps (DEFAULT, GPT_OSS, QWEN3, GLM_52, NO_OFF) based on API testing |
+| `thinkingLevelMap` | [`thinking-levels.ts`](thinking-levels.ts) + [`reasoning.generated.ts`](reasoning.generated.ts) (models.dev reasoning options), with an `off` override table for models that ignore `none` |
 | `input` | `["text", "image"]` if `capabilities` includes `"vision"`, else `["text"]` |
 | `contextWindow` | `model_info.*.context_length` (falls back to 128000) |
 | `maxTokens` | Probed per-model limits from [`limits.generated.ts`](limits.generated.ts), generated by `scripts/generate-limits.ts` (requires `OLLAMA_API_KEY`). Models without a probed limit fall back to 32768. |
@@ -146,17 +146,11 @@ Cache pricing is informational only: the `/pricing` page lists a "Cached input"
 
 ### Thinking level mapping
 
-Pi's thinking levels are mapped to Ollama Cloud's OpenAI-compatible `reasoning_effort` parameter in [`thinking-levels.ts`](thinking-levels.ts). The API accepts `none`, `low`, `medium`, `high`, and `max`. Effects of `max` over `high` vary by model and prompt difficulty - see [`docs/think-experiment.md`](docs/think-experiment.md) for details.
+Pi's thinking levels are mapped to Ollama Cloud's OpenAI-compatible `reasoning_effort` parameter in [`thinking-levels.ts`](thinking-levels.ts). The API accepts `none`, `low`, `medium`, `high`, `xhigh`, and `max`. Effects of `max` over `high` vary by model and prompt difficulty.
 
-| Map | Models | Levels exposed | Notes |
-|---|---|---|---|
-| `DEFAULT` | Most thinking models | off, low, medium, high, xhigh | `minimal` hidden (duplicate of low) |
-| `GPT_OSS` | `gpt-oss*` | low, medium, high | Can't disable thinking, no off or xhigh |
-| `QWEN3` | `qwen3*` (except `qwen3-vl*`) | off, medium | Binary-only (think/nothink), no gradation |
-| `GLM_52` | `glm-5.2` | off, high, xhigh | GLM supports disabled thinking; Ollama's model page confirms `high` and `max` reasoning efforts |
-| `NO_OFF` | `qwen3-vl*`, `kimi-k2-thinking`, `minimax*` | low, medium, high, xhigh | "none" doesn't disable thinking on these models |
+Per-model support is sourced from models.dev: [`scripts/generate-reasoning.ts`](scripts/generate-reasoning.ts) fetches the `ollama-cloud` provider's `reasoning_options` into `reasoning.generated.ts`, and `resolve()` maps each model's effort values onto Pi's levels. Models with `effort` values expose those grades; `toggle`-only models expose a single on/off level. Models with no models.dev entry fall back to `DEFAULT`.
 
-See [docs/think-experiment.md](docs/think-experiment.md) for the testing methodology and results.
+Because the API reports only a boolean `thinking` capability and models.dev does not reliably encode the `none` behavior, the `off` switch is handled via a small override table in `thinking-levels.ts`: it defaults to enabled, and is hidden only for models verified (by live probing) not to honor `reasoning_effort:"none"` - currently `gpt-oss:20b`, `gpt-oss:120b`, and `minimax-m2.7`. The per-model metadata gaps behind the models.dev sourcing are tracked upstream in [ollama/ollama#18385](https://github.com/ollama/ollama/issues/18385).
 
 ## Tools
 

package/limits.generated.ts

@@ -1,10 +1,11 @@
 // Auto-generated by scripts/generate-limits.ts
 // Do not edit manually.
-// Probed models: 19 (0 failed)
+// Entries: 20
 
 export const MODEL_MAX_OUTPUT_TOKENS: Record<string, number> = {
   "deepseek-v4-flash:0731": 65536,
   "deepseek-v4-pro:0813": 65536,
+  "deepseek-v4.1-flash": 393216,
   "gemma4:31b": 262144,
   "glm-5.1": 131072,
   "glm-5.2": 131072,

package/models.generated.ts

@@ -1,7 +1,7 @@
 // Auto-generated by scripts/generate-models.ts
 // Do not edit manually.
-// Generated: 2026-09-03T10:12:02.243Z
-// Model count: 19
+// Generated: 2026-09-12T02:41:57.025Z
+// Model count: 20
 
 import type { ProviderModelConfig } from "@earendil-works/pi-coding-agent";
 
@@ -39,8 +39,8 @@ export const GENERATED_MODELS: ProviderModelConfig[] = [
     reasoning: true,
     thinkingLevelMap: {
       high: "high",
-      low: "low",
-      medium: "medium",
+      low: null,
+      medium: null,
       minimal: null,
       off: "none",
       xhigh: "max",
@@ -77,10 +77,50 @@ export const GENERATED_MODELS: ProviderModelConfig[] = [
     input: ["text"],
     maxTokens: 65536,
     reasoning: true,
+    thinkingLevelMap: {
+      high: "high",
+      low: null,
+      medium: null,
+      minimal: null,
+      off: "none",
+      xhigh: "max",
+    },
+  },
+  {
+    id: "deepseek-v4.1-flash",
+    name: "deepseek-v4.1-flash",
+    compat: {
+      maxTokensField: "max_tokens",
+      openRouterRouting: {},
+      requiresAssistantAfterToolResult: false,
+      requiresReasoningContentOnAssistantMessages: false,
+      requiresThinkingAsText: false,
+      requiresToolResultName: false,
+      sendSessionAffinityHeaders: false,
+      supportsDeveloperRole: false,
+      supportsLongCacheRetention: false,
+      supportsReasoningEffort: true,
+      supportsStore: false,
+      supportsStrictMode: false,
+      supportsUsageInStreaming: true,
+      thinkingFormat: "openai",
+      vercelGatewayRouting: {},
+      zaiToolStream: false,
+    },
+    contextWindow: 1048576,
+    cost: {
+      cacheRead: 0.006,
+      cacheWrite: 0,
+      input: 0.3,
+      output: 1.2,
+    },
+    input: ["text", "image"],
+    maxTokens: 393216,
+    reasoning: true,
     thinkingLevelMap: {
       high: "high",
       low: "low",
-      medium: "medium",
+      medium: null,
       minimal: null,
       off: "none",
       xhigh: "max",
@@ -118,12 +158,12 @@ export const GENERATED_MODELS: ProviderModelConfig[] = [
     maxTokens: 262144,
     reasoning: true,
     thinkingLevelMap: {
-      high: "high",
-      low: "low",
+      high: null,
+      low: null,
       medium: "medium",
       minimal: null,
       off: "none",
-      xhigh: "max",
+      xhigh: null,
     },
   },
   {
@@ -158,12 +198,12 @@ export const GENERATED_MODELS: ProviderModelConfig[] = [
     maxTokens: 131072,
     reasoning: true,
     thinkingLevelMap: {
-      high: "high",
-      low: "low",
+      high: null,
+      low: null,
       medium: "medium",
       minimal: null,
       off: "none",
-      xhigh: "max",
+      xhigh: null,
     },
   },
   {
@@ -240,7 +280,7 @@ export const GENERATED_MODELS: ProviderModelConfig[] = [
     thinkingLevelMap: {
       high: "high",
       low: "low",
-      medium: "medium",
+      medium: null,
       minimal: null,
       off: "none",
       xhigh: "max",
@@ -280,7 +320,7 @@ export const GENERATED_MODELS: ProviderModelConfig[] = [
     thinkingLevelMap: {
       high: "high",
       low: "low",
-      medium: "medium",
+      medium: null,
       minimal: null,
       off: "none",
       xhigh: "max",
@@ -398,12 +438,12 @@ export const GENERATED_MODELS: ProviderModelConfig[] = [
     maxTokens: 262144,
     reasoning: true,
     thinkingLevelMap: {
-      high: "high",
-      low: "low",
+      high: null,
+      low: null,
       medium: "medium",
       minimal: null,
       off: "none",
-      xhigh: "max",
+      xhigh: null,
     },
   },
   {
@@ -438,12 +478,12 @@ export const GENERATED_MODELS: ProviderModelConfig[] = [
     maxTokens: 262144,
     reasoning: true,
     thinkingLevelMap: {
-      high: "high",
-      low: "low",
+      high: null,
+      low: null,
       medium: "medium",
       minimal: null,
       off: "none",
-      xhigh: "max",
+      xhigh: null,
     },
   },
   {
@@ -480,7 +520,7 @@ export const GENERATED_MODELS: ProviderModelConfig[] = [
     thinkingLevelMap: {
       high: "high",
       low: "low",
-      medium: "medium",
+      medium: null,
       minimal: null,
       off: "none",
       xhigh: "max",
@@ -518,12 +558,12 @@ export const GENERATED_MODELS: ProviderModelConfig[] = [
     maxTokens: 131072,
     reasoning: true,
     thinkingLevelMap: {
-      high: "high",
-      low: "low",
+      high: null,
+      low: null,
       medium: "medium",
       minimal: null,
       off: null,
-      xhigh: "max",
+      xhigh: null,
     },
   },
   {
@@ -562,7 +602,7 @@ export const GENERATED_MODELS: ProviderModelConfig[] = [
       low: "low",
       medium: "medium",
       minimal: null,
-      off: null,
+      off: "none",
       xhigh: "max",
     },
   },
@@ -630,12 +670,12 @@ export const GENERATED_MODELS: ProviderModelConfig[] = [
     maxTokens: 131072,
     reasoning: true,
     thinkingLevelMap: {
-      high: "high",
-      low: "low",
+      high: null,
+      low: null,
       medium: "medium",
       minimal: null,
       off: "none",
-      xhigh: "max",
+      xhigh: null,
     },
   },
   {
@@ -670,12 +710,12 @@ export const GENERATED_MODELS: ProviderModelConfig[] = [
     maxTokens: 65536,
     reasoning: true,
     thinkingLevelMap: {
-      high: "high",
-      low: "low",
+      high: null,
+      low: null,
       medium: "medium",
       minimal: null,
       off: "none",
-      xhigh: "max",
+      xhigh: null,
     },
   },
   {
@@ -710,12 +750,12 @@ export const GENERATED_MODELS: ProviderModelConfig[] = [
     maxTokens: 65536,
     reasoning: true,
     thinkingLevelMap: {
-      high: "high",
-      low: "low",
+      high: null,
+      low: null,
       medium: "medium",
       minimal: null,
       off: "none",
-      xhigh: "max",
+      xhigh: null,
     },
   },
   {

package/models.ts

@@ -93,7 +93,7 @@ function buildCompat(): ProviderModelConfig["compat"] {
   return {
     // Ollama uses "system" role, not "developer" (ollama: docs.ollama.com/api/openai-compatibility, pi: types.ts#supportsDeveloperRole).
     supportsDeveloperRole: false,
-    // reasoning_effort works (ollama: docs.ollama.com/api/openai-compatibility, pi: types.ts#supportsReasoningEffort, tested in think-experiment.md).
+    // reasoning_effort works (ollama: docs.ollama.com/api/openai-compatibility, pi: types.ts#supportsReasoningEffort).
     supportsReasoningEffort: true,
     // "store" is not a supported field (ollama: docs.ollama.com/api/openai-compatibility, pi: types.ts#supportsStore).
     supportsStore: false,
@@ -109,7 +109,7 @@ function buildCompat(): ProviderModelConfig["compat"] {
     requiresThinkingAsText: false,
     // DeepSeek-specific, not needed for Ollama (pi: types.ts#requiresReasoningContentOnAssistantMessages).
     requiresReasoningContentOnAssistantMessages: false,
-    // reasoning_effort format works (pi: types.ts#thinkingFormat, tested in think-experiment.md).
+    // reasoning_effort format works (pi: types.ts#thinkingFormat).
     thinkingFormat: "openai",
     // Ollama does not support tool_choice, so strict mode is unavailable (ollama: docs.ollama.com/api/openai-compatibility, pi: types.ts#supportsStrictMode).
     supportsStrictMode: false,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-ollama-cloud",
-  "version": "0.11.0",
+  "version": "0.12.0",
   "type": "module",
   "keywords": [
     "pi-package"
@@ -13,6 +13,7 @@
     "models.ts",
     "models.generated.ts",
     "pricing.generated.ts",
+    "reasoning.generated.ts",
     "thinking-levels.ts",
     "usage.ts",
     "utils.ts",
@@ -32,8 +33,9 @@
     "format": "biome format --write .",
     "test": "vitest run",
     "smoke:web-tools": "tsx scripts/smoke-web-tools.ts",
-    "generate-models": "tsx scripts/generate-pricing.ts && tsx scripts/generate-models.ts && biome format --write models.generated.ts pricing.generated.ts",
-    "generate-limits": "tsx scripts/generate-limits.ts && biome format --write limits.generated.ts"
+    "generate-models": "tsx scripts/generate-pricing.ts && tsx scripts/generate-reasoning.ts && tsx scripts/generate-models.ts && biome format --write models.generated.ts pricing.generated.ts reasoning.generated.ts",
+    "generate-limits": "tsx scripts/generate-limits.ts",
+    "generate-reasoning": "tsx scripts/generate-reasoning.ts && biome format --write reasoning.generated.ts"
   },
   "pi": {
     "extensions": [
@@ -51,6 +53,6 @@
     "@types/node": "^26.1.2",
     "@typescript/native-preview": "7.0.0-dev.20260707.2",
     "tsx": "^4.19.0",
-    "vitest": "^4.1.6"
+    "vitest": "^4.1.11"
   }
 }

package/pricing.generated.ts

@@ -1,7 +1,7 @@
 // Auto-generated by scripts/generate-pricing.ts
 // Do not edit manually.
-// Generated: 2026-09-03T10:12:00.135Z
-// Model count: 19
+// Generated: 2026-09-12T02:41:55.206Z
+// Model count: 20
 
 export interface ModelPrice {
   input: number;
@@ -13,6 +13,7 @@ export interface ModelPrice {
 export const MODEL_PRICING: Record<string, ModelPrice> = {
   "deepseek-v4-flash:0731": { input: 0.44, output: 1.32, cacheRead: 0.014, cacheWrite: 0 },
   "deepseek-v4-pro:0813": { input: 1.32, output: 3.96, cacheRead: 0.044, cacheWrite: 0 },
+  "deepseek-v4.1-flash": { input: 0.3, output: 1.2, cacheRead: 0.006, cacheWrite: 0 },
   "gemma4:31b": { input: 0.14, output: 0.4, cacheRead: 0.05, cacheWrite: 0 },
   "glm-5.1": { input: 1, output: 3.2, cacheRead: 0.2, cacheWrite: 0 },
   "glm-5.2": { input: 1.4, output: 4.4, cacheRead: 0.26, cacheWrite: 0 },

package/reasoning.generated.ts

@@ -0,0 +1,35 @@
+// Auto-generated by scripts/generate-reasoning.ts
+// Do not edit manually.
+// Model count: 22
+
+export type ModelsDevReasoningOption =
+  | { type: "toggle" }
+  | {
+      type: "effort";
+      values: Array<"none" | "minimal" | "low" | "medium" | "high" | "xhigh" | "max" | "ultra" | "default" | null>;
+    };
+
+export const MODEL_REASONING_OPTIONS: Record<string, ModelsDevReasoningOption[]> = {
+  "deepseek-v4-flash": [{ type: "toggle" }, { type: "effort", values: ["high", "max"] }],
+  "deepseek-v4-flash:0731": [{ type: "toggle" }, { type: "effort", values: ["high", "max"] }],
+  "deepseek-v4-pro": [{ type: "toggle" }, { type: "effort", values: ["high", "max"] }],
+  "deepseek-v4.1-flash": [{ type: "toggle" }, { type: "effort", values: ["low", "high", "max"] }],
+  "gemma4:31b": [{ type: "toggle" }],
+  "glm-5.1": [{ type: "toggle" }],
+  "glm-5.2": [{ type: "effort", values: ["high", "max"] }],
+  "glm-5.3": [{ type: "effort", values: ["low", "high", "max"] }],
+  "glm-5.3-flash": [{ type: "effort", values: ["low", "high", "max"] }],
+  "gpt-oss:120b": [{ type: "effort", values: ["low", "medium", "high"] }],
+  "gpt-oss:20b": [{ type: "effort", values: ["low", "medium", "high"] }],
+  "kimi-k2.5": [{ type: "toggle" }],
+  "kimi-k2.6": [{ type: "toggle" }],
+  "kimi-k2.7-code": [{ type: "toggle" }],
+  "kimi-k3": [{ type: "toggle" }, { type: "effort", values: ["low", "high", "max"] }],
+  "minimax-m2.5": [],
+  "minimax-m2.7": [{ type: "toggle" }],
+  "minimax-m3": [{ type: "toggle" }, { type: "effort", values: ["low", "medium", "high", "max"] }],
+  "nemotron-3-nano:30b": [{ type: "toggle" }],
+  "nemotron-3-super": [{ type: "toggle" }],
+  "nemotron-3-ultra": [{ type: "toggle" }],
+  "qwen3.5:397b": [{ type: "toggle" }],
+};

package/thinking-levels.ts

@@ -2,26 +2,36 @@
  * Thinking level mapping for Ollama Cloud models.
  *
  * Maps Pi's thinking levels to Ollama Cloud's OpenAI-compatible
- * `reasoning_effort` values. The API accepts "none", "low", "medium",
- * "high", and "max". On simple prompts, "max" can be a no-op over
- * "high", but on harder prompts it can increase thinking substantially
- * (e.g. deepseek-v4-pro: ~32k tokens on high vs ~55k on max).
+ * `reasoning_effort` values. The API accepts "minimal", "none", "low",
+ * "medium", "high", "xhigh", "ultra", and "max". On simple prompts, "max" can
+ * be a no-op over "high", but on harder prompts it can increase thinking
+ * substantially.
  *
- * A `null` value means the level is hidden in Pi's UI.
+ * Every value EFFORT_TO_LEVEL can send was verified against the live chat
+ * completions API (2026-09-11: "minimal", "xhigh", and "ultra" probed across
+ * gpt-oss, deepseek-v4, glm, minimax, and qwen thinking models, all accepted
+ * with graded reasoning), so a future models.dev row that lists them passes
+ * through a value the endpoint demonstrably accepts.
+ *
+ * The per-model level support comes from models.dev: scripts/generate-reasoning.ts
+ * fetches the `ollama-cloud` provider's `reasoning_options` into
+ * reasoning.generated.ts (the same data source pi uses for its built-in
+ * providers), and resolve() maps a model's effort values onto Pi's levels.
+ * We fall back to models.dev because the Cloud API does not yet expose
+ * per-model supported levels (tracked upstream: https://github.com/ollama/ollama/issues/18385).
  *
- * Model-specific behavior discovered through testing (see docs/think-experiment.md):
- *   - Most models: all levels work, "none" disables thinking
- *   - GPT-OSS: no off mode, only low/medium/high
- *   - Qwen 3.x (non-VL): binary-only (think/nothink) - off works
- *   - Qwen 3 VL: "none" doesn't disable thinking - off is hidden
- *   - GLM 5.2: off/high/max are exposed; low/medium are hidden
- *   - Kimi K2 Thinking: "none" doesn't disable thinking - off is hidden
- *   - MiniMax M2.x: "none" doesn't disable thinking - off is hidden
+ * The API exposes only a boolean `thinking` capability plus a global effort
+ * vocabulary, and models.dev does not reliably encode the `none` behavior, so
+ * the `off` switch is handled separately: it defaults to "none" (a live probe
+ * of the current catalog confirmed every model except the OFF_NULL overrides
+ * below honors it), and models verified not to honor `none` pin it to null
+ * (hidden) via OFF_NULL.
  *
- * Reference: https://docs.ollama.com/api/openai-compatibility
+ * A `null` value means the level is hidden in Pi's UI.
  */
 
 import type { ProviderModelConfig } from "@earendil-works/pi-coding-agent";
+import { MODEL_REASONING_OPTIONS, type ModelsDevReasoningOption } from "./reasoning.generated.ts";
 
 export type ThinkingLevelMap = NonNullable<ProviderModelConfig["thinkingLevelMap"]>;
 
@@ -35,63 +45,101 @@ export const DEFAULT: ThinkingLevelMap = {
   xhigh: "max",
 };
 
-/** GPT-OSS: can't disable thinking, only low/medium/high.
- *  https://ollama.com/library/gpt-oss */
-export const GPT_OSS: ThinkingLevelMap = {
-  off: null,
-  minimal: null,
-  low: "low",
-  medium: "medium",
-  high: "high",
-  xhigh: null,
-};
+/**
+ * Models where a live probe of `reasoning_effort:"none"` still produced
+ * reasoning (i.e. thinking cannot be disabled), so the `off` level is hidden.
+ * Confirmed against the current catalog by scripts/../test probing; the API
+ * and models.dev do not expose this behavior.
+ */
+/**
+ * Models where a live probe of `reasoning_effort:"none"` still produced
+ * reasoning (i.e. thinking cannot be disabled), so the `off` level is hidden.
+ * Confirmed against the current catalog by scripts/../test probing; the API
+ * and models.dev do not expose this behavior.
+ *
+ * Exact ids hide only the named variant; family prefixes cover future model
+ * revisions in the same family. gpt-oss is a family-wide prefix because both
+ * probed variants leak and the behavior is documented for the family.
+ * minimax is NOT matched family-wide here: minimax-m3 was probed to honor
+ * `none`, so only the verified-leaking minimax-m2.7 is pinned by exact id.
+ */
+const OFF_NULL_EXACT = new Set(["minimax-m2.7"]);
+const OFF_NULL_FAMILIES = ["gpt-oss"];
 
-/** Qwen 3.x: binary-only (think/nothink), no gradation.
- *  https://docs.ollama.com/capabilities/thinking */
-export const QWEN3: ThinkingLevelMap = {
-  off: "none",
-  minimal: null,
-  low: null,
-  medium: "medium",
-  high: null,
-  xhigh: null,
-};
+function hidesOff(id: string): boolean {
+  return OFF_NULL_EXACT.has(id) || OFF_NULL_FAMILIES.some((prefix) => id.startsWith(prefix));
+}
 
-/** GLM 5.2: Ollama's model page confirms support for "high" and "max" reasoning efforts.
- *  https://ollama.com/library/glm-5.2 */
-export const GLM_52: ThinkingLevelMap = {
-  off: "none",
-  minimal: null,
-  low: null,
-  medium: null,
-  high: "high",
-  xhigh: "max",
+/**
+ * Map a models.dev `effort` value onto the Pi level key and the reasoning_effort
+ * string to send for it. Ollama's top effort value is "max"; Pi exposes it via
+ * the extra-high level, so "max" (and "xhigh"/"ultra") map to the xhigh key.
+ */
+const EFFORT_TO_LEVEL: Record<string, { key: "minimal" | "low" | "medium" | "high" | "xhigh"; value: string }> = {
+  minimal: { key: "minimal", value: "minimal" },
+  low: { key: "low", value: "low" },
+  medium: { key: "medium", value: "medium" },
+  high: { key: "high", value: "high" },
+  xhigh: { key: "xhigh", value: "xhigh" },
+  max: { key: "xhigh", value: "max" },
+  ultra: { key: "xhigh", value: "ultra" },
 };
+/**
+ * Build a ThinkingLevelMap from models.dev reasoning_options.
+ * Levels come from `effort` values (mapped via EFFORT_TO_LEVEL); `off` defaults
+ * to "none" (probe-derived, see file header) and is hidden only via the
+ * OFF_NULL exact/family sets (see hidesOff).
+ * A toggle-only model is binary (on/off) and exposes a single "medium" level.
+ */
+function buildMap(options: readonly ModelsDevReasoningOption[], id: string): ThinkingLevelMap {
+  const map: ThinkingLevelMap = {
+    off: hidesOff(id) ? null : "none",
+    minimal: null,
+    low: null,
+    medium: null,
+    high: null,
+    xhigh: null,
+  };
 
-/** "none" doesn't disable thinking - off is hidden.
- *  Used by kimi and minimax families. */
-export const NO_OFF: ThinkingLevelMap = {
-  off: null,
-  minimal: null,
-  low: "low",
-  medium: "medium",
-  high: "high",
-  xhigh: "max",
-};
+  if (options.length > 0 && options.every((option) => option.type === "toggle")) {
+    // Binary on/off model: no graded effort, expose a single level.
+    return { ...map, medium: "medium" };
+  }
+
+  const efforts = options.flatMap((option) => (option.type === "effort" ? (option.values ?? []) : []));
+  for (const effort of efforts) {
+    const target = effort !== null && effort !== "default" ? EFFORT_TO_LEVEL[effort] : undefined;
+    if (target) map[target.key] = target.value;
+  }
+  return map;
+}
+
+/**
+ * Read MODEL_REASONING_OPTIONS[id] without tripping over inherited keys (e.g.
+ * "constructor"), which would otherwise resolve to the Object constructor and
+ * crash buildMap.
+ */
+function ownOptions(id: string): ModelsDevReasoningOption[] | undefined {
+  return Object.hasOwn(MODEL_REASONING_OPTIONS, id) ? MODEL_REASONING_OPTIONS[id] : undefined;
+}
 
 /**
  * Resolve the thinking level map for a model.
- * Matches by model ID prefix (case-sensitive, checks first chars).
+ * Looks up the model id (exact, then `:tag` family) in the generated models.dev
+ * table, falling back to DEFAULT for models with no entry. The matched key is
+ * the one passed to buildMap so the OFF_NULL set (keyed on bare family names)
+ * applies to tagged ids that resolve through a family match.
  */
 export function resolve(id: string, capabilities: string[]): ThinkingLevelMap | undefined {
   if (!capabilities.includes("thinking")) return undefined;
 
-  if (id.startsWith("gpt-oss")) return GPT_OSS;
-  if (id === "glm-5.2") return GLM_52;
-  if (id.startsWith("qwen3-vl")) return NO_OFF;
-  if (id.startsWith("qwen3")) return QWEN3;
-  if (id === "kimi-k2-thinking") return NO_OFF;
-  if (id.startsWith("minimax")) return NO_OFF;
-
-  return DEFAULT;
+  const colon = id.lastIndexOf(":");
+  const exact = ownOptions(id);
+  const matchedKey = exact !== undefined ? id : colon > 0 ? id.slice(0, colon) : "";
+  const options = exact ?? ownOptions(matchedKey);
+  // An empty array (e.g. minimax-m2.5) carries no verified options; fall back
+  // to DEFAULT rather than a degenerate map whose only selectable level can
+  // be a leaking off.
+  if (options === undefined || options.length === 0) return DEFAULT;
+  return buildMap(options, matchedKey);
 }