feat: thinking levels for custom models (reasoning detected from /models metadata)

Custom models were pinned reasoning:false, so every one of them offered a
single thinking level ("off") while built-in reasoning models offered the
standard ladder. The endpoint's /models metadata now decides:

- detectReasoning() trusts explicit metadata only — a boolean
  reasoning / supports_reasoning / thinking, or a reasoning token in
  supported_parameters (vLLM, OpenRouter), capabilities, tags or
  architecture.output_modalities. No name guessing: a false positive would
  send reasoning_effort to a server that rejects it.
- CustomModel carries the detected flag (persisted, sanitizer keeps it),
  toPiModel passes it through, and pi-ai does the rest — default
  off/minimal/low/medium/high set, reasoning_effort on the wire.

Verified in the browser against a local OpenAI-compatible mock: the
reasoning-capable model exposes off/minimal/low/medium/high, the plain one
stays at off, and a prompt with medium actually arrives as
reasoning_effort=medium.

src/core/settings.ts

@@ -217,12 +217,16 @@ export interface ModelsSettings {
 export const CUSTOM_MODEL_DEFAULT_CONTEXT_WINDOW = 8192;
 export const CUSTOM_MODEL_DEFAULT_MAX_TOKENS = 4096;
 
-/** One discovered custom model. `image`/`context`/`maxTokens` are detected from
- *  the endpoint's /models metadata; anything it omits falls back to a default. */
+/** One discovered custom model. `image`/`reasoning`/`context`/`maxTokens` are
+ *  detected from the endpoint's /models metadata; anything it omits falls back
+ *  to a default. */
 export interface CustomModel {
 	id: string;
 	/** Accepts image input (Vision-capable). Default false (text-only). */
 	image?: boolean;
+	/** Accepts a reasoning effort (thinking). Default false → the only
+	 *  thinking level is "off", exactly like a non-reasoning built-in. */
+	reasoning?: boolean;
 	/** Context window in tokens, from the endpoint. Absent → the default. */
 	context?: number;
 	/** Max output tokens, from the endpoint. Absent → the default. */
@@ -577,6 +581,7 @@ function sanitizeCustomProvider(raw: Record<string, unknown>): CustomProviderSet
 			if (!id) continue;
 			const model: CustomModel = { id };
 			if (rec.image === true) model.image = true;
+			if (rec.reasoning === true) model.reasoning = true;
 			const context = positiveInt(rec.context);
 			if (context) model.context = context;
 			const maxTokens = positiveInt(rec.maxTokens);

src/server/custom-provider.ts

@@ -35,8 +35,9 @@ const customEndpointAuth: ApiKeyAuth = {
 	},
 };
 
-/** Map a discovered model to a pi-ai entry: image→input, context/maxTokens→
- *  limits (else CUSTOM_MODEL_DEFAULT_*). baseUrl is stamped per-model. */
+/** Map a discovered model to a pi-ai entry: image→input, reasoning→thinking
+ *  levels, context/maxTokens→limits (else CUSTOM_MODEL_DEFAULT_*). baseUrl is
+ *  stamped per-model. */
 function toPiModel(model: CustomModel, baseUrl: string): Model<Api> {
 	return {
 		id: model.id,
@@ -44,7 +45,10 @@ function toPiModel(model: CustomModel, baseUrl: string): Model<Api> {
 		api: "openai-completions",
 		provider: CUSTOM_PROVIDER.id,
 		baseUrl,
-		reasoning: false,
+		// Detected from the endpoint's metadata (see detectReasoning). true makes
+		// pi-ai offer the standard levels and send reasoning_effort on the wire;
+		// no thinkingLevelMap → the default off/minimal/low/medium/high set.
+		reasoning: Boolean(model.reasoning),
 		input: model.image ? ["text", "image"] : ["text"],
 		cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
 		contextWindow: model.context ?? CUSTOM_MODEL_DEFAULT_CONTEXT_WINDOW,
@@ -139,6 +143,25 @@ function detectImage(row: Record<string, unknown>): boolean {
 	return false;
 }
 
+/** Does a /models row advertise reasoning? Metadata only — no name guessing,
+ *  because a false positive sends reasoning_effort an endpoint will reject.
+ *  Trusted signals: an explicit boolean (reasoning / supports_reasoning /
+ *  thinking) or a reasoning token in the capability/parameter lists servers
+ *  publish (vLLM + OpenRouter's supported_parameters, capabilities, tags).
+ *  Absent or unrecognized → false (thinking stays "off"). */
+function detectReasoning(row: Record<string, unknown>): boolean {
+	for (const key of ["reasoning", "supports_reasoning", "thinking"] as const) {
+		if (row[key] === true) return true;
+	}
+	const architecture = isRecord(row.architecture) ? row.architecture : undefined;
+	const candidates = [row.supported_parameters, row.capabilities, row.tags, architecture?.output_modalities];
+	return candidates.some(
+		(list) =>
+			Array.isArray(list) &&
+			list.some((v) => typeof v === "string" && (/reasoning/i.test(v) || v.toLowerCase() === "thinking")),
+	);
+}
+
 /** First positive integer among `keys` on a row (endpoints name limits
  *  inconsistently). */
 function firstPositive(row: Record<string, unknown>, keys: readonly string[]): number | undefined {
@@ -167,6 +190,7 @@ function extractModels(body: unknown): CustomModel[] | null {
 		if (isRecord(row) && typeof row.id === "string" && row.id.trim()) {
 			const model: CustomModel = { id: row.id };
 			if (detectImage(row)) model.image = true;
+			if (detectReasoning(row)) model.reasoning = true;
 			const context = firstPositive(row, CONTEXT_KEYS);
 			if (context) model.context = context;
 			const maxTokens = firstPositive(row, MAX_TOKEN_KEYS);

tests/custom-endpoint.test.ts

@@ -2,8 +2,8 @@
  * Tests for the custom OpenAI-compatible endpoint (custom-endpoint wave).
  * Covers the pure core (normalizeBaseUrl, customProviderUsable, the sanitizer,
  * resolution), the deep server module (provider build + live sync + model
- * discovery incl. image-modality detection), and the catalog federation row.
- * No network: the discovery probe runs against a stubbed global fetch.
+ * discovery incl. image-modality and reasoning detection), and the catalog
+ * federation row. No network: the discovery probe runs against a stubbed fetch.
  */
 
 import assert from "node:assert/strict";
@@ -23,6 +23,7 @@ import {
 	syncCustomProvider,
 } from "../src/server/custom-provider.ts";
 import { piModels } from "../src/agent/models.ts";
+import { getSupportedThinkingLevels } from "@earendil-works/pi-ai";
 
 // ─── normalizeBaseUrl ────────────────────────────────────────────────────────
 describe("normalizeBaseUrl", () => {
@@ -72,6 +73,15 @@ describe("sanitizeUserSettings — customProvider", () => {
 		assert.equal(out?.customProvider?.baseUrl, undefined);
 		assert.equal(out?.customProvider?.models, undefined);
 	});
+	it("keeps reasoning:true and drops a non-true reasoning flag", () => {
+		const out = sanitizeUserSettings({
+			customProvider: {
+				baseUrl: "https://h/v1",
+				models: [{ id: "thinker", reasoning: true }, { id: "plain", reasoning: "yes" }],
+			},
+		});
+		assert.deepEqual(out?.customProvider?.models, [{ id: "thinker", reasoning: true }, { id: "plain" }]);
+	});
 });
 
 // ─── resolution ──────────────────────────────────────────────────────────────
@@ -114,6 +124,19 @@ describe("buildOpenAiCompatibleProvider / syncCustomProvider", () => {
 		assert.equal(big!.maxTokens, 393216);
 	});
 
+	it("gives a reasoning-capable model the standard thinking levels (others stay \"off\")", () => {
+		const provider = buildOpenAiCompatibleProvider({
+			baseUrl: "http://h/v1",
+			models: [{ id: "thinker", reasoning: true }, { id: "plain" }],
+		});
+		const [thinker, plain] = provider.getModels();
+		assert.equal(thinker!.reasoning, true);
+		// No thinkingLevelMap → pi-ai's default set, same as any reasoning model.
+		assert.deepEqual(getSupportedThinkingLevels(thinker!), ["off", "minimal", "low", "medium", "high"]);
+		assert.equal(plain!.reasoning, false);
+		assert.deepEqual(getSupportedThinkingLevels(plain!), ["off"]);
+	});
+
 	it("registers on a usable sync and removes when the base URL clears", () => {
 		syncCustomProvider({ baseUrl: "http://h/v1", models: [{ id: "live" }] });
 		assert.ok(piModels().getProvider(CUSTOM_PROVIDER.id), "provider registered");
@@ -184,6 +207,29 @@ describe("discoverModels", () => {
 		assert.equal(byId["ctx"]!.maxTokens, undefined);
 		assert.equal(byId["plain"]!.context, undefined);
 	});
+	it("detects reasoning from endpoint metadata only — never from the model name", async () => {
+		stubFetch(() => ({
+			status: 200,
+			body: {
+				data: [
+					{ id: "vllm-reasoner", supported_parameters: ["temperature", "reasoning", "reasoning_effort"] },
+					{ id: "bool-reasoner", reasoning: true },
+					{ id: "thinking-flag", thinking: true },
+					{ id: "cap-reasoner", capabilities: ["completion", "reasoning"] },
+					{ id: "deepseek-r1-distill" }, // a reasoning-sounding NAME, no metadata → stays off
+				],
+			},
+		}));
+		const r = await discoverModels("http://h/v1");
+		assert.ok(r.ok);
+		const byId = Object.fromEntries(r.models.map((m) => [m.id, Boolean(m.reasoning)]));
+		assert.equal(byId["vllm-reasoner"], true);
+		assert.equal(byId["bool-reasoner"], true);
+		assert.equal(byId["thinking-flag"], true);
+		assert.equal(byId["cap-reasoner"], true);
+		assert.equal(byId["deepseek-r1-distill"], false, "name guessing would risk sending reasoning_effort to a server that rejects it");
+	});
+
 	it("guards the scheme before any fetch", async () => {
 		let called = false;
 		globalThis.fetch = (async () => {