pi-free 2.0.13 → 2.0.15

package/CHANGELOG.md

@@ -7,6 +7,34 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [2.0.15] - 2026-06-02
+
+### Fixed
+
+- **Qwen 3.7 reasoning compat** — `qwen/qwen3.7-max` on Cline/OpenRouter uses DeepSeek-style `reasoning_content` format. Added `DEEPSEEK_PROXY_COMPAT` so Pi preserves and replays reasoning tokens correctly, preventing plan-mode hangs ([#213](https://github.com/apmantza/pi-free/pull/213)).
+
+- **Kimi K2.6 reasoning compat** — Kimi models on NVIDIA/OpenRouter need `requiresReasoningContentOnAssistantMessages: true` to correctly replay reasoning tokens in assistant messages. Without it, the model gets stuck when trying to call tools or produce output after thinking. Refs [earendil-works/pi#5309](https://github.com/earendil-works/pi/issues/5309) ([#213](https://github.com/apmantza/pi-free/pull/213)).
+
+- **MiniMax reasoning compat** — MiniMax M3 and other MiniMax models now have full DeepSeek-style compat (`thinkingFormat: "deepseek"`, `requiresReasoningContentOnAssistantMessages: true`). Previously, models marked `reasoning: true` without `thinkingFormat` caused Pi to enter plan mode but couldn't parse the reasoning tokens, resulting in hangs ([#212](https://github.com/apmantza/pi-free/pull/212), [#213](https://github.com/apmantza/pi-free/pull/213)).
+
+### Added
+
+- **`/probe-routeway` command** — Tests each Routeway model with a minimal chat request and auto-hides models that return 5xx or 404 errors. Runs lazily on first `session_start` with 24h probe cache TTL. Follows the same pattern as `/probe-nvidia` ([#213](https://github.com/apmantza/pi-free/pull/213)).
+
+## [2.0.14] - 2026-06-02
+
+### Added
+
+- **Routeway provider** — OpenAI-compatible gateway (`api.routeway.ai/v1`) with 219 models, 16 free (`:free` suffix). Set `ROUTEWAY_API_KEY` or add `routeway_api_key` to `~/.pi/free.json`. Toggle with `/toggle-routeway` ([#209](https://github.com/apmantza/pi-free/pull/209)).
+
+### Fixed
+
+- **Cline free model merging** — Free-to-try models (e.g. `qwen3.7-plus`) from Cline's recommended list now appear in the free model picker even when absent from the main catalog ([#209](https://github.com/apmantza/pi-free/pull/209)).
+
+- **`_pricingKnown` / `_freeKnown` authoritatve flag** — Providers can now signal whether pricing data is authoritative via `_pricingKnown`. When `false`, `isFreeModel` falls back to name-based detection. Kilo's `isFree` API flag now flows through as `_freeKnown` ([#209](https://github.com/apmantza/pi-free/pull/209)).
+
+- **MiniMax reasoning compat** — MiniMax M3 and other MiniMax models now have `supportsReasoningEffort: true` compat settings. Previously, models marked `reasoning: true` without compat caused Pi to enter plan mode without knowing the thinking format, resulting in hangs.
+
 ## [2.0.13] - 2026-05-21
 
 ### Added

package/README.md

@@ -64,6 +64,7 @@ Free models are shown by default — look for the provider prefixes:
 - `codestral/` — Codestral via Mistral (free Experiment plan: 2 req/min, 1B tokens/month)
 - `deepinfra/` — DeepInfra inference cloud ($5 one-time trial credit, no credit card)
 - `novita/` — Novita AI (100+ open-source models, OpenAI-compatible, 3 free models)
+- `routeway/` — Routeway AI gateway (OpenAI-compatible, `:free` models)
 
 > **Note:** Paid providers may occasionally offer free models or promotional credits. The `isFreeModel` helper automatically detects free models based on provider pricing data or model names containing "free". For providers that don't expose pricing (like CrofAI), only models with "free" in their names are marked as free.
 
@@ -102,6 +103,7 @@ Want to see paid models too? Run the toggle command for your provider:
 /toggle-sambanova # Toggle SambaNova (🔄 freemium)
 /toggle-llm7      # Toggle LLM7 (✅ free gateway)
 /toggle-novita    # Toggle Novita AI (💳 paid — 3 free models)
+/toggle-routeway  # Toggle Routeway AI (💳 paid — has :free models)
 /toggle-fastrouter # Toggle FastRouter (🔧 dynamic — always discovered)
 ```
 
@@ -132,7 +134,8 @@ Add your API keys to this file:
   "sambanova_api_key": "...",
   "llm7_api_key": "...",
   "zenmux_api_key": "...",
-  "crofai_api_key": "..."
+  "crofai_api_key": "...",
+  "routeway_api_key": "sk-..."
 }
 ```
 
@@ -464,10 +467,11 @@ Each provider has toggle commands to switch between free and all models:
 
 Test models for 404/403 errors and auto-hide broken ones:
 
-| Command         | What it does                                                |
-| --------------- | ----------------------------------------------------------- |
-| `/probe-nvidia` | Test all NVIDIA models, auto-hide 404s in `~/.pi/free.json` |
-| `/probe-ollama` | Test all Ollama models, auto-hide 403s in `~/.pi/free.json` |
+| Command           | What it does                                                |
+| ----------------- | ----------------------------------------------------------- |
+| `/probe-nvidia`   | Test all NVIDIA models, auto-hide 404s in `~/.pi/free.json` |
+| `/probe-ollama`   | Test all Ollama models, auto-hide 403s in `~/.pi/free.json` |
+| `/probe-routeway` | Test all Routeway models, auto-hide 5xx/404s               |
 
 **How it works:**
 

package/config.ts

@@ -17,6 +17,7 @@ export {
 	PROVIDER_MODAL,
 	PROVIDER_NVIDIA,
 	PROVIDER_QWEN,
+	PROVIDER_ROUTEWAY,
 } from "./constants.ts";
 import { createLogger } from "./lib/logger.ts";
 
@@ -33,6 +34,7 @@ interface PiFreeConfig {
 	sambanova_api_key?: string;
 	together_api_key?: string;
 	novita_api_key?: string;
+	routeway_api_key?: string;
 	fastrouter_api_key?: string;
 	kilo_free_only?: boolean;
 	hidden_models?: string[];
@@ -48,6 +50,7 @@ interface PiFreeConfig {
 	sambanova_show_paid?: boolean;
 	together_show_paid?: boolean;
 	novita_show_paid?: boolean;
+	routeway_show_paid?: boolean;
 	fastrouter_show_paid?: boolean;
 	openrouter_show_paid?: boolean;
 	opencode_show_paid?: boolean;
@@ -64,6 +67,7 @@ const CONFIG_TEMPLATE: PiFreeConfig = {
 	sambanova_api_key: "",
 	together_api_key: "",
 	novita_api_key: "",
+	routeway_api_key: "",
 	fastrouter_api_key: "",
 
 	kilo_free_only: false,
@@ -80,6 +84,7 @@ const CONFIG_TEMPLATE: PiFreeConfig = {
 	sambanova_show_paid: false,
 	together_show_paid: false,
 	novita_show_paid: false,
+	routeway_show_paid: false,
 	fastrouter_show_paid: false,
 	openrouter_show_paid: false,
 	opencode_show_paid: false,
@@ -222,6 +227,10 @@ export function getNovitaShowPaid(): boolean {
 	return resolveBool("NOVITA_SHOW_PAID", loadConfigFile().novita_show_paid);
 }
 
+export function getRoutewayShowPaid(): boolean {
+	return resolveBool("ROUTEWAY_SHOW_PAID", loadConfigFile().routeway_show_paid);
+}
+
 export function getFastrouterShowPaid(): boolean {
 	return resolveBool(
 		"FASTROUTER_SHOW_PAID",
@@ -266,6 +275,8 @@ export function getProviderShowPaid(providerId: string): boolean {
 			return getTogetherShowPaid();
 		case "novita":
 			return getNovitaShowPaid();
+		case "routeway":
+			return getRoutewayShowPaid();
 		case "fastrouter":
 			return getFastrouterShowPaid();
 		case "ollama-cloud":
@@ -331,6 +342,10 @@ export function getNovitaApiKey(): string | undefined {
 	return resolve("NOVITA_API_KEY", loadConfigFile().novita_api_key);
 }
 
+export function getRoutewayApiKey(): string | undefined {
+	return resolve("ROUTEWAY_API_KEY", loadConfigFile().routeway_api_key);
+}
+
 export function getFastrouterApiKey(): string | undefined {
 	return resolve("FASTROUTER_API_KEY", loadConfigFile().fastrouter_api_key);
 }

package/constants.ts

@@ -23,6 +23,7 @@ export const PROVIDER_DEEPINFRA = "deepinfra";
 export const PROVIDER_SAMBANOVA = "sambanova";
 export const PROVIDER_TOGETHER = "together";
 export const PROVIDER_NOVITA = "novita";
+export const PROVIDER_ROUTEWAY = "routeway";
 
 export const ALL_UNIQUE_PROVIDERS = [
 	PROVIDER_KILO,
@@ -40,6 +41,7 @@ export const ALL_UNIQUE_PROVIDERS = [
 	PROVIDER_SAMBANOVA,
 	PROVIDER_TOGETHER,
 	PROVIDER_NOVITA,
+	PROVIDER_ROUTEWAY,
 ] as const;
 
 // =============================================================================
@@ -62,6 +64,7 @@ export const BASE_URL_DEEPINFRA = "https://api.deepinfra.com/v1/openai";
 export const BASE_URL_SAMBANOVA = "https://api.sambanova.ai/v1";
 export const BASE_URL_TOGETHER = "https://api.together.xyz/v1";
 export const BASE_URL_NOVITA = "https://api.novita.ai/openai/v1";
+export const BASE_URL_ROUTEWAY = "https://api.routeway.ai/v1";
 
 /** Cline fetches free models from OpenRouter */
 export const BASE_URL_OPENROUTER = "https://openrouter.ai/api/v1";

package/index.ts

@@ -14,6 +14,7 @@
  * - DeepInfra: AI inference cloud ($5 trial credit)
  * - SambaNova: Fast inference on RDU hardware (free tier, no credit card)
  * - Together: Fast inference on 200+ open-source models ($1 trial credit)
+ * - Routeway: OpenAI-compatible gateway with free `:free` models
  * - LLM7: AI gateway (free default/fast selectors)
  */
 
@@ -24,6 +25,13 @@ import {
 	processQuotaResponse,
 	formatQuotaStatus,
 } from "./lib/quota-monitor.ts";
+import {
+	startModelCall,
+	recordModelCall,
+	getAllTelemetry,
+	getTelemetryPath,
+	clearTelemetry,
+} from "./lib/telemetry.ts";
 import {
 	applyGlobalFilter,
 	getGlobalFreeOnly,
@@ -41,6 +49,7 @@ import deepinfra from "./providers/deepinfra/deepinfra.ts";
 import sambanova from "./providers/sambanova/sambanova.ts";
 import together from "./providers/together/together.ts";
 import novita from "./providers/novita/novita.ts";
+import routeway from "./providers/routeway/routeway.ts";
 import nvidia from "./providers/nvidia/nvidia.ts";
 import ollama from "./providers/ollama/ollama.ts";
 import zenmux from "./providers/zenmux/zenmux.ts";
@@ -145,6 +154,64 @@ function setupGlobalCommands(pi: ExtensionAPI) {
 			ctx.ui.notify(lines.join("\n"), "info");
 		},
 	});
+
+	// /telemetry — Show model telemetry data
+	pi.registerCommand("free-telemetry", {
+		description:
+			"Show real-world performance data for free models (tokens/s, latency, success rate)",
+		handler: async (_args, ctx) => {
+			const allTelemetry = getAllTelemetry();
+			const entries = Object.entries(allTelemetry);
+
+			if (entries.length === 0) {
+				ctx.ui.notify(
+					"No telemetry data yet. Use some free models first!",
+					"info",
+				);
+				return;
+			}
+
+			// Sort by total calls descending
+			entries.sort((a, b) => b[1].totalCalls - a[1].totalCalls);
+
+			const lines = ["📊 Model Telemetry:", ""];
+			lines.push(
+				`${`Model`.padEnd(40)} ${`Calls`.padEnd(6)} ${`OK%`.padEnd(6)} ${`Lat`.padEnd(7)} ${`tok/s`.padEnd(7)} ${`Cost`}`,
+			);
+			lines.push(`─`.repeat(75));
+
+			for (const [key, t] of entries.slice(0, 20)) {
+				const name = key.length > 38 ? key.slice(0, 35) + "..." : key;
+				const calls = String(t.totalCalls).padStart(5);
+				const ok = `${t.successRate}%`.padStart(5);
+				const lat =
+					t.avgLatencyMs > 0
+						? `${t.avgLatencyMs}ms`.padStart(6)
+						: "—".padStart(6);
+				const tps =
+					t.avgTokensPerSecond > 0
+						? `${t.avgTokensPerSecond}`.padStart(6)
+						: "—".padStart(6);
+				const cost =
+					t.totalCost > 0
+						? `$${t.totalCost.toFixed(4)}`.padStart(8)
+						: "free".padStart(8);
+				lines.push(`${name.padEnd(40)} ${calls} ${ok} ${lat} ${tps} ${cost}`);
+			}
+
+			lines.push("", `File: ${getTelemetryPath()}`);
+			ctx.ui.notify(lines.join("\n"), "info");
+		},
+	});
+
+	// /clear-free-telemetry — Clear all telemetry data
+	pi.registerCommand("clear-free-telemetry", {
+		description: "Clear all model telemetry data",
+		handler: async (_args, ctx) => {
+			clearTelemetry();
+			ctx.ui.notify("Telemetry data cleared", "info");
+		},
+	});
 }
 
 // =============================================================================
@@ -183,6 +250,70 @@ function setupQuotaMonitoring(pi: ExtensionAPI) {
 	});
 }
 
+// =============================================================================
+// Model Telemetry
+// =============================================================================
+
+function setupTelemetry(pi: ExtensionAPI) {
+	// Only track telemetry for FREE models (uses same isFreeModel logic as model filtering)
+	// eslint-disable-next-line @typescript-eslint/no-explicit-any
+	(pi as any).on("before_agent_start", (_event: any, ctx: any) => {
+		if (!ctx.model) return;
+		if (!isFreeModel(ctx.model as any)) return;
+		const provider = ctx.model?.provider;
+		const model = ctx.model?.id;
+		if (provider && model) {
+			startModelCall(provider, model);
+		}
+	});
+
+	// Record telemetry when a turn completes
+	pi.on("turn_end", (event, ctx) => {
+		if (!ctx.model) return;
+		if (!isFreeModel(ctx.model as any)) return;
+
+		const msg = (
+			event as {
+				message?: {
+					role?: string;
+					model?: string;
+					usage?: {
+						input?: number;
+						output?: number;
+						totalTokens?: number;
+						cost?: { total?: number };
+					};
+					stopReason?: string;
+					errorMessage?: string;
+				};
+			}
+		).message;
+
+		if (msg?.role !== "assistant") return;
+
+		const provider = ctx.model?.provider;
+		const model = msg.model || ctx.model?.id;
+		if (!provider || !model) return;
+
+		const usage = msg.usage;
+		const inputTokens = usage?.input ?? 0;
+		const outputTokens = usage?.output ?? 0;
+		const totalTokens = usage?.totalTokens ?? inputTokens + outputTokens;
+		const cost = usage?.cost?.total ?? 0;
+		const isError = msg.stopReason === "error" || !!msg.errorMessage;
+
+		recordModelCall(
+			provider,
+			model,
+			{ input: inputTokens, output: outputTokens, totalTokens },
+			cost,
+			!isError,
+			msg.stopReason,
+			msg.errorMessage,
+		);
+	});
+}
+
 // =============================================================================
 // Main Entry Point
 // =============================================================================
@@ -197,6 +328,9 @@ export default async function piFreeEntry(pi: ExtensionAPI) {
 	// Setup quota monitoring (passive, no extra API calls)
 	setupQuotaMonitoring(pi);
 
+	// Setup model telemetry (tracks real-world performance)
+	setupTelemetry(pi);
+
 	// Load all unique providers
 	// Each provider will register itself with the global toggle system
 	await Promise.allSettled([
@@ -212,6 +346,7 @@ export default async function piFreeEntry(pi: ExtensionAPI) {
 		sambanova(pi),
 		together(pi),
 		novita(pi),
+		routeway(pi),
 	]);
 
 	// Setup dynamic built-in providers (Mistral, Groq, Cerebras, xAI, Hugging Face,

package/lib/built-in-toggle.ts

@@ -278,9 +278,9 @@ function setupStatusBar(
 
 function getApiKeyEnvForProvider(providerId: string): string {
 	const envMap: Record<string, string> = {
-		opencode: "OPENCODE_API_KEY",
-		"opencode-go": "OPENCODE_API_KEY",
-		openrouter: "OPENROUTER_API_KEY",
+		opencode: "$OPENCODE_API_KEY",
+		"opencode-go": "$OPENCODE_API_KEY",
+		openrouter: "$OPENROUTER_API_KEY",
 	};
-	return envMap[providerId] || `${providerId.toUpperCase()}_API_KEY`;
+	return envMap[providerId] || `$${providerId.toUpperCase()}_API_KEY`;
 }

package/lib/probe-cache.ts

@@ -0,0 +1,86 @@
+/**
+ * Provider model probe cache.
+ *
+ * Stores the last successful accessibility probe per provider/model so
+ * background cleanup can avoid spending quota on the same checks every session.
+ */
+
+import { homedir } from "node:os";
+import { join } from "node:path";
+import { createJSONStore } from "./json-persistence.ts";
+import { createLogger } from "./logger.ts";
+
+const _logger = createLogger("probe-cache");
+
+export const DEFAULT_PROBE_TTL_MS = 24 * 60 * 60 * 1000;
+
+export type ProbeStatus = "ok" | "broken";
+
+export interface ModelProbeResult {
+	modelId: string;
+	status: ProbeStatus;
+}
+
+interface ModelProbeEntry {
+	lastProbedAt: string;
+	status: ProbeStatus;
+}
+
+interface ProviderProbeCache {
+	provider: string;
+	models: Record<string, ModelProbeEntry>;
+}
+
+interface ProbeCacheData {
+	providers: Record<string, ProviderProbeCache>;
+}
+
+const CACHE_FILE = join(homedir(), ".pi", "probe-cache.json");
+const _cache = createJSONStore<ProbeCacheData>(CACHE_FILE, { providers: {} });
+
+export function getModelsDueForProbe(
+	providerId: string,
+	modelIds: string[],
+	ttlMs = DEFAULT_PROBE_TTL_MS,
+): string[] {
+	const provider = _cache.load().providers[providerId];
+	const now = Date.now();
+
+	return modelIds.filter((modelId) => {
+		const entry = provider?.models[modelId];
+		if (!entry) return true;
+
+		// Broken models are normally hidden immediately. If a user later unhides one,
+		// re-check it instead of letting a stale broken cache suppress cleanup.
+		if (entry.status === "broken") return true;
+
+		const lastProbedAt = Date.parse(entry.lastProbedAt);
+		if (!Number.isFinite(lastProbedAt)) return true;
+
+		return now - lastProbedAt >= ttlMs;
+	});
+}
+
+export function recordModelProbeResults(
+	providerId: string,
+	results: ModelProbeResult[],
+): void {
+	if (results.length === 0) return;
+
+	const data = _cache.load();
+	const provider = (data.providers[providerId] ??= {
+		provider: providerId,
+		models: {},
+	});
+	const lastProbedAt = new Date().toISOString();
+
+	for (const result of results) {
+		provider.models[result.modelId] = {
+			lastProbedAt,
+			status: result.status,
+		};
+	}
+
+	_cache.save(data);
+	_logger.debug(`Recorded ${results.length} probe results for ${providerId}`);
+}

package/lib/provider-compat.ts

@@ -23,6 +23,10 @@ export function isLikelyReasoningModel(model: ProviderModelIdentity): boolean {
 	const haystack = `${model.id} ${model.name ?? ""}`.toLowerCase();
 	return (
 		isDeepSeekModel(model) ||
+		haystack.includes("minimax") ||
+		haystack.includes("kimi") ||
+		haystack.includes("qwen3.7") ||
+		haystack.includes("qwen3-7") ||
 		haystack.includes("thinking") ||
 		haystack.includes("reasoning") ||
 		haystack.includes("reasoner") ||
@@ -42,5 +46,34 @@ export function getProxyModelCompat(
 		return DEEPSEEK_PROXY_COMPAT;
 	}
 
+	// MiniMax on OpenRouter/Cline uses reasoning_content (DeepSeek format)
+	if (model.id.toLowerCase().includes("minimax")) {
+		return {
+			supportsStore: false,
+			supportsDeveloperRole: false,
+			supportsReasoningEffort: true,
+			requiresReasoningContentOnAssistantMessages: true,
+			thinkingFormat: "deepseek",
+		};
+	}
+
+	// Qwen 3.7+ on OpenRouter/Cline uses reasoning_content (DeepSeek format)
+	if (
+		model.id.toLowerCase().includes("qwen3.7") ||
+		model.id.toLowerCase().includes("qwen3-7")
+	) {
+		return DEEPSEEK_PROXY_COMPAT;
+	}
+
+	// Kimi K2.6 needs reasoning_content on assistant messages (OpenRouter issue #5309)
+	if (model.id.toLowerCase().includes("kimi")) {
+		return {
+			supportsStore: false,
+			supportsDeveloperRole: false,
+			supportsReasoningEffort: true,
+			requiresReasoningContentOnAssistantMessages: true,
+		};
+	}
+
 	return undefined;
 }

package/lib/registry.ts

@@ -82,7 +82,12 @@ function detectPricingExposed(allModels: ProviderModelConfig[]): boolean {
  * @returns true if the model is definitively free per the provider's API
  */
 export function isFreeModel(
-	model: ProviderModelConfig & { provider?: string; _pricingKnown?: boolean },
+	model: ProviderModelConfig & {
+		provider?: string;
+		_pricingKnown?: boolean;
+		_freeKnown?: boolean;
+		_isFree?: boolean;
+	},
 	allModels?: ProviderModelConfig[],
 ): boolean {
 	return isFreeModelInternal(model, allModels);
@@ -90,9 +95,21 @@ export function isFreeModel(
 
 // Internal implementation to work around TypeScript filter callback issues
 function isFreeModelInternal(
-	model: ProviderModelConfig & { provider?: string; _pricingKnown?: boolean },
+	model: ProviderModelConfig & {
+		provider?: string;
+		_pricingKnown?: boolean;
+		_freeKnown?: boolean;
+		_isFree?: boolean;
+	},
 	allModels: ProviderModelConfig[] | undefined,
 ): boolean {
+	// Some gateways expose an authoritative free/paid flag. Prefer it over
+	// pricing because a few non-chat or preview models can report zero token
+	// prices while still not being offered as free chat models.
+	if (model._freeKnown === true) {
+		return model._isFree === true;
+	}
+
 	// Determine if pricing is exposed
 	let pricingExposed: boolean;
 
@@ -213,7 +230,12 @@ export function applyGlobalFilter(
 
 	for (const [providerId, entry] of providerRegistry) {
 		try {
-			applyFilterToProvider(providerId, entry, freeOnly, options.force === true);
+			applyFilterToProvider(
+				providerId,
+				entry,
+				freeOnly,
+				options.force === true,
+			);
 		} catch (err) {
 			_logger.error(
 				`[pi-free] Failed to apply filter to ${providerId}`,