@linxiraos/pi-catalog 1.4.0 → 1.4.2

package/src/provider-models/descriptors.ts

@@ -40,6 +40,7 @@ import {
 	groqModelManagerOptions,
 	helmcodeModelManagerOptions,
 	huggingfaceModelManagerOptions,
+	inceptionModelManagerOptions,
 	kiloModelManagerOptions,
 	kimiCodeModelManagerOptions,
 	litellmModelManagerOptions,
@@ -126,6 +127,7 @@ const MODEL_MANAGER_FACTORIES: Readonly<Partial<Record<KnownProvider, ModelManag
 	groq: config => groqModelManagerOptions(config),
 	helmcode: config => helmcodeModelManagerOptions(config),
 	huggingface: config => huggingfaceModelManagerOptions(config),
+	inception: config => inceptionModelManagerOptions(config),
 	kilo: config => kiloModelManagerOptions(config),
 	"kimi-code": config => kimiCodeModelManagerOptions(config),
 	litellm: config => litellmModelManagerOptions(config),

package/src/provider-models/openai-compat.d.ts

@@ -419,6 +419,13 @@ export interface TogetherModelManagerConfig {
     fetch?: FetchImpl;
 }
 export declare function togetherModelManagerOptions(config?: TogetherModelManagerConfig): ModelManagerOptions<"openai-completions">;
+/**
+ * Builds Inception's authenticated model-discovery manager. A successful
+ * `/v1/models` fetch is the complete roster, keeping only rows that advertise
+ * `chat.completions` (the shared roster also lists FIM/edit and Decisions-only
+ * models); without an API key the bundled Mercury seeds are served.
+ */
+export declare function inceptionModelManagerOptions(config?: ModelManagerConfig): ModelManagerOptions<"openai-completions">;
 export interface CoreWeaveModelManagerConfig {
     apiKey?: string;
     baseUrl?: string;
@@ -468,8 +475,8 @@ export interface YoloAutoModelManagerConfig {
     fetch?: FetchImpl;
 }
 /**
- * Yolo-Auto model manager: OpenAI-compatible chat completions at the
- * flat-rate `deepseek-flash-v4` endpoint. Live `/v1/models` discovery replaces the
+ * Yolo-Auto model manager: flat-rate OpenAI-compatible chat completions
+ * (`qwen3.8-27b`, `yolo`, ...). Live `/v1/models` discovery replaces the
  * bundled seed once a key is stored or present in `YOLO_AUTO_API_KEY`. Models
  * the provider adds later inherit metadata (reasoning, thinking, context)
  * from the global bundled reference index, so new ids work without a per-id

package/src/provider-models/openai-compat.ts

@@ -569,8 +569,8 @@ function perTokenPricingCost(pricing: unknown): TokenCost {
 	return {
 		input: toPositiveNumber(rates.prompt, 0) * 1_000_000,
 		output: toPositiveNumber(rates.completion, 0) * 1_000_000,
-		cacheRead: toPositiveNumber(rates.input_cache_read, 0) * 1_000_000,
-		cacheWrite: toPositiveNumber(rates.input_cache_write, 0) * 1_000_000,
+		cacheRead: toPositiveNumber(rates.input_cache_read ?? rates.input_cache_reads, 0) * 1_000_000,
+		cacheWrite: toPositiveNumber(rates.input_cache_write ?? rates.input_cache_writes, 0) * 1_000_000,
 	};
 }
 
@@ -4740,6 +4740,48 @@ export function togetherModelManagerOptions(
 	return createSimpleOpenAICompletionsOptions("together", "https://api.together.xyz/v1", config);
 }
 
+// ---------------------------------------------------------------------------
+// Inception
+// ---------------------------------------------------------------------------
+
+/**
+ * Builds Inception's authenticated model-discovery manager. A successful
+ * `/v1/models` fetch is the complete roster, keeping only rows that advertise
+ * `chat.completions` (the shared roster also lists FIM/edit and Decisions-only
+ * models); without an API key the bundled Mercury seeds are served.
+ */
+export function inceptionModelManagerOptions(config?: ModelManagerConfig): ModelManagerOptions<"openai-completions"> {
+	return {
+		...createOpenAICompatibleModelManagerOptions({
+			api: "openai-completions",
+			providerId: "inception",
+			defaultBaseUrl: "https://api.inceptionlabs.ai/v1",
+			config,
+			requireApiKey: true,
+			dynamicModelsAuthoritative: true,
+			mapModel: (entry, defaults, reference) => {
+				// The shared roster also contains FIM/edit and Decisions-only models;
+				// `supported_endpoints` is required, so a row without it is not chat.
+				if (!Array.isArray(entry.supported_endpoints) || !entry.supported_endpoints.includes("chat.completions")) {
+					return null;
+				}
+				const model = mapWithBundledReference(entry, defaults, reference);
+				return {
+					...model,
+					input: Array.isArray(entry.input_modalities) ? toInputCapabilities(entry.input_modalities) : model.input,
+					cost: isRecord(entry.pricing) ? perTokenPricingCost(entry.pricing) : model.cost,
+					contextWindow: toPositiveNumber(entry.context_length, model.contextWindow),
+					maxTokens: toPositiveNumber(entry.max_output_length, model.maxTokens),
+					...(Array.isArray(entry.supported_features) && {
+						supportsTools: entry.supported_features.includes("tools"),
+					}),
+				};
+			},
+		}),
+		staticModels: seedModels<"openai-completions">("inception"),
+	};
+}
+
 // ---------------------------------------------------------------------------
 // 15.5 CoreWeave Serverless Inference
 // ---------------------------------------------------------------------------
@@ -4928,7 +4970,7 @@ export interface MoonshotModelManagerConfig {
  * use the K2.x binary `thinking: { type }` block, so the wire path routes it
  * through OpenAI-style `reasoning_effort` (see `buildOpenAICompat`).
  */
-const MOONSHOT_KIMI_K3_COST = { input: 3, output: 15, cacheRead: 0.3, cacheWrite: 0 } as const;
+const MOONSHOT_KIMI_K3_COST = { input: 3, output: 15, cacheRead: 0.3, cacheWrite: 3 } as const;
 export function moonshotModelManagerOptions(
 	config?: MoonshotModelManagerConfig,
 ): ModelManagerOptions<"openai-completions"> {
@@ -5271,8 +5313,8 @@ function mapYoloAutoModel(
 	};
 }
 /**
- * Yolo-Auto model manager: OpenAI-compatible chat completions at the
- * flat-rate `deepseek-flash-v4` endpoint. Live `/v1/models` discovery replaces the
+ * Yolo-Auto model manager: flat-rate OpenAI-compatible chat completions
+ * (`qwen3.8-27b`, `yolo`, ...). Live `/v1/models` discovery replaces the
  * bundled seed once a key is stored or present in `YOLO_AUTO_API_KEY`. Models
  * the provider adds later inherit metadata (reasoning, thinking, context)
  * from the global bundled reference index, so new ids work without a per-id

package/src/types.d.ts

@@ -688,8 +688,8 @@ export interface AnthropicCompat {
     fastMode?: boolean;
 }
 /**
- * Compatibility settings for Bedrock Converse prompt caching. Cache pricing is
- * deliberately not used to infer these request-shape capabilities.
+ * Compatibility settings for Bedrock Converse requests and native streaming
+ * side-compaction. Cache pricing does not imply these capabilities.
  */
 export interface BedrockCompat {
     /** Explicit disabled-thinking wire form; unset preserves the provider's existing behavior. */
@@ -737,8 +737,13 @@ export interface BedrockCompat {
      * local prefix-binding safeguards. Default: true.
      */
     supportsThinkingBindingControls?: boolean;
+    /**
+     * Whether the model supports a compaction-only Bedrock Runtime
+     * InvokeModelWithResponseStream request. Ordinary turns still use Converse.
+     */
+    supportsServerCompaction?: boolean;
 }
-/** Fully-resolved Bedrock Converse prompt-cache capabilities, materialized once by `buildModel`. */
+/** Fully-resolved Bedrock Converse and side-compaction capabilities, materialized once by `buildModel`. */
 export interface ResolvedBedrockCompat {
     /** See {@link BedrockCompat.disabledThinking}. */
     disabledThinking?: BedrockCompat["disabledThinking"];
@@ -753,6 +758,8 @@ export interface ResolvedBedrockCompat {
     supportsForcedToolChoice: boolean;
     /** See {@link BedrockCompat.supportsThinkingBindingControls}. */
     supportsThinkingBindingControls: boolean;
+    /** See {@link BedrockCompat.supportsServerCompaction}; false unless the Bedrock model-card policy opts in. */
+    supportsServerCompaction: boolean;
     /**
      * Stream-watchdog idle-timeout fallback in ms for hosts with no keepalive
      * events; 0 disables the idle watchdog. Undefined defers to

package/src/types.ts

@@ -810,8 +810,8 @@ export interface AnthropicCompat {
 }
 
 /**
- * Compatibility settings for Bedrock Converse prompt caching. Cache pricing is
- * deliberately not used to infer these request-shape capabilities.
+ * Compatibility settings for Bedrock Converse requests and native streaming
+ * side-compaction. Cache pricing does not imply these capabilities.
  */
 export interface BedrockCompat {
 	/** Explicit disabled-thinking wire form; unset preserves the provider's existing behavior. */
@@ -859,9 +859,14 @@ export interface BedrockCompat {
 	 * local prefix-binding safeguards. Default: true.
 	 */
 	supportsThinkingBindingControls?: boolean;
+	/**
+	 * Whether the model supports a compaction-only Bedrock Runtime
+	 * InvokeModelWithResponseStream request. Ordinary turns still use Converse.
+	 */
+	supportsServerCompaction?: boolean;
 }
 
-/** Fully-resolved Bedrock Converse prompt-cache capabilities, materialized once by `buildModel`. */
+/** Fully-resolved Bedrock Converse and side-compaction capabilities, materialized once by `buildModel`. */
 export interface ResolvedBedrockCompat {
 	/** See {@link BedrockCompat.disabledThinking}. */
 	disabledThinking?: BedrockCompat["disabledThinking"];
@@ -876,6 +881,8 @@ export interface ResolvedBedrockCompat {
 	supportsForcedToolChoice: boolean;
 	/** See {@link BedrockCompat.supportsThinkingBindingControls}. */
 	supportsThinkingBindingControls: boolean;
+	/** See {@link BedrockCompat.supportsServerCompaction}; false unless the Bedrock model-card policy opts in. */
+	supportsServerCompaction: boolean;
 	/**
 	 * Stream-watchdog idle-timeout fallback in ms for hosts with no keepalive
 	 * events; 0 disables the idle watchdog. Undefined defers to

package/src/wire/codex.d.ts

@@ -78,8 +78,10 @@ export declare function getCodexAccountId(accessToken: string): string | undefin
  * needed: `chatgpt_data_residency` is the authoritative claim, with
  * `chatgpt_compute_residency` as the fallback for tokens that only carry that.
  *
- * Returns undefined for a non-JWT token, or for the (common) accounts whose
- * claims omit residency entirely — those workspaces are not region-pinned.
+ * Returns undefined for a non-JWT token, or for the (common) accounts that are
+ * not region-pinned: their claims omit residency or carry the
+ * {@link CODEX_UNCONSTRAINED_RESIDENCY} sentinel. The backend answers 404 to a
+ * declared `no_constraint`, and codex-rs never sends it.
  */
 export declare function getCodexResidency(accessToken: string): string | undefined;
 /**

package/src/wire/codex.ts

@@ -105,6 +105,9 @@ export function getCodexAccountId(accessToken: string): string | undefined {
 	}
 }
 
+/** Residency claim value of an account that is not region-pinned. */
+const CODEX_UNCONSTRAINED_RESIDENCY = "no_constraint";
+
 /**
  * Extract the account's data residency from a Codex JWT access token.
  *
@@ -115,8 +118,10 @@ export function getCodexAccountId(accessToken: string): string | undefined {
  * needed: `chatgpt_data_residency` is the authoritative claim, with
  * `chatgpt_compute_residency` as the fallback for tokens that only carry that.
  *
- * Returns undefined for a non-JWT token, or for the (common) accounts whose
- * claims omit residency entirely — those workspaces are not region-pinned.
+ * Returns undefined for a non-JWT token, or for the (common) accounts that are
+ * not region-pinned: their claims omit residency or carry the
+ * {@link CODEX_UNCONSTRAINED_RESIDENCY} sentinel. The backend answers 404 to a
+ * declared `no_constraint`, and codex-rs never sends it.
  */
 export function getCodexResidency(accessToken: string): string | undefined {
 	try {
@@ -130,7 +135,10 @@ export function getCodexResidency(accessToken: string): string | undefined {
 		for (const claim of [auth?.chatgpt_data_residency, auth?.chatgpt_compute_residency]) {
 			if (typeof claim !== "string") continue;
 			const residency = claim.trim();
-			if (residency.length > 0) return residency;
+			if (residency.length === 0) continue;
+			// The first present claim is authoritative, so an unconstrained data
+			// residency is not overridden by a compute residency.
+			return residency.toLowerCase() === CODEX_UNCONSTRAINED_RESIDENCY ? undefined : residency;
 		}
 		return undefined;
 	} catch {

package/src/wire/devin.d.ts

@@ -6,23 +6,9 @@
  */
 /** Base host for Codeium/Windsurf's Cascade API (Connect protocol over HTTP/1.1). */
 export declare const DEVIN_DEFAULT_BASE_URL = "https://server.codeium.com";
-/** Session token as the wire format carries it: the scheme prefix is required. */
-export declare function normalizeDevinSessionToken(apiKey: string | undefined): string;
-/** Released-CLI metadata with credential bytes already encoded for the wire. */
-export declare function devinWireMetadata(apiKey: string | undefined, userJwt?: string): {
-    ideName: "devin-cli";
-    ideType: "chisel";
-    ideVersion: "3000.11.3";
-    extensionName: "chisel";
-    extensionVersion: "3000.11.3";
-    locale: "en";
-    os: "darwin" | "linux" | "windows";
-    apiKey: string;
-    userJwt: string;
-};
 /**
  * Fields for `Metadata` on released-CLI calls (`GetUserJwt`, `AssignModel`,
- * `GetChatMessage`, `GetUserStatus`) authenticated by a Devin session token.
+ * `GetChatMessage`, `GetUserStatus`).
  */
 export declare function devinCliMetadata(apiKey: string | undefined, userJwt?: string): {
     ideName: "devin-cli";

package/src/wire/devin.ts

@@ -17,7 +17,7 @@ const DEVIN_LOCALE = "en";
 /**
  * Released Devin CLI request identity. The backend gates behavior on this
  * tuple: `ideType: "chisel"` is what unlocks router assignment (`AssignModel`)
- * and the CLI model surface, which the older Windsurf identity does not reach.
+ * and the CLI model surface.
  */
 const DEVIN_CLI_METADATA = {
 	ideName: "devin-cli",
@@ -31,9 +31,8 @@ const DEVIN_CLI_METADATA = {
 
 /**
  * Native discovery identity. The Devin CLI announces itself as the `chisel`
- * client on its dev channel for `GetCliModelConfigs`; that identity — not the
- * released chat identity or the legacy Windsurf one — unlocks the full native
- * config set.
+ * client on its dev channel for `GetCliModelConfigs`; that identity, not the
+ * released chat identity, unlocks the full native config set.
  */
 const DEVIN_DISCOVERY_METADATA = {
 	ideName: "chisel",
@@ -44,33 +43,30 @@ const DEVIN_DISCOVERY_METADATA = {
 	os: DEVIN_OS,
 } as const;
 
-/** Session token as the wire format carries it: the scheme prefix is required. */
-export function normalizeDevinSessionToken(apiKey: string | undefined): string {
-	if (!apiKey) return "";
-	return apiKey.startsWith(DEVIN_SESSION_TOKEN_PREFIX) ? apiKey : `${DEVIN_SESSION_TOKEN_PREFIX}${apiKey}`;
-}
-
-/** Released-CLI metadata with credential bytes already encoded for the wire. */
-export function devinWireMetadata(apiKey: string | undefined, userJwt = "") {
-	return {
-		apiKey: apiKey ?? "",
-		userJwt,
-		...DEVIN_CLI_METADATA,
-	};
+/** Prefixes bare JWT-shaped session tokens; every other key, legacy `sk-ws-` ones included, goes out as stored. */
+function encodeDevinCredential(apiKey: string | undefined): string {
+	const key = apiKey?.trim();
+	if (!key) return "";
+	if (key.startsWith(DEVIN_SESSION_TOKEN_PREFIX)) return key;
+	return key.split(".").length === 3 ? `${DEVIN_SESSION_TOKEN_PREFIX}${key}` : key;
 }
 
 /**
  * Fields for `Metadata` on released-CLI calls (`GetUserJwt`, `AssignModel`,
- * `GetChatMessage`, `GetUserStatus`) authenticated by a Devin session token.
+ * `GetChatMessage`, `GetUserStatus`).
  */
 export function devinCliMetadata(apiKey: string | undefined, userJwt = "") {
-	return devinWireMetadata(normalizeDevinSessionToken(apiKey), userJwt);
+	return {
+		apiKey: encodeDevinCredential(apiKey),
+		userJwt,
+		...DEVIN_CLI_METADATA,
+	};
 }
 
 /** Fields for `Metadata` on the dev-channel `GetCliModelConfigs` call. */
 export function devinDiscoveryMetadata(apiKey: string | undefined) {
 	return {
-		apiKey: normalizeDevinSessionToken(apiKey),
+		apiKey: encodeDevinCredential(apiKey),
 		...DEVIN_DISCOVERY_METADATA,
 	};
 }