@vib-rato/ai 0.16.0

package/src/provider-models/special.ts

@@ -0,0 +1,185 @@
+import { UNK_CONTEXT_WINDOW, UNK_MAX_TOKENS } from "@vib-rato/ai";
+import { once, sanitizeText } from "@vib-rato/utils";
+
+import type { ModelManagerOptions } from "../model-manager";
+import { buildZCodeSourceHeaders, resolveGlmZcodeAnthropicBaseUrl } from "../providers/anthropic";
+import { fetchKiroApiModels, isKiroApiKey, kiroApiStaticModels } from "../providers/kiro-api-key";
+import { fetchOpenCodexModels, OPENCODEX_MODEL_CACHE_TTL_MS } from "../providers/openai-opencodex-responses";
+import { fetchCodexModels } from "../utils/discovery/codex";
+import { fetchOpenAICompatibleModels } from "../utils/discovery/openai-compatible";
+import { createBundledReferenceMap } from "./bundled-references";
+export function openCodexModelManagerOptions(): ModelManagerOptions<"openai-responses"> {
+	return {
+		providerId: "opencodex",
+		cacheTtlMs: OPENCODEX_MODEL_CACHE_TTL_MS,
+		fetchDynamicModels: fetchOpenCodexModels,
+	};
+}
+
+// ---------------------------------------------------------------------------
+// OpenAI code provider
+// ---------------------------------------------------------------------------
+
+export interface OpenAICodexModelManagerConfig {
+	accessToken?: string;
+	accountId?: string;
+	clientVersion?: string;
+}
+
+export function openaiCodexModelManagerOptions(
+	config: OpenAICodexModelManagerConfig = {},
+): ModelManagerOptions<"openai-codex-responses"> {
+	const { accessToken, accountId, clientVersion } = config;
+	return {
+		providerId: "openai-codex",
+		...(accessToken
+			? {
+					fetchDynamicModels: async () => {
+						const result = await fetchCodexModels({ accessToken, accountId, clientVersion });
+						return result?.models ?? null;
+					},
+				}
+			: undefined),
+	};
+}
+
+// ---------------------------------------------------------------------------
+// Cursor
+// ---------------------------------------------------------------------------
+
+export interface CursorModelManagerConfig {
+	apiKey?: string;
+	baseUrl?: string;
+	clientVersion?: string;
+}
+
+export function cursorModelManagerOptions(config: CursorModelManagerConfig = {}): ModelManagerOptions<"cursor-agent"> {
+	const { apiKey, baseUrl, clientVersion } = config;
+	return {
+		providerId: "cursor",
+		...(apiKey
+			? {
+					fetchDynamicModels: async () => {
+						const { fetchCursorUsableModels } = await cursorDiscovery();
+						return fetchCursorUsableModels({ apiKey, baseUrl, clientVersion });
+					},
+				}
+			: undefined),
+	};
+}
+
+const cursorDiscovery = once(() => import("../utils/discovery/cursor"));
+
+// ---------------------------------------------------------------------------
+// Zai
+// ---------------------------------------------------------------------------
+
+export interface ZaiModelManagerConfig {}
+
+export function zaiModelManagerOptions(_config: ZaiModelManagerConfig = {}): ModelManagerOptions<"anthropic-messages"> {
+	return { providerId: "zai" };
+}
+
+// ---------------------------------------------------------------------------
+// GLM ZCode (unofficial Z.AI OAuth)
+// ---------------------------------------------------------------------------
+
+export interface GlmZcodeModelManagerConfig {
+	apiKey?: string;
+	baseUrl?: string;
+}
+
+export function glmZcodeModelManagerOptions(
+	config: GlmZcodeModelManagerConfig = {},
+): ModelManagerOptions<"anthropic-messages"> {
+	const apiKey = config.apiKey;
+	const baseUrl = resolveGlmZcodeAnthropicBaseUrl();
+	const providerRefs = createBundledReferenceMap<"anthropic-messages">("glm-zcode");
+	// Same-family GLM references: the thin `glm-zcode` slice bundles only the
+	// newest model, while the `zai` slice carries the rest of the GLM family
+	// with real capabilities (reasoning, thinking, limits). Resolving through
+	// both keeps a newly selectable GLM model from degrading to generic
+	// unknown metadata; the provider/base are rewritten below.
+	const familyRefs = createBundledReferenceMap<"anthropic-messages">("zai");
+	const resolveReference = (modelId: string) => providerRefs.get(modelId) ?? familyRefs.get(modelId);
+
+	return {
+		providerId: "glm-zcode",
+		...(apiKey ? { cacheDynamicModelProvenance: `${Bun.hash(apiKey).toString(36)}\0${baseUrl}` } : undefined),
+		...(apiKey
+			? {
+					fetchDynamicModels: () =>
+						fetchOpenAICompatibleModels({
+							api: "anthropic-messages",
+							provider: "glm-zcode",
+							baseUrl: baseUrl.endsWith("/v1") ? baseUrl : `${baseUrl}/v1`,
+							apiKey,
+							headers: {
+								...buildZCodeSourceHeaders(),
+								"anthropic-version": "2023-06-01",
+								"anthropic-dangerous-direct-browser-access": "true",
+							},
+							mapModel: (entry, defaults) => {
+								// The remote catalog is provider-controlled text that ends up
+								// rendered in the TUI model selector; strip ANSI/OSC and other
+								// control sequences at the discovery boundary.
+								const remoteName =
+									typeof entry.name === "string" && entry.name.length > 0
+										? sanitizeText(entry.name).replace(/\s+/g, " ").trim().slice(0, 200)
+										: "";
+								const reference = resolveReference(defaults.id);
+								if (!reference) {
+									return { ...defaults, name: remoteName || defaults.id };
+								}
+								return {
+									...reference,
+									id: defaults.id,
+									provider: "glm-zcode",
+									name: remoteName || reference.name,
+									baseUrl,
+									contextWindow:
+										defaults.contextWindow === UNK_CONTEXT_WINDOW
+											? reference.contextWindow
+											: defaults.contextWindow,
+									maxTokens: defaults.maxTokens === UNK_MAX_TOKENS ? reference.maxTokens : defaults.maxTokens,
+								};
+							},
+						}),
+				}
+			: undefined),
+	};
+}
+// ---------------------------------------------------------------------------
+// JetBrains Junie (JetBrains AI Service, Ingrazzio gateway)
+// ---------------------------------------------------------------------------
+
+export interface JetBrainsJunieModelManagerConfig {}
+
+export function jetbrainsJunieModelManagerOptions(
+	_config: JetBrainsJunieModelManagerConfig = {},
+): ModelManagerOptions<"anthropic-messages"> {
+	return { providerId: "jetbrains-junie" };
+}
+
+// ---------------------------------------------------------------------------
+// Kiro (Amazon Q Developer / CodeWhisperer)
+// ---------------------------------------------------------------------------
+
+export interface KiroModelManagerConfig {
+	apiKey?: string;
+}
+
+export function kiroModelManagerOptions(
+	config: KiroModelManagerConfig = {},
+): ModelManagerOptions<"kiro-codewhisperer-stream"> {
+	const apiKey = config.apiKey;
+	return {
+		providerId: "kiro",
+		...(isKiroApiKey(apiKey)
+			? {
+					staticModels: kiroApiStaticModels(),
+					fetchDynamicModels: () => fetchKiroApiModels(apiKey),
+				}
+			: undefined),
+	};
+}

package/src/providers/amazon-bedrock.d.ts

@@ -0,0 +1,60 @@
+/**
+ * Amazon Bedrock Converse Stream provider.
+ *
+ * Talks directly to `bedrock-runtime.{region}.amazonaws.com` over HTTPS with
+ * SigV4 signing and decodes the `application/vnd.amazon.eventstream` response.
+ * No `@aws-sdk/*`, no `@smithy/*`, no `proxy-agent`. Proxies are honored via
+ * Bun's native `HTTPS_PROXY` support.
+ */
+import type { Effort } from "../model-thinking";
+import type { StreamFunction, StreamOptions, ThinkingBudgets, Tool, ToolChoice } from "../types";
+export type BedrockThinkingDisplay = "summarized" | "omitted";
+export interface BedrockOptions extends StreamOptions {
+    region?: string;
+    profile?: string;
+    toolChoice?: ToolChoice;
+    reasoning?: Effort;
+    thinkingBudgets?: ThinkingBudgets;
+    interleavedThinking?: boolean;
+    /**
+     * Controls how Anthropic model returns thinking content in Bedrock responses.
+     * - `"summarized"`: thinking blocks include human-readable summaries (default here).
+     * - `"omitted"`: thinking content is suppressed; the encrypted signature still
+     *   travels back for multi-turn continuity.
+     *
+     * Starting with Anthropic model Opus 4.7 the Anthropic API default is `"omitted"`, which
+     * leaves callers waiting on a silent stream during long reasoning runs (issue
+     * #1373). We default to `"summarized"` so adaptive-thinking models that accept
+     * the field keep producing visible thinking deltas. Older adaptive-thinking
+     * models (Opus 4.6, Sonnet 4.6+) reject the field, so we omit it for them.
+     */
+    thinkingDisplay?: BedrockThinkingDisplay;
+}
+interface WireToolSpec {
+    toolSpec: {
+        name: string;
+        description: string;
+        inputSchema: {
+            json: unknown;
+        };
+    };
+}
+interface WireToolChoice {
+    auto?: Record<string, never>;
+    any?: Record<string, never>;
+    tool?: {
+        name: string;
+    };
+}
+interface WireToolConfig {
+    tools: WireToolSpec[];
+    toolChoice?: WireToolChoice;
+}
+export declare const streamBedrock: StreamFunction<"bedrock-converse-stream">;
+export declare function stripBedrockForcedToolChoiceForRetry<T extends {
+    toolConfig?: {
+        toolChoice?: unknown;
+    };
+}>(body: T): T;
+export declare function convertToolConfig(tools: Tool[] | undefined, toolChoice: BedrockOptions["toolChoice"]): WireToolConfig | undefined;
+export {};