@rigelbuild/omp-catalog 18.0.3

package/src/compat/anthropic.ts

@@ -0,0 +1,179 @@
+/**
+ * Anthropic-messages compat builder — the anthropic-side analogue of
+ * `./openai`. Runs exactly once per model (from `buildModel`); detect-time
+ * defaults come from provider ids, strict host checks, and model-id
+ * classification, with explicit spec overrides assigned on top.
+ */
+import { hostMatchesUrl, modelMatchesHost } from "../hosts";
+import {
+	hasOpus47ApiRestrictions,
+	isAnthropicFableOrMythosModel,
+	isKimiK3ModelId,
+	supportsMidConversationSystemMessages,
+} from "../identity/family";
+import type { ModelSpec, ResolvedAnthropicCompat } from "../types";
+import { applyCompatOverrides } from "./apply";
+
+const OFFICIAL_ANTHROPIC_URL = "https://api.anthropic.com";
+
+/**
+ * Official first-party Anthropic API. A missing baseUrl is official on purpose:
+ * request dispatch falls back to `https://api.anthropic.com`. This is the one
+ * auth-sensitive host check — OAuth credentials are attached based on it — so
+ * it requires the exact origin or a path boundary (`/`) after it; a bare
+ * prefix check would accept lookalikes like `https://api.anthropic.com.evil.com`.
+ */
+export function isOfficialAnthropicApiUrl(baseUrl?: string): boolean {
+	if (!baseUrl) return true;
+	const lower = baseUrl.toLowerCase();
+	return lower === OFFICIAL_ANTHROPIC_URL || lower.startsWith(`${OFFICIAL_ANTHROPIC_URL}/`);
+}
+
+/**
+ * Kimi models whose native endpoints (api.kimi.com, api.moonshot.ai) keep
+ * thinking enabled server-side no matter what the request carries: they
+ * reject disabled thinking (#3852) AND reject forced tool_choice
+ * (`tool_choice 'specified' is incompatible with thinking enabled`), so compat
+ * must keep thinking on and downgrade forced choices to `auto`. Covers the
+ * K2.7 Code family — public id, its Highspeed variant, and the kimi-code
+ * `kimi-for-coding[-highspeed]` aliases — plus K3 (`kimi-k3`, bare `k3` on
+ * kimi-code). Broader than the `compat/openai.ts` mirror: the OpenAI surface
+ * has its own per-dialect policies (K3 forced-choice guard, explicit disable
+ * shapes), while this surface has no working way to force a tool.
+ * Native-only host gating is the caller's responsibility.
+ */
+const KIMI_K27_CODE_MODEL_PATTERN = /(?:^|\/)kimi[-._]?k2(?:[._-]?|p)7[-._]?code(?:[-._]?highspeed)?$/i;
+
+function matchesKimiMandatoryThinkingModel(spec: ModelSpec<"anthropic-messages">): boolean {
+	if (KIMI_K27_CODE_MODEL_PATTERN.test(spec.id)) return true;
+	if (spec.id === "kimi-for-coding" || spec.id === "kimi-for-coding-highspeed") return true;
+	return isKimiK3ModelId(spec.id) || spec.id === "k3";
+}
+
+const CLOUDFLARE_ANTHROPIC_GATEWAY_URL_MARKER = /gateway\.ai\.cloudflare\.com\/.+\/anthropic(?:\/|$)/i;
+const VERTEX_ANTHROPIC_URL_MARKER = /aiplatform\.googleapis\.com\/.+\/publishers\/anthropic\//i;
+const BEDROCK_ANTHROPIC_URL_MARKER = /(?:^|\/\/|\.)bedrock-runtime\.[a-z0-9-]+\.amazonaws\.com/i;
+const AZURE_ANTHROPIC_URL_MARKER = /(?:^|\/\/|\.)[a-z0-9-]+\.(?:inference|services)\.ai\.azure\.com/i;
+
+/**
+ * Cloudflare AI Gateway's `/anthropic` route forwards to signature-enforcing
+ * Anthropic (same failure class as GitHub Copilot #2851 / ZenMux #4192).
+ * Detection is by baseUrl marker rather than provider id: users routinely
+ * declare `provider: "custom"` (or other free-form ids) in `models.yml` for
+ * their own Cloudflare gateway account.
+ */
+function isCloudflareAnthropicGateway(baseUrl?: string): boolean {
+	return baseUrl !== undefined && CLOUDFLARE_ANTHROPIC_GATEWAY_URL_MARKER.test(baseUrl);
+}
+
+/**
+ * Google Vertex's `publishers/anthropic/models/…:streamRawPredict` route
+ * proxies Claude through Google's identity layer and returns full thinking
+ * signatures, so it is a SIGNING endpoint.
+ */
+function isVertexAnthropicRoute(baseUrl?: string): boolean {
+	return baseUrl !== undefined && VERTEX_ANTHROPIC_URL_MARKER.test(baseUrl);
+}
+
+/**
+ * AWS Bedrock's Anthropic route (`bedrock-runtime.<region>.amazonaws.com`)
+ * forwards Claude requests through Anthropic's signature protocol. Users can
+ * front Bedrock with a custom `anthropic-messages` provider entry in
+ * `models.yml`; the URL marker makes those signing by default without
+ * requiring a provider-id list.
+ */
+function isBedrockAnthropicRoute(baseUrl?: string): boolean {
+	return baseUrl !== undefined && BEDROCK_ANTHROPIC_URL_MARKER.test(baseUrl);
+}
+
+/**
+ * Azure AI Inference / Foundry Anthropic route
+ * (`<resource>.inference.ai.azure.com`, `<resource>.services.ai.azure.com`).
+ * Fronts Claude behind Azure identity and enforces Anthropic signatures on
+ * replay.
+ */
+function isAzureAnthropicRoute(baseUrl?: string): boolean {
+	return baseUrl !== undefined && AZURE_ANTHROPIC_URL_MARKER.test(baseUrl);
+}
+
+/**
+ * Known non-official URLs that enforce Anthropic thinking signatures on replay.
+ *
+ * Runtime routing calls this with the effective URL because a model's resolved
+ * compat can be stale after Foundry or a provider base-URL override reroutes it.
+ */
+export function isAnthropicSigningProxyUrl(baseUrl?: string): boolean {
+	return (
+		hostMatchesUrl(baseUrl, "githubCopilot") ||
+		hostMatchesUrl(baseUrl, "zenmux") ||
+		isCloudflareAnthropicGateway(baseUrl) ||
+		isVertexAnthropicRoute(baseUrl) ||
+		isBedrockAnthropicRoute(baseUrl) ||
+		isAzureAnthropicRoute(baseUrl)
+	);
+}
+
+/** Build the resolved anthropic-messages compat record for a model spec. */
+export function buildAnthropicCompat(spec: ModelSpec<"anthropic-messages">): ResolvedAnthropicCompat {
+	const baseUrl = spec.baseUrl;
+	const official = isOfficialAnthropicApiUrl(baseUrl);
+	// Z.AI's Anthropic-compatible proxy lives at `api.z.ai/api/anthropic`.
+	const isZai = modelMatchesHost(spec, "zai");
+	// GitHub Copilot's `anthropic-messages` proxy (api.githubcopilot.com/v1/messages)
+	// rejects the per-tool `eager_input_streaming` field with
+	// `tools.0.custom.eager_input_streaming: Extra inputs are not permitted` and
+	// doesn't whitelist the `fine-grained-tool-streaming-2025-05-14` beta either
+	// (issue #2558), so eager tool-input streaming is unavailable on this host.
+	const isCopilot = modelMatchesHost(spec, "githubCopilot");
+	// ZenMux's `anthropic-messages` route (zenmux.ai/api/anthropic) forwards to
+	// signature-enforcing Anthropic — same failure class as GitHub Copilot #2851
+	// (issue #4192).
+	const isZenmux = modelMatchesHost(spec, "zenmux");
+	const requiresThinkingEnabled = modelMatchesHost(spec, "moonshotNative") && matchesKimiMandatoryThinkingModel(spec);
+	const isAzure = isAzureAnthropicRoute(baseUrl);
+	const signingEndpoint = official || isCopilot || isZenmux || isAnthropicSigningProxyUrl(baseUrl);
+	const compat: ResolvedAnthropicCompat = {
+		officialEndpoint: official,
+		signingEndpoint,
+		disableStrictTools: isAzure,
+		disableAdaptiveThinking: false,
+		allowAnthropicHeaderOverrides: false,
+		supportsEagerToolInputStreaming: official,
+		// Long cache retention is only sent to the official API by default;
+		// proxies opt in explicitly via `compat.supportsLongCacheRetention: true`.
+		supportsLongCacheRetention: official,
+		// First-party Claude API only. Bedrock/Vertex/Foundry and other
+		// Anthropic-compatible gateways reject mid-conversation system roles, so
+		// detection requires the canonical api.anthropic.com host plus a
+		// supported model id.
+		supportsMidConversationSystem: official && supportsMidConversationSystemMessages(spec.id),
+		supportsForcedToolChoice: !requiresThinkingEnabled && !isAnthropicFableOrMythosModel(spec.id),
+		// Opus 4.7+ and Fable/Mythos reject temperature/top_p/top_k with a 400.
+		supportsSamplingParams: !hasOpus47ApiRestrictions(spec.id),
+		// Z.AI workaround (issue #814): its proxy deserializes tool_result blocks
+		// into a class that reads `.id`.
+		requiresToolResultId: isZai,
+		requiresThinkingEnabled,
+		// Official Anthropic and Anthropic-compatible signing proxies enforce
+		// signature-based thinking-chain integrity, so unsigned thinking blocks
+		// must stay text there. Every other `anthropic-messages` reasoning
+		// endpoint replays unsigned thinking natively so the reasoning chain
+		// survives continuation and doesn't destabilize the next tool-call
+		// arguments (#2005, #2257, #2265, #3288, #3433, #3434). Opaque custom
+		// signing proxies remain the reporter's responsibility to mark with
+		// `compat.replayUnsignedThinking: false`; the transport surfaces a
+		// pointed remediation the first time the signing 400 fires (#4297).
+		//
+		// Known signing Anthropic-messages hosts (Copilot, ZenMux, Cloudflare
+		// AI Gateway `/anthropic`, Google Vertex `publishers/anthropic`, AWS
+		// Bedrock `bedrock-runtime.<region>.amazonaws.com`, and Azure
+		// AI Inference / Foundry `<res>.(inference|services).ai.azure.com`)
+		// are excluded automatically because they can be recognised by provider
+		// id or baseUrl marker.
+		replayUnsignedThinking: !signingEndpoint && (Boolean(spec.reasoning) || modelMatchesHost(spec, "deepseekFamily")),
+		escapeBuiltinToolNames: modelMatchesHost(spec, "umans"),
+		streamIdleTimeoutMs: spec.compat?.streamIdleTimeoutMs,
+	};
+	applyCompatOverrides(compat, spec.compat);
+	return compat;
+}

package/src/compat/apply.ts

@@ -0,0 +1,15 @@
+/**
+ * Assign defined override values onto a freshly-built resolved compat record,
+ * in place. Keys the record doesn't declare are ignored (loosely-typed config
+ * may carry junk). `buildModel` is the only intended caller — the record being
+ * mutated is the single per-model allocation; nothing here runs per request.
+ */
+export function applyCompatOverrides(compat: object, overrides: object | undefined): void {
+	if (!overrides) return;
+	for (const key in overrides) {
+		const value = (overrides as Record<string, unknown>)[key];
+		if (value !== undefined && key in compat) {
+			(compat as Record<string, unknown>)[key] = value;
+		}
+	}
+}

package/src/compat/bedrock.ts

@@ -0,0 +1,153 @@
+import { supportsAdaptiveThinkingDisplay } from "../identity/family";
+import type { ModelSpec, ResolvedBedrockCompat } from "../types";
+import { applyCompatOverrides } from "./apply";
+
+const NO_EXPLICIT_CHECKPOINTS: ResolvedBedrockCompat = {
+	promptCacheMode: "none",
+	supportsLongPromptCacheRetention: false,
+	promptCacheMinimumTokens: 0,
+	promptCacheMaximumCheckpoints: 0,
+};
+const EXPLICIT_CHECKPOINTS_1024_5M: ResolvedBedrockCompat = {
+	promptCacheMode: "explicit",
+	supportsLongPromptCacheRetention: false,
+	promptCacheMinimumTokens: 1024,
+	promptCacheMaximumCheckpoints: 4,
+};
+
+const EXPLICIT_CHECKPOINTS_1024_1H: ResolvedBedrockCompat = {
+	promptCacheMode: "explicit",
+	supportsLongPromptCacheRetention: true,
+	promptCacheMinimumTokens: 1024,
+	promptCacheMaximumCheckpoints: 4,
+};
+
+const EXPLICIT_CHECKPOINTS_2048_5M: ResolvedBedrockCompat = {
+	promptCacheMode: "explicit",
+	supportsLongPromptCacheRetention: false,
+	promptCacheMinimumTokens: 2048,
+	promptCacheMaximumCheckpoints: 4,
+};
+
+const EXPLICIT_CHECKPOINTS_4096_5M: ResolvedBedrockCompat = {
+	promptCacheMode: "explicit",
+	supportsLongPromptCacheRetention: false,
+	promptCacheMinimumTokens: 4096,
+	promptCacheMaximumCheckpoints: 4,
+};
+
+const EXPLICIT_CHECKPOINTS_4096_1H: ResolvedBedrockCompat = {
+	promptCacheMode: "explicit",
+	supportsLongPromptCacheRetention: true,
+	promptCacheMinimumTokens: 4096,
+	promptCacheMaximumCheckpoints: 4,
+};
+
+// AWS モデルカード: 512 トークン、最大 4 個のキャッシュチェックポイント、5 分と 1 時間の TTL。
+// https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-opus-5.html
+const EXPLICIT_CHECKPOINTS_512_1H: ResolvedBedrockCompat = {
+	promptCacheMode: "explicit",
+	supportsLongPromptCacheRetention: true,
+	promptCacheMinimumTokens: 512,
+	promptCacheMaximumCheckpoints: 4,
+};
+
+/**
+ * Explicit Nova cache points complement Bedrock's automatic prefix caching:
+ * AWS recommends them for consistent cache hits and input-cost savings. Keep
+ * these exact documented model and inference-profile IDs conservative rather
+ * than treating arbitrary Nova-like application profiles as checkpoint-capable.
+ */
+function detectedBedrockCompat(modelId: string): ResolvedBedrockCompat {
+	const id = modelId.toLowerCase();
+
+	if (
+		id === "amazon.nova-lite-v1:0" ||
+		id === "us.amazon.nova-lite-v1:0" ||
+		id === "amazon.nova-micro-v1:0" ||
+		id === "us.amazon.nova-micro-v1:0" ||
+		id === "amazon.nova-pro-v1:0" ||
+		id === "us.amazon.nova-pro-v1:0" ||
+		id === "amazon.nova-premier-v1:0" ||
+		id === "us.amazon.nova-premier-v1:0" ||
+		id === "amazon.nova-2-lite-v1:0" ||
+		id === "us.amazon.nova-2-lite-v1:0" ||
+		id === "eu.amazon.nova-2-lite-v1:0" ||
+		id === "jp.amazon.nova-2-lite-v1:0" ||
+		id === "global.amazon.nova-2-lite-v1:0"
+	) {
+		return EXPLICIT_CHECKPOINTS_1024_5M;
+	}
+
+	// https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html
+	// This list is deliberately sourced from AWS model cards, not cache pricing:
+	// https://docs.aws.amazon.com/bedrock/latest/userguide/model-cards.html
+	if (
+		id.includes("anthropic.claude-opus-4-5") ||
+		id.includes("anthropic.claude-sonnet-4-5") ||
+		id.includes("anthropic.claude-haiku-4-5") ||
+		id.includes("anthropic.claude-opus-4-7") ||
+		id.includes("anthropic.claude-opus-4-8") ||
+		id.includes("anthropic.claude-sonnet-5")
+	) {
+		return EXPLICIT_CHECKPOINTS_4096_1H;
+	}
+	if (id.includes("anthropic.claude-opus-5")) {
+		return EXPLICIT_CHECKPOINTS_512_1H;
+	}
+	if (id.includes("anthropic.claude-opus-4-6")) {
+		return EXPLICIT_CHECKPOINTS_4096_5M;
+	}
+	if (id.includes("anthropic.claude-3-5-haiku")) {
+		return EXPLICIT_CHECKPOINTS_2048_5M;
+	}
+	if (id.includes("anthropic.claude-fable-5")) {
+		return EXPLICIT_CHECKPOINTS_1024_1H;
+	}
+
+	if (
+		id.includes("anthropic.claude-opus-4-1") ||
+		id.includes("anthropic.claude-opus-4-20250514") ||
+		id.includes("anthropic.claude-sonnet-4-20250514") ||
+		id.includes("anthropic.claude-sonnet-4-6") ||
+		id.includes("anthropic.claude-3-7-sonnet") ||
+		id.includes("anthropic.claude-3-5-sonnet-20241022-v2")
+	) {
+		return EXPLICIT_CHECKPOINTS_1024_5M;
+	}
+
+	return NO_EXPLICIT_CHECKPOINTS;
+}
+
+/**
+ * Bedrock ConverseStream sends no ping/keepalive events, so a reasoning model
+ * that goes quiet mid-thinking (summarized-display gaps, `omitted` thinking,
+ * or the wedged long tool-call generation of issue #4900) reads as a dead
+ * stream to the generic 300s idle watchdog and dies with "Provider stream
+ * stalled while waiting for the next event" (issue #4758's Bedrock variant).
+ * Widen the floor to 600s for reasoning models, mirroring the GLM coding-plan
+ * floor; explicit `spec.compat.streamIdleTimeoutMs` overrides still win.
+ */
+const BEDROCK_REASONING_STREAM_IDLE_TIMEOUT_MS = 600_000;
+/**
+ * Adaptive-thinking Claude (Opus 4.7+, Sonnet/Opus 5, Fable/Mythos 5) reasons
+ * for much longer stretches, and starting with Opus 4.7 / Fable 5 the
+ * Anthropic-side display default is `omitted` (issue #1373), so quiet gaps run
+ * longest on exactly this family — Fable 5 being the worst offender in the
+ * field. Direct Anthropic keeps these streams alive with ping keepalives and
+ * tolerates up to 3x the 300s idle budget of real-event silence (#4900);
+ * pingless Bedrock needs the same 900s tolerance in the raw idle floor.
+ */
+const BEDROCK_ADAPTIVE_THINKING_STREAM_IDLE_TIMEOUT_MS = 900_000;
+
+/** Resolve Bedrock Converse prompt-cache and stream-watchdog compat once per model. */
+export function buildBedrockCompat(spec: ModelSpec<"bedrock-converse-stream">): ResolvedBedrockCompat {
+	const compat = { ...detectedBedrockCompat(spec.id) };
+	compat.streamIdleTimeoutMs = spec.reasoning
+		? supportsAdaptiveThinkingDisplay(spec.id)
+			? BEDROCK_ADAPTIVE_THINKING_STREAM_IDLE_TIMEOUT_MS
+			: BEDROCK_REASONING_STREAM_IDLE_TIMEOUT_MS
+		: undefined;
+	applyCompatOverrides(compat, spec.compat);
+	return compat;
+}

package/src/compat/devin.ts

@@ -0,0 +1,12 @@
+import type { ModelSpec, ResolvedDevinCompat } from "../types";
+
+/**
+ * Resolve devin-agent (Codeium Cascade) compat. Cascade has no wire
+ * reasoning/effort field — effort is selected by routing to a sibling model id
+ * (the `thinking.effortRouting` baked by variant-collapse). So the thinking
+ * deriver must never fabricate an effort ladder from identity for these models;
+ * only explicit routed metadata counts.
+ */
+export function buildDevinCompat(_spec: ModelSpec<"devin-agent">): ResolvedDevinCompat {
+	return { trustExplicitThinkingOnly: true };
+}