@oh-my-pi/pi-catalog 16.1.21 → 16.1.23

package/CHANGELOG.md

@@ -2,6 +2,19 @@
 
 ## [Unreleased]
 
+## [16.1.23] - 2026-06-26
+
+### Added
+
+- Added `OpenAICompat.qwenPreserveThinking` — auto-enabled when the resolved `thinkingFormat` is `"qwen"` or `"qwen-chat-template"` AND `replayReasoningContent` is on (i.e. the four built-in local OpenAI-compatible providers, or a custom provider pointed at a loopback / RFC1918 / `*.local` baseUrl). Pairs with the chat-completions encoder change so the request body carries `preserve_thinking: true` (twin top-level + `chat_template_kwargs` emission), keeping Qwen3.6+ from stripping `<think>...</think>` off older assistant turns and breaking the local slot's KV cache between user messages. Non-Qwen chat templates ignore the parameter, so the flag stays a no-op outside the Qwen path; users on a cloud Qwen host (Alibaba Dashscope / Qwen Portal) can opt in with `compat.qwenPreserveThinking: true`. ([#3541](https://github.com/can1357/oh-my-pi/issues/3541))
+- Added CoreWeave Serverless Inference as an OpenAI-compatible provider with models.dev-backed bundled catalog metadata.
+
+## [16.1.22] - 2026-06-26
+
+### Added
+
+- Added `OpenAICompat.replayReasoningContent` — auto-enabled for the built-in local OpenAI-compatible providers (`llama.cpp`, `lm-studio`, `vllm`, `ollama` on `openai-completions`) and for any provider pointed at a loopback / RFC1918 / `*.local` baseUrl. NOT gated on `spec.reasoning`: the runtime discovery paths for `llama.cpp` / `lm-studio` / `openai-models-list` hardcode `reasoning: false` because the upstream `/models` endpoints don't advertise the capability, while the stream parser still records incoming `reasoning_content` deltas as thinking blocks — gating on the spec flag would leave every discovered local Qwen / DeepSeek model re-triggering #3528. The encoder only writes `reasoning_content` when a thinking block actually exists on the turn, so the flag is a no-op on pure-text histories. Built-in proxy providers (currently `litellm`) are excluded from both checks because they forward to an unrelated upstream that gains no KV-cache benefit and may 400 on the extra field; users running a custom proxy in front of a llama.cpp-style backend can opt in via the sparse `compat.replayReasoningContent: true` override. Signals to the `openai-completions` encoder that preserved `thinking` blocks must be re-emitted as `reasoning_content` on every assistant turn so chat templates that reconstruct `<think>…</think>` from the field (Qwen3, DeepSeek-R1, GLM-5.x) keep the prior turn's tokens byte-stable and llama.cpp's prefix KV cache survives. ([#3528](https://github.com/can1357/oh-my-pi/issues/3528))
+
 ## [16.1.20] - 2026-06-25
 
 ### Fixed

package/dist/types/index.d.ts

@@ -12,5 +12,6 @@ export * from "./types";
 export * from "./utils";
 export * from "./variant-collapse";
 export * from "./wire/codex";
+export * from "./wire/coreweave";
 export * from "./wire/gemini-headers";
 export * from "./wire/github-copilot";

package/dist/types/provider-models/descriptors.d.ts

@@ -339,6 +339,14 @@ export declare const CATALOG_PROVIDERS: readonly [{
         readonly label: "Wafer Serverless";
         readonly oauthProvider: "wafer-serverless";
     };
+}, {
+    readonly id: "coreweave";
+    readonly defaultModel: "openai/gpt-oss-120b";
+    readonly envVars: readonly ["COREWEAVE_API_KEY", "WANDB_API_KEY"];
+    readonly createModelManagerOptions: (config: ModelManagerConfig) => import("..").ModelManagerOptions<"openai-completions", unknown>;
+    readonly catalogDiscovery: {
+        readonly label: "CoreWeave Serverless Inference";
+    };
 }, {
     readonly id: "xai";
     readonly defaultModel: "grok-4-fast-non-reasoning";

package/dist/types/provider-models/openai-compat.d.ts

@@ -34,10 +34,12 @@ export interface ModelsDevModel {
  * by the generator's policy pass (scripts/generated-policies.ts).
  */
 export declare const ANTHROPIC_CURATED_FALLBACK_MODELS: readonly ModelSpec<"anthropic-messages">[];
+type SimpleProviderDiscoveryHeaders = Record<string, string> | (() => Record<string, string> | undefined);
 type SimpleProviderConfig = {
     apiKey?: string;
     baseUrl?: string;
     fetch?: FetchImpl;
+    headers?: SimpleProviderDiscoveryHeaders;
 };
 export declare function createSimpleOpenAICompletionsOptions(providerId: Parameters<typeof getBundledModels>[0], defaultBaseUrl: string, config?: SimpleProviderConfig): ModelManagerOptions<"openai-completions">;
 export interface UmansModelManagerConfig {
@@ -307,6 +309,12 @@ export interface TogetherModelManagerConfig {
     fetch?: FetchImpl;
 }
 export declare function togetherModelManagerOptions(config?: TogetherModelManagerConfig): ModelManagerOptions<"openai-completions">;
+export interface CoreWeaveModelManagerConfig {
+    apiKey?: string;
+    baseUrl?: string;
+    fetch?: FetchImpl;
+}
+export declare function coreWeaveModelManagerOptions(config?: CoreWeaveModelManagerConfig): ModelManagerOptions<"openai-completions">;
 export interface MoonshotModelManagerConfig {
     apiKey?: string;
     baseUrl?: string;

package/dist/types/types.d.ts

@@ -185,6 +185,42 @@ export interface OpenAICompat {
     requiresReasoningContentForAllAssistantTurns?: boolean;
     /** Whether the provider accepts a synthetic placeholder (e.g. ".") for missing reasoning_content on tool-call turns. Default: true. Set to false for providers like DeepSeek that validate the exact reasoning_content value. */
     allowsSyntheticReasoningContentForToolCalls?: boolean;
+    /**
+     * Replay preserved thinking blocks as `reasoning_content` (or the configured
+     * `reasoningContentField`) on EVERY assistant turn that carried reasoning,
+     * regardless of whether the upstream provider validates the field. Local
+     * llama.cpp-style servers (llama.cpp, LM Studio, vLLM, sglang, Ollama in
+     * openai-completions mode) re-tokenize the full chat-template prompt every
+     * request; Qwen3 / DeepSeek-R1 / GLM templates reconstruct the `<think>`
+     * block from `reasoning_content`. Dropping the field re-renders the
+     * assistant turn without `<think>`, diverging from the slot's KV cache state
+     * and forcing full prompt re-processing (#3528). Default: auto-detected
+     * (loopback/private baseUrl or local provider id with thinking-enabled
+     * models).
+     */
+    replayReasoningContent?: boolean;
+    /**
+     * Send `preserve_thinking: true` so the Qwen3.6+ chat template renders
+     * `<think>...</think>` markup for EVERY assistant turn (not just turns
+     * after the last user message). Without it, the template strips the think
+     * block from older assistant turns:
+     *
+     * ```jinja
+     * {%- if (preserve_thinking is defined and preserve_thinking is true)
+     *        or (loop.index0 > ns.last_query_index) %}
+     *   <|im_start|>assistant\n<think>\n{rc}\n</think>\n\n{content}
+     * {%- else %}
+     *   <|im_start|>assistant\n{content}
+     * ```
+     *
+     * The cache from the original generation has `<think>...</think>` tokens,
+     * so once a new user message arrives the prior assistant turns become
+     * "older" and the stripped re-render diverges — full prompt re-processing
+     * on SWA models (#3541). Default: auto-detected (Qwen thinking format on
+     * a local llama.cpp-style backend, paired with `replayReasoningContent`).
+     * Non-Qwen templates ignore the flag, so the auto-detection is safe.
+     */
+    qwenPreserveThinking?: boolean;
     /** Whether assistant tool-call messages must include non-empty content. Default: false. */
     requiresAssistantContentForToolCalls?: boolean;
     /** Whether the provider supports the `tool_choice` parameter. Default: true. */
@@ -391,6 +427,8 @@ export interface ResolvedOpenAISharedCompat {
     requiresReasoningContentForToolCalls: boolean;
     requiresReasoningContentForAllAssistantTurns: boolean;
     allowsSyntheticReasoningContentForToolCalls: boolean;
+    replayReasoningContent: boolean;
+    qwenPreserveThinking: boolean;
     requiresThinkingAsText: boolean;
     requiresMistralToolIds: boolean;
     requiresToolResultName: boolean;
@@ -418,7 +456,7 @@ export interface ResolvedOpenAISharedCompat {
  * `buildModel`; request handlers read fields and never detect, resolve, or
  * allocate.
  */
-export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat & Required<Omit<OpenAICompat, "supportsDeveloperRole" | "supportsReasoningEffort" | "reasoningEffortMap" | "supportsReasoningParams" | "thinkingFormat" | "reasoningDisableMode" | "omitReasoningEffort" | "includeEncryptedReasoning" | "filterReasoningHistory" | "disableReasoningOnForcedToolChoice" | "disableReasoningOnToolChoice" | "supportsToolChoice" | "supportsForcedToolChoice" | "reasoningContentField" | "requiresReasoningContentForToolCalls" | "requiresReasoningContentForAllAssistantTurns" | "allowsSyntheticReasoningContentForToolCalls" | "requiresThinkingAsText" | "requiresMistralToolIds" | "requiresToolResultName" | "requiresAssistantAfterToolResult" | "requiresAssistantContentForToolCalls" | "stripDeepseekSpecialTokens" | "streamMarkupHealingPattern" | "reasoningDeltasMayBeCumulative" | "emptyLengthFinishIsContextError" | "usesOpenAIToolCallIdLimit" | "promptCacheSessionHeader" | "openRouterRouting" | "isOpenRouterHost" | "supportsStrictMode" | "supportsLongPromptCacheRetention" | "alwaysSendMaxTokens" | "wireModelIdMode" | "vercelGatewayRouting" | "extraBody" | "toolStrictMode" | "toolSchemaFlavor" | "streamIdleTimeoutMs" | "cacheControlFormat" | "thinkingKeep" | "strictResponsesPairing" | "supportsImageDetailOriginal" | "requiresJuiceZeroHack" | "enableGeminiThinkingLoopGuard" | "whenThinking">> & {
+export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat & Required<Omit<OpenAICompat, "supportsDeveloperRole" | "supportsReasoningEffort" | "reasoningEffortMap" | "supportsReasoningParams" | "thinkingFormat" | "reasoningDisableMode" | "omitReasoningEffort" | "includeEncryptedReasoning" | "filterReasoningHistory" | "disableReasoningOnForcedToolChoice" | "disableReasoningOnToolChoice" | "supportsToolChoice" | "supportsForcedToolChoice" | "reasoningContentField" | "requiresReasoningContentForToolCalls" | "requiresReasoningContentForAllAssistantTurns" | "allowsSyntheticReasoningContentForToolCalls" | "replayReasoningContent" | "qwenPreserveThinking" | "requiresThinkingAsText" | "requiresMistralToolIds" | "requiresToolResultName" | "requiresAssistantAfterToolResult" | "requiresAssistantContentForToolCalls" | "stripDeepseekSpecialTokens" | "streamMarkupHealingPattern" | "reasoningDeltasMayBeCumulative" | "emptyLengthFinishIsContextError" | "usesOpenAIToolCallIdLimit" | "promptCacheSessionHeader" | "openRouterRouting" | "isOpenRouterHost" | "supportsStrictMode" | "supportsLongPromptCacheRetention" | "alwaysSendMaxTokens" | "wireModelIdMode" | "vercelGatewayRouting" | "extraBody" | "toolStrictMode" | "toolSchemaFlavor" | "streamIdleTimeoutMs" | "cacheControlFormat" | "thinkingKeep" | "strictResponsesPairing" | "supportsImageDetailOriginal" | "requiresJuiceZeroHack" | "enableGeminiThinkingLoopGuard" | "whenThinking">> & {
     vercelGatewayRouting?: OpenAICompat["vercelGatewayRouting"];
     extraBody?: OpenAICompat["extraBody"];
     cacheControlFormat?: OpenAICompat["cacheControlFormat"];

package/dist/types/wire/coreweave.d.ts

@@ -0,0 +1,11 @@
+export declare const COREWEAVE_PROJECT_HEADER: "OpenAI-Project";
+export interface CoreWeaveProjectEnv {
+    [key: string]: string | undefined;
+    COREWEAVE_PROJECT?: string;
+    WANDB_INFERENCE_PROJECT?: string;
+    WANDB_ENTITY?: string;
+    WANDB_PROJECT?: string;
+}
+export declare function resolveCoreWeaveProject(env: CoreWeaveProjectEnv): string | undefined;
+export declare function coreWeaveProjectHeaders(env: CoreWeaveProjectEnv): Record<string, string> | undefined;
+export declare function hasCoreWeaveProjectHeader(headers: Record<string, string>): boolean;

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "16.1.21",
+	"version": "16.1.23",
 	"description": "Model catalog for omp: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -34,12 +34,12 @@
 	},
 	"dependencies": {
 		"@bufbuild/protobuf": "^2.12.0",
-		"@oh-my-pi/pi-utils": "16.1.21",
+		"@oh-my-pi/pi-utils": "16.1.23",
 		"arktype": "^2.2.0",
 		"zod": "^4"
 	},
 	"devDependencies": {
-		"@oh-my-pi/pi-ai": "16.1.21",
+		"@oh-my-pi/pi-ai": "16.1.23",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/compat/openai.ts

@@ -166,6 +166,53 @@ function detectStrictModeSupport(provider: string, baseUrl: string): boolean {
 	);
 }
 
+/**
+ * Local OpenAI-compatible inference servers whose chat templates re-tokenize
+ * the entire prompt every request — llama.cpp prefix-KV-cache reuse only
+ * survives when the rendered tokens stay byte-identical across turns. The
+ * runtime auto-enables {@link OpenAICompat.replayReasoningContent} for these
+ * providers (and for any provider pointed at a loopback / RFC1918 baseUrl) so
+ * Qwen3 / DeepSeek-R1 / GLM templates can reconstruct the prior assistant
+ * turn's `<think>` block from `reasoning_content` (#3528).
+ */
+const LOCAL_OPENAI_COMPAT_PROVIDERS = new Set(["llama.cpp", "lm-studio", "vllm", "ollama"]);
+
+/**
+ * Local proxy providers that share the loopback-default baseUrl but forward
+ * to an unrelated upstream (OpenAI, Anthropic, …) rather than running a
+ * chat-template renderer themselves — `replayReasoningContent` would push
+ * `reasoning_content` to the upstream, which gains no KV-cache benefit and
+ * may 400 on the extra field. Excluded from BOTH the provider check above
+ * and the loopback heuristic below; users who want the replay on a custom
+ * proxy setup can opt in via the sparse `compat.replayReasoningContent`
+ * override.
+ */
+const PROXY_OPENAI_COMPAT_PROVIDERS = new Set(["litellm"]);
+
+function hasLocalLoopbackBaseUrl(baseUrl: string | undefined): boolean {
+	if (!baseUrl) return false;
+	let hostname: string;
+	try {
+		hostname = new URL(baseUrl).hostname.toLowerCase();
+	} catch {
+		return false;
+	}
+	if (
+		hostname === "localhost" ||
+		hostname === "127.0.0.1" ||
+		hostname === "0.0.0.0" ||
+		hostname === "::1" ||
+		hostname === "[::1]"
+	) {
+		return true;
+	}
+	if (/^10\./.test(hostname)) return true;
+	if (/^192\.168\./.test(hostname)) return true;
+	if (/^172\.(1[6-9]|2[0-9]|3[01])\./.test(hostname)) return true;
+	if (hostname.endsWith(".local")) return true;
+	return false;
+}
+
 /**
  * Build the resolved chat-completions compat record for a model spec.
  * Provider takes precedence over URL-based detection since it's explicitly configured.
@@ -400,6 +447,41 @@ export function buildOpenAICompat(spec: ModelSpec<"openai-completions">): Resolv
 		// DeepSeek V4 and Xiaomi MiMo reject synthetic reasoning_content placeholders (".") on tool-call turns.
 		// Kimi and OpenRouter accept them when actual reasoning is unavailable.
 		allowsSyntheticReasoningContentForToolCalls: (!isDeepseekFamily || !spec.reasoning) && !isXiaomiMimo,
+		// Local llama.cpp-style servers re-tokenize the entire chat-template
+		// prompt each request; Qwen3 / DeepSeek-R1 / GLM templates reconstruct
+		// the prior assistant turn's `<think>` block from `reasoning_content`,
+		// so dropping the field re-renders the assistant turn without thinking
+		// content and forces full prompt re-processing (#3528). The
+		// `requires*ReasoningContent*` flags above stay off for these hosts —
+		// they accept but don't validate the field — so the encoder needs a
+		// distinct opt-in to replay on every reasoning turn. NOT gated on
+		// `spec.reasoning`: the runtime discovery paths for `llama.cpp` /
+		// `lm-studio` / `openai-models-list` hardcode `reasoning: false`
+		// because the upstream `/models` endpoints don't advertise the
+		// capability, but the OpenAI stream parser still records incoming
+		// `reasoning_content` deltas as thinking blocks. Gating on the spec
+		// flag would leave every discovered local Qwen / DeepSeek model
+		// re-triggering #3528. The encoder only writes `reasoning_content`
+		// when a thinking block actually exists on the turn
+		// (`nonEmptyThinkingBlocks.length > 0`), so the flag is a no-op on
+		// pure-text histories.
+		replayReasoningContent:
+			!PROXY_OPENAI_COMPAT_PROVIDERS.has(provider) &&
+			(LOCAL_OPENAI_COMPAT_PROVIDERS.has(provider) || hasLocalLoopbackBaseUrl(baseUrl)),
+		// `preserve_thinking: true` makes the Qwen3.6+ chat template render
+		// `<think>...</think>` for older assistant turns too, instead of
+		// stripping it the moment a new user message moves them past
+		// `last_query_index`. Without it, the slot's KV cache (which holds the
+		// raw `<think>X</think>` tokens emitted during generation) diverges
+		// from the next-turn render and llama.cpp falls back to full prompt
+		// re-processing — the exact symptom reported in #3541. Auto-enabled
+		// for Qwen thinking dialects on local llama.cpp-style backends (paired
+		// with `replayReasoningContent` above). Non-Qwen templates ignore the
+		// parameter, so the flag stays a no-op outside the Qwen path.
+		qwenPreserveThinking:
+			(thinkingFormat === "qwen" || thinkingFormat === "qwen-chat-template") &&
+			!PROXY_OPENAI_COMPAT_PROVIDERS.has(provider) &&
+			(LOCAL_OPENAI_COMPAT_PROVIDERS.has(provider) || hasLocalLoopbackBaseUrl(baseUrl)),
 		requiresAssistantContentForToolCalls: isKimiModel || isDirectDeepseekReasoning,
 		cacheControlFormat: isOpenRouter && spec.id.startsWith("anthropic/") ? "anthropic" : undefined,
 		openRouterRouting: undefined,
@@ -517,6 +599,13 @@ export function buildOpenAIResponsesCompat(spec: OpenAIResponsesSpecLike): Resol
 			reasoningCapable,
 		requiresReasoningContentForAllAssistantTurns: isDeepseekFamily && reasoningCapable && !isOpenRouter,
 		allowsSyntheticReasoningContentForToolCalls: !isDeepseekFamily || !reasoningCapable,
+		// The Responses API replays reasoning through encrypted `summary` items,
+		// not via a top-level `reasoning_content` field — this flag is
+		// chat-completions-only.
+		replayReasoningContent: false,
+		// Responses-only; the Qwen `preserve_thinking` template knob lives on
+		// the chat-completions wire shape, never on Responses.
+		qwenPreserveThinking: false,
 		requiresThinkingAsText: false,
 		requiresMistralToolIds: false,
 		requiresToolResultName: false,

package/src/index.ts

@@ -12,5 +12,6 @@ export * from "./types";
 export * from "./utils";
 export * from "./variant-collapse";
 export * from "./wire/codex";
+export * from "./wire/coreweave";
 export * from "./wire/gemini-headers";
 export * from "./wire/github-copilot";

package/src/models.json

@@ -13487,6 +13487,802 @@
 			}
 		}
 	},
+	"coreweave": {
+		"deepseek-ai/DeepSeek-V3.1": {
+			"id": "deepseek-ai/DeepSeek-V3.1",
+			"name": "DeepSeek V3.1",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": false,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0.55,
+				"output": 1.65,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 161000,
+			"maxTokens": 161000
+		},
+		"deepseek-ai/DeepSeek-V4-Flash": {
+			"id": "deepseek-ai/DeepSeek-V4-Flash",
+			"name": "DeepSeek V4 Flash",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": true,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0,
+				"output": 0,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 1048576,
+			"maxTokens": 384000,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"minimal",
+					"low",
+					"medium",
+					"high",
+					"xhigh"
+				],
+				"effortMap": {
+					"minimal": "high",
+					"low": "high",
+					"medium": "high",
+					"high": "high",
+					"xhigh": "max"
+				}
+			}
+		},
+		"deepseek-ai/DeepSeek-V4-Pro": {
+			"id": "deepseek-ai/DeepSeek-V4-Pro",
+			"name": "DeepSeek V4 Pro",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": true,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0,
+				"output": 0,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 1048576,
+			"maxTokens": 393216,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"minimal",
+					"low",
+					"medium",
+					"high",
+					"xhigh"
+				],
+				"effortMap": {
+					"minimal": "high",
+					"low": "high",
+					"medium": "high",
+					"high": "high",
+					"xhigh": "max"
+				}
+			}
+		},
+		"google/gemma-4-31B-it": {
+			"id": "google/gemma-4-31B-it",
+			"name": "Gemma 4 31B Instruct",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": true,
+			"input": [
+				"text",
+				"image"
+			],
+			"cost": {
+				"input": 0,
+				"output": 0,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 262144,
+			"maxTokens": 131072,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"minimal",
+					"low",
+					"medium",
+					"high",
+					"xhigh"
+				]
+			}
+		},
+		"ibm-granite/granite-4.1-8b": {
+			"id": "ibm-granite/granite-4.1-8b",
+			"name": "Granite 4.1 8B",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": false,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0,
+				"output": 0,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 131072,
+			"maxTokens": 131072
+		},
+		"JetBrains/Mellum2-12B-A2.5B-Instruct": {
+			"id": "JetBrains/Mellum2-12B-A2.5B-Instruct",
+			"name": "JetBrains/Mellum2-12B-A2.5B-Instruct",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": false,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0,
+				"output": 0,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": null,
+			"maxTokens": null
+		},
+		"meta-llama/Llama-3.1-70B-Instruct": {
+			"id": "meta-llama/Llama-3.1-70B-Instruct",
+			"name": "Llama 3.1 70B",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": false,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0.8,
+				"output": 0.8,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 128000,
+			"maxTokens": 128000
+		},
+		"meta-llama/Llama-3.1-8B-Instruct": {
+			"id": "meta-llama/Llama-3.1-8B-Instruct",
+			"name": "Meta-Llama-3.1-8B-Instruct",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": true,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0.22,
+				"output": 0.22,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 128000,
+			"maxTokens": 128000,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"minimal",
+					"low",
+					"medium",
+					"high",
+					"xhigh"
+				]
+			}
+		},
+		"meta-llama/Llama-3.3-70B-Instruct": {
+			"id": "meta-llama/Llama-3.3-70B-Instruct",
+			"name": "Llama-3.3-70B-Instruct",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": true,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0.71,
+				"output": 0.71,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 128000,
+			"maxTokens": 128000,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"minimal",
+					"low",
+					"medium",
+					"high",
+					"xhigh"
+				]
+			}
+		},
+		"meta-llama/Llama-4-Scout-17B-16E-Instruct": {
+			"id": "meta-llama/Llama-4-Scout-17B-16E-Instruct",
+			"name": "Llama 4 Scout 17B 16E Instruct",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": true,
+			"input": [
+				"text",
+				"image"
+			],
+			"cost": {
+				"input": 0.17,
+				"output": 0.66,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 64000,
+			"maxTokens": 64000,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"minimal",
+					"low",
+					"medium",
+					"high",
+					"xhigh"
+				]
+			}
+		},
+		"microsoft/Phi-4-mini-instruct": {
+			"id": "microsoft/Phi-4-mini-instruct",
+			"name": "Phi-4-mini-instruct",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": true,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0.08,
+				"output": 0.35,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 128000,
+			"maxTokens": 128000,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"minimal",
+					"low",
+					"medium",
+					"high",
+					"xhigh"
+				]
+			}
+		},
+		"MiniMaxAI/MiniMax-M2.5": {
+			"id": "MiniMaxAI/MiniMax-M2.5",
+			"name": "MiniMax M2.5",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": false,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0.3,
+				"output": 1.2,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 196608,
+			"maxTokens": 196608
+		},
+		"moonshotai/Kimi-K2.5": {
+			"id": "moonshotai/Kimi-K2.5",
+			"name": "Kimi K2.5",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": true,
+			"input": [
+				"text",
+				"image"
+			],
+			"cost": {
+				"input": 0.5,
+				"output": 2.85,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 262144,
+			"maxTokens": 262144,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"minimal",
+					"low",
+					"medium",
+					"high",
+					"xhigh"
+				]
+			}
+		},
+		"moonshotai/Kimi-K2.6": {
+			"id": "moonshotai/Kimi-K2.6",
+			"name": "Kimi-K2.6",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": true,
+			"input": [
+				"text",
+				"image"
+			],
+			"cost": {
+				"input": 0,
+				"output": 0,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 262144,
+			"maxTokens": 262144,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"minimal",
+					"low",
+					"medium",
+					"high",
+					"xhigh"
+				]
+			}
+		},
+		"moonshotai/Kimi-K2.7-Code": {
+			"id": "moonshotai/Kimi-K2.7-Code",
+			"name": "Kimi K2.7 Code",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": true,
+			"input": [
+				"text",
+				"image"
+			],
+			"cost": {
+				"input": 0,
+				"output": 0,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 262144,
+			"maxTokens": 262144,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"minimal",
+					"low",
+					"medium",
+					"high",
+					"xhigh"
+				]
+			}
+		},
+		"nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8": {
+			"id": "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8",
+			"name": "NVIDIA Nemotron 3 Super 120B",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": false,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0.2,
+				"output": 0.8,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 262144,
+			"maxTokens": 262144
+		},
+		"nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B": {
+			"id": "nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B",
+			"name": "nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": false,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0,
+				"output": 0,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": null,
+			"maxTokens": null
+		},
+		"openai/gpt-oss-120b": {
+			"id": "openai/gpt-oss-120b",
+			"name": "gpt-oss-120b",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": true,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0.15,
+				"output": 0.6,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 131072,
+			"maxTokens": 131072,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"low",
+					"medium",
+					"high"
+				]
+			}
+		},
+		"openai/gpt-oss-20b": {
+			"id": "openai/gpt-oss-20b",
+			"name": "gpt-oss-20b",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": true,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0.05,
+				"output": 0.2,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 131072,
+			"maxTokens": 131072,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"low",
+					"medium",
+					"high"
+				]
+			}
+		},
+		"OpenPipe/Qwen3-14B-Instruct": {
+			"id": "OpenPipe/Qwen3-14B-Instruct",
+			"name": "OpenPipe Qwen3 14B Instruct",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": false,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0.05,
+				"output": 0.22,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 32768,
+			"maxTokens": 32768
+		},
+		"Qwen/Qwen3-235B-A22B-Instruct-2507": {
+			"id": "Qwen/Qwen3-235B-A22B-Instruct-2507",
+			"name": "Qwen3 235B A22B Instruct 2507",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": false,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0.1,
+				"output": 0.1,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 262144,
+			"maxTokens": 262144
+		},
+		"Qwen/Qwen3-235B-A22B-Thinking-2507": {
+			"id": "Qwen/Qwen3-235B-A22B-Thinking-2507",
+			"name": "Qwen3-235B-A22B-Thinking-2507",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": true,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0.1,
+				"output": 0.1,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 262144,
+			"maxTokens": 262144,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"minimal",
+					"low",
+					"medium",
+					"high"
+				],
+				"requiresEffort": true
+			}
+		},
+		"Qwen/Qwen3-30B-A3B-Instruct-2507": {
+			"id": "Qwen/Qwen3-30B-A3B-Instruct-2507",
+			"name": "Qwen3 30B A3B Instruct 2507",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": false,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0.1,
+				"output": 0.3,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 262144,
+			"maxTokens": 262144
+		},
+		"Qwen/Qwen3-Coder-480B-A35B-Instruct": {
+			"id": "Qwen/Qwen3-Coder-480B-A35B-Instruct",
+			"name": "Qwen3-Coder-480B-A35B-Instruct",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": false,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 1,
+				"output": 1.5,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 262144,
+			"maxTokens": 262144
+		},
+		"Qwen/Qwen3.5-27B": {
+			"id": "Qwen/Qwen3.5-27B",
+			"name": "Qwen3.5 27B",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": true,
+			"input": [
+				"text",
+				"image"
+			],
+			"cost": {
+				"input": 0,
+				"output": 0,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 262144,
+			"maxTokens": 65536,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"minimal",
+					"low",
+					"medium",
+					"high"
+				]
+			}
+		},
+		"Qwen/Qwen3.5-35B-A3B": {
+			"id": "Qwen/Qwen3.5-35B-A3B",
+			"name": "Qwen3.5 35B-A3B",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": true,
+			"input": [
+				"text",
+				"image"
+			],
+			"cost": {
+				"input": 0,
+				"output": 0,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 262144,
+			"maxTokens": 65536,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"minimal",
+					"low",
+					"medium",
+					"high"
+				]
+			}
+		},
+		"Qwen/Qwen3.6-27B": {
+			"id": "Qwen/Qwen3.6-27B",
+			"name": "Qwen3.6 27B",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": true,
+			"input": [
+				"text",
+				"image"
+			],
+			"cost": {
+				"input": 0,
+				"output": 0,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 262144,
+			"maxTokens": 65536,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"minimal",
+					"low",
+					"medium",
+					"high"
+				]
+			}
+		},
+		"Qwen/Qwen3.6-35B-A3B": {
+			"id": "Qwen/Qwen3.6-35B-A3B",
+			"name": "Qwen3.6 35B-A3B",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": true,
+			"input": [
+				"text",
+				"image"
+			],
+			"cost": {
+				"input": 0,
+				"output": 0,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 262144,
+			"maxTokens": 65536,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"minimal",
+					"low",
+					"medium",
+					"high"
+				]
+			}
+		},
+		"zai-org/GLM-5-FP8": {
+			"id": "zai-org/GLM-5-FP8",
+			"name": "GLM 5",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": false,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 1,
+				"output": 3.2,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 200000,
+			"maxTokens": 200000
+		},
+		"zai-org/GLM-5.1": {
+			"id": "zai-org/GLM-5.1",
+			"name": "GLM-5.1",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": true,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 1.4,
+				"output": 4.4,
+				"cacheRead": 0.26,
+				"cacheWrite": 0
+			},
+			"contextWindow": 200000,
+			"maxTokens": 131072,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"minimal",
+					"low",
+					"medium",
+					"high",
+					"xhigh"
+				]
+			}
+		},
+		"zai-org/GLM-5.2": {
+			"id": "zai-org/GLM-5.2",
+			"name": "GLM-5.2",
+			"api": "openai-completions",
+			"provider": "coreweave",
+			"baseUrl": "https://api.inference.wandb.ai/v1",
+			"reasoning": true,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0,
+				"output": 0,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 262144,
+			"maxTokens": 164000,
+			"thinking": {
+				"mode": "effort",
+				"efforts": [
+					"minimal",
+					"low",
+					"medium",
+					"high",
+					"xhigh"
+				]
+			}
+		}
+	},
 	"cursor": {
 		"claude-4.5-opus-high": {
 			"id": "claude-4.5-opus-high",
@@ -87658,4 +88454,4 @@
 			}
 		}
 	}
-}
+}

package/src/provider-models/descriptors.ts

@@ -14,6 +14,7 @@ import {
 	anthropicModelManagerOptions,
 	cerebrasModelManagerOptions,
 	cloudflareAiGatewayModelManagerOptions,
+	coreWeaveModelManagerOptions,
 	deepseekModelManagerOptions,
 	firepassModelManagerOptions,
 	fireworksModelManagerOptions,
@@ -375,6 +376,13 @@ export const CATALOG_PROVIDERS = [
 			oauthProvider: "wafer-serverless",
 		},
 	},
+	{
+		id: "coreweave",
+		defaultModel: "openai/gpt-oss-120b",
+		envVars: ["COREWEAVE_API_KEY", "WANDB_API_KEY"],
+		createModelManagerOptions: (config: ModelManagerConfig) => coreWeaveModelManagerOptions(config),
+		catalogDiscovery: { label: "CoreWeave Serverless Inference" },
+	},
 	{
 		id: "xai",
 		defaultModel: "grok-4-fast-non-reasoning",

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

@@ -10,6 +10,7 @@ import type { ModelManagerOptions } from "../model-manager";
 import { getBundledModels } from "../models";
 import type { Api, FetchImpl, Model, ModelSpec, Provider, ThinkingConfig } from "../types";
 import { isAnthropicOAuthToken, isRecord, toBoolean, toNumber, toPositiveNumber } from "../utils";
+import { coreWeaveProjectHeaders } from "../wire/coreweave";
 import {
 	COPILOT_API_HEADERS,
 	getGitHubCopilotBaseUrl,
@@ -470,7 +471,19 @@ function isLikelyNanoGptTextModelId(id: string): boolean {
 	return !NANO_GPT_NON_TEXT_MODEL_TOKENS.some(token => normalized.includes(token));
 }
 
-type SimpleProviderConfig = { apiKey?: string; baseUrl?: string; fetch?: FetchImpl };
+type SimpleProviderDiscoveryHeaders = Record<string, string> | (() => Record<string, string> | undefined);
+type SimpleProviderConfig = {
+	apiKey?: string;
+	baseUrl?: string;
+	fetch?: FetchImpl;
+	headers?: SimpleProviderDiscoveryHeaders;
+};
+
+function resolveSimpleProviderHeaders(
+	headers: SimpleProviderDiscoveryHeaders | undefined,
+): Record<string, string> | undefined {
+	return typeof headers === "function" ? headers() : headers;
+}
 
 export function createSimpleOpenAICompletionsOptions(
 	providerId: Parameters<typeof getBundledModels>[0],
@@ -489,6 +502,7 @@ export function createSimpleOpenAICompletionsOptions(
 					provider: providerId,
 					baseUrl,
 					apiKey,
+					headers: resolveSimpleProviderHeaders(config?.headers),
 					mapModel: (entry, defaults) => {
 						const reference = references.get(defaults.id);
 						return mapWithBundledReference(entry, defaults, reference);
@@ -516,6 +530,7 @@ function createSimpleOpenAIResponsesOptions(
 					provider: providerId,
 					baseUrl,
 					apiKey,
+					headers: resolveSimpleProviderHeaders(config?.headers),
 					mapModel: (entry, defaults) => {
 						const reference = references.get(defaults.id);
 						return mapWithBundledReference(entry, defaults, reference);
@@ -2480,6 +2495,25 @@ export function togetherModelManagerOptions(
 	return createSimpleOpenAICompletionsOptions("together", "https://api.together.xyz/v1", config);
 }
 
+// ---------------------------------------------------------------------------
+// 15.5 CoreWeave Serverless Inference
+// ---------------------------------------------------------------------------
+
+export interface CoreWeaveModelManagerConfig {
+	apiKey?: string;
+	baseUrl?: string;
+	fetch?: FetchImpl;
+}
+
+export function coreWeaveModelManagerOptions(
+	config?: CoreWeaveModelManagerConfig,
+): ModelManagerOptions<"openai-completions"> {
+	return createSimpleOpenAICompletionsOptions("coreweave", "https://api.inference.wandb.ai/v1", {
+		...config,
+		headers: () => coreWeaveProjectHeaders(Bun.env),
+	});
+}
+
 // ---------------------------------------------------------------------------
 // 16. Moonshot
 // ---------------------------------------------------------------------------
@@ -3653,6 +3687,19 @@ const MODELS_DEV_PROVIDER_DESCRIPTORS_CORE: readonly ModelsDevProviderDescriptor
 	openAiCompletionsDescriptor("cerebras", "cerebras", "https://api.cerebras.ai/v1"),
 	// --- Together ---
 	openAiCompletionsDescriptor("togetherai", "together", "https://api.together.xyz/v1"),
+	// --- CoreWeave Serverless Inference ---
+	openAiCompletionsDescriptor("wandb", "coreweave", "https://api.inference.wandb.ai/v1", {
+		transformModel: model => {
+			if (!model.id.startsWith("openai/gpt-oss-")) {
+				return model;
+			}
+			return {
+				...model,
+				reasoning: true,
+				thinking: { mode: "effort", efforts: [Effort.Low, Effort.Medium, Effort.High] },
+			};
+		},
+	}),
 	// --- NVIDIA ---
 	openAiCompletionsDescriptor("nvidia", "nvidia", "https://integrate.api.nvidia.com/v1", {
 		defaultContextWindow: 131072,

package/src/types.ts

@@ -222,6 +222,42 @@ export interface OpenAICompat {
 	requiresReasoningContentForAllAssistantTurns?: boolean;
 	/** Whether the provider accepts a synthetic placeholder (e.g. ".") for missing reasoning_content on tool-call turns. Default: true. Set to false for providers like DeepSeek that validate the exact reasoning_content value. */
 	allowsSyntheticReasoningContentForToolCalls?: boolean;
+	/**
+	 * Replay preserved thinking blocks as `reasoning_content` (or the configured
+	 * `reasoningContentField`) on EVERY assistant turn that carried reasoning,
+	 * regardless of whether the upstream provider validates the field. Local
+	 * llama.cpp-style servers (llama.cpp, LM Studio, vLLM, sglang, Ollama in
+	 * openai-completions mode) re-tokenize the full chat-template prompt every
+	 * request; Qwen3 / DeepSeek-R1 / GLM templates reconstruct the `<think>`
+	 * block from `reasoning_content`. Dropping the field re-renders the
+	 * assistant turn without `<think>`, diverging from the slot's KV cache state
+	 * and forcing full prompt re-processing (#3528). Default: auto-detected
+	 * (loopback/private baseUrl or local provider id with thinking-enabled
+	 * models).
+	 */
+	replayReasoningContent?: boolean;
+	/**
+	 * Send `preserve_thinking: true` so the Qwen3.6+ chat template renders
+	 * `<think>...</think>` markup for EVERY assistant turn (not just turns
+	 * after the last user message). Without it, the template strips the think
+	 * block from older assistant turns:
+	 *
+	 * ```jinja
+	 * {%- if (preserve_thinking is defined and preserve_thinking is true)
+	 *        or (loop.index0 > ns.last_query_index) %}
+	 *   <|im_start|>assistant\n<think>\n{rc}\n</think>\n\n{content}
+	 * {%- else %}
+	 *   <|im_start|>assistant\n{content}
+	 * ```
+	 *
+	 * The cache from the original generation has `<think>...</think>` tokens,
+	 * so once a new user message arrives the prior assistant turns become
+	 * "older" and the stripped re-render diverges — full prompt re-processing
+	 * on SWA models (#3541). Default: auto-detected (Qwen thinking format on
+	 * a local llama.cpp-style backend, paired with `replayReasoningContent`).
+	 * Non-Qwen templates ignore the flag, so the auto-detection is safe.
+	 */
+	qwenPreserveThinking?: boolean;
 	/** Whether assistant tool-call messages must include non-empty content. Default: false. */
 	requiresAssistantContentForToolCalls?: boolean;
 	/** Whether the provider supports the `tool_choice` parameter. Default: true. */
@@ -433,6 +469,8 @@ export interface ResolvedOpenAISharedCompat {
 	requiresReasoningContentForToolCalls: boolean;
 	requiresReasoningContentForAllAssistantTurns: boolean;
 	allowsSyntheticReasoningContentForToolCalls: boolean;
+	replayReasoningContent: boolean;
+	qwenPreserveThinking: boolean;
 	requiresThinkingAsText: boolean;
 	requiresMistralToolIds: boolean;
 	requiresToolResultName: boolean;
@@ -482,6 +520,8 @@ export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat &
 			| "requiresReasoningContentForToolCalls"
 			| "requiresReasoningContentForAllAssistantTurns"
 			| "allowsSyntheticReasoningContentForToolCalls"
+			| "replayReasoningContent"
+			| "qwenPreserveThinking"
 			| "requiresThinkingAsText"
 			| "requiresMistralToolIds"
 			| "requiresToolResultName"

package/src/wire/coreweave.ts

@@ -0,0 +1,42 @@
+export const COREWEAVE_PROJECT_HEADER = "OpenAI-Project" as const;
+
+export interface CoreWeaveProjectEnv {
+	[key: string]: string | undefined;
+	COREWEAVE_PROJECT?: string;
+	WANDB_INFERENCE_PROJECT?: string;
+	WANDB_ENTITY?: string;
+	WANDB_PROJECT?: string;
+}
+
+function cleanEnvValue(value: string | undefined): string | undefined {
+	const trimmed = value?.trim();
+	return trimmed ? trimmed : undefined;
+}
+
+export function resolveCoreWeaveProject(env: CoreWeaveProjectEnv): string | undefined {
+	const explicitProject = cleanEnvValue(env.COREWEAVE_PROJECT) ?? cleanEnvValue(env.WANDB_INFERENCE_PROJECT);
+	if (explicitProject) {
+		return explicitProject;
+	}
+
+	const wandbProject = cleanEnvValue(env.WANDB_PROJECT);
+	if (!wandbProject) {
+		return undefined;
+	}
+	if (wandbProject.includes("/")) {
+		return wandbProject;
+	}
+
+	const wandbEntity = cleanEnvValue(env.WANDB_ENTITY);
+	return wandbEntity ? `${wandbEntity}/${wandbProject}` : undefined;
+}
+
+export function coreWeaveProjectHeaders(env: CoreWeaveProjectEnv): Record<string, string> | undefined {
+	const project = resolveCoreWeaveProject(env);
+	return project ? { [COREWEAVE_PROJECT_HEADER]: project } : undefined;
+}
+
+export function hasCoreWeaveProjectHeader(headers: Record<string, string>): boolean {
+	const normalized = COREWEAVE_PROJECT_HEADER.toLowerCase();
+	return Object.keys(headers).some(header => header.toLowerCase() === normalized);
+}