@oh-my-pi/pi-catalog 16.1.22 → 16.2.0

package/CHANGELOG.md

@@ -2,6 +2,31 @@
 
 ## [Unreleased]
 
+## [16.2.0] - 2026-06-27
+
+### Added
+
+- Added GitLab Duo Agent catalog discovery, including namespace selection, live model mapping, and a bundled fallback model for fresh installs.
+- Added OpenAICompat.supportsNamedToolChoice to support forced tool use on string-only OpenAI-compatible chat servers without emitting the named function-object tool_choice shape.
+- Added model metadata support for provider-native remote compaction and compaction-only model selection.
+
+### Changed
+
+- Disabled the thinking-effort selector for GitLab Duo Agent models since the underlying platform parameters are server-fixed.
+
+### Fixed
+
+- Improved GitLab Duo Agent and Duo Workflow namespace and project discovery to robustly handle paginated groups, SSH remotes with custom ports, Git worktrees, self-managed GitLab instances with relative paths, and configuration via GITLAB_DUO_PROJECT_PATH or GITLAB_DUO_PROJECT_ID.
+- Fixed built-in LiteLLM discovery to prefer rich proxy metadata from management endpoints and avoid caching stale capability data.
+- Fixed GitLab Duo Workflow model specifications to resolve correct static context windows, enabling accurate context usage tracking and auto-compaction.
+
+## [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

package/README.md

@@ -24,7 +24,7 @@ Import from subpaths (`@oh-my-pi/pi-catalog/<module>`) or the root barrel.
 Never edit `src/models.json` by hand — it is produced from upstream sources (models.dev, provider catalog discovery, OpenCode docs) by `scripts/generate-models.ts` and the resolvers in `src/provider-models/`. Regenerate with:
 
 ```sh
-bun --cwd=packages/catalog run generate-models
+bun run gen:models
 ```
 
 To change an entry, fix the source: resolver overrides in `provider-models/openai-compat.ts`, provider entries in `provider-models/descriptors.ts`, generator fixups in `scripts/generate-models.ts`, or thinking policies in `model-thinking.ts`.

package/dist/types/discovery/gitlab-duo-workflow.d.ts

@@ -0,0 +1,30 @@
+import type { FetchImpl, ModelSpec } from "../types";
+type GitLabDuoWorkflowCandidateSource = "override" | "project" | "remote" | "group";
+export interface GitLabDuoWorkflowModelRef {
+    name: string;
+    ref: string;
+}
+/**
+ * GitLab Duo Workflow model/namespace discovery configuration.
+ */
+export interface GitLabDuoWorkflowDiscoveryConfig {
+    apiKey: string;
+    baseUrl?: string;
+    fetch?: FetchImpl;
+    namespaceId?: string;
+    projectId?: string;
+    projectPath?: string;
+    cwd?: string;
+}
+export interface GitLabDuoWorkflowNamespaceSelection {
+    rootNamespaceId: string;
+    namespacePath?: string;
+    projectPath?: string;
+    source: GitLabDuoWorkflowCandidateSource;
+}
+export declare function discoverGitLabDuoWorkflowNamespace(config: GitLabDuoWorkflowDiscoveryConfig): Promise<GitLabDuoWorkflowNamespaceSelection>;
+export declare function discoverGitLabDuoWorkflowRuntimeNamespace(config: GitLabDuoWorkflowDiscoveryConfig): Promise<GitLabDuoWorkflowNamespaceSelection>;
+export declare function fetchGitLabDuoWorkflowModels(config: GitLabDuoWorkflowDiscoveryConfig): Promise<readonly ModelSpec<"gitlab-duo-agent">[] | null>;
+export declare function buildGitLabDuoWorkflowModelSpec(model: GitLabDuoWorkflowModelRef, baseUrl?: string, rootNamespaceId?: string): ModelSpec<"gitlab-duo-agent">;
+export declare function buildGitLabDuoWorkflowFallbackModel(id?: string, name?: string, baseUrl?: string): ModelSpec<"gitlab-duo-agent">;
+export {};

package/dist/types/discovery/index.d.ts

@@ -1,4 +1,5 @@
 export * from "./antigravity";
 export * from "./codex";
 export * from "./gemini";
+export * from "./gitlab-duo-workflow";
 export * from "./openai-compatible";

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/descriptor-types.d.ts

@@ -50,7 +50,7 @@ export declare function allowsUnauthenticatedCatalogDiscovery(descriptor: Catalo
  * - Every entry is a member of `KnownProvider`.
  * - `createModelManagerOptions` present (and not `specialModelManager`) ⇒
  *   appears in `PROVIDER_DESCRIPTORS` for runtime model discovery.
- * - `catalogDiscovery` present ⇒ participates in `generate-models.ts`.
+ * - \`catalogDiscovery\` present ⇒ participates in \`generate-models.ts\`.
  */
 export interface ProviderCatalogEntry {
     readonly id: string;

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

@@ -101,6 +101,12 @@ export declare const CATALOG_PROVIDERS: readonly [{
     readonly id: "gitlab-duo";
     readonly defaultModel: "duo-chat-opus-4-6";
     readonly envVars: readonly ["GITLAB_TOKEN"];
+}, {
+    readonly id: "gitlab-duo-agent";
+    readonly defaultModel: "claude_sonnet_4_6_vertex";
+    readonly envVars: readonly ["GITLAB_TOKEN"];
+    readonly createModelManagerOptions: (config: ModelManagerConfig) => import("..").ModelManagerOptions<"gitlab-duo-agent", unknown>;
+    readonly dynamicModelsAuthoritative: true;
 }, {
     readonly id: "google";
     readonly defaultModel: "gemini-3.1-pro-preview";
@@ -339,6 +345,14 @@ 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 {
@@ -116,7 +118,7 @@ export declare const XAI_OAUTH_CURATED_MODELS: readonly XAICuratedModel[];
  * Single source of truth for the curated to Model fan-in, consumed by both
  * - {@link xaiOAuthModelManagerOptions} (runtime static seed handed to the model
  *   manager so the picker is populated on a fresh login), and
- * - `packages/ai/scripts/generate-models.ts` (bundles the same entries into
+ * - \`packages/catalog/scripts/generate-models.ts\` (bundles the same entries into
  *   `models.json`, so the synchronous `ModelRegistry.#loadModels()` boot path
  *   sees `xai-oauth` without waiting for a refresh — fixes the boot-time
  *   default-model reset when `modelRoles.default = "xai-oauth/<id>"`).
@@ -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;
@@ -355,6 +363,20 @@ export interface LiteLLMModelManagerConfig {
     baseUrl?: string;
     fetch?: FetchImpl;
 }
+export interface FetchLiteLLMRichModelsOptions<TApi extends Api> {
+    api: TApi;
+    provider: Provider;
+    baseUrl: string;
+    apiKey?: string;
+    headers?: Record<string, string>;
+    fetch?: FetchImpl;
+    signal?: AbortSignal;
+    referenceResolver?: (modelId: string) => ModelSpec<TApi> | undefined;
+}
+export declare const OPENAI_COMPAT_DISCOVERY_DEFAULT_CONTEXT_WINDOW = 128000;
+export declare const OPENAI_COMPAT_DISCOVERY_DEFAULT_MAX_TOKENS = 32768;
+export declare function normalizeLiteLLMManagementBaseUrl(baseUrl: string): string;
+export declare function fetchLiteLLMRichModels<TApi extends Api>(options: FetchLiteLLMRichModelsOptions<TApi>): Promise<ModelSpec<TApi>[] | null>;
 export declare function litellmModelManagerOptions(config?: LiteLLMModelManagerConfig): ModelManagerOptions<"openai-completions">;
 export interface VllmModelManagerConfig {
     apiKey?: string;

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

@@ -1,5 +1,6 @@
 import type { DevinModelDiscoveryOptions } from "../discovery/devin";
 import type { ModelManagerOptions } from "../model-manager";
+import type { FetchImpl } from "../types";
 export interface OpenAICodexModelManagerConfig {
     accessToken?: string;
     accountId?: string;
@@ -12,6 +13,15 @@ export interface CursorModelManagerConfig {
     clientVersion?: string;
 }
 export declare function cursorModelManagerOptions(config?: CursorModelManagerConfig): ModelManagerOptions<"cursor-agent">;
+export interface GitLabDuoWorkflowModelManagerConfig {
+    apiKey?: string;
+    baseUrl?: string;
+    fetch?: FetchImpl;
+    namespaceId?: string;
+    projectId?: string;
+    cwd?: string;
+}
+export declare function gitLabDuoWorkflowModelManagerOptions(config?: GitLabDuoWorkflowModelManagerConfig): ModelManagerOptions<"gitlab-duo-agent">;
 export interface DevinModelManagerConfig {
     apiKey?: string;
     baseUrl?: string;

package/dist/types/types.d.ts

@@ -1,6 +1,6 @@
 import type { Effort } from "./effort";
 export type { KnownProvider } from "./provider-models/descriptors";
-export type KnownApi = "openai-completions" | "openai-responses" | "openrouter" | "openai-codex-responses" | "azure-openai-responses" | "anthropic-messages" | "bedrock-converse-stream" | "google-generative-ai" | "google-gemini-cli" | "google-vertex" | "ollama-chat" | "cursor-agent" | "devin-agent";
+export type KnownApi = "openai-completions" | "openai-responses" | "openrouter" | "openai-codex-responses" | "azure-openai-responses" | "anthropic-messages" | "bedrock-converse-stream" | "google-generative-ai" | "google-gemini-cli" | "google-vertex" | "ollama-chat" | "cursor-agent" | "gitlab-duo-agent" | "devin-agent";
 export type Api = KnownApi | (string & {});
 /** Canonical thinking transport used by a model. */
 export type ThinkingControlMode = "effort" | "budget" | "google-level" | "anthropic-adaptive" | "anthropic-budget-effort";
@@ -199,6 +199,28 @@ export interface OpenAICompat {
      * 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. */
@@ -209,6 +231,14 @@ export interface OpenAICompat {
      * to provider-default auto selection. Default: true.
      */
     supportsForcedToolChoice?: boolean;
+    /**
+     * Whether the chat-completions endpoint accepts the object form that pins one
+     * named function (`{ type: "function", function: { name } }`). Some
+     * OpenAI-compatible hosts such as llama.cpp only accept string
+     * `tool_choice` values; request builders downgrade a named force to
+     * `"required"` when this is false. Default: true.
+     */
+    supportsNamedToolChoice?: boolean;
     /**
      * Drop reasoning fields (`reasoning_effort`, OpenRouter `reasoning`) for
      * the request when `tool_choice` forces a tool call. Mirrors the Anthropic
@@ -401,11 +431,13 @@ export interface ResolvedOpenAISharedCompat {
     disableReasoningOnToolChoice: boolean;
     supportsToolChoice: boolean;
     supportsForcedToolChoice: boolean;
+    supportsNamedToolChoice: boolean;
     reasoningContentField?: OpenAICompat["reasoningContentField"];
     requiresReasoningContentForToolCalls: boolean;
     requiresReasoningContentForAllAssistantTurns: boolean;
     allowsSyntheticReasoningContentForToolCalls: boolean;
     replayReasoningContent: boolean;
+    qwenPreserveThinking: boolean;
     requiresThinkingAsText: boolean;
     requiresMistralToolIds: boolean;
     requiresToolResultName: boolean;
@@ -433,7 +465,7 @@ 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" | "replayReasoningContent" | "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" | "supportsNamedToolChoice" | "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"];
@@ -494,6 +526,17 @@ export type ResolvedDevinCompat = Required<DevinCompat>;
 export type CompatConfigOf<TApi extends Api> = TApi extends "openai-completions" | "openrouter" | "openai-responses" | "azure-openai-responses" | "openai-codex-responses" ? OpenAICompat : TApi extends "anthropic-messages" ? AnthropicCompat : TApi extends "devin-agent" ? DevinCompat : undefined;
 /** Resolved compat for a given API: complete record, materialized once by `buildModel`. */
 export type CompatOf<TApi extends Api> = TApi extends "openrouter" ? ResolvedOpenRouterCompat : TApi extends "openai-completions" ? ResolvedOpenAICompat : TApi extends "openai-responses" | "azure-openai-responses" | "openai-codex-responses" ? ResolvedOpenAIResponsesCompat : TApi extends "anthropic-messages" ? ResolvedAnthropicCompat : TApi extends "devin-agent" ? ResolvedDevinCompat : undefined;
+/** Provider-native compaction endpoint configuration for one model. */
+export interface RemoteCompactionConfig<TApi extends Api = Api> {
+    /** Enables provider-native compaction for providers not enabled by built-in policy. */
+    enabled?: boolean;
+    /** Adapter family used by the configured compaction endpoint. */
+    api?: TApi;
+    /** Absolute compact endpoint URL; when omitted, the adapter derives it from the model base URL. */
+    endpoint?: string;
+    /** Model id sent to the compaction endpoint when it differs from the active model id. */
+    model?: string;
+}
 export interface Model<TApi extends Api = Api> {
     id: string;
     /**
@@ -523,6 +566,8 @@ export interface Model<TApi extends Api = Api> {
      * reports that native tool calling is unsupported.
      */
     supportsTools?: boolean;
+    /** GitLab Duo Workflow root namespace selected during catalog discovery. */
+    gitlabDuoWorkflowRootNamespaceId?: string;
     cost: {
         input: number;
         output: number;
@@ -565,6 +610,10 @@ export interface Model<TApi extends Api = Api> {
     preferWebsockets?: boolean;
     /** Preferred model to switch to when context promotion is triggered (model id or provider/id). */
     contextPromotionTarget?: string;
+    /** Preferred model to use only for compaction (model id or provider/id); the active session model is unchanged. */
+    compactionModel?: string;
+    /** Provider-native compaction endpoint configuration. */
+    remoteCompaction?: RemoteCompactionConfig<TApi>;
     /** Provider-assigned priority value (lower = higher priority). */
     priority?: number;
     /** Canonical thinking capability metadata for this model. */

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": "16.1.22",
+	"version": "16.2.0",
 	"description": "Model catalog for omp: bundled model database, provider discovery descriptors, model identity, classification, and equivalence",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -30,16 +30,16 @@
 		"test": "bun test --parallel",
 		"fix": "biome check --write --unsafe .",
 		"fmt": "biome format --write .",
-		"generate-models": "bun scripts/generate-models.ts"
+		"gen:models": "bun scripts/generate-models.ts"
 	},
 	"dependencies": {
 		"@bufbuild/protobuf": "^2.12.0",
-		"@oh-my-pi/pi-utils": "16.1.22",
+		"@oh-my-pi/pi-utils": "16.2.0",
 		"arktype": "^2.2.0",
 		"zod": "^4"
 	},
 	"devDependencies": {
-		"@oh-my-pi/pi-ai": "16.1.22",
+		"@oh-my-pi/pi-ai": "16.2.0",
 		"@types/bun": "^1.3.14"
 	},
 	"engines": {

package/src/compat/openai.ts

@@ -403,6 +403,7 @@ export function buildOpenAICompat(spec: ModelSpec<"openai-completions">): Resolv
 		disableReasoningOnToolChoice: isDeepseekFamily && Boolean(spec.reasoning) && !isOpenRouter,
 		supportsToolChoice: !isDirectDeepseekReasoning,
 		supportsForcedToolChoice: true,
+		supportsNamedToolChoice: provider !== "llama.cpp",
 		maxTokensField: useMaxTokens ? "max_tokens" : "max_completion_tokens",
 		requiresToolResultName: isMistral,
 		requiresAssistantAfterToolResult: isMistral,
@@ -468,6 +469,20 @@ export function buildOpenAICompat(spec: ModelSpec<"openai-completions">): Resolv
 		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,
@@ -579,6 +594,7 @@ export function buildOpenAIResponsesCompat(spec: OpenAIResponsesSpecLike): Resol
 		disableReasoningOnToolChoice: isDeepseekFamily && reasoningCapable && !isOpenRouter,
 		supportsToolChoice: true,
 		supportsForcedToolChoice: true,
+		supportsNamedToolChoice: true,
 		reasoningContentField: "reasoning_content",
 		requiresReasoningContentForToolCalls:
 			(isKimiModel || (isDeepseekFamily && reasoningCapable) || (isOpenRouter && reasoningCapable)) &&
@@ -589,6 +605,9 @@ export function buildOpenAIResponsesCompat(spec: OpenAIResponsesSpecLike): Resol
 		// 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,