@oh-my-pi/pi-catalog 16.1.21 → 16.1.22

package/CHANGELOG.md

@@ -2,6 +2,12 @@
 
 ## [Unreleased]
 
+## [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/types.d.ts

@@ -185,6 +185,20 @@ 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;
     /** 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 +405,7 @@ export interface ResolvedOpenAISharedCompat {
     requiresReasoningContentForToolCalls: boolean;
     requiresReasoningContentForAllAssistantTurns: boolean;
     allowsSyntheticReasoningContentForToolCalls: boolean;
+    replayReasoningContent: boolean;
     requiresThinkingAsText: boolean;
     requiresMistralToolIds: boolean;
     requiresToolResultName: boolean;
@@ -418,7 +433,7 @@ 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" | "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/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-catalog",
-	"version": "16.1.21",
+	"version": "16.1.22",
 	"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": "16.1.22",
 		"arktype": "^2.2.0",
 		"zod": "^4"
 	},
 	"devDependencies": {
-		"@oh-my-pi/pi-ai": "16.1.21",
+		"@oh-my-pi/pi-ai": "16.1.22",
 		"@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 +447,27 @@ 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)),
 		requiresAssistantContentForToolCalls: isKimiModel || isDirectDeepseekReasoning,
 		cacheControlFormat: isOpenRouter && spec.id.startsWith("anthropic/") ? "anthropic" : undefined,
 		openRouterRouting: undefined,
@@ -517,6 +585,10 @@ 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,
 		requiresThinkingAsText: false,
 		requiresMistralToolIds: false,
 		requiresToolResultName: false,

package/src/types.ts

@@ -222,6 +222,20 @@ 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;
 	/** 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 +447,7 @@ export interface ResolvedOpenAISharedCompat {
 	requiresReasoningContentForToolCalls: boolean;
 	requiresReasoningContentForAllAssistantTurns: boolean;
 	allowsSyntheticReasoningContentForToolCalls: boolean;
+	replayReasoningContent: boolean;
 	requiresThinkingAsText: boolean;
 	requiresMistralToolIds: boolean;
 	requiresToolResultName: boolean;
@@ -482,6 +497,7 @@ export type ResolvedOpenAICompat = ResolvedOpenAISharedCompat &
 			| "requiresReasoningContentForToolCalls"
 			| "requiresReasoningContentForAllAssistantTurns"
 			| "allowsSyntheticReasoningContentForToolCalls"
+			| "replayReasoningContent"
 			| "requiresThinkingAsText"
 			| "requiresMistralToolIds"
 			| "requiresToolResultName"