@oh-my-pi/pi-ai 16.1.21 → 16.1.22

package/CHANGELOG.md

@@ -2,6 +2,12 @@
 
 ## [Unreleased]
 
+## [16.1.22] - 2026-06-26
+
+### Fixed
+
+- Fixed llama.cpp / LM Studio / vLLM (and any local OpenAI-compatible server on a loopback or RFC1918 baseUrl) re-processing the full prompt on every assistant continuation when the prior turn produced `reasoning_content`: the `openai-completions` encoder dropped the preserved `thinking` block on re-serialization for compat profiles without `requiresReasoningContentForToolCalls` / `thinkingFormat: "zai"`, so the chat template re-rendered the assistant turn without `<think>…</think>` and the rendered tokens diverged from the slot's KV cache state. The auto-learn capture-at-stop nudge made it reproduce on every turn. The encoder now replays preserved thinking as `reasoning_content` (honoring the streamed signature when it identifies a recognized wire field — `reasoning_content` / `reasoning` / `reasoning_text` — and falling back to the configured `reasoningContentField` for opaque signatures) whenever the new `compat.replayReasoningContent` flag is set, and the cross-API `transformMessages` predicate (`openAICompletionsReplaysUnsignedThinking`) honors the same flag ahead of the `model.reasoning` gate so a switch into a discovered local target (where the spec carries `reasoning: false` because the upstream `/models` endpoints don't advertise the capability) still preserves the prior turn's thinking block as signature-stripped reasoning instead of demoting it to conversation text. The chat-template-rendered prefix stays byte-stable across turns and llama.cpp's prefix KV cache survives. ([#3528](https://github.com/can1357/oh-my-pi/issues/3528))
+
 ## [16.1.21] - 2026-06-26
 
 ### Fixed

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-ai",
-	"version": "16.1.21",
+	"version": "16.1.22",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"homepage": "https://omp.sh",
 	"author": "Can Boluk",
@@ -38,9 +38,9 @@
 	},
 	"dependencies": {
 		"@bufbuild/protobuf": "^2.12.0",
-		"@oh-my-pi/pi-catalog": "16.1.21",
-		"@oh-my-pi/pi-utils": "16.1.21",
-		"@oh-my-pi/pi-wire": "16.1.21",
+		"@oh-my-pi/pi-catalog": "16.1.22",
+		"@oh-my-pi/pi-utils": "16.1.22",
+		"@oh-my-pi/pi-wire": "16.1.22",
 		"arktype": "^2.2.0",
 		"zod": "^4"
 	},

package/src/providers/openai-completions.ts

@@ -1805,6 +1805,29 @@ export function convertMessages(
 					// signature on cross-API replays before the block reaches us.
 					const reasoningField = compat.reasoningContentField ?? "reasoning_content";
 					assistantMsg[reasoningField] = nonEmptyThinkingBlocks.map(b => b.thinking).join("\n");
+				} else if (compat.replayReasoningContent) {
+					// Local llama.cpp-style servers (llama.cpp, LM Studio, vLLM, Ollama
+					// in openai-completions mode, custom providers pointed at a
+					// loopback baseUrl) re-tokenize the entire prompt every request.
+					// Qwen3 / DeepSeek-R1 / GLM chat templates reconstruct the prior
+					// assistant turn's `<think>` block from `reasoning_content`; if we
+					// drop the field the template re-renders the assistant turn
+					// without thinking content, the rendered tokens diverge from the
+					// slot's existing KV cache, and llama.cpp falls back to full
+					// prompt re-processing (#3528). Honor the streamed signature when
+					// it identifies a recognized wire field so a model that emitted
+					// `reasoning` (some llama.cpp builds) round-trips to the same
+					// field; otherwise fall back to the configured
+					// `reasoningContentField`. Gated by the new compat flag rather
+					// than the existing `requires*` flags because local servers
+					// accept but don't validate the field — they just need it to
+					// preserve cache locality.
+					const signature = nonEmptyThinkingBlocks[0].thinkingSignature;
+					const reasoningField: OpenAICompletionsReasoningField =
+						signature === "reasoning_content" || signature === "reasoning" || signature === "reasoning_text"
+							? signature
+							: (compat.reasoningContentField ?? "reasoning_content");
+					assistantMsg[reasoningField] = nonEmptyThinkingBlocks.map(b => b.thinking).join("\n");
 				}
 			}
 

package/src/providers/transform-messages.ts

@@ -253,9 +253,20 @@ function isAnthropicMessagesModel(model: Model): model is Model<"anthropic-messa
  * branch.
  */
 function openAICompletionsReplaysUnsignedThinking(model: Model, compat: Model["compat"]): boolean {
-	if (model.api !== "openai-completions" || !model.reasoning) return false;
+	if (model.api !== "openai-completions") return false;
 	if (compat === undefined || !("requiresReasoningContentForToolCalls" in compat)) return false;
 	if (compat.requiresThinkingAsText) return false;
+	// Local llama.cpp-style servers (`replayReasoningContent`) need the replay
+	// for KV-cache prefix reuse — Qwen3 / DeepSeek-R1 / GLM chat templates
+	// reconstruct the prior turn's `<think>` block from `reasoning_content`
+	// (#3528). Checked BEFORE the `model.reasoning` gate: the runtime discovery
+	// paths for `llama.cpp` / `lm-studio` / `openai-models-list` hardcode
+	// `reasoning: false` even when the upstream actually emits reasoning, so
+	// gating on the spec flag here would let a cross-API switch into such a
+	// target demote the prior `thinking` block to text and lose the
+	// cache-stable prefix `replayReasoningContent` is meant to preserve.
+	if (compat.replayReasoningContent) return true;
+	if (!model.reasoning) return false;
 	// Hosts that REQUIRE `reasoning_content` on tool-call turns (DeepSeek
 	// reasoning, Kimi, OpenRouter reasoning, OpenCode thinking-on) already
 	// accept the replay; Z.AI-format hosts (Z.AI, Zhipu, Moonshot Kimi native,