@oh-my-pi/pi-ai 14.2.1 → 14.3.0

package/CHANGELOG.md

@@ -2,6 +2,19 @@
 
 ## [Unreleased]
 
+## [14.3.0] - 2026-04-25
+
+### Added
+
+- Added support for Claude Opus 4.7 (`claude-opus-4-7`) model ([#726](https://github.com/can1357/oh-my-pi/issues/726))
+  - Suppresses sampling parameters (temperature/top_p/top_k) that Opus 4.7 rejects
+  - Enables `display: "summarized"` for adaptive thinking to restore visible thinking content
+
+### Fixed
+
+- Fixed Cursor provider losing conversation history on follow-up turns (model responding "this appears to be the start of our session") by populating `ConversationStateStructure.rootPromptMessagesJson` with JSON blob IDs for the system prompt plus prior user/assistant/tool-result messages. Cursor's server builds the model prompt from `rootPromptMessagesJson`, not from the protobuf `turns[]` tree, so sending only the system prompt there caused prior turns to be dropped
+- Fixed Cursor provider multi-turn conversations failing with `Connect error internal: Blob not found` on the second message by storing `ConversationStateStructure.turns`, `AgentConversationTurnStructure.user_message`, and `AgentConversationTurnStructure.steps` as content-addressed blob IDs in the KV store (matching the existing handling for `rootPromptMessagesJson`) rather than sending the raw serialized bytes inline ([#678](https://github.com/can1357/oh-my-pi/issues/678))
+
 ## [14.2.1] - 2026-04-24
 
 ### Fixed

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-ai",
-	"version": "14.2.1",
+	"version": "14.3.0",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"homepage": "https://github.com/can1357/oh-my-pi",
 	"author": "Can Boluk",
@@ -41,24 +41,24 @@
 		"generate-models": "bun scripts/generate-models.ts"
 	},
 	"dependencies": {
-		"@anthropic-ai/sdk": "^0.78",
-		"@aws-sdk/client-bedrock-runtime": "^3",
-		"@aws-sdk/credential-provider-node": "^3",
-		"@bufbuild/protobuf": "^2.11",
-		"@google/genai": "^1.43",
-		"@oh-my-pi/pi-natives": "14.2.1",
-		"@oh-my-pi/pi-utils": "14.2.1",
-		"@sinclair/typebox": "^0.34",
-		"@smithy/node-http-handler": "^4.4",
-		"ajv": "^8.18",
-		"ajv-formats": "^3.0",
-		"openai": "^6.25",
-		"partial-json": "^0.1",
-		"proxy-agent": "^6.5",
+		"@anthropic-ai/sdk": "^0.91.1",
+		"@aws-sdk/client-bedrock-runtime": "^3.1037.0",
+		"@aws-sdk/credential-provider-node": "^3.972.36",
+		"@bufbuild/protobuf": "^2.12.0",
+		"@google/genai": "^1.50.1",
+		"@oh-my-pi/pi-natives": "14.3.0",
+		"@oh-my-pi/pi-utils": "14.3.0",
+		"@sinclair/typebox": "^0.34.49",
+		"@smithy/node-http-handler": "^4.6.1",
+		"ajv": "^8.20.0",
+		"ajv-formats": "^3.0.1",
+		"openai": "^6.34.0",
+		"partial-json": "^0.1.7",
+		"proxy-agent": "^8.0.1",
 		"zod": "4.3.6"
 	},
 	"devDependencies": {
-		"@types/bun": "^1.3"
+		"@types/bun": "^1.3.13"
 	},
 	"engines": {
 		"bun": ">=1.3.7"

package/src/model-thinking.ts

@@ -154,19 +154,27 @@ export function applyGeneratedModelPolicies(models: ApiModel<Api>[]): void {
 }
 
 /**
- * Link `-spark` model variants to their base models for context promotion.
+ * Link OpenAI model variants to their context promotion targets.
  *
- * When a spark model's context is exhausted, the agent can promote to the
- * corresponding full model. This sets `contextPromotionTarget` on each
- * spark variant that has a matching base model.
+ * When a model's context is exhausted, the agent can promote to a sibling
+ * model with a larger context window on the same provider:
+ * - `-spark` variants promote to `gpt-5.5`.
+ * - `gpt-5.5` (270K input) promotes to `gpt-5.4` (1M input).
  */
-export function linkSparkPromotionTargets(models: ApiModel<Api>[]): void {
+export function linkOpenAIPromotionTargets(models: ApiModel<Api>[]): void {
 	for (const candidate of models) {
 		const parsedCandidate = parseKnownModel(candidate.id);
-		if (parsedCandidate.family !== "openai" || parsedCandidate.variant !== "codex-spark") continue;
-		const baseId = candidate.id.slice(0, -"-spark".length);
+		if (parsedCandidate.family !== "openai") continue;
+		let targetId: string | undefined;
+		if (parsedCandidate.variant === "codex-spark") {
+			targetId = "gpt-5.5";
+		} else if (parsedCandidate.variant === "base" && semverEqual(parsedCandidate.version, "5.5")) {
+			targetId = "gpt-5.4";
+		} else {
+			continue;
+		}
 		const fallback = models.find(
-			model => model.provider === candidate.provider && model.api === candidate.api && model.id === baseId,
+			model => model.provider === candidate.provider && model.api === candidate.api && model.id === targetId,
 		);
 		if (!fallback) continue;
 		candidate.contextPromotionTarget = `${fallback.provider}/${fallback.id}`;
@@ -283,6 +291,17 @@ export function mapEffortToAnthropicAdaptiveEffort<TApi extends Api>(
 	}
 }
 
+/**
+ * Returns true for Anthropic models with Opus 4.7 API restrictions:
+ * - Sampling parameters (temperature/top_p/top_k) return 400 error
+ * - Thinking content is omitted by default (needs display: "summarized")
+ */
+export function hasOpus47ApiRestrictions(modelId: string): boolean {
+	const parsed = parseAnthropicModel(getCanonicalModelId(modelId));
+	if (!parsed) return false;
+	return semverGte(parsed.version, "4.7") && parsed.kind === "opus";
+}
+
 function anthropicModelHasRealXHighEffort<TApi extends Api>(model: ApiModel<TApi>): boolean {
 	if (model.api !== "anthropic-messages") return false;
 	const parsedModel = parseKnownModel(model.id);

package/src/models.json

@@ -16931,7 +16931,7 @@
 			},
 			"contextWindow": 128000,
 			"maxTokens": 128000,
-			"contextPromotionTarget": "litellm/gpt-5.3-codex",
+			"contextPromotionTarget": "litellm/gpt-5.5",
 			"thinking": {
 				"mode": "effort",
 				"minLevel": "low",
@@ -17011,7 +17011,8 @@
 				"mode": "effort",
 				"minLevel": "low",
 				"maxLevel": "xhigh"
-			}
+			},
+			"contextPromotionTarget": "litellm/gpt-5.4"
 		},
 		"gpt-image-2": {
 			"id": "gpt-image-2",
@@ -32938,7 +32939,7 @@
 				"maxLevel": "xhigh"
 			},
 			"applyPatchToolType": "freeform",
-			"contextPromotionTarget": "openai/gpt-5.3-codex"
+			"contextPromotionTarget": "openai/gpt-5.5"
 		},
 		"gpt-5.4": {
 			"id": "gpt-5.4",
@@ -33068,7 +33069,8 @@
 				"minLevel": "low",
 				"maxLevel": "xhigh"
 			},
-			"applyPatchToolType": "freeform"
+			"applyPatchToolType": "freeform",
+			"contextPromotionTarget": "openai/gpt-5.4"
 		},
 		"o1": {
 			"id": "o1",
@@ -33597,7 +33599,7 @@
 			},
 			"contextWindow": 128000,
 			"maxTokens": 128000,
-			"contextPromotionTarget": "openai-codex/gpt-5.3-codex",
+			"contextPromotionTarget": "openai-codex/gpt-5.5",
 			"thinking": {
 				"mode": "effort",
 				"minLevel": "low",
@@ -33715,7 +33717,8 @@
 				"minLevel": "low",
 				"maxLevel": "xhigh"
 			},
-			"applyPatchToolType": "freeform"
+			"applyPatchToolType": "freeform",
+			"contextPromotionTarget": "openai-codex/gpt-5.4"
 		}
 	},
 	"opencode": {
@@ -33765,8 +33768,7 @@
 				"mode": "effort",
 				"minLevel": "low",
 				"maxLevel": "xhigh"
-			},
-			"contextPromotionTarget": "opencode/gpt-5.3-codex"
+			}
 		},
 		"gpt-5.4": {
 			"id": "gpt-5.4",
@@ -34828,8 +34830,7 @@
 				"mode": "effort",
 				"minLevel": "low",
 				"maxLevel": "xhigh"
-			},
-			"contextPromotionTarget": "opencode-zen/gpt-5.3-codex"
+			}
 		},
 		"gpt-5.4": {
 			"id": "gpt-5.4",

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

@@ -246,26 +246,64 @@ async function fetchOllamaNativeModels(baseUrl: string): Promise<Model<"openai-r
 	}
 	const payload = (await response.json()) as { models?: Array<{ name?: string; model?: string }> };
 	const entries = payload.models ?? [];
-	const models: Model<"openai-responses">[] = [];
-	for (const entry of entries) {
-		const id = entry.model ?? entry.name;
-		if (!id) {
-			continue;
-		}
-		models.push({
-			id,
-			name: entry.name ?? id,
-			api: "openai-responses",
-			provider: "ollama",
-			baseUrl,
-			reasoning: false,
-			input: ["text"],
-			cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
-			contextWindow: 128000,
-			maxTokens: 8192,
+	const resolved = await Promise.all(
+		entries.map(async (entry): Promise<Model<"openai-responses"> | null> => {
+			const id = entry.model ?? entry.name;
+			if (!id) return null;
+			const { contextWindow, maxTokens } = await fetchOllamaModelLimits(nativeBaseUrl, id);
+			return {
+				id,
+				name: entry.name ?? id,
+				api: "openai-responses",
+				provider: "ollama",
+				baseUrl,
+				reasoning: false,
+				input: ["text"],
+				cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0 },
+				contextWindow,
+				maxTokens,
+			};
+		}),
+	);
+	const models: Model<"openai-responses">[] = resolved.filter((m): m is Model<"openai-responses"> => m !== null);
+	return models.sort((left, right) => left.id.localeCompare(right.id));
+}
+
+/** Ollama's default `num_ctx` when the runtime request does not override it. */
+const OLLAMA_DEFAULT_CONTEXT_WINDOW = 4096;
+/** Cap max output tokens at a value that matches OMP's other openai-responses defaults. */
+const OLLAMA_DEFAULT_MAX_TOKENS = 8192;
+
+/**
+ * Query Ollama's `/api/show` endpoint for a single model and pull its native
+ * context length out of `model_info.<arch>.context_length`. Falls back to
+ * Ollama's default context window when the endpoint or field is unavailable
+ * so discovery still succeeds against older Ollama builds.
+ */
+async function fetchOllamaModelLimits(
+	nativeBaseUrl: string,
+	modelId: string,
+): Promise<{ contextWindow: number; maxTokens: number }> {
+	try {
+		const response = await fetch(`${nativeBaseUrl}/api/show`, {
+			method: "POST",
+			headers: { "Content-Type": "application/json", Accept: "application/json" },
+			body: JSON.stringify({ model: modelId }),
 		});
+		if (!response.ok) {
+			return { contextWindow: OLLAMA_DEFAULT_CONTEXT_WINDOW, maxTokens: OLLAMA_DEFAULT_MAX_TOKENS };
+		}
+		const payload = (await response.json()) as { model_info?: Record<string, unknown> };
+		const info = payload.model_info ?? {};
+		for (const [key, value] of Object.entries(info)) {
+			if (key.endsWith(".context_length") && typeof value === "number" && value > 0) {
+				return { contextWindow: value, maxTokens: OLLAMA_DEFAULT_MAX_TOKENS };
+			}
+		}
+	} catch {
+		// fall through to default
 	}
-	return models.sort((left, right) => left.id.localeCompare(right.id));
+	return { contextWindow: OLLAMA_DEFAULT_CONTEXT_WINDOW, maxTokens: OLLAMA_DEFAULT_MAX_TOKENS };
 }
 
 const OPENAI_NON_RESPONSES_PREFIXES = [

package/src/providers/anthropic.ts

@@ -8,7 +8,7 @@ import type {
 	MessageParam,
 } from "@anthropic-ai/sdk/resources/messages";
 import { $env, abortableSleep, isEnoent } from "@oh-my-pi/pi-utils";
-import { mapEffortToAnthropicAdaptiveEffort } from "../model-thinking";
+import { hasOpus47ApiRestrictions, mapEffortToAnthropicAdaptiveEffort } from "../model-thinking";
 import { calculateCost } from "../models";
 import { getEnvApiKey, OUTPUT_FALLBACK_BUFFER } from "../stream";
 import type {
@@ -1421,6 +1421,13 @@ function buildParams(
 		params.top_k = options.topK;
 	}
 
+	// Opus 4.7+ rejects non-default sampling parameters with 400 error.
+	if (hasOpus47ApiRestrictions(model.id)) {
+		delete params.temperature;
+		delete (params as AnthropicSamplingParams).top_p;
+		delete (params as AnthropicSamplingParams).top_k;
+	}
+
 	if (context.tools) {
 		params.tools = convertTools(context.tools, isOAuthToken);
 	}

package/src/providers/cursor.ts

@@ -2109,10 +2109,86 @@ function extractAssistantMessageText(msg: Message): string {
 }
 
 /**
- * Convert context.messages to Cursor's serialized ConversationTurn format.
+ * Derive a stable, UUID-formatted `message_id` from a content key.
+ * Ensures identical historical messages hash to the same blob IDs across
+ * requests, so `conversationBlobStores` does not grow unboundedly and
+ * unchanged history reuses existing blob IDs.
+ */
+function deterministicMessageId(key: string): string {
+	const hex = createHash("sha256").update(key).digest("hex");
+	return `${hex.slice(0, 8)}-${hex.slice(8, 12)}-${hex.slice(12, 16)}-${hex.slice(16, 20)}-${hex.slice(20, 32)}`;
+}
+
+/**
+ * Index of the last user/developer message in `messages`, or -1 if none.
+ * Used to exclude the current user turn from history builders — it goes in
+ * `ConversationActionSchema.userMessageAction`, not in history structures.
+ */
+function findLastUserMessageIndex(messages: Message[]): number {
+	for (let i = messages.length - 1; i >= 0; i--) {
+		const role = messages[i].role;
+		if (role === "user" || role === "developer") {
+			return i;
+		}
+	}
+	return -1;
+}
+
+/**
+ * Build `ConversationStateStructure.rootPromptMessagesJson` blob IDs for the
+ * system prompt plus prior conversation history, as JSON blobs matching
+ * Cursor's internal Vercel-AI-SDK-shaped message format.
+ *
+ * Cursor's server uses `rootPromptMessagesJson` (not `turns[]`) to build the
+ * actual model prompt. `turns[]` is UI/display metadata. Without populating
+ * this field, multi-turn conversations lose prior context — the model sees
+ * only an empty placeholder where historical user turns should be.
+ * The last user message is excluded because it is sent in the action.
+ */
+function buildRootPromptMessagesJson(
+	messages: Message[],
+	systemPromptId: Uint8Array,
+	blobStore: Map<string, Uint8Array>,
+): Uint8Array[] {
+	const entries: Uint8Array[] = [systemPromptId];
+	const lastUserIdx = findLastUserMessageIndex(messages);
+
+	const pushJson = (obj: unknown) => {
+		const bytes = new TextEncoder().encode(JSON.stringify(obj));
+		entries.push(storeCursorBlob(blobStore, bytes));
+	};
+
+	for (let i = 0; i < messages.length; i++) {
+		if (i === lastUserIdx) break;
+		const msg = messages[i];
+		if (msg.role === "user" || msg.role === "developer") {
+			const text = extractUserMessageText(msg);
+			if (!text) continue;
+			pushJson({ role: "user", content: [{ type: "text", text }] });
+		} else if (msg.role === "assistant") {
+			const text = extractAssistantMessageText(msg);
+			if (!text) continue;
+			pushJson({ role: "assistant", content: [{ type: "text", text }] });
+		} else if (msg.role === "toolResult") {
+			const text = toolResultToText(msg);
+			if (!text) continue;
+			pushJson({
+				role: "user",
+				content: [{ type: "text", text: `[Tool Result]\n${text}` }],
+			});
+		}
+	}
+
+	return entries;
+}
+
+/**
+ * Convert context.messages to Cursor's ConversationTurnStructure blob IDs.
  * Groups messages into turns: each turn is a user message followed by the assistant's response.
  * Excludes the last user message (which goes in the action).
- * Returns blob IDs for ConversationStateStructure.turns field.
+ *
+ * Each `AgentConversationTurnStructure.user_message`, `steps[]`, and the outer
+ * `ConversationStateStructure.turns[]` entry is a blob ID into `blobStore`.
  */
 function buildConversationTurns(messages: Message[], blobStore: Map<string, Uint8Array>): Uint8Array[] {
 	const turns: Uint8Array[] = [];
@@ -2149,10 +2225,10 @@ function buildConversationTurns(messages: Message[], blobStore: Map<string, Uint
 
 		const userMessage = create(UserMessageSchema, {
 			text: userText,
-			messageId: crypto.randomUUID(),
+			messageId: deterministicMessageId(`u:${turns.length}:${userText}`),
 		});
 		const userMessageBytes = toBinary(UserMessageSchema, userMessage);
-		const userMessageId = storeCursorBlob(blobStore, userMessageBytes);
+		const userMessageBlobId = storeCursorBlob(blobStore, userMessageBytes);
 
 		// Collect and serialize steps until next user message
 		const stepBlobIds: Uint8Array[] = [];
@@ -2189,9 +2265,10 @@ function buildConversationTurns(messages: Message[], blobStore: Map<string, Uint
 			i++;
 		}
 
-		// Cursor stores turn parts in the KV blob channel; these fields carry blob IDs.
+		// Create the serialized turn using Structure types. The bytes fields
+		// (user_message, steps) are blob IDs resolved through the KV store.
 		const agentTurn = create(AgentConversationTurnStructureSchema, {
-			userMessage: userMessageId,
+			userMessage: userMessageBlobId,
 			steps: stepBlobIds,
 		});
 		const turn = create(ConversationTurnStructureSchema, {
@@ -2254,15 +2331,21 @@ function buildGrpcRequest(
 		},
 	});
 
-	// Build conversation turns from prior messages (excluding the last user message)
+	// Build conversation turns from prior messages (excluding the last user message).
+	// This populates the UI-side history view (`turns[]`).
 	const turns = buildConversationTurns(context.messages, blobStore);
 
+	// Build `rootPromptMessagesJson` from prior messages. Cursor's server uses this
+	// field (not `turns[]`) to construct the actual model prompt; if we only send the
+	// system prompt here, multi-turn conversations lose prior context and the model
+	// sees only the current user message.
+	const rootPromptMessagesJson = buildRootPromptMessagesJson(context.messages, systemPromptId, blobStore);
+
+	// Preserve cached non-history state fields (todos, file states, summaries, etc.)
+	// when the system prompt is unchanged; otherwise start fresh.
 	const hasMatchingPrompt = state.conversationState?.rootPromptMessagesJson?.some(entry =>
 		Buffer.from(entry).equals(systemPromptId),
 	);
-
-	// Use cached state if available and system prompt matches, but always update turns
-	// from context.messages to ensure full conversation history is sent
 	const baseState =
 		state.conversationState && hasMatchingPrompt
 			? state.conversationState
@@ -2281,10 +2364,13 @@ function buildGrpcRequest(
 					readPaths: [],
 				});
 
-	// Always populate turns from context.messages to ensure Cursor sees full conversation
+	// Always override `rootPromptMessagesJson` and `turns` with content freshly built from
+	// `context.messages`. The server-echoed checkpoint replaces historical user entries
+	// with empty placeholders, so we cannot rely on the cached `rootPromptMessagesJson`.
 	const conversationState = create(ConversationStateStructureSchema, {
 		...baseState,
-		turns: turns.length > 0 ? turns : baseState.turns,
+		rootPromptMessagesJson,
+		turns,
 	});
 
 	const modelDetails = create(ModelDetailsSchema, {