@oh-my-pi/pi-ai 14.2.1 → 14.4.0

package/CHANGELOG.md

@@ -2,6 +2,40 @@
 
 ## [Unreleased]
 
+## [14.4.0] - 2026-04-26
+
+### Added
+
+- Added an `examples` option to `StringEnum` to include example values in the generated schema
+
+### Changed
+
+- Changed Anthropic tool schema generation to strip unsupported schema fields (including `patternProperties`), add `additionalProperties: false` for object types, and apply Anthropic strict-mode limits when marking tools as strict
+- Changed Anthropic strict tool planning to cap strict `tools` at twenty entries and convert excess optional/union parameters to nullable schemas to stay within provider constraints
+
+### Fixed
+
+- Fixed Anthropic tool schema compilation failures by keeping the `write` tool out of the strict-tool allowlist when the full coding-agent tool set is active
+- Fixed Anthropic 400 `tools.*.custom: For 'object' type, property 'minItems' is not supported` by stripping `minItems` from object-shaped JSON schema nodes (array nodes still keep supported `minItems` values)
+- Fixed Anthropic tool schemas that used tuple-style arrays by stripping unsupported `maxItems` and only preserving provider-supported `minItems` values
+- Fixed Anthropic and OpenRouter Anthropic tool calls that previously failed with `compiled grammar is too large` by retrying automatically without strict tool schemas and reusing non-strict mode for subsequent requests in the same provider session
+- Fixed parsing of JSON tool arguments containing raw control characters inside string values (such as embedded newlines) by escaping them before JSON parsing
+- Fixed `validateToolArguments` to accept stringified objects and arrays that include literal control characters inside string fields
+- Fixed OpenAI Codex Spark OAuth selection to fall back to non-Pro accounts when no ChatGPT Pro account is connected, so users without a Pro account can still attempt Spark requests in case the server permits access.
+
+## [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
@@ -2097,4 +2131,4 @@ _Dedicated to Peter's shoulder ([@steipete](https://twitter.com/steipete))_
 
 ## [0.9.4] - 2025-11-26
 
-Initial release with multi-provider LLM support.
+Initial release with multi-provider LLM support.

package/package.json

@@ -1,7 +1,7 @@
 {
 	"type": "module",
 	"name": "@oh-my-pi/pi-ai",
-	"version": "14.2.1",
+	"version": "14.4.0",
 	"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": "14.4.0",
+		"@oh-my-pi/pi-utils": "14.4.0",
+		"@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/auth-storage.ts

@@ -1711,6 +1711,11 @@ export class AuthStorage {
 			}),
 		);
 
+		// Skip the Pro-plan filter when no candidate is confirmed Pro, so users with only
+		// non-Pro accounts can still attempt Spark requests (e.g. trial/grandfathered access).
+		const enforceProRequirement =
+			requiresProModel && candidates.some(candidate => hasOpenAICodexProPlan(candidate.usage));
+
 		const fallback = candidates[0];
 
 		for (const candidate of candidates) {
@@ -1719,6 +1724,7 @@ export class AuthStorage {
 				allowBlocked: false,
 				prefetchedUsage: candidate.usage,
 				usagePrechecked: candidate.usageChecked,
+				enforceProRequirement,
 			});
 			if (apiKey) return apiKey;
 		}
@@ -1729,6 +1735,7 @@ export class AuthStorage {
 				allowBlocked: true,
 				prefetchedUsage: fallback.usage,
 				usagePrechecked: fallback.usageChecked,
+				enforceProRequirement,
 			});
 		}
 
@@ -1774,14 +1781,22 @@ export class AuthStorage {
 			allowBlocked: boolean;
 			prefetchedUsage?: UsageReport | null;
 			usagePrechecked?: boolean;
+			enforceProRequirement?: boolean;
 		},
 	): Promise<string | undefined> {
-		const { checkUsage, allowBlocked, prefetchedUsage = null, usagePrechecked = false } = usageOptions;
+		const {
+			checkUsage,
+			allowBlocked,
+			prefetchedUsage = null,
+			usagePrechecked = false,
+			enforceProRequirement,
+		} = usageOptions;
 		if (!allowBlocked && this.#isCredentialBlocked(providerKey, selection.index)) {
 			return undefined;
 		}
 
 		const requiresProModel = requiresOpenAICodexProModel(provider, options?.modelId);
+		const applyProFilter = enforceProRequirement ?? requiresProModel;
 		let usage: UsageReport | null = null;
 		let usageChecked = false;
 
@@ -1796,7 +1811,7 @@ export class AuthStorage {
 				});
 				usageChecked = true;
 			}
-			if (requiresProModel && !hasOpenAICodexProPlan(usage)) {
+			if (applyProFilter && !hasOpenAICodexProPlan(usage)) {
 				return undefined;
 			}
 			if (checkUsage && !allowBlocked && usage && this.#isUsageLimitReached(usage)) {
@@ -1846,7 +1861,7 @@ export class AuthStorage {
 					});
 					usageChecked = true;
 				}
-				if (requiresProModel && !hasOpenAICodexProPlan(usage)) {
+				if (applyProFilter && !hasOpenAICodexProPlan(usage)) {
 					return undefined;
 				}
 				if (checkUsage && !allowBlocked && usage && this.#isUsageLimitReached(usage)) {

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

@@ -4751,6 +4751,56 @@
 			}
 		}
 	},
+	"deepseek": {
+		"deepseek-v4-flash": {
+			"id": "deepseek-v4-flash",
+			"name": "DeepSeek V4 Flash",
+			"api": "openai-completions",
+			"provider": "deepseek",
+			"baseUrl": "https://api.deepseek.com",
+			"reasoning": true,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 0.14,
+				"output": 0.28,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 1000000,
+			"maxTokens": 384000,
+			"thinking": {
+				"mode": "effort",
+				"minLevel": "minimal",
+				"maxLevel": "high"
+			}
+		},
+		"deepseek-v4-pro": {
+			"id": "deepseek-v4-pro",
+			"name": "DeepSeek V4 Pro",
+			"api": "openai-completions",
+			"provider": "deepseek",
+			"baseUrl": "https://api.deepseek.com",
+			"reasoning": true,
+			"input": [
+				"text"
+			],
+			"cost": {
+				"input": 1.74,
+				"output": 3.48,
+				"cacheRead": 0,
+				"cacheWrite": 0
+			},
+			"contextWindow": 1000000,
+			"maxTokens": 384000,
+			"thinking": {
+				"mode": "effort",
+				"minLevel": "minimal",
+				"maxLevel": "high"
+			}
+		}
+	},
 	"github-copilot": {
 		"claude-haiku-4.5": {
 			"id": "claude-haiku-4.5",
@@ -16931,7 +16981,7 @@
 			},
 			"contextWindow": 128000,
 			"maxTokens": 128000,
-			"contextPromotionTarget": "litellm/gpt-5.3-codex",
+			"contextPromotionTarget": "litellm/gpt-5.5",
 			"thinking": {
 				"mode": "effort",
 				"minLevel": "low",
@@ -17011,7 +17061,8 @@
 				"mode": "effort",
 				"minLevel": "low",
 				"maxLevel": "xhigh"
-			}
+			},
+			"contextPromotionTarget": "litellm/gpt-5.4"
 		},
 		"gpt-image-2": {
 			"id": "gpt-image-2",
@@ -32938,7 +32989,7 @@
 				"maxLevel": "xhigh"
 			},
 			"applyPatchToolType": "freeform",
-			"contextPromotionTarget": "openai/gpt-5.3-codex"
+			"contextPromotionTarget": "openai/gpt-5.5"
 		},
 		"gpt-5.4": {
 			"id": "gpt-5.4",
@@ -33068,7 +33119,8 @@
 				"minLevel": "low",
 				"maxLevel": "xhigh"
 			},
-			"applyPatchToolType": "freeform"
+			"applyPatchToolType": "freeform",
+			"contextPromotionTarget": "openai/gpt-5.4"
 		},
 		"o1": {
 			"id": "o1",
@@ -33597,7 +33649,7 @@
 			},
 			"contextWindow": 128000,
 			"maxTokens": 128000,
-			"contextPromotionTarget": "openai-codex/gpt-5.3-codex",
+			"contextPromotionTarget": "openai-codex/gpt-5.5",
 			"thinking": {
 				"mode": "effort",
 				"minLevel": "low",
@@ -33715,7 +33767,8 @@
 				"minLevel": "low",
 				"maxLevel": "xhigh"
 			},
-			"applyPatchToolType": "freeform"
+			"applyPatchToolType": "freeform",
+			"contextPromotionTarget": "openai-codex/gpt-5.4"
 		}
 	},
 	"opencode": {
@@ -33765,8 +33818,7 @@
 				"mode": "effort",
 				"minLevel": "low",
 				"maxLevel": "xhigh"
-			},
-			"contextPromotionTarget": "opencode/gpt-5.3-codex"
+			}
 		},
 		"gpt-5.4": {
 			"id": "gpt-5.4",
@@ -34828,8 +34880,7 @@
 				"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 = [