@oh-my-pi/pi-ai 14.3.0 → 14.4.1

package/CHANGELOG.md

@@ -2,6 +2,27 @@
 
 ## [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
@@ -2110,4 +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": "14.3.0",
+	"version": "14.4.1",
 	"description": "Unified LLM API with automatic model discovery and provider configuration",
 	"homepage": "https://github.com/can1357/oh-my-pi",
 	"author": "Can Boluk",
@@ -46,8 +46,8 @@
 		"@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",
+		"@oh-my-pi/pi-natives": "14.4.1",
+		"@oh-my-pi/pi-utils": "14.4.1",
 		"@sinclair/typebox": "^0.34.49",
 		"@smithy/node-http-handler": "^4.6.1",
 		"ajv": "^8.20.0",

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/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",

package/src/providers/anthropic.ts

@@ -19,6 +19,7 @@ import type {
 	ImageContent,
 	Message,
 	Model,
+	ProviderSessionState,
 	RedactedThinkingContent,
 	SimpleStreamOptions,
 	StopReason,
@@ -31,7 +32,7 @@ import type {
 	ToolResultMessage,
 	Usage,
 } from "../types";
-import { isAnthropicOAuthToken, normalizeToolCallId, resolveCacheRetention } from "../utils";
+import { isAnthropicOAuthToken, isRecord, normalizeToolCallId, resolveCacheRetention } from "../utils";
 import { createAbortSourceTracker } from "../utils/abort";
 import { AssistantMessageEventStream } from "../utils/event-stream";
 import { isFoundryEnabled } from "../utils/foundry";
@@ -39,7 +40,8 @@ import { finalizeErrorMessage, type RawHttpRequestDump, rewriteCopilotError } fr
 import { createWatchdog, getStreamFirstEventTimeoutMs } from "../utils/idle-iterator";
 import { parseStreamingJson } from "../utils/json-parse";
 import { parseGitHubCopilotApiKey } from "../utils/oauth/github-copilot";
-import { isCopilotRetryableError } from "../utils/retry";
+import { extractHttpStatusFromError, isCopilotRetryableError } from "../utils/retry";
+import { COMBINATOR_KEYS, NO_STRICT } from "../utils/schema";
 import {
 	buildCopilotDynamicHeaders,
 	hasCopilotVisionInput,
@@ -82,7 +84,6 @@ export function buildBetaHeader(baseBetas: string[], extraBetas: string[]): stri
 const claudeCodeBetaDefaults = [
 	"claude-code-20250219",
 	"oauth-2025-04-20",
-	"interleaved-thinking-2025-05-14",
 	"context-management-2025-06-27",
 	"prompt-caching-scope-2026-01-05",
 ];
@@ -180,6 +181,58 @@ function supportsAdaptiveThinkingDisplay(modelId: string): boolean {
 	const minor = Number(match[2]);
 	return major > 4 || (major === 4 && minor >= 7);
 }
+
+const ANTHROPIC_PROVIDER_SESSION_STATE_KEY = "anthropic-messages";
+
+type AnthropicProviderSessionState = ProviderSessionState & {
+	strictToolsDisabled: boolean;
+};
+
+function createAnthropicProviderSessionState(): AnthropicProviderSessionState {
+	const state: AnthropicProviderSessionState = {
+		strictToolsDisabled: false,
+		close: () => {
+			state.strictToolsDisabled = false;
+		},
+	};
+	return state;
+}
+
+function getAnthropicProviderSessionState(
+	providerSessionState: Map<string, ProviderSessionState> | undefined,
+): AnthropicProviderSessionState | undefined {
+	if (!providerSessionState) return undefined;
+	const existing = providerSessionState.get(ANTHROPIC_PROVIDER_SESSION_STATE_KEY) as
+		| AnthropicProviderSessionState
+		| undefined;
+	if (existing) return existing;
+	const created = createAnthropicProviderSessionState();
+	providerSessionState.set(ANTHROPIC_PROVIDER_SESSION_STATE_KEY, created);
+	return created;
+}
+
+function isAnthropicStrictGrammarTooLargeError(error: unknown): boolean {
+	if (extractHttpStatusFromError(error) !== 400) return false;
+	const message = error instanceof Error ? error.message : String(error);
+	const isStrictGrammarTooLarge = /compiled grammar/i.test(message) && /too large/i.test(message);
+	const isSchemaCompilationTooComplex =
+		/schema/i.test(message) && /too complex/i.test(message) && /compil/i.test(message);
+	return /invalid_request_error/i.test(message) && (isStrictGrammarTooLarge || isSchemaCompilationTooComplex);
+}
+
+function hasStrictAnthropicTools(params: MessageCreateParamsStreaming): boolean {
+	const tools = params.tools as Array<{ strict?: unknown }> | undefined;
+	return tools?.some(tool => tool.strict === true) ?? false;
+}
+
+function dropAnthropicStrictTools(params: MessageCreateParamsStreaming): void {
+	const tools = params.tools as Array<{ strict?: unknown }> | undefined;
+	if (!tools) return;
+	for (const tool of tools) {
+		delete tool.strict;
+	}
+}
+
 function getCacheControl(
 	baseUrl: string,
 	cacheRetention?: CacheRetention,
@@ -705,19 +758,29 @@ export const streamAnthropic: StreamFunction<"anthropic-messages"> = (
 			const baseUrl =
 				resolveAnthropicBaseUrl(model, options?.apiKey ?? getEnvApiKey(model.provider) ?? "") ??
 				"https://api.anthropic.com";
-			let params = buildParams(model, baseUrl, context, isOAuthToken, options);
-			const replacementPayload = await options?.onPayload?.(params, model);
-			if (replacementPayload !== undefined) {
-				params = replacementPayload as typeof params;
-			}
-			rawRequestDump = {
-				provider: model.provider,
-				api: output.api,
-				model: model.id,
-				method: "POST",
-				url: `${baseUrl}/v1/messages`,
-				body: params,
+			const providerSessionState = getAnthropicProviderSessionState(options?.providerSessionState);
+			let disableStrictTools = providerSessionState?.strictToolsDisabled ?? false;
+			let strictFallbackErrorMessage: string | undefined;
+			const prepareParams = async (): Promise<MessageCreateParamsStreaming> => {
+				let nextParams = buildParams(model, baseUrl, context, isOAuthToken, options, disableStrictTools);
+				const replacementPayload = await options?.onPayload?.(nextParams, model);
+				if (replacementPayload !== undefined) {
+					nextParams = replacementPayload as typeof nextParams;
+				}
+				if (disableStrictTools) {
+					dropAnthropicStrictTools(nextParams);
+				}
+				rawRequestDump = {
+					provider: model.provider,
+					api: output.api,
+					model: model.id,
+					method: "POST",
+					url: `${baseUrl}/v1/messages`,
+					body: nextParams,
+				};
+				return nextParams;
 			};
+			let params = await prepareParams();
 
 			type Block = (
 				| ThinkingContent
@@ -954,6 +1017,28 @@ export const streamAnthropic: StreamFunction<"anthropic-messages"> = (
 					break;
 				} catch (streamError) {
 					const streamFailure = activeAbortTracker.getLocalAbortReason() ?? streamError;
+					if (
+						!disableStrictTools &&
+						firstTokenTime === undefined &&
+						hasStrictAnthropicTools(params) &&
+						isAnthropicStrictGrammarTooLargeError(streamFailure)
+					) {
+						strictFallbackErrorMessage = await finalizeErrorMessage(streamFailure, rawRequestDump);
+						output.errorMessage = strictFallbackErrorMessage;
+						if (providerSessionState) {
+							providerSessionState.strictToolsDisabled = true;
+						}
+						disableStrictTools = true;
+						params = await prepareParams();
+						providerRetryAttempt = 0;
+						output.content.length = 0;
+						output.responseId = undefined;
+						output.providerPayload = undefined;
+						output.usage = createEmptyUsage(copilotDynamicHeaders?.premiumRequests);
+						output.stopReason = "stop";
+						firstTokenTime = undefined;
+						continue;
+					}
 					const isTransientEnvelopeFailure =
 						isTransientStreamParseError(streamFailure) || isTransientStreamEnvelopeError(streamFailure);
 					const canRetryTransientEnvelopeFailure = isTransientEnvelopeFailure && !streamedReplayUnsafeContent;
@@ -971,7 +1056,7 @@ export const streamAnthropic: StreamFunction<"anthropic-messages"> = (
 					await abortableSleep(delayMs, options?.signal);
 					output.content.length = 0;
 					output.responseId = undefined;
-					output.errorMessage = undefined;
+					output.errorMessage = strictFallbackErrorMessage;
 					output.providerPayload = undefined;
 					output.usage = createEmptyUsage(copilotDynamicHeaders?.premiumRequests);
 					output.stopReason = "stop";
@@ -1074,6 +1159,7 @@ export function buildAnthropicClientOptions(args: AnthropicClientOptionsArgs): A
 		dynamicHeaders,
 		isOAuth,
 	} = args;
+	const needsInterleavedBeta = interleavedThinking && !supportsAdaptiveThinkingDisplay(model.id);
 	const oauthToken = isOAuth ?? isAnthropicOAuthToken(apiKey);
 	const baseUrl = resolveAnthropicBaseUrl(model, apiKey);
 	const foundryCustomHeaders = resolveAnthropicCustomHeaders(model);
@@ -1107,7 +1193,7 @@ export function buildAnthropicClientOptions(args: AnthropicClientOptionsArgs): A
 	}
 
 	const betaFeatures = [...extraBetas];
-	if (interleavedThinking) {
+	if (needsInterleavedBeta) {
 		betaFeatures.push("interleaved-thinking-2025-05-14");
 	}
 
@@ -1402,6 +1488,7 @@ function buildParams(
 	context: Context,
 	isOAuthToken: boolean,
 	options?: AnthropicOptions,
+	disableStrictTools = false,
 ): MessageCreateParamsStreaming {
 	const { cacheControl } = getCacheControl(baseUrl, options?.cacheRetention);
 	const params: AnthropicSamplingParams = {
@@ -1411,9 +1498,6 @@ function buildParams(
 		stream: true,
 	};
 
-	if (options?.temperature !== undefined) {
-		params.temperature = options.temperature;
-	}
 	if (options?.topP !== undefined) {
 		params.top_p = options.topP;
 	}
@@ -1423,13 +1507,16 @@ function buildParams(
 
 	// 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;
+		delete params.top_p;
+		delete params.top_k;
 	}
 
 	if (context.tools) {
-		params.tools = convertTools(context.tools, isOAuthToken);
+		params.tools = convertTools(
+			context.tools,
+			isOAuthToken,
+			disableStrictTools || model.provider === "github-copilot",
+		);
 	}
 
 	if (options?.thinkingEnabled && model.reasoning && model.provider !== "github-copilot") {
@@ -1667,20 +1754,309 @@ export function convertAnthropicMessages(
 	return params;
 }
 
-function convertTools(tools: Tool[], isOAuthToken: boolean): Anthropic.Messages.Tool[] {
-	if (!tools) return [];
+const ANTHROPIC_UNSUPPORTED_TOOL_SCHEMA_FIELDS = new Set(["maxItems", "patternProperties"]);
+const ANTHROPIC_STRICT_TOOL_ALLOWLIST = new Set(["bash", "python", "edit", "find"]);
+const MAX_ANTHROPIC_STRICT_TOOLS = 20;
+const MAX_ANTHROPIC_STRICT_OPTIONAL_PARAMETERS = 24;
+const MAX_ANTHROPIC_STRICT_UNION_PARAMETERS = 16;
+
+/** `minItems` / `maxItems` apply to arrays; Anthropic rejects them on `type: "object"` (including `minItems: 0`/`1`). */
+function isJsonSchemaArrayNode(schema: Record<string, unknown>): boolean {
+	const t = schema.type;
+	if (t === "array") return true;
+	if (Array.isArray(t) && t.includes("array") && !t.includes("object")) return true;
+	return false;
+}
+
+function isJsonSchemaObjectNode(schema: Record<string, unknown>): boolean {
+	if (isJsonSchemaArrayNode(schema)) return false;
+	if (schema.type === "object") return true;
+	if (Array.isArray(schema.type) && schema.type.includes("object")) return true;
+	if (isRecord(schema.properties)) return true;
+	return false;
+}
+
+function normalizeAnthropicToolSchema(
+	schema: unknown,
+	cache: WeakMap<Record<string, unknown>, Record<string, unknown>> = new WeakMap(),
+): unknown {
+	if (!isRecord(schema)) return schema;
+
+	const cached = cache.get(schema);
+	if (cached) return cached;
+
+	const result = Object.fromEntries(
+		Object.entries(schema).filter(([key]) => !ANTHROPIC_UNSUPPORTED_TOOL_SCHEMA_FIELDS.has(key)),
+	);
+	cache.set(schema, result);
+	if (isJsonSchemaObjectNode(result)) {
+		delete result.minItems;
+	} else {
+		const minItems = result.minItems;
+		if (typeof minItems === "number" && minItems !== 0 && minItems !== 1) {
+			delete result.minItems;
+		}
+	}
+
+	const type = result.type;
+	const canBeObject =
+		type === "object" || (Array.isArray(type) && type.includes("object")) || isRecord(result.properties);
+	if (canBeObject) {
+		result.additionalProperties = false;
+	}
+
+	if (isRecord(result.properties)) {
+		result.properties = Object.fromEntries(
+			Object.entries(result.properties).map(([propertyName, propertySchema]) => [
+				propertyName,
+				normalizeAnthropicToolSchema(propertySchema, cache),
+			]),
+		);
+	}
+
+	if (Array.isArray(result.items)) {
+		result.items = result.items.map(item => normalizeAnthropicToolSchema(item, cache));
+	} else if (isRecord(result.items)) {
+		result.items = normalizeAnthropicToolSchema(result.items, cache);
+	}
+
+	for (const key of COMBINATOR_KEYS) {
+		const variants = result[key];
+		if (Array.isArray(variants)) {
+			result[key] = variants.map(variant => normalizeAnthropicToolSchema(variant, cache));
+		}
+	}
+
+	for (const defsKey of ["$defs", "definitions"] as const) {
+		const definitions = result[defsKey];
+		if (!isRecord(definitions)) continue;
+		result[defsKey] = Object.fromEntries(
+			Object.entries(definitions).map(([definitionName, definitionSchema]) => [
+				definitionName,
+				normalizeAnthropicToolSchema(definitionSchema, cache),
+			]),
+		);
+	}
+
+	return result;
+}
+
+type AnthropicToolInputSchema = Anthropic.Messages.Tool["input_schema"];
+
+type AnthropicToolSchemaPlan = {
+	inputSchema: AnthropicToolInputSchema;
+	strict: boolean;
+};
+
+type AnthropicStrictBudget = {
+	optionalRemaining: number;
+	unionRemaining: number;
+	optionalCount: number;
+	unionCount: number;
+};
+
+function hasAnthropicUnionType(schema: Record<string, unknown>): boolean {
+	return Array.isArray(schema.type) || Array.isArray(schema.anyOf);
+}
+
+function hasNullVariant(schema: Record<string, unknown>): boolean {
+	if (Array.isArray(schema.type) && schema.type.includes("null")) return true;
+	return Array.isArray(schema.anyOf) && schema.anyOf.some(variant => isRecord(variant) && variant.type === "null");
+}
+
+function makeAnthropicNullableSchema(schema: unknown, budget: AnthropicStrictBudget): unknown | undefined {
+	if (isRecord(schema)) {
+		if (hasNullVariant(schema)) return schema;
+		if (Array.isArray(schema.anyOf)) {
+			return { ...schema, anyOf: [...schema.anyOf, { type: "null" }] };
+		}
+		if (Array.isArray(schema.type)) {
+			return { ...schema, type: [...schema.type, "null"] };
+		}
+	}
+
+	if (budget.unionRemaining <= 0) return undefined;
+	budget.unionRemaining--;
+	budget.unionCount++;
+	return { anyOf: [schema, { type: "null" }] };
+}
+
+function normalizeAnthropicStrictSchemaNode(
+	schema: unknown,
+	budget: AnthropicStrictBudget,
+	cache: WeakMap<Record<string, unknown>, Record<string, unknown>>,
+): unknown | undefined {
+	if (Array.isArray(schema)) {
+		const result: unknown[] = [];
+		for (const entry of schema) {
+			const normalized = normalizeAnthropicStrictSchemaNode(entry, budget, cache);
+			if (normalized === undefined) return undefined;
+			result.push(normalized);
+		}
+		return result;
+	}
+
+	if (!isRecord(schema)) return schema;
+
+	const cached = cache.get(schema);
+	if (cached) return cached;
+
+	const result: Record<string, unknown> = { ...schema };
+	cache.set(schema, result);
+
+	if (hasAnthropicUnionType(result)) {
+		if (budget.unionRemaining <= 0) return undefined;
+		budget.unionRemaining--;
+		budget.unionCount++;
+	}
+
+	if (isRecord(result.properties)) {
+		const originalRequired = new Set(
+			Array.isArray(result.required)
+				? result.required.filter((entry): entry is string => typeof entry === "string")
+				: [],
+		);
+		const properties: Record<string, unknown> = {};
+		const required: string[] = [];
+
+		for (const [propertyName, propertySchema] of Object.entries(result.properties)) {
+			const normalizedProperty = normalizeAnthropicStrictSchemaNode(propertySchema, budget, cache);
+			if (normalizedProperty === undefined) return undefined;
+
+			if (originalRequired.has(propertyName)) {
+				properties[propertyName] = normalizedProperty;
+				required.push(propertyName);
+				continue;
+			}
+
+			if (budget.optionalRemaining > 0) {
+				budget.optionalRemaining--;
+				budget.optionalCount++;
+				properties[propertyName] = normalizedProperty;
+				continue;
+			}
+
+			const nullableProperty = makeAnthropicNullableSchema(normalizedProperty, budget);
+			if (nullableProperty === undefined) return undefined;
+			properties[propertyName] = nullableProperty;
+			required.push(propertyName);
+		}
+
+		result.properties = properties;
+		result.required = required;
+	}
+
+	if (Array.isArray(result.items)) {
+		const items = normalizeAnthropicStrictSchemaNode(result.items, budget, cache);
+		if (items === undefined) return undefined;
+		result.items = items;
+	} else if (isRecord(result.items)) {
+		const items = normalizeAnthropicStrictSchemaNode(result.items, budget, cache);
+		if (items === undefined) return undefined;
+		result.items = items;
+	}
+
+	for (const key of COMBINATOR_KEYS) {
+		const variants = result[key];
+		if (!Array.isArray(variants)) continue;
+		const normalizedVariants = normalizeAnthropicStrictSchemaNode(variants, budget, cache);
+		if (normalizedVariants === undefined) return undefined;
+		result[key] = normalizedVariants;
+	}
 
-	return tools.map(tool => {
-		const jsonSchema = tool.parameters as any; // TypeBox already generates JSON Schema
+	for (const defsKey of ["$defs", "definitions"] as const) {
+		const definitions = result[defsKey];
+		if (!isRecord(definitions)) continue;
+		const normalizedDefinitions: Record<string, unknown> = {};
+		for (const [definitionName, definitionSchema] of Object.entries(definitions)) {
+			const normalizedDefinition = normalizeAnthropicStrictSchemaNode(definitionSchema, budget, cache);
+			if (normalizedDefinition === undefined) return undefined;
+			normalizedDefinitions[definitionName] = normalizedDefinition;
+		}
+		result[defsKey] = normalizedDefinitions;
+	}
 
+	return result;
+}
+
+function normalizeAnthropicStrictSchema(
+	schema: Record<string, unknown>,
+	optionalRemaining: number,
+	unionRemaining: number,
+): { schema: Record<string, unknown>; optionalCount: number; unionCount: number } | undefined {
+	const budget: AnthropicStrictBudget = {
+		optionalRemaining,
+		unionRemaining,
+		optionalCount: 0,
+		unionCount: 0,
+	};
+	const normalized = normalizeAnthropicStrictSchemaNode(schema, budget, new WeakMap());
+	if (!isRecord(normalized)) return undefined;
+	return { schema: normalized, optionalCount: budget.optionalCount, unionCount: budget.unionCount };
+}
+
+function buildAnthropicBaseToolInputSchema(tool: Tool): Record<string, unknown> {
+	const jsonSchema = tool.parameters as Record<string, unknown>;
+	return normalizeAnthropicToolSchema({
+		...jsonSchema,
+		type: "object",
+		properties: isRecord(jsonSchema.properties) ? jsonSchema.properties : {},
+		required: Array.isArray(jsonSchema.required)
+			? jsonSchema.required.filter((entry): entry is string => typeof entry === "string")
+			: [],
+	}) as Record<string, unknown>;
+}
+
+function buildAnthropicToolSchemaPlans(tools: Tool[], disableStrictTools = false): AnthropicToolSchemaPlan[] {
+	const plans = tools.map(
+		(tool): AnthropicToolSchemaPlan => ({
+			inputSchema: buildAnthropicBaseToolInputSchema(tool) as AnthropicToolInputSchema,
+			strict: false,
+		}),
+	);
+	if (NO_STRICT || disableStrictTools) return plans;
+
+	const candidateIndexes = tools.flatMap((tool, index) => {
+		if (!ANTHROPIC_STRICT_TOOL_ALLOWLIST.has(tool.name)) return [];
+		return tool.strict === false ? [] : [index];
+	});
+
+	let strictToolCount = 0;
+	let strictOptionalParameterCount = 0;
+	let strictUnionParameterCount = 0;
+	for (const index of candidateIndexes) {
+		if (strictToolCount >= MAX_ANTHROPIC_STRICT_TOOLS) break;
+
+		const strictResult = normalizeAnthropicStrictSchema(
+			plans[index].inputSchema as Record<string, unknown>,
+			MAX_ANTHROPIC_STRICT_OPTIONAL_PARAMETERS - strictOptionalParameterCount,
+			MAX_ANTHROPIC_STRICT_UNION_PARAMETERS - strictUnionParameterCount,
+		);
+		if (!strictResult) continue;
+
+		plans[index] = {
+			inputSchema: strictResult.schema as AnthropicToolInputSchema,
+			strict: true,
+		};
+		strictToolCount++;
+		strictOptionalParameterCount += strictResult.optionalCount;
+		strictUnionParameterCount += strictResult.unionCount;
+	}
+
+	return plans;
+}
+
+function convertTools(tools: Tool[], isOAuthToken: boolean, disableStrictTools = false): Anthropic.Messages.Tool[] {
+	if (!tools) return [];
+	const schemaPlans = buildAnthropicToolSchemaPlans(tools, disableStrictTools);
+
+	return tools.map((tool, index) => {
+		const plan = schemaPlans[index];
 		return {
 			name: isOAuthToken ? applyClaudeToolPrefix(tool.name) : tool.name,
 			description: tool.description || "",
-			input_schema: {
-				type: "object" as const,
-				properties: jsonSchema.properties || {},
-				required: jsonSchema.required || [],
-			},
+			input_schema: plan.inputSchema,
+			...(plan.strict ? { strict: true } : {}),
 		};
 	});
 }

package/src/providers/azure-openai-responses.ts

@@ -10,11 +10,11 @@ import {
 	type Api,
 	type AssistantMessage,
 	type Context,
-	isSpecialServiceTier,
 	type Model,
 	type ServiceTier,
 	type StreamFunction,
 	type StreamOptions,
+	shouldSendServiceTier,
 	type Tool,
 	type ToolChoice,
 } from "../types";
@@ -298,7 +298,7 @@ function buildParams(
 	if (options?.repetitionPenalty !== undefined) {
 		params.repetition_penalty = options.repetitionPenalty;
 	}
-	if (isSpecialServiceTier(options?.serviceTier)) {
+	if (shouldSendServiceTier(options?.serviceTier, model.provider)) {
 		params.service_tier = options.serviceTier;
 	}
 

package/src/providers/openai-codex-responses.ts

@@ -19,12 +19,12 @@ import {
 	type Api,
 	type AssistantMessage,
 	type Context,
-	isSpecialServiceTier,
 	type Model,
 	type ProviderSessionState,
 	type ServiceTier,
 	type StreamFunction,
 	type StreamOptions,
+	shouldSendServiceTier,
 	type TextContent,
 	type ThinkingContent,
 	type Tool,
@@ -75,27 +75,6 @@ export interface OpenAICodexResponsesOptions extends StreamOptions {
 	serviceTier?: ServiceTier;
 }
 
-export const CODEX_INSTRUCTIONS = `You are an expert coding assistant operating inside pi, a coding agent harness.`;
-
-export interface CodexSystemPrompt {
-	instructions: string;
-	developerMessages: string[];
-}
-
-export function buildCodexSystemPrompt(args: { userSystemPrompt?: string }): CodexSystemPrompt {
-	const { userSystemPrompt } = args;
-	const developerMessages: string[] = [];
-
-	if (userSystemPrompt && userSystemPrompt.trim().length > 0) {
-		developerMessages.push(userSystemPrompt.trim());
-	}
-
-	return {
-		instructions: CODEX_INSTRUCTIONS,
-		developerMessages,
-	};
-}
-
 const CODEX_DEBUG = $flag("PI_CODEX_DEBUG");
 const CODEX_MAX_RETRIES = 5;
 const CODEX_RETRYABLE_STATUS = new Set([408, 429, 500, 502, 503, 504]);
@@ -514,7 +493,7 @@ async function buildTransformedCodexRequestBody(
 	if (options?.repetitionPenalty !== undefined) {
 		params.repetition_penalty = options.repetitionPenalty;
 	}
-	if (isSpecialServiceTier(options?.serviceTier)) {
+	if (shouldSendServiceTier(options?.serviceTier, model.provider)) {
 		params.service_tier = options.serviceTier;
 	}
 	if (context.tools && context.tools.length > 0) {
@@ -537,8 +516,7 @@ async function buildTransformedCodexRequestBody(
 		}
 	}
 
-	const systemPrompt = buildCodexSystemPrompt({ userSystemPrompt: context.systemPrompt });
-	params.instructions = systemPrompt.instructions;
+	params.instructions = context.systemPrompt;
 
 	const codexOptions: CodexRequestOptions = {
 		reasoningEffort: options?.reasoning,
@@ -547,7 +525,7 @@ async function buildTransformedCodexRequestBody(
 		include: options?.include,
 	};
 
-	return transformRequestBody(params, model, codexOptions, systemPrompt);
+	return transformRequestBody(params, model, codexOptions);
 }
 
 async function openInitialCodexEventStream(

package/src/providers/openai-completions.ts

@@ -14,14 +14,15 @@ import { getEnvApiKey } from "../stream";
 import {
 	type AssistantMessage,
 	type Context,
-	isSpecialServiceTier,
 	type Message,
 	type MessageAttribution,
 	type Model,
+	type ProviderSessionState,
 	type ServiceTier,
 	type StopReason,
 	type StreamFunction,
 	type StreamOptions,
+	shouldSendServiceTier,
 	type TextContent,
 	type ThinkingContent,
 	type Tool,
@@ -139,6 +140,56 @@ type BuiltOpenAICompletionTools = {
 	toolStrictMode: AppliedToolStrictMode;
 };
 
+const OPENAI_COMPLETIONS_PROVIDER_SESSION_STATE_PREFIX = "openai-completions:";
+
+type OpenAICompletionsProviderSessionState = ProviderSessionState & {
+	strictToolsDisabled: boolean;
+};
+
+function createOpenAICompletionsProviderSessionState(): OpenAICompletionsProviderSessionState {
+	const state: OpenAICompletionsProviderSessionState = {
+		strictToolsDisabled: false,
+		close: () => {
+			state.strictToolsDisabled = false;
+		},
+	};
+	return state;
+}
+
+function getOpenAICompletionsProviderSessionState(
+	model: Model<"openai-completions">,
+	baseUrl: string | undefined,
+	providerSessionState: Map<string, ProviderSessionState> | undefined,
+): OpenAICompletionsProviderSessionState | undefined {
+	if (!providerSessionState) return undefined;
+	const key = `${OPENAI_COMPLETIONS_PROVIDER_SESSION_STATE_PREFIX}${model.provider}:${baseUrl ?? ""}:${model.id}`;
+	const existing = providerSessionState.get(key) as OpenAICompletionsProviderSessionState | undefined;
+	if (existing) return existing;
+	const created = createOpenAICompletionsProviderSessionState();
+	providerSessionState.set(key, created);
+	return created;
+}
+
+function isOpenRouterAnthropicModel(model: Model<"openai-completions">): boolean {
+	return model.provider === "openrouter" && model.id.toLowerCase().startsWith("anthropic/");
+}
+
+function isCompiledGrammarTooLargeStrictError(
+	error: unknown,
+	capturedErrorResponse: CapturedHttpErrorResponse | undefined,
+): boolean {
+	const status = extractHttpStatusFromError(error) ?? capturedErrorResponse?.status;
+	if (status !== 400) return false;
+	const messageParts = [error instanceof Error ? error.message : undefined, capturedErrorResponse?.bodyText]
+		.filter((value): value is string => typeof value === "string" && value.trim().length > 0)
+		.join("\n");
+	return (
+		/invalid_request_error/i.test(messageParts) &&
+		/compiled grammar/i.test(messageParts) &&
+		/too large/i.test(messageParts)
+	);
+}
+
 // LIMITATION: The think tag parser uses naive string matching for <think>/<thinking> tags.
 // If MiniMax models output these literal strings in code blocks, XML examples, or explanations,
 // they will be incorrectly consumed as thinking delimiters, truncating visible output.
@@ -227,9 +278,23 @@ export const streamOpenAICompletions: StreamFunction<"openai-completions"> = (
 			} = await createClient(model, context, apiKey, options?.headers, options?.initiatorOverride);
 			getCapturedErrorResponse = captureErrorResponse;
 			let appliedToolStrictMode: AppliedToolStrictMode = "mixed";
+			const providerSessionState = getOpenAICompletionsProviderSessionState(
+				model,
+				baseUrl,
+				options?.providerSessionState,
+			);
+			let disableStrictTools = providerSessionState?.strictToolsDisabled ?? false;
+			let strictFallbackErrorMessage: string | undefined;
 			const createCompletionsStream = async (toolStrictModeOverride?: ToolStrictModeOverride) => {
 				clearCapturedErrorResponse();
-				const { params, toolStrictMode } = buildParams(model, context, options, baseUrl, toolStrictModeOverride);
+				const effectiveToolStrictModeOverride = disableStrictTools ? "none" : toolStrictModeOverride;
+				const { params, toolStrictMode } = buildParams(
+					model,
+					context,
+					options,
+					baseUrl,
+					effectiveToolStrictModeOverride,
+				);
 				appliedToolStrictMode = toolStrictMode;
 				options?.onPayload?.(params);
 				rawRequestDump = {
@@ -251,10 +316,24 @@ export const streamOpenAICompletions: StreamFunction<"openai-completions"> = (
 				});
 			} catch (error) {
 				const capturedErrorResponse = getCapturedErrorResponse();
-				if (!shouldRetryWithoutStrictTools(error, capturedErrorResponse, appliedToolStrictMode, context.tools)) {
-					throw error;
+				if (
+					isOpenRouterAnthropicModel(model) &&
+					!disableStrictTools &&
+					isCompiledGrammarTooLargeStrictError(error, capturedErrorResponse)
+				) {
+					strictFallbackErrorMessage = await finalizeErrorMessage(error, rawRequestDump, capturedErrorResponse);
+					output.errorMessage = strictFallbackErrorMessage;
+					if (providerSessionState) {
+						providerSessionState.strictToolsDisabled = true;
+					}
+					disableStrictTools = true;
+					openaiStream = await createCompletionsStream("none");
+				} else {
+					if (!shouldRetryWithoutStrictTools(error, capturedErrorResponse, appliedToolStrictMode, context.tools)) {
+						throw error;
+					}
+					openaiStream = await createCompletionsStream("none");
 				}
-				openaiStream = await createCompletionsStream("none");
 			}
 			const firstEventWatchdog = createWatchdog(
 				options?.streamFirstEventTimeoutMs ?? getStreamFirstEventTimeoutMs(idleTimeoutMs),
@@ -541,6 +620,7 @@ export const streamOpenAICompletions: StreamFunction<"openai-completions"> = (
 				throw new Error(output.errorMessage || "Provider returned an error stop reason");
 			}
 
+			output.errorMessage = strictFallbackErrorMessage;
 			output.duration = Date.now() - startTime;
 			if (firstTokenTime) output.ttft = firstTokenTime - startTime;
 			stream.push({ type: "done", reason: output.stopReason, message: output });
@@ -614,6 +694,16 @@ async function createClient(
 		copilotPremiumRequests = copilot.premiumRequests;
 		baseUrl = resolveGitHubCopilotBaseUrl(model.baseUrl, rawApiKey) ?? model.baseUrl;
 	}
+	// Azure OpenAI requires /deployments/{id}/chat/completions?api-version=YYYY-MM-DD.
+	// The generic openai-completions path adds neither, producing silent 404s.
+	let azureDefaultQuery: Record<string, string> | undefined;
+	if (baseUrl?.includes(".openai.azure.com")) {
+		const apiVersion = $env.AZURE_OPENAI_API_VERSION || "2024-10-21";
+		if (!baseUrl.includes("/deployments/")) {
+			baseUrl = `${baseUrl}/deployments/${model.id}`;
+		}
+		azureDefaultQuery = { "api-version": apiVersion };
+	}
 	let capturedErrorResponse: CapturedHttpErrorResponse | undefined;
 	const wrappedFetch = Object.assign(
 		async (input: string | URL | Request, init?: RequestInit): Promise<Response> => {
@@ -649,6 +739,7 @@ async function createClient(
 			dangerouslyAllowBrowser: true,
 			maxRetries: 5,
 			defaultHeaders: headers,
+			defaultQuery: azureDefaultQuery,
 			fetch: wrappedFetch,
 		}),
 		copilotPremiumRequests,
@@ -720,7 +811,7 @@ function buildParams(
 	if (options?.repetitionPenalty !== undefined) {
 		params.repetition_penalty = options.repetitionPenalty;
 	}
-	if (isSpecialServiceTier(options?.serviceTier)) {
+	if (shouldSendServiceTier(options?.serviceTier, model.provider)) {
 		params.service_tier = options.serviceTier;
 	}
 
@@ -1095,6 +1186,10 @@ export function convertMessages(
 					(assistantMsg as any).reasoning_details = reasoningDetails;
 				}
 			}
+			// DeepSeek requires non-null content when reasoning_content is present
+			if (assistantMsg.content === null && hasReasoningField) {
+				assistantMsg.content = "";
+			}
 			// Skip assistant messages that have no content, no tool calls, and no reasoning payload.
 			// Some OpenAI-compatible backends require replaying reasoning-only assistant turns
 			// so follow-up requests preserve the provider-specific reasoning field name.

package/src/providers/openai-responses.ts

@@ -11,13 +11,13 @@ import {
 	type AssistantMessage,
 	type CacheRetention,
 	type Context,
-	isSpecialServiceTier,
 	type MessageAttribution,
 	type Model,
 	type ProviderSessionState,
 	type ServiceTier,
 	type StreamFunction,
 	type StreamOptions,
+	shouldSendServiceTier,
 	type Tool,
 	type ToolChoice,
 } from "../types";
@@ -384,7 +384,7 @@ function buildParams(
 	if (options?.repetitionPenalty !== undefined) {
 		params.repetition_penalty = options.repetitionPenalty;
 	}
-	if (isSpecialServiceTier(options?.serviceTier)) {
+	if (shouldSendServiceTier(options?.serviceTier, model.provider)) {
 		params.service_tier = options.serviceTier;
 	}
 

package/src/stream.ts

@@ -72,6 +72,7 @@ const serviceProviderMap: Record<string, KeyResolver> = {
 	"opencode-go": "OPENCODE_API_KEY",
 	"opencode-zen": "OPENCODE_API_KEY",
 	cursor: "CURSOR_ACCESS_TOKEN",
+	deepseek: "DEEPSEEK_API_KEY",
 	"openai-codex": "OPENAI_CODEX_OAUTH_TOKEN",
 	"azure-openai-responses": "AZURE_OPENAI_API_KEY",
 	exa: "EXA_API_KEY",

package/src/types.ts

@@ -100,6 +100,7 @@ export type KnownProvider =
 	| "github-copilot"
 	| "gitlab-duo"
 	| "cursor"
+	| "deepseek"
 	| "xai"
 	| "groq"
 	| "cerebras"
@@ -151,7 +152,13 @@ export type CacheRetention = "none" | "short" | "long";
 /** OpenAI service tier for processing priority. Only applies to OpenAI-compatible APIs. */
 export type ServiceTier = "auto" | "default" | "flex" | "scale" | "priority";
 
-export function isSpecialServiceTier(serviceTier?: ServiceTier | null): serviceTier is "flex" | "scale" | "priority" {
+export function shouldSendServiceTier(
+	serviceTier?: ServiceTier | null,
+	provider?: Provider,
+): serviceTier is "flex" | "scale" | "priority" {
+	if (provider !== "openai" && provider !== "openai-codex") {
+		return false;
+	}
 	return serviceTier === "flex" || serviceTier === "scale" || serviceTier === "priority";
 }
 

package/src/utils/schema/strict-mode.ts

@@ -17,13 +17,14 @@ import { isJsonObject } from "./types";
  */
 export function StringEnum<const T extends readonly string[]>(
 	values: T,
-	options?: { description?: string; default?: T[number] },
+	options?: { description?: string; default?: T[number]; examples?: readonly T[number][] },
 ): TUnsafe<T[number]> {
 	return Type.Unsafe<T[number]>({
 		type: "string",
 		enum: values as unknown as string[],
 		...(options?.description && { description: options.description }),
 		...(options?.default && { default: options.default }),
+		...(options?.examples && { examples: options.examples }),
 	});
 }
 

package/src/utils/validation.ts

@@ -141,6 +141,13 @@ function tryParseLeadingJsonContainer(value: string): unknown | undefined {
 					return JSON.parse(cleaned) as unknown;
 				} catch {}
 			}
+			// Try escaping raw control chars that appear inside string literals.
+			const escapedControls = escapeRawControlsInJsonStrings(prefix);
+			if (escapedControls !== prefix) {
+				try {
+					return JSON.parse(escapedControls) as unknown;
+				} catch {}
+			}
 			// Also try single-char healing on the extracted prefix.
 			return tryHealMalformedJson(prefix);
 		}
@@ -191,6 +198,72 @@ function cleanLiteralEscapes(value: string): string {
 	}
 	return result;
 }
+/**
+ * Escape raw control characters (0x00–0x1F) that appear *inside* JSON string
+ * literals. LLMs sometimes emit literal newlines/tabs/etc. inside string
+ * content instead of `\n` / `\t` escape sequences, which `JSON.parse` rejects
+ * even though the surrounding structure is valid.
+ *
+ * This function only rewrites characters while inside a string; structural
+ * whitespace outside of strings is preserved unchanged.
+ */
+function escapeRawControlsInJsonStrings(value: string): string {
+	let result = "";
+	let inString = false;
+	let escaped = false;
+	let changed = false;
+	for (let i = 0; i < value.length; i += 1) {
+		const ch = value[i];
+		if (inString) {
+			if (escaped) {
+				result += ch;
+				escaped = false;
+				continue;
+			}
+			if (ch === "\\") {
+				result += ch;
+				escaped = true;
+				continue;
+			}
+			if (ch === '"') {
+				result += ch;
+				inString = false;
+				continue;
+			}
+			const code = ch.charCodeAt(0);
+			if (code < 0x20) {
+				changed = true;
+				switch (ch) {
+					case "\n":
+						result += "\\n";
+						break;
+					case "\r":
+						result += "\\r";
+						break;
+					case "\t":
+						result += "\\t";
+						break;
+					case "\b":
+						result += "\\b";
+						break;
+					case "\f":
+						result += "\\f";
+						break;
+					default:
+						result += `\\u${code.toString(16).padStart(4, "0")}`;
+				}
+				continue;
+			}
+			result += ch;
+			continue;
+		}
+		if (ch === '"') {
+			inString = true;
+		}
+		result += ch;
+	}
+	return changed ? result : value;
+}
 
 /** Maximum single-character edits to attempt when healing malformed JSON. */
 const MAX_HEAL_DISTANCE = 3;
@@ -286,6 +359,17 @@ function tryParseJsonForTypes(value: string, expectedTypes: string[]): { value:
 		}
 	} catch {
 		if (looksJsonObject || looksJsonArray) {
+			// Try escaping raw control chars inside string literals (LLMs sometimes
+			// emit literal newlines/tabs inside string content rather than `\n`/`\t`).
+			const escapedControls = escapeRawControlsInJsonStrings(trimmed);
+			if (escapedControls !== trimmed) {
+				try {
+					const parsed = JSON.parse(escapedControls) as unknown;
+					if (matchesExpectedType(parsed, expectedTypes)) {
+						return { value: parsed, changed: true };
+					}
+				} catch {}
+			}
 			// Try extracting a valid JSON prefix (handles trailing junk after balanced container)
 			const leading = tryParseLeadingJsonContainer(trimmed);
 			if (leading !== undefined && matchesExpectedType(leading, expectedTypes)) {
@@ -501,6 +585,26 @@ function normalizeOptionalNullsForSchema(schema: unknown, value: unknown): { val
 		nextValue[key] = normalized.value;
 	}
 
+	// Strip unknown keys with null/"null" values when the schema forbids extras.
+	// LLMs sometimes hallucinate verbs alongside valid ones (e.g. `split: null`,
+	// `original: null`). Rejecting the entire tool call wastes a turn; treating
+	// these the same as null on known optional fields is a safer fallback. Keys
+	// with non-null unknown values are left intact so genuine schema mistakes
+	// still surface as validation errors.
+	if (schemaObject.additionalProperties === false) {
+		const knownKeys = new Set(Object.keys(properties));
+		for (const key of Object.keys(nextValue)) {
+			if (knownKeys.has(key)) continue;
+			const v = nextValue[key];
+			if (v !== null && v !== "null") continue;
+			if (!changed) {
+				nextValue = { ...nextValue };
+				changed = true;
+			}
+			delete nextValue[key];
+		}
+	}
+
 	return { value: changed ? nextValue : value, changed };
 }