@psnext/slingcli 2.5.20260724-1 → 2.5.20260727-1

package/node_modules/@earendil-works/pi-ai/dist/api/openai-codex-responses.js

@@ -16,6 +16,7 @@ import { AssistantMessageEventStream } from "../utils/event-stream.js";
 import { headersToRecord } from "../utils/headers.js";
 import { resolveHttpProxyUrlForTarget } from "../utils/node-http-proxy.js";
 import { uuidv7 } from "../utils/uuid.js";
+import { createGrammarToolInputProperties } from "./constrained-sampling.js";
 import { clampOpenAIPromptCacheKey } from "./openai-prompt-cache.js";
 import { convertResponsesMessages, convertResponsesTools, processResponsesStream } from "./openai-responses-shared.js";
 import { buildBaseOptions } from "./simple-options.js";
@@ -34,6 +35,7 @@ const REQUEST_COMPRESSION_ZSTD_LEVEL = 3;
 const CODEX_TOOL_CALL_PROVIDERS = new Set(["openai", "openai-codex", "opencode"]);
 const WEBSOCKET_MESSAGE_TOO_BIG_CLOSE_CODE = 1009;
 const WEBSOCKET_CONNECTION_LIMIT_REACHED_CODE = "websocket_connection_limit_reached";
+const PREVIOUS_RESPONSE_NOT_FOUND_CODE = "previous_response_not_found";
 const CODEX_RESPONSE_STATUSES = new Set([
     "completed",
     "incomplete",
@@ -79,9 +81,14 @@ function getRetryAfterDelayMs(headers) {
     }
     return undefined;
 }
-function capRetryDelayMs(delayMs, options) {
+class RetryDelayExceededError extends Error {
+}
+function validateRetryDelayMs(delayMs, options) {
     const maxRetryDelayMs = options?.maxRetryDelayMs ?? DEFAULT_MAX_RETRY_DELAY_MS;
-    return maxRetryDelayMs > 0 ? Math.min(delayMs, maxRetryDelayMs) : delayMs;
+    if (maxRetryDelayMs > 0 && delayMs > maxRetryDelayMs) {
+        throw new RetryDelayExceededError(`Server requested ${Math.ceil(delayMs / 1000)}s retry delay (max: ${Math.ceil(maxRetryDelayMs / 1000)}s)`);
+    }
+    return delayMs;
 }
 function sleep(ms, signal) {
     return new Promise((resolve, reject) => {
@@ -157,12 +164,14 @@ export const stream = (model, context, options) => {
                 throw new Error(`No API key for provider: ${model.provider}`);
             }
             const accountId = extractAccountId(apiKey);
-            let body = buildRequestBody(model, context, options);
+            const grammarToolInputProperties = createGrammarToolInputProperties(context.tools, model.compat?.supportsOpenAIGrammarTools ?? false);
+            const cacheSessionId = options?.cacheRetention === "none" ? undefined : options?.sessionId;
+            const codexSessionId = clampOpenAIPromptCacheKey(cacheSessionId);
+            let body = buildRequestBody(model, context, options, codexSessionId, grammarToolInputProperties);
             const nextBody = await options?.onPayload?.(body, model);
             if (nextBody !== undefined) {
                 body = nextBody;
             }
-            const codexSessionId = clampOpenAIPromptCacheKey(options?.sessionId);
             const websocketRequestId = codexSessionId || uuidv7();
             const sseHeaders = buildSSEHeaders(model.headers, options?.headers, accountId, apiKey, codexSessionId);
             const websocketHeaders = buildWebSocketHeaders(model.headers, options?.headers, accountId, apiKey, websocketRequestId);
@@ -170,19 +179,25 @@ export const stream = (model, context, options) => {
             const httpTimeoutMs = normalizeTimeoutMs(options?.timeoutMs);
             const websocketConnectTimeoutMs = normalizeTimeoutMs(options?.websocketConnectTimeoutMs);
             const transport = options?.transport || "auto";
-            const websocketDisabledForSession = transport !== "sse" && isWebSocketSseFallbackActive(options?.sessionId);
+            let startEmitted = false;
+            const websocketDisabledForSession = transport !== "sse" && isWebSocketSseFallbackActive(cacheSessionId);
             if (websocketDisabledForSession) {
-                recordWebSocketSseFallback(options?.sessionId);
+                recordWebSocketSseFallback(cacheSessionId);
             }
             if (transport !== "sse" && !websocketDisabledForSession) {
                 let websocketStarted = false;
                 let retriedWebSocketConnectionLimit = false;
+                let retriedMissingWebSocketContinuation = false;
                 while (true) {
                     websocketStarted = false;
                     try {
                         await processWebSocketStream(resolveCodexWebSocketUrl(model.baseUrl), body, websocketHeaders, output, stream, model, () => {
                             websocketStarted = true;
-                        }, httpTimeoutMs, websocketConnectTimeoutMs, options);
+                            if (!startEmitted) {
+                                startEmitted = true;
+                                stream.push({ type: "start", partial: output });
+                            }
+                        }, httpTimeoutMs, websocketConnectTimeoutMs, cacheSessionId, grammarToolInputProperties, options);
                         if (options?.signal?.aborted) {
                             throw new Error("Request was aborted");
                         }
@@ -197,6 +212,11 @@ export const stream = (model, context, options) => {
                     catch (error) {
                         const aborted = options?.signal?.aborted;
                         const connectionLimitBeforeStart = !websocketStarted && isWebSocketConnectionLimitReachedError(error);
+                        const previousResponseNotFound = isPreviousResponseNotFoundError(error);
+                        if (!aborted && previousResponseNotFound && !retriedMissingWebSocketContinuation) {
+                            retriedMissingWebSocketContinuation = true;
+                            continue;
+                        }
                         if (!aborted && connectionLimitBeforeStart && !retriedWebSocketConnectionLimit) {
                             retriedWebSocketConnectionLimit = true;
                             continue;
@@ -211,11 +231,11 @@ export const stream = (model, context, options) => {
                             phase: websocketStarted ? "after_message_stream_start" : "before_message_stream_start",
                             requestBytes: new TextEncoder().encode(bodyJson).byteLength,
                         }));
-                        recordWebSocketFailure(options?.sessionId, error);
+                        recordWebSocketFailure(cacheSessionId, error);
                         if (websocketStarted) {
                             throw error;
                         }
-                        recordWebSocketSseFallback(options?.sessionId);
+                        recordWebSocketSseFallback(cacheSessionId);
                         break;
                     }
                 }
@@ -265,9 +285,7 @@ export const stream = (model, context, options) => {
                         const retryAfterDelayMs = getRetryAfterDelayMs(response.headers);
                         const delayMs = retryAfterDelayMs === undefined
                             ? BASE_DELAY_MS * 2 ** attempt
-                            : response.status === 429
-                                ? capRetryDelayMs(retryAfterDelayMs, options)
-                                : retryAfterDelayMs;
+                            : validateRetryDelayMs(retryAfterDelayMs, options);
                         await sleep(delayMs, options?.signal);
                         continue;
                     }
@@ -287,7 +305,9 @@ export const stream = (model, context, options) => {
                     }
                     lastError = error instanceof Error ? error : new Error(String(error));
                     // Network errors are retryable
-                    if (attempt < maxRetries && !lastError.message.includes("usage limit")) {
+                    if (attempt < maxRetries &&
+                        !(lastError instanceof RetryDelayExceededError) &&
+                        !lastError.message.includes("usage limit")) {
                         const delayMs = BASE_DELAY_MS * 2 ** attempt;
                         await sleep(delayMs, options?.signal);
                         continue;
@@ -301,8 +321,11 @@ export const stream = (model, context, options) => {
             if (!response.body) {
                 throw new Error("No response body");
             }
-            stream.push({ type: "start", partial: output });
-            await processStream(response, output, stream, model, options);
+            if (!startEmitted) {
+                startEmitted = true;
+                stream.push({ type: "start", partial: output });
+            }
+            await processStream(response, output, stream, model, grammarToolInputProperties, options);
             if (options?.signal?.aborted) {
                 throw new Error("Request was aborted");
             }
@@ -311,8 +334,9 @@ export const stream = (model, context, options) => {
         }
         catch (error) {
             for (const block of output.content) {
-                // partialJson is only a streaming scratch buffer; never persist it.
+                // Streaming scratch buffers are only used during parsing; never persist them.
                 delete block.partialJson;
+                delete block.customInput;
             }
             output.stopReason = options?.signal?.aborted ? "aborted" : "error";
             output.errorMessage = formatProviderError(normalizeProviderError(error));
@@ -338,11 +362,19 @@ export const streamSimple = (model, context, options) => {
 // ============================================================================
 // Request Building
 // ============================================================================
-function buildRequestBody(model, context, options) {
+function buildRequestBody(model, context, options, cacheSessionId, grammarToolInputProperties = createGrammarToolInputProperties(context.tools, model.compat?.supportsOpenAIGrammarTools ?? false)) {
+    const supportsStrictMode = model.compat?.supportsStrictMode ?? true;
+    const supportsOpenAIGrammarTools = model.compat?.supportsOpenAIGrammarTools ?? false;
     const toolPlacement = splitDeferredTools(context, model.compat?.supportsToolSearch ?? false);
     const messages = convertResponsesMessages(model, context, CODEX_TOOL_CALL_PROVIDERS, {
         includeSystemPrompt: false,
+        grammarToolInputProperties,
         deferredTools: toolPlacement.deferred,
+        toolOptions: {
+            strict: null,
+            supportsStrictMode,
+            supportsOpenAIGrammarTools,
+        },
     });
     const body = {
         model: model.id,
@@ -352,7 +384,7 @@ function buildRequestBody(model, context, options) {
         input: messages,
         text: { verbosity: options?.textVerbosity || "low" },
         include: ["reasoning.encrypted_content"],
-        prompt_cache_key: clampOpenAIPromptCacheKey(options?.sessionId),
+        prompt_cache_key: cacheSessionId,
         tool_choice: options?.toolChoice ?? "auto",
         parallel_tool_calls: true,
     };
@@ -363,7 +395,11 @@ function buildRequestBody(model, context, options) {
         body.service_tier = options.serviceTier;
     }
     if (toolPlacement.immediate.length > 0) {
-        body.tools = convertResponsesTools(toolPlacement.immediate, { strict: null });
+        body.tools = convertResponsesTools(toolPlacement.immediate, {
+            strict: null,
+            supportsStrictMode,
+            supportsOpenAIGrammarTools,
+        });
     }
     if (options?.reasoningEffort !== undefined) {
         const effort = options.reasoningEffort === "none"
@@ -424,9 +460,10 @@ function resolveCodexWebSocketUrl(baseUrl) {
 // ============================================================================
 // Response Processing
 // ============================================================================
-async function processStream(response, output, stream, model, options) {
+async function processStream(response, output, stream, model, grammarToolInputProperties, options) {
     await processResponsesStream(mapCodexEvents(parseSSE(response, options?.signal)), output, stream, model, {
         serviceTier: options?.serviceTier,
+        grammarToolInputProperties,
         resolveServiceTier: resolveCodexServiceTier,
         applyServiceTierPricing: (usage, serviceTier) => applyServiceTierPricing(usage, serviceTier, model),
     });
@@ -457,6 +494,9 @@ function isCodexNonTransportError(error) {
 function isWebSocketConnectionLimitReachedError(error) {
     return error instanceof CodexApiError && error.code === WEBSOCKET_CONNECTION_LIMIT_REACHED_CODE;
 }
+function isPreviousResponseNotFoundError(error) {
+    return error instanceof CodexApiError && error.code === PREVIOUS_RESPONSE_NOT_FOUND_CODE;
+}
 function extractCodexEventError(event) {
     const nested = event.error && typeof event.error === "object" ? event.error : undefined;
     return {
@@ -1062,26 +1102,25 @@ function buildCachedWebSocketRequestBody(entry, body) {
         input: delta,
     };
 }
-async function* startWebSocketOutputOnFirstEvent(events, output, stream, onStart) {
+async function* startWebSocketOutputOnFirstEvent(events, onStart) {
     let started = false;
     for await (const event of events) {
         if (!started) {
             started = true;
             onStart();
-            stream.push({ type: "start", partial: output });
         }
         yield event;
     }
 }
-async function processWebSocketStream(url, body, headers, output, stream, model, onStart, idleTimeoutMs, websocketConnectTimeoutMs, options) {
-    const { socket, entry, reused, release } = await acquireWebSocket(url, headers, options?.sessionId, options?.signal, websocketConnectTimeoutMs, options?.env);
+async function processWebSocketStream(url, body, headers, output, stream, model, onStart, idleTimeoutMs, websocketConnectTimeoutMs, cacheSessionId, grammarToolInputProperties, options) {
+    const { socket, entry, reused, release } = await acquireWebSocket(url, headers, cacheSessionId, options?.signal, websocketConnectTimeoutMs, options?.env);
     let keepConnection = true;
     const useCachedContext = options?.transport === "websocket-cached" || options?.transport === "auto";
     // ChatGPT Codex Responses rejects `store: true` ("Store must be set to false").
     // WebSocket continuation still works via connection-scoped previous_response_id state.
     const fullBody = body;
     const requestBody = useCachedContext && entry ? buildCachedWebSocketRequestBody(entry, fullBody) : fullBody;
-    const stats = options?.sessionId ? getOrCreateWebSocketDebugStats(options.sessionId) : undefined;
+    const stats = cacheSessionId ? getOrCreateWebSocketDebugStats(cacheSessionId) : undefined;
     if (stats) {
         stats.requests++;
         if (reused)
@@ -1106,8 +1145,9 @@ async function processWebSocketStream(url, body, headers, output, stream, model,
     }
     try {
         socket.send(JSON.stringify({ type: "response.create", ...requestBody }));
-        await processResponsesStream(startWebSocketOutputOnFirstEvent(mapCodexEvents(parseWebSocket(socket, options?.signal, idleTimeoutMs)), output, stream, onStart), output, stream, model, {
+        await processResponsesStream(startWebSocketOutputOnFirstEvent(mapCodexEvents(parseWebSocket(socket, options?.signal, idleTimeoutMs)), onStart), output, stream, model, {
             serviceTier: options?.serviceTier,
+            grammarToolInputProperties,
             resolveServiceTier: resolveCodexServiceTier,
             applyServiceTierPricing: (usage, serviceTier) => applyServiceTierPricing(usage, serviceTier, model),
         });
@@ -1117,7 +1157,8 @@ async function processWebSocketStream(url, body, headers, output, stream, model,
         else if (useCachedContext && entry && output.responseId) {
             const responseItems = convertResponsesMessages(model, { messages: [output] }, CODEX_TOOL_CALL_PROVIDERS, {
                 includeSystemPrompt: false,
-            }).filter((item) => item.type !== "function_call_output");
+                grammarToolInputProperties,
+            }).filter((item) => item.type !== "function_call_output" && item.type !== "custom_tool_call_output");
             entry.continuation = {
                 lastRequestBody: fullBody,
                 lastResponseId: output.responseId,

package/node_modules/@earendil-works/pi-ai/dist/api/openai-completions.js

@@ -6,7 +6,9 @@ import { shortHash } from "../utils/hash.js";
 import { headersToRecord } from "../utils/headers.js";
 import { parseStreamingJson } from "../utils/json-parse.js";
 import { getProviderEnvValue } from "../utils/provider-env.js";
+import { retryProviderRequest } from "../utils/provider-retry.js";
 import { sanitizeSurrogates } from "../utils/sanitize-unicode.js";
+import { appendGrammarToolInputJsonDelta, createGrammarToolInputProperties, getGrammarToolInput, resolveGrammarConstrainedSampling, resolveJsonSchemaStrictSampling, } from "./constrained-sampling.js";
 import { buildCopilotDynamicHeaders, hasCopilotVisionInput } from "./github-copilot-headers.js";
 import { clampOpenAIPromptCacheKey } from "./openai-prompt-cache.js";
 import { buildBaseOptions } from "./simple-options.js";
@@ -120,10 +122,11 @@ export const stream = (model, context, options) => {
         try {
             const apiKey = getClientApiKey(model.provider, options?.apiKey, options?.headers);
             const compat = getCompat(model);
+            const grammarToolInputProperties = createGrammarToolInputProperties(context.tools, compat.supportsOpenAIGrammarTools);
             const cacheRetention = resolveCacheRetention(options?.cacheRetention, options?.env);
             const cacheSessionId = cacheRetention === "none" ? undefined : options?.sessionId;
             const client = createClient(model, context, apiKey, options?.headers, cacheSessionId, compat);
-            let params = buildParams(model, context, options, compat, cacheRetention);
+            let params = buildParams(model, context, options, compat, cacheRetention, grammarToolInputProperties);
             const nextParams = await options?.onPayload?.(params, model);
             if (nextParams !== undefined) {
                 params = nextParams;
@@ -131,11 +134,13 @@ export const stream = (model, context, options) => {
             const requestOptions = {
                 ...(options?.signal ? { signal: options.signal } : {}),
                 ...(options?.timeoutMs !== undefined ? { timeout: options.timeoutMs } : {}),
-                maxRetries: options?.maxRetries ?? 0,
+                maxRetries: 0,
             };
-            const { data: openaiStream, response } = await client.chat.completions
-                .create(params, requestOptions)
-                .withResponse();
+            const { data: openaiStream, response } = await retryProviderRequest(() => client.chat.completions.create(params, requestOptions).withResponse(), {
+                maxRetries: options?.maxRetries,
+                maxRetryDelayMs: options?.maxRetryDelayMs,
+                signal: options?.signal,
+            });
             await options?.onResponse?.({ status: response.status, headers: headersToRecord(response.headers) }, model);
             stream.push({ type: "start", partial: output });
             let textBlock = null;
@@ -146,6 +151,21 @@ export const stream = (model, context, options) => {
             const pendingReasoningDetailsByToolCallId = new Map();
             const blocks = output.content;
             const getContentIndex = (block) => blocks.indexOf(block);
+            const getCustomToolCallInput = (block) => {
+                const property = block.customInput?.property;
+                if (property === undefined)
+                    return "";
+                const value = block.arguments[property];
+                return typeof value === "string" ? value : "";
+            };
+            const appendCustomToolCallInput = (block, nextInput, close) => {
+                const customInput = block.customInput;
+                if (!customInput)
+                    return undefined;
+                const delta = appendGrammarToolInputJsonDelta(customInput.jsonBuffer, customInput.property, nextInput, close);
+                block.arguments = { [customInput.property]: nextInput };
+                return delta;
+            };
             const finishBlock = (block) => {
                 const contentIndex = getContentIndex(block);
                 if (contentIndex === -1) {
@@ -168,10 +188,24 @@ export const stream = (model, context, options) => {
                     });
                 }
                 else if (block.type === "toolCall") {
-                    block.arguments = parseStreamingJson(block.partialArgs);
+                    if (block.customInput) {
+                        const delta = appendCustomToolCallInput(block, getCustomToolCallInput(block), true);
+                        if (delta !== undefined) {
+                            stream.push({
+                                type: "toolcall_delta",
+                                contentIndex,
+                                delta,
+                                partial: output,
+                            });
+                        }
+                    }
+                    else {
+                        block.arguments = parseStreamingJson(block.partialArgs);
+                    }
                     // Finalize in-place and strip the scratch buffers so replay only
                     // carries parsed arguments.
                     delete block.partialArgs;
+                    delete block.customInput;
                     delete block.streamIndex;
                     stream.push({
                         type: "toolcall_end",
@@ -213,17 +247,27 @@ export const stream = (model, context, options) => {
             };
             const ensureToolCallBlock = (toolCall) => {
                 const streamIndex = typeof toolCall.index === "number" ? toolCall.index : undefined;
+                const name = toolCall.function?.name ?? toolCall.custom?.name ?? "";
                 let block = streamIndex !== undefined ? toolCallBlocksByIndex.get(streamIndex) : undefined;
                 if (!block && toolCall.id) {
                     block = toolCallBlocksById.get(toolCall.id);
                 }
                 if (!block) {
+                    // Note: the "input" fallback here should/must not be taken.  in case the LLM makes up
+                    // a tool we don't knwo about, we at least have a place to stash our stuff.
+                    const customInputProperty = toolCall.custom
+                        ? (grammarToolInputProperties.get(name) ?? "input")
+                        : undefined;
+                    const hasCustomInput = customInputProperty !== undefined;
                     block = {
                         type: "toolCall",
                         id: toolCall.id || "",
-                        name: toolCall.function?.name || "",
-                        arguments: {},
-                        partialArgs: "",
+                        name,
+                        arguments: hasCustomInput ? { [customInputProperty]: "" } : {},
+                        partialArgs: hasCustomInput ? undefined : "",
+                        customInput: hasCustomInput
+                            ? { property: customInputProperty, jsonBuffer: { input: "", started: false, closed: false } }
+                            : undefined,
                         streamIndex,
                     };
                     if (streamIndex !== undefined) {
@@ -246,6 +290,18 @@ export const stream = (model, context, options) => {
                 if (toolCall.id) {
                     toolCallBlocksById.set(toolCall.id, block);
                 }
+                if (!block.name && name) {
+                    block.name = name;
+                }
+                if (toolCall.custom && !block.customInput) {
+                    const customInputProperty = grammarToolInputProperties.get(block.name) ?? "input";
+                    block.arguments = { [customInputProperty]: "" };
+                    block.customInput = {
+                        property: customInputProperty,
+                        jsonBuffer: { input: "", started: false, closed: false },
+                    };
+                    delete block.partialArgs;
+                }
                 applyPendingReasoningDetail(block);
                 return block;
             };
@@ -327,8 +383,9 @@ export const stream = (model, context, options) => {
                                 block.id = toolCall.id;
                                 toolCallBlocksById.set(toolCall.id, block);
                             }
-                            if (!block.name && toolCall.function?.name) {
-                                block.name = toolCall.function.name;
+                            const name = toolCall.function?.name ?? toolCall.custom?.name;
+                            if (!block.name && name) {
+                                block.name = name;
                             }
                             let delta = "";
                             if (toolCall.function?.arguments) {
@@ -336,6 +393,10 @@ export const stream = (model, context, options) => {
                                 block.partialArgs = (block.partialArgs ?? "") + toolCall.function.arguments;
                                 block.arguments = parseStreamingJson(block.partialArgs);
                             }
+                            else if (toolCall.custom?.input) {
+                                const nextInput = getCustomToolCallInput(block) + toolCall.custom.input;
+                                delta = appendCustomToolCallInput(block, nextInput, false) ?? "";
+                            }
                             stream.push({
                                 type: "toolcall_delta",
                                 contentIndex: getContentIndex(block),
@@ -384,6 +445,7 @@ export const stream = (model, context, options) => {
                 delete block.index;
                 // Streaming scratch buffers are only used during parsing; never persist them.
                 delete block.partialArgs;
+                delete block.customInput;
                 delete block.streamIndex;
             }
             output.stopReason = options?.signal?.aborted ? "aborted" : "error";
@@ -447,8 +509,8 @@ function createClient(model, context, apiKey, optionsHeaders, sessionId, compat
         defaultHeaders: headers,
     });
 }
-function buildParams(model, context, options, compat = getCompat(model), cacheRetention = resolveCacheRetention(options?.cacheRetention, options?.env)) {
-    const messages = convertMessages(model, context, compat);
+function buildParams(model, context, options, compat = getCompat(model), cacheRetention = resolveCacheRetention(options?.cacheRetention, options?.env), grammarToolInputProperties = createGrammarToolInputProperties(context.tools, compat.supportsOpenAIGrammarTools)) {
+    const messages = convertMessages(model, context, compat, { grammarToolInputProperties });
     const cacheControl = getCompatCacheControl(compat, cacheRetention);
     const params = {
         model: model.id,
@@ -642,7 +704,7 @@ function addCacheControlToSystemPrompt(messages, cacheControl) {
 function addCacheControlToLastConversationMessage(messages, cacheControl) {
     for (let i = messages.length - 1; i >= 0; i--) {
         const message = messages[i];
-        if (message.role === "user" || message.role === "assistant") {
+        if (message.role === "user" || message.role === "assistant" || message.role === "tool") {
             if (addCacheControlToMessage(message, cacheControl)) {
                 return;
             }
@@ -660,7 +722,7 @@ function addCacheControlToInstructionMessage(message, cacheControl) {
     return addCacheControlToTextContent(message, cacheControl);
 }
 function addCacheControlToMessage(message, cacheControl) {
-    if (message.role === "user" || message.role === "assistant") {
+    if (message.role === "user" || message.role === "assistant" || message.role === "tool") {
         return addCacheControlToTextContent(message, cacheControl);
     }
     return false;
@@ -693,7 +755,7 @@ function addCacheControlToTextContent(message, cacheControl) {
     }
     return false;
 }
-export function convertMessages(model, context, compat) {
+export function convertMessages(model, context, compat, options) {
     const params = [];
     const normalizeToolCallId = (id) => {
         // Handle pipe-separated IDs from OpenAI Responses API
@@ -823,14 +885,27 @@ export function convertMessages(model, context, compat) {
             }
             const toolCalls = msg.content.filter(isToolCallBlock);
             if (toolCalls.length > 0) {
-                assistantMsg.tool_calls = toolCalls.map((tc) => ({
-                    id: tc.id,
-                    type: "function",
-                    function: {
-                        name: tc.name,
-                        arguments: JSON.stringify(tc.arguments),
-                    },
-                }));
+                assistantMsg.tool_calls = toolCalls.map((tc) => {
+                    const customInputProperty = options?.grammarToolInputProperties?.get(tc.name);
+                    if (customInputProperty !== undefined) {
+                        return {
+                            id: tc.id,
+                            type: "custom",
+                            custom: {
+                                name: tc.name,
+                                input: sanitizeSurrogates(getGrammarToolInput(tc.name, tc.arguments, customInputProperty)),
+                            },
+                        };
+                    }
+                    return {
+                        id: tc.id,
+                        type: "function",
+                        function: {
+                            name: tc.name,
+                            arguments: JSON.stringify(tc.arguments),
+                        },
+                    };
+                });
                 const reasoningDetails = toolCalls
                     .filter((tc) => tc.thoughtSignature)
                     .map((tc) => {
@@ -948,16 +1023,36 @@ export function convertMessages(model, context, compat) {
     return params;
 }
 function convertTools(tools, compat) {
-    return tools.map((tool) => ({
-        type: "function",
-        function: {
-            name: tool.name,
-            description: tool.description,
-            parameters: tool.parameters, // TypeBox already generates JSON Schema
-            // Only include strict if provider supports it. Some reject unknown fields.
-            ...(compat.supportsStrictMode !== false && { strict: false }),
-        },
-    }));
+    return tools.map((tool) => {
+        const grammar = resolveGrammarConstrainedSampling(tool, compat.supportsOpenAIGrammarTools);
+        if (grammar) {
+            return {
+                type: "custom",
+                custom: {
+                    name: tool.name,
+                    description: tool.description,
+                    format: {
+                        type: "grammar",
+                        grammar: {
+                            syntax: grammar.format,
+                            definition: grammar.definition,
+                        },
+                    },
+                },
+            };
+        }
+        const strict = resolveJsonSchemaStrictSampling(tool, compat.supportsStrictMode !== false);
+        return {
+            type: "function",
+            function: {
+                name: tool.name,
+                description: tool.description,
+                parameters: tool.parameters, // TypeBox already generates JSON Schema
+                // Only include strict if provider supports it. Some reject unknown fields.
+                ...(compat.supportsStrictMode !== false && { strict: strict ?? false }),
+            },
+        };
+    });
 }
 function parseChunkUsage(rawUsage, model) {
     const promptTokens = rawUsage.prompt_tokens || 0;
@@ -1074,6 +1169,7 @@ function detectCompat(model) {
         chatTemplateKwargs: {},
         zaiToolStream: false,
         supportsStrictMode: !isMoonshot && !isTogether && !isCloudflareAiGateway && !isNvidia,
+        supportsOpenAIGrammarTools: false,
         cacheControlFormat,
         sendSessionAffinityHeaders: false,
         deferredToolsMode: undefined,
@@ -1110,6 +1206,7 @@ function getCompat(model) {
         chatTemplateKwargs: model.compat.chatTemplateKwargs ?? detected.chatTemplateKwargs,
         zaiToolStream: model.compat.zaiToolStream ?? detected.zaiToolStream,
         supportsStrictMode: model.compat.supportsStrictMode ?? detected.supportsStrictMode,
+        supportsOpenAIGrammarTools: model.compat.supportsOpenAIGrammarTools ?? detected.supportsOpenAIGrammarTools,
         cacheControlFormat: model.compat.cacheControlFormat ?? detected.cacheControlFormat,
         sendSessionAffinityHeaders: model.compat.sendSessionAffinityHeaders ?? detected.sendSessionAffinityHeaders,
         deferredToolsMode: model.compat.deferredToolsMode ?? detected.deferredToolsMode,