@ryuu-reinzz/ryuu-cli 0.0.0-stage → 0.4.0

package/src/providers/nvidia.js

@@ -0,0 +1,671 @@
+import OpenAI from 'openai';
+import { CancelledError, ProviderError, isRecoverableError } from '../utils/errors.js';
+import { redact } from '../utils/redaction.js';
+import { toToolArgumentsString } from '../utils/validation.js';
+import { MODEL_ID_PATTERN, MAX_MODEL_ID_LENGTH, PROVIDER_REQUEST_FIELDS } from './catalog.js';
+import { applyRequestExtras } from './vendor-extras.js';
+
+const RETRY_BASE_MS = 500;
+const RETRY_MAX_MS = 8000;
+
+const sleep = (ms) => new Promise((resolve) => setTimeout(resolve, ms));
+
+function abortableSleep(ms, signal) {
+  if (!signal) return sleep(ms);
+  return new Promise((resolve, reject) => {
+    const timer = setTimeout(() => {
+      signal.removeEventListener('abort', onAbort);
+      resolve();
+    }, ms);
+    const onAbort = () => {
+      clearTimeout(timer);
+      reject(new CancelledError());
+    };
+    signal.addEventListener('abort', onAbort, { once: true });
+  });
+}
+
+function statusOf(error) {
+  return error?.status ?? error?.response?.status ?? null;
+}
+
+const TIMEOUT_CODES = new Set(['ETIMEDOUT', 'UND_ERR_CONNECT_TIMEOUT', 'UND_ERR_HEADERS_TIMEOUT']);
+const NETWORK_CODES = new Set([
+  'ECONNREFUSED',
+  'ENOTFOUND',
+  'EAI_AGAIN',
+  'ECONNRESET',
+  'ECONNABORTED',
+  'EPIPE',
+  'EHOSTUNREACH',
+  'ENETUNREACH',
+  'UND_ERR_SOCKET',
+]);
+
+/**
+ * True when a provider failure is worth retrying: the request may succeed on
+ * the next attempt (network blip, rate limit, server error) or the failure is
+ * known to be repairable locally (a poisoned tool-call payload).
+ */
+export function isRecoverableProviderError(error) {
+  return isRecoverableError(error);
+}
+
+/**
+ * Walk error.cause (fetch/undici nesting) looking for a known code.
+ * The OpenAI SDK v7 wraps connection failures so the code can live two
+ * or three levels below the top-level error, with no top-level `code`.
+ */
+function findCauseCode(error, codes, depth = 0) {
+  if (!error || depth > 6) return '';
+  const code = String(error.code || '');
+  if (codes.has(code)) return code;
+  return findCauseCode(error.cause, codes, depth + 1);
+}
+
+function causeDetail(error) {
+  const messages = [];
+  for (let current = error?.cause, depth = 0; current && depth < 4; current = current.cause, depth += 1) {
+    if (typeof current.message === 'string' && current.message) messages.push(current.message);
+  }
+  return messages.join(' / ').slice(0, 300);
+}
+
+function retryAfterMs(error) {
+  const headers = error?.headers ?? error?.response?.headers;
+  let value = null;
+  if (headers?.get) value = headers.get('retry-after');
+  else if (headers && typeof headers === 'object') value = headers['retry-after'];
+  if (!value) return null;
+  const seconds = Number(value);
+  if (Number.isFinite(seconds)) return Math.max(0, seconds * 1000);
+  const date = Date.parse(value);
+  if (Number.isFinite(date)) return Math.max(0, date - Date.now());
+  return null;
+}
+
+/** Map raw SDK/network failures onto actionable ProviderError instances. */
+export function mapProviderError(error, {
+  model,
+  providerName = 'NVIDIA',
+  apiKeyPath = 'providers.nvidia.apiKey',
+  docsUrl = null,
+} = {}) {
+  if (error instanceof ProviderError) return error;
+  if (error instanceof CancelledError) return error;
+  if (error?.name === 'AbortError' || error?.code === 'ABORT_ERR') return new CancelledError('Request cancelled.');
+
+  const message = redact(String(error?.message || error || 'Unknown provider error')).slice(0, 800);
+  const status = statusOf(error);
+  const code = String(error?.code || '');
+  const errorName = String(error?.constructor?.name || error?.name || '');
+  const timeoutCode = findCauseCode(error, TIMEOUT_CODES);
+  const networkCode = findCauseCode(error, NETWORK_CODES);
+  const lower = message.toLowerCase();
+  const modelLabel = model ? `"${model}"` : 'the configured model';
+  const where = docsUrl ? ` Verify access on ${docsUrl}.` : '';
+
+  if (errorName === 'APIConnectionTimeoutError' || timeoutCode || /timed? ?out/.test(lower)) {
+    return new ProviderError(`The ${providerName} request timed out. ${message}`, {
+      code: 'PROVIDER_TIMEOUT',
+      retryable: true,
+      status,
+      cause: error,
+    });
+  }
+  if (errorName === 'APIConnectionError' || networkCode) {
+    const detail = causeDetail(error);
+    return new ProviderError(
+      `Cannot reach the ${providerName} API (${networkCode || 'connection error'}). ${message}${detail ? ` ${detail}` : ''}`,
+      { code: 'PROVIDER_NETWORK', retryable: true, status, cause: error },
+    );
+  }
+  if (status === 401) {
+    return new ProviderError(
+      `${providerName} API authentication failed (401). The key is invalid, revoked, or not loaded. Check ${apiKeyPath} in config.json and run /doctor.`,
+      { code: 'AUTH_INVALID', status },
+    );
+  }
+  if (status === 403) {
+    return new ProviderError(
+      `${providerName} API returned 403 Forbidden. Your key may not have access to ${modelLabel} or the endpoint.${where} ${message}`,
+      { code: 'AUTH_FORBIDDEN', status },
+    );
+  }
+  if (status === 404) {
+    return new ProviderError(
+      `${providerName} API returned 404. ${modelLabel} or the endpoint path may be wrong. Check the model id and base URL (expected .../v1). ${message}`,
+      { code: 'MODEL_NOT_FOUND', status },
+    );
+  }
+  if (status === 429) {
+    return new ProviderError(
+      `${providerName} API rate limit reached (429). Provider-side quotas may apply; wait and retry, or reduce request size. ${message}`,
+      { code: 'RATE_LIMIT', status, retryable: true },
+    );
+  }
+  if (status !== null && status >= 500) {
+    return new ProviderError(`${providerName} API server error (${status}). This is usually transient. ${message}`, {
+      code: 'SERVER_ERROR',
+      status,
+      retryable: true,
+    });
+  }
+  if (status === 400 || status === 422) {
+    if (/context|too many tokens|maximum.*token|token limit|input length/.test(lower)) {
+      return new ProviderError(
+        `The request exceeded the model context window. Start a new session or reduce context. ${message}`,
+        { code: 'CONTEXT_LENGTH', status },
+      );
+    }
+    // The provider validates every `function.arguments` string it receives.
+    // A single truncated tool call poisons the whole conversation, so this is
+    // recoverable: the history can be repaired and the request replayed.
+    if (/function\.arguments|must be a valid JSON object string|tool_calls\[/.test(lower)) {
+      return new ProviderError(
+        `The provider rejected a tool call in the conversation history because its arguments were not valid JSON. The history has been repaired; the request is being retried. ${message}`,
+        { code: 'INVALID_TOOL_ARGUMENTS', status, retryable: true },
+      );
+    }
+    if (/tool|function/.test(lower) && /not (support|available)|unsupported|invalid/.test(lower)) {
+      return new ProviderError(
+        `${modelLabel} may not support OpenAI-style tool calling on this endpoint. Try a model with tool-call support or use a smaller task. ${message}`,
+        { code: 'TOOLS_UNSUPPORTED', status },
+      );
+    }
+    return new ProviderError(`${providerName} API rejected the request (${status}). ${message}`, { code: 'BAD_REQUEST', status });
+  }
+  return new ProviderError(`Provider request failed: ${message}`, { code: 'PROVIDER_ERROR', status, retryable: false, cause: error });
+}
+
+/** Provider-exposed reasoning text (NVIDIA NIM uses `reasoning_content`). */
+export function extractReasoning(message = {}) {
+  for (const key of ['reasoning_content', 'reasoning']) {
+    const value = message?.[key];
+    if (typeof value === 'string' && value) return value;
+  }
+  return '';
+}
+
+export function normalizeResponse(response, { durationMs = null } = {}) {
+  const choice = response?.choices?.[0];
+  if (!choice) {
+    throw new ProviderError('Provider returned no choices (empty response body).', {
+      code: 'EMPTY_RESPONSE',
+      retryable: true,
+    });
+  }
+  const message = choice.message || {};
+  const toolCalls = (message.tool_calls || []).map((call, index) => ({
+    id: call.id || `call_synthetic_${index}_${Date.now()}`,
+    name: call.function?.name || '',
+    argumentsText: typeof call.function?.arguments === 'string' ? call.function.arguments : JSON.stringify(call.function?.arguments ?? {}),
+    type: call.type || 'function',
+  })).filter((call) => call.name);
+
+  const usage = response.usage || {};
+  return {
+    id: response.id || null,
+    model: response.model || null,
+    content: typeof message.content === 'string' ? message.content : '',
+    reasoning: extractReasoning(message),
+    toolCalls,
+    finishReason: choice.finish_reason || null,
+    usage: {
+      prompt: usage.prompt_tokens ?? 0,
+      completion: usage.completion_tokens ?? 0,
+      total: usage.total_tokens ?? (usage.prompt_tokens ?? 0) + (usage.completion_tokens ?? 0),
+    },
+    durationMs,
+  };
+}
+
+/**
+ * Accumulate OpenAI-compatible stream chunks. Tool-call names/arguments can be
+ * split across chunks and are keyed by their `index`.
+ */
+export function createStreamAccumulator() {
+  return {
+    id: null,
+    model: null,
+    content: '',
+    reasoning: '',
+    finishReason: null,
+    usage: { prompt: 0, completion: 0, total: 0 },
+    toolCalls: new Map(),
+    choiceChunks: 0,
+    emitted: false,
+  };
+}
+
+export function absorbStreamChunk(acc, chunk, { onReasoning = null, onContent = null } = {}) {
+  if (chunk?.id) acc.id = chunk.id;
+  if (chunk?.model) acc.model = chunk.model;
+  if (chunk?.usage) {
+    acc.usage.prompt = chunk.usage.prompt_tokens ?? acc.usage.prompt;
+    acc.usage.completion = chunk.usage.completion_tokens ?? acc.usage.completion;
+    acc.usage.total = chunk.usage.total_tokens ?? acc.usage.total;
+  }
+  const choice = chunk?.choices?.[0];
+  if (!choice) return;
+  acc.choiceChunks += 1;
+  if (choice.finish_reason) acc.finishReason = choice.finish_reason;
+  const delta = choice.delta || {};
+  const reasoningDelta = typeof delta.reasoning_content === 'string'
+    ? delta.reasoning_content
+    : typeof delta.reasoning === 'string'
+      ? delta.reasoning
+      : '';
+  if (reasoningDelta) {
+    acc.reasoning += reasoningDelta;
+    acc.emitted = true;
+    onReasoning?.(reasoningDelta);
+  }
+  if (typeof delta.content === 'string' && delta.content) {
+    acc.content += delta.content;
+    acc.emitted = true;
+    onContent?.(delta.content);
+  }
+  for (const call of delta.tool_calls || []) {
+    const index = Number.isInteger(call.index) ? call.index : acc.toolCalls.size;
+    const entry = acc.toolCalls.get(index) || { id: null, name: '', argumentsText: '', type: 'function' };
+    if (call.id) entry.id = call.id;
+    if (call.type) entry.type = call.type;
+    if (call.function?.name) entry.name += call.function.name;
+    if (typeof call.function?.arguments === 'string') entry.argumentsText += call.function.arguments;
+    else if (call.function?.arguments) entry.argumentsText += JSON.stringify(call.function.arguments);
+    acc.toolCalls.set(index, entry);
+  }
+}
+
+export function finalizeStream(acc, { durationMs = null } = {}) {
+  if (acc.choiceChunks === 0) {
+    throw new ProviderError('Provider stream ended without any choices.', { code: 'EMPTY_RESPONSE', retryable: true });
+  }
+  const toolCalls = [...acc.toolCalls.entries()]
+    .sort(([a], [b]) => a - b)
+    .map(([index, call]) => ({
+      id: call.id || `call_synthetic_${index}_${Date.now()}`,
+      name: call.name,
+      argumentsText: call.argumentsText || '{}',
+      type: call.type || 'function',
+    }))
+    .filter((call) => call.name);
+  return {
+    id: acc.id,
+    model: acc.model,
+    content: acc.content,
+    reasoning: acc.reasoning,
+    toolCalls,
+    finishReason: acc.finishReason,
+    usage: acc.usage,
+    durationMs,
+  };
+}
+
+/**
+ * Client for any OpenAI-compatible Chat Completions endpoint.
+ *
+ * The class name is historical (NVIDIA was the first provider); every provider
+ * in the catalog uses it with its own descriptor. `providerName` and
+ * `apiKeyPath` only affect user-facing error text — credentials are never
+ * printed, and a provider is never called without its own key.
+ */
+export class NvidiaProvider {
+  constructor(config, { client = null, logger = null, fetchImpl = null, sleepImpl = null } = {}) {
+    if (!config?.apiKey) {
+      const path = config?.apiKeyPath || `providers.${config?.providerId || 'nvidia'}.apiKey`;
+      throw new ProviderError(
+        `Provider ${config?.providerId || 'nvidia'} has no API key. Set ${path} in config.json (\`ryuu --settings\`) before using it.`,
+        { code: 'AUTH_MISSING' },
+      );
+    }
+    this.config = config;
+    this.logger = logger;
+    this.sleep = sleepImpl || abortableSleep;
+    this.fetchImpl = fetchImpl;
+    // An injected client is used verbatim (tests); otherwise the SDK client is
+    // rebuilt whenever the selection changes.
+    this.externalClient = Boolean(client);
+    this.client = client || this.#createClient();
+  }
+
+  #createClient() {
+    return new OpenAI({
+      apiKey: this.config.apiKey,
+      baseURL: this.config.baseURL,
+      timeout: this.config.requestTimeoutMs,
+      maxRetries: 0,
+      ...(this.fetchImpl ? { fetch: this.fetchImpl } : {}),
+    });
+  }
+
+  get model() {
+    return this.config.model;
+  }
+
+  /** Declared provider capabilities; unknown means "assume the usual". */
+  get capabilities() {
+    return this.config.capabilities || null;
+  }
+
+  /**
+   * Extra top-level fields to merge into the request body before it is sent.
+   *
+   * The OpenAI Chat Completions protocol defines no vendor fields, so the
+   * builder comes from `vendor-extras.js`, where a provider module registers
+   * itself by id (River's `reasoning_effort`, for example). Registration is by
+   * id because one instance is reused across `/provider` switches.
+   */
+  requestExtras(context = {}) {
+    return applyRequestExtras(this.config, { ...context, logger: this.logger });
+  }
+
+  get providerId() {
+    return this.config.providerId || null;
+  }
+
+  get providerName() {
+    return this.config.providerName || 'NVIDIA';
+  }
+
+  get apiKeyPath() {
+    return this.config.apiKeyPath || `providers.${this.providerId || 'nvidia'}.apiKey`;
+  }
+
+  /**
+   * Repoint this provider at another provider or model without rebuilding the
+   * agent (conversation memory and the worklog stay intact).
+   */
+  updateSelection({ providerId, providerName, apiKeyPath, docsUrl = null, baseURL, apiKey, model, capabilities = null, ...vendor } = {}) {
+    if (!apiKey) {
+      const path = apiKeyPath || `providers.${providerId || 'unknown'}.apiKey`;
+      throw new ProviderError(
+        `Provider ${providerId || 'unknown'} has no API key. Set ${path} in config.json before switching.`,
+        { code: 'AUTH_MISSING' },
+      );
+    }
+    if (model && (!MODEL_ID_PATTERN.test(model) || model.length > MAX_MODEL_ID_LENGTH)) {
+      throw new ProviderError(`Invalid model id: ${JSON.stringify(model)}`, { code: 'INVALID_MODEL' });
+    }
+    // Mutate in place: the app shares this object as its effective config.
+    Object.assign(this.config, {
+      providerId: providerId || this.config.providerId,
+      providerName: providerName || this.config.providerName,
+      apiKeyPath: apiKeyPath || this.config.apiKeyPath,
+      docsUrl,
+      baseURL: baseURL || this.config.baseURL,
+      apiKey,
+      model: model || this.config.model,
+      capabilities: capabilities || this.config.capabilities || null,
+    });
+    // Vendor request fields (River's `reasoningEffort`, a pass-through
+    // `extraBody`, ...) belong to the provider that is now active. The values
+    // of the provider left behind are dropped, so a request never carries a
+    // field meant for another endpoint.
+    for (const field of PROVIDER_REQUEST_FIELDS) {
+      delete this.config[field];
+      if (vendor[field] !== undefined) this.config[field] = vendor[field];
+    }
+    if (!this.externalClient) this.client = this.#createClient();
+    return this;
+  }
+
+  /**
+   * Send a chat completion request with bounded exponential backoff.
+   * Never logs the API key or request headers.
+   */
+  async chat({
+    messages,
+    tools = [],
+    signal = null,
+    maxTokens = null,
+    stream = false,
+    onReasoning = null,
+    onContent = null,
+    onFirstToken = null,
+    stallTimeoutMs = 0,
+    timeoutMs = null,
+  } = {}) {
+    if (!Array.isArray(messages) || messages.length === 0) {
+      throw new ProviderError('Cannot call the provider without messages.', { code: 'BAD_REQUEST' });
+    }
+    // Validate the selection before any bytes leave the machine.
+    const selectedModel = String(this.config.model || '');
+    if (!selectedModel || !MODEL_ID_PATTERN.test(selectedModel) || selectedModel.length > MAX_MODEL_ID_LENGTH) {
+      throw new ProviderError(
+        `Invalid model id for provider ${this.config.providerId || 'unknown'}: ${JSON.stringify(this.config.model)}. Pick one with /model (or set the provider's model variable).`,
+        { code: 'INVALID_MODEL' },
+      );
+    }
+    const capabilities = this.config.capabilities || null;
+    const supportsTools = capabilities?.tools !== false;
+    const streaming = Boolean(stream) && capabilities?.streaming !== false;
+    if (tools.length > 0 && !supportsTools) {
+      this.logger?.warn('provider does not support tool calling; tools are not sent', {
+        provider: this.config.providerId,
+      });
+    }
+    // Never let a broken tool-call payload reach the provider: one invalid
+    // `function.arguments` string fails the whole request with a 400.
+    const sanitized = sanitizeMessagesForRequest(messages);
+    if (sanitized.repaired > 0) {
+      this.logger?.warn('repaired invalid tool-call arguments before sending', { count: sanitized.repaired });
+    }
+    const body = {
+      model: selectedModel,
+      messages: sanitized.messages,
+      max_tokens: Math.max(64, Math.min(maxTokens ?? this.config.maxOutputTokens, 32768)),
+      stream: streaming,
+    };
+    if (tools.length > 0 && supportsTools) {
+      body.tools = tools;
+      body.tool_choice = 'auto';
+    }
+    // Vendor extensions (e.g. River's `reasoning_effort`, which turns thinking
+    // on for models whose default is off). The base client adds nothing; a
+    // provider-specific subclass overrides requestExtras().
+    const extras = this.requestExtras({ capabilities, streaming });
+    if (extras && typeof extras === 'object') Object.assign(body, extras);
+
+    const startedAt = Date.now();
+    const effectiveTimeout = Math.max(1000, Number(timeoutMs ?? this.config.requestTimeoutMs) || 1000);
+    let attempt = 0;
+    for (;;) {
+      if (signal?.aborted) throw new CancelledError();
+      let emitted = false;
+      let streamedReasoning = '';
+      // Per-attempt stream state. Declared outside the inner try so the catch
+      // can tell a provider stall apart from a cancellation or a real error.
+      let controller = null;
+      let stalled = false;
+      let stallTimer = null;
+      let detachAbort = () => {};
+      const clearStall = () => {
+        if (stallTimer) {
+          clearTimeout(stallTimer);
+          stallTimer = null;
+        }
+      };
+      try {
+        let normalized;
+        if (streaming) {
+          const request = {
+            ...body,
+            stream: true,
+            stream_options: { include_usage: true },
+          };
+          // A stream can stay open indefinitely: the request timeout only
+          // covers the request, not a server that dribbles bytes forever. This
+          // controller adds a per-chunk idle watchdog on top of the timeout.
+          controller = new AbortController();
+          if (signal) {
+            if (signal.aborted) controller.abort();
+            else {
+              const onAbort = () => controller.abort();
+              signal.addEventListener('abort', onAbort, { once: true });
+              detachAbort = () => signal.removeEventListener('abort', onAbort);
+            }
+          }
+          const idleMs = Math.max(0, Number(stallTimeoutMs) || 0);
+          const armStall = () => {
+            if (!idleMs) return;
+            clearStall();
+            stallTimer = setTimeout(() => {
+              stalled = true;
+              controller.abort();
+            }, idleMs);
+            stallTimer.unref?.();
+          };
+          armStall();
+          const iterator = await this.client.chat.completions.create(request, {
+            signal: controller.signal,
+            timeout: effectiveTimeout,
+          });
+          const acc = createStreamAccumulator();
+          let firstTokenSeen = false;
+          for await (const chunk of iterator) {
+            if (signal?.aborted) throw new CancelledError();
+            armStall();
+            absorbStreamChunk(acc, chunk, {
+              onReasoning: (text) => {
+                emitted = true;
+                streamedReasoning += text;
+                onReasoning?.(text);
+              },
+              onContent: (text) => {
+                emitted = true;
+                onContent?.(text);
+              },
+            });
+            if (!firstTokenSeen && (acc.emitted || acc.toolCalls.size > 0)) {
+              firstTokenSeen = true;
+              onFirstToken?.({ latencyMs: Date.now() - startedAt });
+            }
+          }
+          normalized = finalizeStream(acc, { durationMs: Date.now() - startedAt });
+        } else {
+          const response = await this.client.chat.completions.create(body, {
+            signal: signal || undefined,
+            timeout: effectiveTimeout,
+          });
+          normalized = normalizeResponse(response, { durationMs: Date.now() - startedAt });
+        }
+        this.logger?.debug('provider response', {
+          provider: this.config.providerId,
+          model: normalized.model,
+          stream: streaming,
+          toolCalls: normalized.toolCalls.length,
+          reasoningChars: (normalized.reasoning || '').length,
+          finishReason: normalized.finishReason,
+          usage: normalized.usage,
+          durationMs: normalized.durationMs,
+          attempt,
+        });
+        return normalized;
+      } catch (raw) {
+        clearStall();
+        detachAbort();
+        if (signal?.aborted) throw new CancelledError();
+        if (stalled) {
+          const seconds = Math.round((Number(stallTimeoutMs) || 0) / 1000);
+          this.logger?.warn('provider stream stalled', { attempt, idleSeconds: seconds, emitted });
+          throw new ProviderError(
+            `The response stream stopped sending data for ${seconds}s and was aborted. The model or endpoint is not responding; nothing was executed.`,
+            { code: 'STREAM_STALLED', retryable: !emitted },
+          );
+        }
+        const error = mapProviderError(raw, {
+          model: this.config.model,
+          providerName: this.providerName,
+          apiKeyPath: this.apiKeyPath,
+          docsUrl: this.config.docsUrl || null,
+        });
+        if (emitted) {
+          // Retrying after partial output would duplicate reasoning/content.
+          this.logger?.warn('provider stream interrupted', { code: error.code, attempt });
+          throw new ProviderError(
+            `The response stream was interrupted after partial output (${streamedReasoning.length} reasoning chars received); it was not retried to avoid duplicate text. ${error.message}`,
+            { code: 'STREAM_INTERRUPTED', retryable: false, status: error.status, cause: error },
+          );
+        }
+        const canRetry = error.retryable && attempt < this.config.maxRetries;
+        if (!canRetry) {
+          this.logger?.warn('provider request failed', { code: error.code, status: error.status, attempt });
+          throw error;
+        }
+        const backoff = Math.min(RETRY_MAX_MS, RETRY_BASE_MS * 2 ** attempt);
+        const jittered = Math.round(backoff * (0.7 + Math.random() * 0.6));
+        const delay = Math.max(jittered, retryAfterMs(raw) ?? 0);
+        this.logger?.warn('provider request retry', {
+          code: error.code,
+          status: error.status,
+          attempt: attempt + 1,
+          maxRetries: this.config.maxRetries,
+          delayMs: delay,
+        });
+        await this.sleep(delay, signal);
+        attempt += 1;
+      }
+    }
+  }
+
+  /** Safe connectivity check. Returns structured status; never prints the key. */
+  async ping({ signal = null } = {}) {
+    const startedAt = Date.now();
+    try {
+      const response = await this.chat({
+        messages: [
+          { role: 'system', content: 'You are a connectivity probe. Reply with exactly: OK' },
+          { role: 'user', content: 'Reply with OK.' },
+        ],
+        maxTokens: 16,
+        signal,
+      });
+      return {
+        ok: true,
+        model: response.model || this.config.model,
+        latencyMs: Date.now() - startedAt,
+        sample: (response.content || '').trim().slice(0, 40),
+        usage: response.usage,
+      };
+    } catch (error) {
+      return {
+        ok: false,
+        model: this.config.model,
+        latencyMs: Date.now() - startedAt,
+        error: { code: error.code || 'PROVIDER_ERROR', message: error.message, status: error.status ?? null },
+      };
+    }
+  }
+}
+
+export function createProvider(config, deps = {}) {
+  return new NvidiaProvider(config, deps);
+}
+
+/**
+ * Guarantee that every tool call in the request carries a valid JSON object
+ * string.
+ *
+ * The provider validates the whole conversation on every request, so one
+ * truncated payload anywhere in the history makes the request fail with a 400
+ * and the task dies. Repairing the payload here is the last line of defence:
+ * the agent already sanitizes what it stores, but a session resumed from disk
+ * or a plugin-provided message could still carry a broken payload.
+ */
+export function sanitizeMessagesForRequest(messages = []) {
+  let repaired = 0;
+  const sanitized = messages.map((message) => {
+    if (!message || !Array.isArray(message.tool_calls) || message.tool_calls.length === 0) return message;
+    const toolCalls = message.tool_calls.map((call) => {
+      const fn = call?.function || {};
+      const original = typeof fn.arguments === 'string' ? fn.arguments : JSON.stringify(fn.arguments ?? {});
+      const fixed = toToolArgumentsString(original);
+      if (fixed !== original) repaired += 1;
+      return { ...call, function: { ...fn, arguments: fixed } };
+    });
+    return { ...message, tool_calls: toolCalls };
+  });
+  return { messages: sanitized, repaired };
+}