@pellux/goodvibes-tui 0.18.11 → 0.18.13

package/src/providers/discovered-factory.ts

@@ -1,61 +0,0 @@
-import type { DiscoveredServer } from '@pellux/goodvibes-sdk/platform/discovery/scanner';
-import type { LLMProvider } from '@pellux/goodvibes-sdk/platform/providers/interface';
-import { getDiscoveredTraits } from '@pellux/goodvibes-sdk/platform/providers/discovered-traits';
-import { LocalAIProvider, TGIProvider, VLLMProvider } from '@pellux/goodvibes-sdk/platform/providers/discovered-compat';
-import { LlamaCppProvider } from '@pellux/goodvibes-sdk/platform/providers/llama-cpp';
-import { LMStudioProvider } from '@pellux/goodvibes-sdk/platform/providers/lm-studio';
-import { OpenAICompatProvider } from '@pellux/goodvibes-sdk/platform/providers/openai-compat';
-import { OllamaProvider } from '@pellux/goodvibes-sdk/platform/providers/ollama';
-
-export function createDiscoveredProvider(server: DiscoveredServer): LLMProvider {
-  const traits = getDiscoveredTraits(server.serverType);
-  const common = {
-    name: server.name,
-    baseURL: server.baseURL,
-    apiKey: '',
-    defaultModel: server.models[0]!,
-    models: server.models,
-    capabilities: traits.providerCapabilities,
-  };
-
-  switch (traits.adapter) {
-    case 'lm-studio':
-      return new LMStudioProvider(common);
-    case 'ollama':
-      return new OllamaProvider({
-        ...common,
-        reasoningFormat: traits.reasoningFormat,
-      });
-    case 'vllm':
-      return new VLLMProvider({
-        ...common,
-        reasoningFormat: traits.reasoningFormat,
-      });
-    case 'llamacpp':
-      return new LlamaCppProvider({
-        ...common,
-        reasoningFormat: traits.reasoningFormat,
-      });
-    case 'tgi':
-      return new TGIProvider({
-        ...common,
-        reasoningFormat: traits.reasoningFormat,
-      });
-    case 'localai':
-      return new LocalAIProvider({
-        ...common,
-        reasoningFormat: traits.reasoningFormat,
-      });
-    default:
-      return new OpenAICompatProvider({
-        ...common,
-        reasoningFormat: traits.reasoningFormat,
-      });
-  }
-}
-
-export function getDiscoveredReasoningFormat(
-  serverType: DiscoveredServer['serverType'],
-): 'llamacpp' | 'none' {
-  return getDiscoveredTraits(serverType).reasoningFormat;
-}

package/src/providers/discovered-traits.ts

@@ -1,138 +0,0 @@
-import type { ServerType } from '@pellux/goodvibes-sdk/platform/discovery/scanner';
-import type { ProviderCapability } from '@pellux/goodvibes-sdk/platform/providers/capabilities';
-
-export interface DiscoveredServerTraits {
-  readonly adapter:
-    | 'lm-studio'
-    | 'ollama'
-    | 'vllm'
-    | 'llamacpp'
-    | 'tgi'
-    | 'localai'
-    | 'compat';
-  readonly reasoningFormat: 'llamacpp' | 'none';
-  readonly providerCapabilities?: Partial<ProviderCapability>;
-  readonly modelCapabilities: {
-    toolCalling: boolean;
-    codeEditing: boolean;
-    reasoning: boolean;
-    multimodal: boolean;
-  };
-  readonly reasoningEffort?: string[];
-}
-
-const DEFAULT_MODEL_CAPABILITIES = {
-  toolCalling: true,
-  codeEditing: true,
-  reasoning: false,
-  multimodal: false,
-} as const;
-
-const HIGH_TIMEOUT_MS = 300_000;
-
-export function getDiscoveredTraits(serverType: ServerType): DiscoveredServerTraits {
-  switch (serverType) {
-    case 'lm-studio':
-      return {
-        adapter: 'lm-studio',
-        reasoningFormat: 'none',
-        providerCapabilities: {
-          streaming: true,
-          toolCalling: true,
-          parallelTools: true,
-          jsonMode: false,
-          reasoningControls: true,
-          timeoutMs: HIGH_TIMEOUT_MS,
-        },
-        modelCapabilities: {
-          ...DEFAULT_MODEL_CAPABILITIES,
-          reasoning: true,
-        },
-        reasoningEffort: ['instant', 'low', 'medium', 'high'],
-      };
-    case 'ollama':
-      return {
-        adapter: 'ollama',
-        reasoningFormat: 'llamacpp',
-        providerCapabilities: {
-          streaming: true,
-          toolCalling: true,
-          parallelTools: false,
-          jsonMode: true,
-          reasoningControls: true,
-          timeoutMs: HIGH_TIMEOUT_MS,
-        },
-        modelCapabilities: {
-          ...DEFAULT_MODEL_CAPABILITIES,
-          reasoning: true,
-        },
-        reasoningEffort: ['instant', 'low', 'medium', 'high'],
-      };
-    case 'vllm':
-      return {
-        adapter: 'vllm',
-        reasoningFormat: 'none',
-        providerCapabilities: {
-          streaming: true,
-          toolCalling: true,
-          parallelTools: false,
-          jsonMode: true,
-          reasoningControls: false,
-          timeoutMs: HIGH_TIMEOUT_MS,
-        },
-        modelCapabilities: DEFAULT_MODEL_CAPABILITIES,
-      };
-    case 'llamacpp':
-      return {
-        adapter: 'llamacpp',
-        reasoningFormat: 'llamacpp',
-        providerCapabilities: {
-          streaming: true,
-          toolCalling: true,
-          parallelTools: false,
-          jsonMode: true,
-          reasoningControls: true,
-          timeoutMs: HIGH_TIMEOUT_MS,
-        },
-        modelCapabilities: {
-          ...DEFAULT_MODEL_CAPABILITIES,
-          reasoning: true,
-        },
-        reasoningEffort: ['instant', 'low', 'medium', 'high'],
-      };
-    case 'tgi':
-      return {
-        adapter: 'tgi',
-        reasoningFormat: 'none',
-        providerCapabilities: {
-          streaming: true,
-          toolCalling: true,
-          parallelTools: false,
-          jsonMode: true,
-          reasoningControls: false,
-          timeoutMs: HIGH_TIMEOUT_MS,
-        },
-        modelCapabilities: DEFAULT_MODEL_CAPABILITIES,
-      };
-    case 'localai':
-      return {
-        adapter: 'localai',
-        reasoningFormat: 'none',
-        providerCapabilities: {
-          streaming: true,
-          toolCalling: true,
-          parallelTools: false,
-          jsonMode: true,
-          reasoningControls: false,
-          timeoutMs: HIGH_TIMEOUT_MS,
-        },
-        modelCapabilities: DEFAULT_MODEL_CAPABILITIES,
-      };
-    default:
-      return {
-        adapter: 'compat',
-        reasoningFormat: 'none',
-        modelCapabilities: DEFAULT_MODEL_CAPABILITIES,
-      };
-  }
-}

package/src/providers/gemini.ts

@@ -1,462 +0,0 @@
-import type {
-  LLMProvider,
-  ChatRequest,
-  ChatResponse,
-  ProviderEmbeddingRequest,
-  ProviderEmbeddingResult,
-  ProviderRuntimeMetadata,
-  ProviderRuntimeMetadataDeps,
-} from '@pellux/goodvibes-sdk/platform/providers/interface';
-import { REASONING_BUDGET_MAP } from '@pellux/goodvibes-sdk/platform/providers/interface';
-import { ProviderError } from '@pellux/goodvibes-sdk/platform/types/errors';
-import { withRetry } from '@pellux/goodvibes-sdk/platform/utils/retry';
-import { logger } from '@pellux/goodvibes-sdk/platform/utils/logger';
-import {
-  toGeminiFunctionDeclarations,
-  toGeminiContents,
-  fromGeminiParts,
-} from '@pellux/goodvibes-sdk/platform/providers/tool-formats';
-import type { GeminiPart } from '@pellux/goodvibes-sdk/platform/providers/tool-formats';
-import type { CacheHitTracker } from '@pellux/goodvibes-sdk/platform/providers/cache-strategy';
-import { summarizeError, toProviderError } from '@pellux/goodvibes-sdk/platform/utils/error-display';
-
-const GEMINI_API_BASE = 'https://generativelanguage.googleapis.com/v1beta';
-const GEMINI_CACHE_TTL_SECONDS = 3600;
-const NOOP_CACHE_HIT_TRACKER: Pick<CacheHitTracker, 'recordTurn'> = {
-  recordTurn: () => {},
-};
-
-interface GeminiCandidate {
-  content: { parts: GeminiPart[]; role: string };
-  finishReason: string;
-}
-
-interface GeminiResponseBody {
-  candidates?: GeminiCandidate[];
-  usageMetadata?: {
-    promptTokenCount?: number;
-    candidatesTokenCount?: number;
-    cachedContentTokenCount?: number;
-  };
-}
-
-/**
- * GeminiProvider — calls the Gemini generateContent API directly via fetch.
- * Tools are `functionDeclarations` inside a `tools` array.
- * Tool calls come as `functionCall` parts; results as `functionResponse` parts.
- * Uses streamGenerateContent for real-time token delivery when onDelta is provided.
- */
-export class GeminiProvider implements LLMProvider {
-  readonly name = 'gemini';
-  /** Maps function call name → thoughtSignature for the current turn. */
-  private thoughtSignatures = new Map<string, string>();
-  readonly models: string[] = [];
-
-  private readonly apiKey: string;
-  private readonly embeddingModel = 'gemini-embedding-001';
-  private readonly cacheHitTracker: Pick<CacheHitTracker, 'recordTurn'>;
-
-  /** Active cached content resource name (e.g., "cachedContents/abc123") */
-  private cachedContentName: string | null = null;
-  /** Hash of the content that was cached (systemPrompt + tools + model) */
-  private cachedContentHash: string | null = null;
-  /** When the cache expires (epoch ms) */
-  private cachedContentExpiry: number = 0;
-  /** Hashes known to be below the 32K cache minimum — skip API call */
-  private uncacheableHashes = new Set<string>();
-
-  constructor(apiKey: string, cacheHitTracker: Pick<CacheHitTracker, 'recordTurn'> = NOOP_CACHE_HIT_TRACKER) {
-    this.apiKey = apiKey;
-    this.cacheHitTracker = cacheHitTracker;
-  }
-
-  private computeCacheHash(
-    systemPrompt: string | undefined,
-    tools: import('./interface.ts').ChatRequest['tools'],
-    model: string,
-  ): string {
-    const raw = (systemPrompt ?? '') + JSON.stringify(tools ?? []) + model;
-    const hasher = new Bun.CryptoHasher('sha256');
-    hasher.update(raw);
-    return hasher.digest('hex');
-  }
-
-  private async ensureCachedContent(
-    systemPrompt: string | undefined,
-    tools: import('./interface.ts').ChatRequest['tools'],
-    model: string,
-  ): Promise<string | null> {
-    // Skip if no system prompt and no tools
-    if (!systemPrompt && (!tools || tools.length === 0)) return null;
-
-    const hash = this.computeCacheHash(systemPrompt, tools, model);
-
-    // Skip if previously determined to be below 32K threshold
-    if (this.uncacheableHashes.has(hash)) return null;
-
-    // Reuse existing cache if hash matches and not expired (with 60s buffer)
-    if (
-      this.cachedContentName &&
-      this.cachedContentHash === hash &&
-      this.cachedContentExpiry > Date.now() + 60_000
-    ) {
-      return this.cachedContentName;
-    }
-
-    // Estimate tokens — skip if below 28K (conservative buffer below 32K minimum)
-    const estimatedChars = (systemPrompt?.length ?? 0) + JSON.stringify(tools ?? []).length;
-    if (estimatedChars / 3 < 28_000) {
-      if (this.uncacheableHashes.size >= 50) this.uncacheableHashes.clear();
-      this.uncacheableHashes.add(hash);
-      logger.debug('[Gemini] Content below 32K cache threshold, skipping cache', {
-        estimatedTokens: Math.round(estimatedChars / 3),
-      });
-      return null;
-    }
-
-    // Delete old cache if hash changed (fire-and-forget)
-    if (this.cachedContentName && this.cachedContentHash !== hash) {
-      const oldName = this.cachedContentName;
-      fetch(`${GEMINI_API_BASE}/${oldName}`, {
-        method: 'DELETE',
-        headers: { 'x-goog-api-key': this.apiKey },
-      }).catch(err => logger.warn('[Gemini] Failed to delete old cache', { error: summarizeError(err) }));
-    }
-
-    // Create new cached content
-    try {
-      const cacheBody: Record<string, unknown> = {
-        model: `models/${model}`,
-        ttl: `${GEMINI_CACHE_TTL_SECONDS}s`,
-      };
-
-      if (systemPrompt) {
-        cacheBody['systemInstruction'] = { parts: [{ text: systemPrompt }] };
-      }
-
-      if (tools && tools.length > 0) {
-        cacheBody['tools'] = [{ functionDeclarations: toGeminiFunctionDeclarations(tools) }];
-      }
-
-      const res = await fetch(`${GEMINI_API_BASE}/cachedContents`, {
-        method: 'POST',
-        headers: {
-          'Content-Type': 'application/json',
-          'x-goog-api-key': this.apiKey,
-        },
-        body: JSON.stringify(cacheBody),
-      });
-
-      if (!res.ok) {
-        const text = await res.text().catch(() => '');
-        if (text.includes('too few tokens') || text.includes('minimum')) {
-          if (this.uncacheableHashes.size >= 50) this.uncacheableHashes.clear();
-          this.uncacheableHashes.add(hash);
-          logger.debug('[Gemini] Content below cache minimum', { status: res.status, error: text.slice(0, 200) });
-        } else {
-          logger.debug('[Gemini] Cache creation failed', { status: res.status, error: text.slice(0, 200) });
-        }
-        return null;
-      }
-
-      const data = await res.json() as { name: string; expireTime: string };
-      this.cachedContentName = data.name;
-      this.cachedContentHash = hash;
-      this.cachedContentExpiry = new Date(data.expireTime).getTime();
-
-      logger.info(`[Gemini] Created cache: ${data.name} (expires ${data.expireTime})`);
-      return data.name;
-    } catch (err) {
-      logger.debug('[Gemini] Cache creation error', { error: summarizeError(err) });
-      return null;
-    }
-  }
-
-  async chat(params: ChatRequest): Promise<ChatResponse> {
-    const { messages, tools, model, maxTokens, signal, systemPrompt, onDelta, reasoningEffort } = params;
-
-    return withRetry(async () => {
-      const { contents, systemInstruction } = toGeminiContents(messages, systemPrompt);
-
-      // Inject thoughtSignatures into both model functionCall parts and user functionResponse parts
-      // (Gemini thinking models require the signature on both sides of the round-trip)
-      for (const c of contents) {
-        for (const part of c.parts) {
-          const p = part as Record<string, unknown>;
-          if (p.functionCall) {
-            const fc = p.functionCall as { name: string };
-            const sig = this.thoughtSignatures.get(fc.name);
-            if (sig) p.thoughtSignature = sig;
-          }
-          if (p.functionResponse) {
-            const fr = p.functionResponse as { name: string };
-            const sig = this.thoughtSignatures.get(fr.name);
-            if (sig) p.thoughtSignature = sig;
-          }
-        }
-      }
-
-      const body: Record<string, unknown> = { contents };
-
-      const cachedName = await this.ensureCachedContent(systemPrompt, tools, model);
-
-      if (cachedName) {
-        // Cached content already contains systemInstruction and tools — do NOT resend them
-        body['cachedContent'] = cachedName;
-      } else {
-        if (systemInstruction) {
-          body['systemInstruction'] = systemInstruction;
-        }
-
-        if (tools && tools.length > 0) {
-          body['tools'] = [{
-            functionDeclarations: toGeminiFunctionDeclarations(tools),
-          }];
-        }
-      }
-
-      if (maxTokens) {
-        body['generationConfig'] = { maxOutputTokens: maxTokens };
-      }
-
-      if (reasoningEffort) {
-        const budget = REASONING_BUDGET_MAP[reasoningEffort];
-        if (budget !== undefined) {
-          body['generationConfig'] = {
-            ...(body['generationConfig'] as Record<string, unknown> ?? {}),
-            thinking_config: { thinking_budget: budget },
-          };
-        }
-      }
-
-      // Always use streaming endpoint; parse NDJSON chunks
-      const url = `${GEMINI_API_BASE}/models/${model}:streamGenerateContent?alt=sse`;
-
-      let res: Response;
-      try {
-        res = await fetch(url, {
-          method: 'POST',
-          headers: {
-            'Content-Type': 'application/json',
-            'x-goog-api-key': this.apiKey,
-          },
-          body: JSON.stringify(body),
-          signal,
-        });
-      } catch (err: unknown) {
-        throw toProviderError(err, {
-          provider: this.name,
-          operation: 'chat',
-          phase: 'request',
-        });
-      }
-
-      if (!res.ok) {
-        const text = await res.text().catch(() => 'unknown error');
-        throw new ProviderError(`Gemini API error ${res.status}: ${text}`, {
-          statusCode: res.status,
-          provider: this.name,
-          operation: 'chat',
-          phase: 'request',
-        });
-      }
-
-      // Accumulate state from streaming chunks
-      const allParts: GeminiPart[] = [];
-      let inputTokens = 0;
-      let outputTokens = 0;
-      let cacheReadTokens = 0;
-      let lastFinishReason = '';
-      let streamedText = '';
-
-      const reader = res.body?.getReader();
-      if (!reader) {
-        throw new ProviderError('Gemini chat returned no response body.', {
-          statusCode: 502,
-          provider: this.name,
-          operation: 'chat',
-          phase: 'response',
-        });
-      }
-
-      const decoder = new TextDecoder();
-      let buffer = '';
-
-      try {
-        while (true) {
-          const { done, value } = await reader.read();
-          if (done) break;
-
-          buffer += decoder.decode(value, { stream: true });
-          const lines = buffer.split('\n');
-          buffer = lines.pop() ?? '';
-
-          for (const line of lines) {
-            if (!line.startsWith('data: ')) continue;
-            const data = line.slice(6).trim();
-            if (!data || data === '[DONE]') continue;
-
-            let chunk: GeminiResponseBody;
-            try {
-              chunk = JSON.parse(data) as GeminiResponseBody;
-            } catch {
-              logger.debug('Gemini SSE: failed to parse JSON chunk', { data });
-              continue;
-            }
-
-            const candidate = chunk.candidates?.[0];
-            if (candidate) {
-              const parts = candidate.content?.parts ?? [];
-              for (const part of parts) {
-                allParts.push(part);
-                if (part.text && onDelta) {
-                  streamedText += part.text;
-                  onDelta({ content: part.text });
-                }
-                if (part.functionCall) {
-                  // Capture thoughtSignature if present (Gemini thinking models)
-                  if ((part as Record<string, unknown>).thoughtSignature) {
-                    this.thoughtSignatures.set(part.functionCall.name, (part as Record<string, unknown>).thoughtSignature as string);
-                  }
-                  if (onDelta) {
-                    onDelta({ toolCalls: [{ index: 0, name: part.functionCall.name, arguments: JSON.stringify(part.functionCall.args) }] });
-                  }
-                }
-              }
-              if (candidate.finishReason) {
-                lastFinishReason = candidate.finishReason;
-              }
-            }
-
-            if (chunk.usageMetadata) {
-              inputTokens = chunk.usageMetadata.promptTokenCount ?? inputTokens;
-              outputTokens = chunk.usageMetadata.candidatesTokenCount ?? outputTokens;
-              cacheReadTokens = chunk.usageMetadata.cachedContentTokenCount ?? cacheReadTokens;
-            }
-          }
-        }
-      } finally {
-        reader.releaseLock();
-      }
-
-      // Use streamed text directly if available (avoids re-parsing duplicated text parts)
-      const { text: parsedText, toolCalls } = fromGeminiParts(allParts);
-      // Prefer streamedText for content; fall back to parsed if no streaming happened
-      const text = streamedText || parsedText;
-
-      let stopReason: ChatResponse['stopReason'] = 'end';
-      if (lastFinishReason === 'MAX_TOKENS') stopReason = 'max_tokens';
-      else if (toolCalls.length > 0) stopReason = 'tool_use';
-
-      // Clear old signatures — new ones were captured from this response's functionCall parts
-      // (kept across calls within a tool-use loop, cleared when no new functionCalls arrive)
-      if (toolCalls.length === 0) {
-        this.thoughtSignatures.clear();
-      }
-
-      this.cacheHitTracker.recordTurn({
-        inputTokens,
-        cacheReadTokens,
-      });
-
-      return {
-        content: text,
-        toolCalls,
-        usage: {
-          inputTokens,
-          outputTokens,
-          ...(cacheReadTokens > 0 ? { cacheReadTokens } : {}),
-          // cacheWriteTokens is omitted: Gemini does not charge separately for cache writes
-        },
-        stopReason,
-      };
-    });
-  }
-
-  async embed(request: ProviderEmbeddingRequest): Promise<ProviderEmbeddingResult> {
-    const model = request.model ?? this.embeddingModel;
-    const body: Record<string, unknown> = {
-      content: { parts: [{ text: request.text }] },
-    };
-    if (request.dimensions) {
-      body['config'] = { outputDimensionality: request.dimensions };
-    }
-
-    let res: Response;
-    try {
-      res = await fetch(`${GEMINI_API_BASE}/models/${model}:embedContent`, {
-        method: 'POST',
-        headers: {
-          'Content-Type': 'application/json',
-          'x-goog-api-key': this.apiKey,
-        },
-        body: JSON.stringify(body),
-        signal: request.signal,
-      });
-    } catch (error: unknown) {
-      throw toProviderError(error, {
-        provider: this.name,
-        operation: 'embed',
-        phase: 'request',
-      });
-    }
-
-    if (!res.ok) {
-      const text = await res.text().catch(() => 'unknown error');
-      throw new ProviderError(`Gemini embeddings API error ${res.status}: ${text}`, {
-        statusCode: res.status,
-        provider: this.name,
-        operation: 'embed',
-        phase: 'request',
-      });
-    }
-
-    const data = await res.json() as { embedding?: { values?: number[] } };
-    const values = data.embedding?.values ?? [];
-    return {
-      vector: Float32Array.from(values),
-      dimensions: values.length,
-      modelId: model,
-      metadata: {
-        usage: request.usage,
-        provider: this.name,
-      },
-    };
-  }
-
-  async describeRuntime(deps: ProviderRuntimeMetadataDeps): Promise<ProviderRuntimeMetadata> {
-    const { buildStandardProviderAuthRoutes } = await import('./runtime-metadata.ts');
-    const authRoutes = await buildStandardProviderAuthRoutes({
-      providerId: 'gemini',
-      apiKeyEnvVars: ['GEMINI_API_KEY', 'GOOGLE_API_KEY', 'GOOGLE_GEMINI_API_KEY'],
-      serviceNames: ['gemini'],
-    }, deps);
-    return {
-      auth: {
-        mode: 'api-key',
-        configured: Boolean(this.apiKey),
-        detail: this.apiKey ? 'Gemini API key available' : 'Gemini API key is not configured',
-        envVars: ['GEMINI_API_KEY', 'GOOGLE_API_KEY', 'GOOGLE_GEMINI_API_KEY'],
-        routes: authRoutes,
-      },
-      models: {
-        models: this.models,
-        embeddingModel: this.embeddingModel,
-        embeddingDimensions: 384,
-      },
-      usage: {
-        streaming: true,
-        toolCalling: true,
-        parallelTools: true,
-        notes: ['Embeddings use Gemini embedContent with reduced output dimensionality when requested.'],
-      },
-      policy: {
-        local: false,
-        streamProtocol: 'gemini-sse',
-        reasoningMode: 'thinking_budget',
-        supportedReasoningEfforts: ['instant', 'low', 'medium', 'high'],
-        cacheStrategy: 'gemini-cached-content',
-      },
-    };
-  }
-}