@geravant/sinain 1.0.18 → 1.1.0

package/sinain-core/src/agent/analyzer.ts

@@ -0,0 +1,366 @@
+import type { AgentConfig, AgentResult, ContextWindow, RecorderStatus, RecordCommand } from "../types.js";
+import { normalizeAppName } from "./context-window.js";
+import { log, error } from "../log.js";
+import { levelFor, applyLevel } from "../privacy/index.js";
+
+const TAG = "agent";
+
+/**
+ * Model-specific timeouts in milliseconds.
+ * Only increases timeouts for slow models to avoid false timeouts.
+ * Default 15s is kept for fast models.
+ */
+const MODEL_TIMEOUTS: Record<string, number> = {
+  'google/gemini-2.5-flash-lite': 15000,
+  'google/gemini-2.5-flash': 15000,
+  'google/gemini-2.0-flash': 15000,
+  'anthropic/claude-3-opus': 60000,
+  'anthropic/claude-3.5-sonnet': 30000,
+  'anthropic/claude-3-haiku': 15000,
+  'default': 15000,
+};
+
+/** Get timeout for a specific model. */
+function getModelTimeout(model: string): number {
+  return MODEL_TIMEOUTS[model] ?? MODEL_TIMEOUTS['default'];
+}
+
+/** Message part for multimodal API calls. */
+type ContentPart =
+  | { type: "text"; text: string }
+  | { type: "image_url"; image_url: { url: string; detail: "low" } };
+
+/**
+ * Build recorder status section for the prompt.
+ */
+function buildRecorderSection(status: RecorderStatus | null): string {
+  if (!status) return "";
+  if (!status.recording) return "\nRecorder: idle (not recording)";
+
+  const label = status.label ? ` "${status.label}"` : "";
+  const durationSec = Math.round(status.durationMs / 1000);
+  return `\nRecorder: RECORDING${label} (${durationSec}s, ${status.segments} segments)`;
+}
+
+/**
+ * Static system prompt (cached as module constant).
+ * Contains rules, output format, and behavioral instructions.
+ * Previously allocated ~3KB per tick; now zero-allocation.
+ */
+const SYSTEM_PROMPT = `You are an AI monitoring a user's screen and audio in real-time.
+You produce outputs as JSON.
+
+Respond ONLY with valid JSON. No markdown, no code fences, no explanation.
+Your entire response must be parseable by JSON.parse().
+
+{"hud":"...","digest":"...","record":{"command":"start"|"stop","label":"..."},"task":"..."}
+
+Output fields:
+- "hud" (required): max 60 words describing what user is doing NOW
+- "digest" (required): 5-8 sentences with detailed activity description
+- "record" (optional): control recording — {"command":"start","label":"Meeting name"} or {"command":"stop"}
+- "task" (optional): natural language instruction to spawn a background task
+
+When to use "record":
+- START when user begins a meeting, call, lecture, YouTube video, or important audio content
+- STOP when the content ends or user navigates away
+- Provide descriptive labels like "Team standup", "Client call", "YouTube: [video title from OCR]"
+- For YouTube/video content: extract video title from screen OCR for the label
+
+When to use "task":
+- User explicitly asks for research, lookup, or action
+- Something needs external search or processing that isn't a real-time response
+- Example: "Search for React 19 migration guide", "Find docs for this API"
+
+When to spawn "task" for video content:
+- If user watches a YouTube video for 2+ minutes AND no task has been spawned for this video yet, spawn: "Summarize YouTube video: [title or URL from OCR]"
+- ONLY spawn ONCE per video - do not repeat spawn for the same video in subsequent ticks
+- Extract video title or URL from screen OCR to include in the task
+
+When to spawn "task" for coding problems:
+- If user is actively working on a coding problem/challenge for 1+ minutes:
+  - Spawn: "Solve coding problem: [problem description/title from OCR]"
+- This includes LeetCode, HackerRank, interviews, coding assessments, or any visible coding challenge
+- Look for problem signals: "Input:", "Output:", "Example", "Constraints:", problem titles, test cases
+- Include as much context as possible from the screen OCR (problem description, examples, constraints)
+- ONLY spawn ONCE per distinct problem - do not repeat for the same problem
+- The spawned task should provide a complete solution with code and explanation
+
+Audio sources: [\ud83d\udd0a]=system/speaker audio, [\ud83c\udf99]=microphone (user's voice).
+Treat [\ud83c\udf99] as direct user speech. Treat [\ud83d\udd0a] as external audio.
+
+Rules:
+- "hud" is for a minimal overlay display. Example: "Editing hud-relay.mjs in IDEA"
+- "digest" is for an AI assistant to understand the full situation and offer help.
+- If nothing is happening, hud="Idle" and digest explains what was last seen.
+- Include specific filenames, URLs, error messages, UI text from OCR in digest.
+- Do NOT suggest actions in digest — just describe the situation factually.
+- Only include "record" or "task" when genuinely appropriate — most responses won't have them.
+- CRITICAL: Output ONLY the JSON object, nothing else.`;
+
+/**
+ * Build the dynamic user prompt (changes every tick).
+ * Contains the current context data: screen OCR, audio transcripts, app state.
+ */
+function buildUserPrompt(ctx: ContextWindow, recorderStatus: RecorderStatus | null = null): string {
+  const now = Date.now();
+
+  // Privacy gating: check levels for openrouter destination
+  let screenLines: string;
+  try {
+    const ocrLevel = levelFor("screen_ocr", "openrouter");
+    const titlesLevel = levelFor("window_titles", "openrouter");
+    screenLines = ctx.screen
+      .map(e => {
+        const app = normalizeAppName(e.meta.app);
+        const ago = Math.round((now - (e.ts || now)) / 1000);
+        const rawOcr = e.ocr ? e.ocr.replace(/\n/g, " ").slice(0, ctx.preset.maxOcrChars) : "(no text)";
+        const ocr = e.ocr ? applyLevel(rawOcr, ocrLevel, "ocr") : "(no text)";
+        const title = e.meta.windowTitle ? applyLevel(e.meta.windowTitle, titlesLevel, "titles") : "";
+        const titlePart = title ? ` [${title}]` : "";
+        return `[${ago}s ago] [${app}]${titlePart} ${ocr || "(no text)"}`;
+      })
+      .join("\n");
+  } catch {
+    // Privacy not yet initialized — use full text
+    screenLines = ctx.screen
+      .map(e => {
+        const app = normalizeAppName(e.meta.app);
+        const ago = Math.round((now - (e.ts || now)) / 1000);
+        const ocr = e.ocr ? e.ocr.replace(/\n/g, " ").slice(0, ctx.preset.maxOcrChars) : "(no text)";
+        return `[${ago}s ago] [${app}] ${ocr}`;
+      })
+      .join("\n");
+  }
+
+  let audioLines: string;
+  try {
+    const audioLevel = levelFor("audio_transcript", "openrouter");
+    audioLines = ctx.audio
+      .map(e => {
+        const ago = Math.round((now - (e.ts || now)) / 1000);
+        const text = applyLevel(e.text.slice(0, ctx.preset.maxTranscriptChars), audioLevel, "audio");
+        return `[${ago}s ago] ${text}`;
+      })
+      .join("\n");
+  } catch {
+    audioLines = ctx.audio
+      .map(e => {
+        const ago = Math.round((now - (e.ts || now)) / 1000);
+        return `[${ago}s ago] ${e.text.slice(0, ctx.preset.maxTranscriptChars)}`;
+      })
+      .join("\n");
+  }
+
+  const appSwitches = ctx.appHistory
+    .map(a => normalizeAppName(a.app))
+    .join(" \u2192 ");
+
+  const recorderSection = buildRecorderSection(recorderStatus);
+
+  // Gate images based on privacy level
+  let imagesForPrompt = ctx.images;
+  try {
+    const imgLevel = levelFor("screen_images", "openrouter");
+    if (imgLevel === "none") {
+      imagesForPrompt = [];
+    }
+  } catch { /* privacy not initialized, keep images */ }
+
+  const hasImages = imagesForPrompt && imagesForPrompt.length > 0;
+  const imageNote = hasImages ? `\n\nScreen screenshots (${imagesForPrompt!.length}) are attached below.` : "";
+
+  return `Active app: ${normalizeAppName(ctx.currentApp)}
+App history: ${appSwitches || "(none)"}${recorderSection}
+
+Screen (OCR text, newest first):
+${screenLines || "(no screen data)"}
+
+Audio transcript (newest first, \ud83d\udd0a=system, \ud83c\udf99=mic):
+${audioLines || "(silence)"}${imageNote}`;
+}
+
+/**
+ * Parse record command from LLM response.
+ */
+function parseRecord(parsed: any): RecordCommand | undefined {
+  if (!parsed.record || typeof parsed.record !== "object") return undefined;
+  const cmd = parsed.record.command;
+  if (cmd !== "start" && cmd !== "stop") return undefined;
+  return {
+    command: cmd,
+    label: typeof parsed.record.label === "string" ? parsed.record.label : undefined,
+  };
+}
+
+/**
+ * Parse task from LLM response.
+ */
+function parseTask(parsed: any): string | undefined {
+  if (typeof parsed.task !== "string" || !parsed.task.trim()) return undefined;
+  return parsed.task.trim();
+}
+
+/**
+ * Call the LLM (OpenRouter) to analyze the context window.
+ * Supports model chain: primary + fallbacks.
+ * When images are present, auto-upgrades to the vision model.
+ */
+export async function analyzeContext(
+  contextWindow: ContextWindow,
+  config: AgentConfig,
+  recorderStatus: RecorderStatus | null = null,
+  traitSystemPrompt?: string,
+): Promise<AgentResult> {
+  const userPrompt = buildUserPrompt(contextWindow, recorderStatus);
+  // Apply privacy gating for images sent to OpenRouter
+  let images = contextWindow.images || [];
+  try {
+    const imgLevel = levelFor("screen_images", "openrouter");
+    if (imgLevel === "none") {
+      images = [];
+    }
+  } catch { /* privacy not initialized, keep images */ }
+  const systemPrompt = traitSystemPrompt ?? SYSTEM_PROMPT;
+
+  const models = [config.model, ...config.fallbackModels];
+
+  // Auto-upgrade: use vision model when images are present
+  if (images.length > 0 && config.visionModel) {
+    // Insert vision model at the front if not already there
+    if (!models.includes(config.visionModel)) {
+      models.unshift(config.visionModel);
+    }
+  }
+
+  let lastError: Error | null = null;
+
+  for (const model of models) {
+    try {
+      return await callModel(systemPrompt, userPrompt, images, model, config);
+    } catch (err: any) {
+      lastError = err;
+      log(TAG, `model ${model} failed: ${err.message || err}, trying next...`);
+    }
+  }
+
+  throw lastError || new Error("all models failed");
+}
+
+async function callModel(
+  systemPrompt: string,
+  userPrompt: string,
+  images: ContextWindow["images"],
+  model: string,
+  config: AgentConfig,
+): Promise<AgentResult> {
+  const start = Date.now();
+  const controller = new AbortController();
+  const timeoutMs = getModelTimeout(model);
+  const timeout = setTimeout(() => controller.abort(), timeoutMs);
+
+  try {
+    // Build user message content: text + optional images
+    let userContent: string | ContentPart[];
+    if (images && images.length > 0) {
+      const parts: ContentPart[] = [{ type: "text", text: userPrompt }];
+      for (const img of images) {
+        parts.push({
+          type: "image_url",
+          image_url: {
+            url: `data:image/jpeg;base64,${img.data}`,
+            detail: "low",
+          },
+        });
+      }
+      userContent = parts;
+    } else {
+      userContent = userPrompt;
+    }
+
+    const imageCount = images?.length || 0;
+
+    const response = await fetch("https://openrouter.ai/api/v1/chat/completions", {
+      method: "POST",
+      headers: {
+        "Authorization": `Bearer ${config.openrouterApiKey}`,
+        "Content-Type": "application/json",
+      },
+      body: JSON.stringify({
+        model,
+        messages: [
+          { role: "system", content: systemPrompt },
+          { role: "user", content: userContent },
+        ],
+        max_tokens: config.maxTokens,
+        temperature: config.temperature,
+      }),
+      signal: controller.signal,
+    });
+
+    if (!response.ok) {
+      const body = await response.text().catch(() => "");
+      throw new Error(`HTTP ${response.status}: ${body.slice(0, 200)}`);
+    }
+
+    const data = await response.json() as any;
+    const latencyMs = Date.now() - start;
+    const raw = data.choices?.[0]?.message?.content?.trim() || "";
+
+    if (imageCount > 0) {
+      log(TAG, `multimodal call: model=${model}, images=${imageCount}`);
+    }
+
+    // Parse JSON response — try direct parse, then extract embedded JSON, then fallback
+    try {
+      const jsonStr = raw.replace(/^```\w*\s*\n?/, "").replace(/\n?\s*```\s*$/, "").trim();
+      const parsed = JSON.parse(jsonStr);
+      return {
+        hud: parsed.hud || "\u2014",
+        digest: parsed.digest || "\u2014",
+        record: parseRecord(parsed),
+        task: parseTask(parsed),
+        latencyMs,
+        tokensIn: data.usage?.prompt_tokens || 0,
+        tokensOut: data.usage?.completion_tokens || 0,
+        model,
+        parsedOk: true,
+      };
+    } catch {
+      // Second chance: extract embedded JSON object
+      const match = raw.match(/\{[\s\S]*\}/);
+      if (match) {
+        try {
+          const parsed = JSON.parse(match[0]);
+          if (parsed.hud) {
+            return {
+              hud: parsed.hud,
+              digest: parsed.digest || "\u2014",
+              record: parseRecord(parsed),
+              task: parseTask(parsed),
+              latencyMs,
+              tokensIn: data.usage?.prompt_tokens || 0,
+              tokensOut: data.usage?.completion_tokens || 0,
+              model,
+              parsedOk: true,
+            };
+          }
+        } catch { /* fall through */ }
+      }
+
+      // Final fallback: use raw text
+      log(TAG, `JSON parse failed (model=${model}), raw: "${raw.slice(0, 120)}"`);
+      return {
+        hud: raw.slice(0, 160) || "\u2014",
+        digest: raw || "\u2014",
+        latencyMs,
+        tokensIn: data.usage?.prompt_tokens || 0,
+        tokensOut: data.usage?.completion_tokens || 0,
+        model,
+        parsedOk: false,
+      };
+    }
+  } finally {
+    clearTimeout(timeout);
+  }
+}

package/sinain-core/src/agent/context-window.ts

@@ -0,0 +1,172 @@
+import type { FeedBuffer } from "../buffers/feed-buffer.js";
+import type { SenseBuffer } from "../buffers/sense-buffer.js";
+import type { ContextWindow, ContextRichness, RichnessPreset } from "../types.js";
+
+/**
+ * Track recently sent image hashes to avoid sending duplicates to vision model.
+ * Uses simple content hash: length + first 1000 chars.
+ */
+const recentlySentImageHashes = new Set<string>();
+const MAX_IMAGE_HASH_CACHE = 20;
+let imageHashCacheOrder: string[] = [];
+
+/**
+ * Richness presets — control how much context goes into agent analysis and escalation.
+ *
+ * lean:     For selective mode. Minimal context, fast + cheap.
+ * standard: For focus mode. Moderate detail.
+ * rich:     Full context. Maximum detail for thorough agent analysis.
+ */
+export const RICHNESS_PRESETS: Record<ContextRichness, RichnessPreset> = {
+  lean:     { maxScreenEvents: 10, maxAudioEntries: 5,  maxOcrChars: 400,  maxTranscriptChars: 400,  maxImages: 0 },
+  standard: { maxScreenEvents: 20, maxAudioEntries: 10, maxOcrChars: 1000, maxTranscriptChars: 800,  maxImages: 1 },
+  rich:     { maxScreenEvents: 50, maxAudioEntries: 30, maxOcrChars: 4000, maxTranscriptChars: 2000, maxImages: 2 },
+} as const;
+
+/** App name normalization map (consistent display names). */
+const APP_NAMES: Record<string, string> = {
+  "idea": "IntelliJ IDEA",
+  "code": "VS Code",
+  "code - insiders": "VS Code Insiders",
+  "webstorm": "WebStorm",
+  "pycharm": "PyCharm",
+  "datagrip": "DataGrip",
+  "google chrome": "Chrome",
+  "firefox": "Firefox",
+  "safari": "Safari",
+  "telegram lite": "Telegram",
+  "telegram": "Telegram",
+  "iterm2": "iTerm",
+  "terminal": "Terminal",
+  "finder": "Finder",
+  "audio midi setup": "Audio MIDI Setup",
+};
+
+export function normalizeAppName(app: string): string {
+  return APP_NAMES[app.toLowerCase()] || app;
+}
+
+/** Short app names for overlay feed (compact display). */
+const APP_SHORT_NAMES: Record<string, string> = {
+  "IntelliJ IDEA": "IDEA",
+  "IntelliJ IDEA Ultimate": "IDEA",
+  "idea": "IDEA",
+  "Google Chrome": "Chrome",
+  "Visual Studio Code": "Code",
+  "Code - Insiders": "Code",
+  "iTerm2": "iTerm",
+  "Terminal": "Term",
+  "Telegram": "TG",
+  "WebStorm": "WS",
+  "PyCharm": "PyCharm",
+  "DataGrip": "DG",
+  "Finder": "Finder",
+};
+
+export function shortAppName(app: string): string {
+  if (APP_SHORT_NAMES[app]) return APP_SHORT_NAMES[app];
+  const lower = app.toLowerCase();
+  for (const [key, value] of Object.entries(APP_SHORT_NAMES)) {
+    if (key.toLowerCase() === lower) return value;
+  }
+  return app;
+}
+
+/**
+ * Build a unified context window from in-process buffers.
+ * Replaces both relay's buildContextWindow() and bridge's ContextManager.
+ *
+ * No HTTP round-trips — direct access to feed and sense buffers.
+ */
+export function buildContextWindow(
+  feedBuffer: FeedBuffer,
+  senseBuffer: SenseBuffer,
+  richness: ContextRichness = "standard",
+  maxAgeMs = 120_000,
+): ContextWindow {
+  const preset = RICHNESS_PRESETS[richness];
+  const cutoff = Date.now() - maxAgeMs;
+
+  // Audio: extract transcript text from feed items tagged as 'audio'
+  const audioItems = feedBuffer.queryBySource("audio", cutoff)
+    .slice(-preset.maxAudioEntries);
+
+  // Screen: get sense events within the time window
+  const screenEvents = senseBuffer.queryByTime(cutoff);
+
+  // Current app
+  const latestSense = screenEvents[screenEvents.length - 1];
+  const currentApp = latestSense?.meta.app || "unknown";
+
+  // Deduplicate OCR text (consecutive identical OCR is noise)
+  const dedupedScreen = [];
+  let lastOcr = "";
+  for (const e of screenEvents) {
+    if (e.ocr && e.ocr !== lastOcr) {
+      dedupedScreen.push(e);
+      lastOcr = e.ocr;
+    } else if (!e.ocr && e.type === "context") {
+      dedupedScreen.push(e);
+    }
+  }
+
+  // App transition timeline
+  const appHistory = senseBuffer.appHistory(cutoff);
+
+  // Limit to preset maximums, newest first for recency weighting
+  const sortedScreen = dedupedScreen.slice(-preset.maxScreenEvents).reverse();
+
+  // Compute newest event timestamp
+  const newestEventTs = Math.max(
+    audioItems[audioItems.length - 1]?.ts || 0,
+    screenEvents[screenEvents.length - 1]?.ts || 0
+  );
+
+  // Extract recent images for multimodal vision (with content-based deduplication)
+  let images: { data: string; app: string; ts: number }[] | undefined;
+  if (preset.maxImages > 0) {
+    const rawImages = senseBuffer.recentImages(preset.maxImages);
+    images = [];
+
+    for (const e of rawImages) {
+      if (!e.imageData) continue;
+
+      // Simple content hash: length + first 1000 chars
+      const hash = `${e.imageData.length}:${e.imageData.slice(0, 1000)}`;
+
+      // Skip if recently sent to vision model (avoid duplicate API calls)
+      if (recentlySentImageHashes.has(hash)) {
+        continue;
+      }
+
+      // Track this hash (LRU eviction)
+      recentlySentImageHashes.add(hash);
+      imageHashCacheOrder.push(hash);
+      while (imageHashCacheOrder.length > MAX_IMAGE_HASH_CACHE) {
+        const oldest = imageHashCacheOrder.shift()!;
+        recentlySentImageHashes.delete(oldest);
+      }
+
+      images.push({
+        data: e.imageData,
+        app: e.meta.app || "unknown",
+        ts: e.ts,
+      });
+    }
+
+    if (images.length === 0) images = undefined;
+  }
+
+  return {
+    audio: audioItems,
+    screen: sortedScreen,
+    images,
+    currentApp,
+    appHistory,
+    audioCount: audioItems.length,
+    screenCount: screenEvents.length,
+    windowMs: maxAgeMs,
+    newestEventTs,
+    preset,
+  };
+}