Add research overlap nudges

backend/src/agent/podman.ts

@@ -2,6 +2,7 @@ import { RoomEvent, type Room } from '@livekit/rtc-node';
 import type { EngineerContext, Collision, Intervention, DataMessage } from '@podman/shared';
 import { analyzeFrame } from '../vision/gemini.js';
 import { detectCollisions } from '../collision/detector.js';
+import { detectResearchOverlaps } from '../collision/research.js';
 import { getGithubState } from '../github/client.js';
 import {
   recordObservation,
@@ -97,7 +98,10 @@ export class PodMan {
     }
 
     const github = await getGithubState(); // cached
-    const collisions = detectCollisions([...this.contexts.values()], github, gitStates);
+    const contexts = [...this.contexts.values()];
+    const fileCollisions = detectCollisions(contexts, github, gitStates);
+    const researchCollisions = await detectResearchOverlaps(contexts, gitStates);
+    const collisions = [...fileCollisions, ...researchCollisions];
 
     // Re-arm: any conflict we previously voiced that is no longer present has
     // resolved, so allow it to alert again if it recurs.
@@ -109,7 +113,9 @@ export class PodMan {
     // Capture git ground-truth overlap now, while engineer_states are fresh, so
     // the outcome-time verifier never depends on a stale sidecar or a late click.
     for (const collision of collisions) {
-      collision.gitOverlap = engineersOverlapOnFile(collision, gitStates);
+      if (collision.overlapKind !== 'research') {
+        collision.gitOverlap = engineersOverlapOnFile(collision, gitStates);
+      }
     }
 
     for (const collision of collisions) await this.handle(collision);
@@ -122,7 +128,7 @@ export class PodMan {
    * basename.
    */
   private conflictKey(collision: Collision): string {
-    return comparableBasename(collision.file);
+    return `${collision.overlapKind ?? 'file'}:${comparableBasename(collision.file)}`;
   }
 
   private async handle(collision: Collision): Promise<void> {
@@ -143,8 +149,41 @@ export class PodMan {
     this.activeConflicts.add(key); // claim now we're alerting; re-armed in onScreenFrame on resolution
     await recordCollision(collision);
     const action = preferredAction(collision, prior);
-    const names = collision.engineers.join(' + ');
     const shortFile = collision.file.split('/').pop() ?? collision.file;
+    const isResearchOverlap = collision.overlapKind === 'research';
+
+    if (isResearchOverlap) {
+      const researcher = collision.researcher ?? collision.engineers[1] ?? 'A teammate';
+      const editor = collision.editor ?? collision.engineers[0] ?? 'a teammate';
+      const topic = collision.researchTopic ?? 'the same area';
+      const source = collision.researchSource ? ` (${collision.researchSource})` : '';
+      const message = `🤝 ${researcher} is researching ${topic}${source} while ${editor} edits ${shortFile} — sync up before duplicating effort.`;
+      const voiceLine = `${researcher} is researching ${topic} while ${editor} works on ${shortFile}. Worth a quick sync.`;
+      const intervention: Intervention = {
+        id: `int_${Date.now()}`,
+        collisionId: collision.id,
+        podId: this.podId,
+        kind: 'card',
+        message,
+        suggestedAction: {
+          kind: 'ping_teammate',
+          params: {
+            file: collision.file,
+            summary: message,
+            engineers: collision.engineers,
+            researchTopic: collision.researchTopic,
+            researchSource: collision.researchSource,
+          },
+        },
+        status: 'pending',
+        createdAt: new Date().toISOString(),
+      };
+      await recordIntervention(intervention);
+      await publishHermesIntervention(this.room, collision, intervention, voiceLine);
+      return;
+    }
+
+    const names = collision.engineers.join(' + ');
 
     // Terse, demo-centered alert — short and direct, not chatty AI prose.
     const message =

backend/src/collision/research.ts

@@ -0,0 +1,141 @@
+import type { Collision, EngineerContext } from '@podman/shared';
+import { env } from '../env.js';
+import type { GitState } from '../memory/db.js';
+import { semanticSimilarity } from '../memory/vectors.js';
+
+export interface ResearchOpts {
+  similarity?: (a: string, b: string) => Promise<number | null>;
+  threshold?: number;
+}
+
+interface EditorFile {
+  engineerId: string;
+  file: string;
+  symbol?: string;
+  activity?: string;
+}
+
+interface Candidate {
+  collision: Collision;
+  score: number;
+}
+
+function stripGitPrefix(raw: string): string {
+  return raw.trim().replace(/^(\?\?|[MADRCU!]{1,2})\s+/, '');
+}
+
+function fileStem(raw: string): string {
+  const base = stripGitPrefix(raw).split(/[\\/]/).pop()?.trim().toLowerCase() ?? '';
+  return base.replace(/\.[^.]+$/, '');
+}
+
+function words(raw: string): string[] {
+  return raw
+    .toLowerCase()
+    .replace(/[^a-z0-9]+/g, ' ')
+    .split(/\s+/)
+    .filter((word) => word.length >= 3);
+}
+
+function uniqueTokens(raw: string): Set<string> {
+  const tokens = new Set(words(raw));
+  for (const word of [...tokens]) {
+    if (word.endsWith('kit')) tokens.add(word.replace(/kit$/, ''));
+    if (word.endsWith('s')) tokens.add(word.slice(0, -1));
+  }
+  return tokens;
+}
+
+function fallbackMatches(researchText: string, fileText: string): boolean {
+  const research = uniqueTokens(researchText);
+  const file = uniqueTokens(fileText);
+  for (const token of file) {
+    if (research.has(token)) return true;
+  }
+  return false;
+}
+
+function collectEditorFiles(
+  contexts: EngineerContext[],
+  gitStates: Map<string, GitState> | undefined,
+): EditorFile[] {
+  const files = new Map<string, EditorFile>();
+  const add = (editor: EditorFile): void => {
+    const stem = fileStem(editor.file);
+    if (!stem) return;
+    files.set(`${editor.engineerId}:${stripGitPrefix(editor.file)}`, editor);
+  };
+
+  for (const [engineerId, git] of gitStates ?? []) {
+    for (const changed of git.changedFiles) {
+      add({ engineerId, file: changed });
+    }
+  }
+
+  for (const context of contexts) {
+    if (context.mode !== 'research' && context.currentFile) {
+      add({
+        engineerId: context.engineerId,
+        file: context.currentFile,
+        symbol: context.currentSymbol,
+        activity: context.activity,
+      });
+    }
+  }
+
+  return [...files.values()];
+}
+
+export async function detectResearchOverlaps(
+  contexts: EngineerContext[],
+  gitStates: Map<string, GitState> | undefined,
+  opts: ResearchOpts = {},
+): Promise<Collision[]> {
+  const similarity = opts.similarity ?? semanticSimilarity;
+  const threshold = opts.threshold ?? env.RESEARCH_OVERLAP_THRESHOLD;
+  const researchers = contexts.filter((c) => c.mode === 'research' && c.researchTopic);
+  const editorFiles = collectEditorFiles(contexts, gitStates);
+  const bestByResearcher = new Map<string, Candidate>();
+
+  for (const researcher of researchers) {
+    const topic = researcher.researchTopic?.trim();
+    if (!topic) continue;
+    const source = researcher.researchSource?.trim();
+    const researchText = [topic, source].filter(Boolean).join(' ');
+
+    for (const editor of editorFiles) {
+      if (editor.engineerId === researcher.engineerId) continue;
+
+      const stem = fileStem(editor.file);
+      if (!stem) continue;
+      const fileText = [stem, editor.symbol, editor.activity].filter(Boolean).join(' ');
+      const score = await similarity(researchText, fileText);
+      const matched = score === null ? fallbackMatches(researchText, fileText) : score >= threshold;
+      if (!matched) continue;
+
+      const rank = score ?? 1;
+      const existing = bestByResearcher.get(researcher.engineerId);
+      if (existing && existing.score >= rank) continue;
+
+      bestByResearcher.set(researcher.engineerId, {
+        score: rank,
+        collision: {
+          id: `col_research_${stem}_${Date.now()}`,
+          podId: researcher.podId,
+          file: editor.file,
+          symbol: editor.symbol,
+          engineers: [editor.engineerId, researcher.engineerId],
+          severity: 'warn',
+          overlapKind: 'research',
+          researchTopic: topic,
+          ...(source ? { researchSource: source } : {}),
+          researcher: researcher.engineerId,
+          editor: editor.engineerId,
+          detectedAt: new Date().toISOString(),
+        },
+      });
+    }
+  }
+
+  return [...bestByResearcher.values()].map((candidate) => candidate.collision);
+}

backend/src/env.ts

@@ -22,7 +22,10 @@ export const env = {
   LIVEKIT_API_KEY: req('LIVEKIT_API_KEY'),
   LIVEKIT_API_SECRET: req('LIVEKIT_API_SECRET'),
   LIVEKIT_AGENT_NAME: opt('LIVEKIT_AGENT_NAME'),
-  LIVEKIT_CONVERSATION_AGENT_NAME: opt('LIVEKIT_CONVERSATION_AGENT_NAME', 'podman-live-conversation'),
+  LIVEKIT_CONVERSATION_AGENT_NAME: opt(
+    'LIVEKIT_CONVERSATION_AGENT_NAME',
+    'podman-live-conversation',
+  ),
   // Gemini
   GEMINI_API_KEY: reqAny('GEMINI_API_KEY', ['GOOGLE_API_KEY', 'GOOGLE_GENERATIVE_AI_API_KEY']),
   GEMINI_VISION_MODEL: opt('GEMINI_VISION_MODEL', 'gemini-2.0-flash'),
@@ -40,6 +43,7 @@ export const env = {
   // Server
   PORT: Number(opt('PORT', '8787')),
   NUDGE_COOLDOWN_MS: Number(opt('NUDGE_COOLDOWN_MS', '180000')),
+  RESEARCH_OVERLAP_THRESHOLD: Number(opt('RESEARCH_OVERLAP_THRESHOLD', '0.6')),
   INTERNAL_AGENT_TOKEN: opt('INTERNAL_AGENT_TOKEN'),
 } as const;
 

backend/src/memory/vectors.ts

@@ -49,7 +49,7 @@ function memoryText(collision: Collision): string {
     .join('\n');
 }
 
-function cosine(a: number[], b: number[]): number {
+export function cosine(a: number[], b: number[]): number {
   const n = Math.min(a.length, b.length);
   let dot = 0;
   let aNorm = 0;
@@ -69,6 +69,12 @@ async function embed(text: string, inputType: 'document' | 'query'): Promise<num
   return (await embedWithVoyage(text, inputType)) ?? embedWithGemini(text, inputType);
 }
 
+export async function semanticSimilarity(a: string, b: string): Promise<number | null> {
+  const [va, vb] = await Promise.all([embed(a, 'query'), embed(b, 'document')]);
+  if (!va || !vb) return null;
+  return cosine(va, vb);
+}
+
 async function embedWithVoyage(
   text: string,
   inputType: 'document' | 'query',

backend/src/vision/gemini.ts

@@ -7,6 +7,11 @@ const ai = new GoogleGenAI({ apiKey: env.GEMINI_API_KEY });
 const SCHEMA = {
   type: Type.OBJECT,
   properties: {
+    mode: {
+      type: Type.STRING,
+      description:
+        'editing when an IDE/editor/terminal is primary; research for browser docs/SDK pages',
+    },
     currentFile: {
       type: Type.STRING,
       description: 'open file path if visible, e.g. src/auth/session.ts',
@@ -20,13 +25,24 @@ const SCHEMA = {
       type: Type.BOOLEAN,
       description: 'dirty git gutter / modified markers visible',
     },
+    researchTopic: {
+      type: Type.STRING,
+      description: 'topic being researched when mode is research, e.g. LiveKit agents setup',
+    },
+    researchSource: {
+      type: Type.STRING,
+      description: 'source domain when mode is research, e.g. docs.livekit.io',
+    },
     confidence: { type: Type.NUMBER, description: '0..1 confidence in this read' },
   },
   propertyOrdering: [
+    'mode',
     'currentFile',
     'currentSymbol',
     'activity',
     'hasUnpushedChanges',
+    'researchTopic',
+    'researchSource',
     'confidence',
   ],
 } as const;
@@ -43,7 +59,10 @@ export async function analyzeFrame(
         role: 'user',
         parts: [
           {
-            text: "You are PodMan watching an engineer's screen. Identify what file/symbol they are working on and whether there are uncommitted edits. JSON only.",
+            text:
+              "You are PodMan watching an engineer's screen. Return JSON only. " +
+              "If the primary window is an IDE/editor/terminal, set mode='editing' and identify the file, symbol, activity, and whether uncommitted edits are visible. " +
+              "If the primary window is a browser/docs/SDK/reference page, set mode='research', leave currentFile empty unless a file path is clearly visible, and extract researchTopic plus researchSource as the source domain.",
           },
           { inlineData: { mimeType: 'image/jpeg', data: jpeg.toString('base64') } },
         ],
@@ -63,6 +82,9 @@ export async function analyzeFrame(
     currentFile: parsed.currentFile,
     currentSymbol: parsed.currentSymbol,
     activity: parsed.activity,
+    mode: parsed.mode,
+    researchTopic: parsed.researchTopic,
+    researchSource: parsed.researchSource,
     hasUnpushedChanges: parsed.hasUnpushedChanges,
     confidence: parsed.confidence ?? 0.5,
     observedAt: new Date().toISOString(),