PR5 v0.5.0 - researchDoc + meetingCapture PatternAdapters

package.json

@@ -1,6 +1,6 @@
 {
   "name": "hermes-orchestrator",
-  "version": "0.4.0",
+  "version": "0.5.0",
   "description": "Supervisor framework for AI agents - classify, spawn, watch, intervene, learn. The Hermes pattern (Claude CLI subprocess) generalised.",
   "type": "module",
   "main": "./dist/index.js",
@@ -25,6 +25,14 @@
     "./patterns/hermes-bug-fix": {
       "types": "./dist/patterns/hermes-bug-fix.d.ts",
       "import": "./dist/patterns/hermes-bug-fix.js"
+    },
+    "./patterns/research-doc": {
+      "types": "./dist/patterns/research-doc.d.ts",
+      "import": "./dist/patterns/research-doc.js"
+    },
+    "./patterns/meeting-capture": {
+      "types": "./dist/patterns/meeting-capture.d.ts",
+      "import": "./dist/patterns/meeting-capture.js"
     }
   },
   "files": ["dist", "README.md", "LICENSE", "docs"],

src/index.ts

@@ -15,6 +15,8 @@ export { HermesRunner, type HermesRunnerOptions } from './adapters/hermes-runner
 export { BonfireMemory, type BonfireMemoryOptions } from './adapters/bonfire-memory.js'
 export { FileMemory, type FileMemoryOptions } from './adapters/file-memory.js'
 export { hermesBugFix } from './patterns/hermes-bug-fix.js'
+export { researchDoc } from './patterns/research-doc.js'
+export { meetingCapture } from './patterns/meeting-capture.js'
 
 export type {
   Task,

src/patterns/meeting-capture.ts

@@ -0,0 +1,74 @@
+import type { PatternAdapter, Task, MemoryHit, RunnerInput } from '../types.js'
+
+const KEYWORDS = [
+  'meeting',
+  'transcribe',
+  'transcript',
+  'recap',
+  'process this call',
+  'process the call',
+  'extract todos',
+  'extract action items',
+  'voice memo',
+  'recording',
+  'standup',
+]
+
+const FILE_EXTENSIONS = ['.mp4', '.m4a', '.mov', '.mp3', '.wav', '.opus', '.flac']
+
+const SYSTEM_PROMPT = `You are a meeting-capture agent. Turn one recording or transcript into one durable recap.
+
+Workflow:
+1. Acquire the transcript: if a media file, transcribe it (local mlx-whisper preferred). If a paste, use it directly.
+2. Run multi-pass extraction (NOT a single monolithic prompt):
+   A. Metadata: date, duration, title, attendees, platform.
+   B. Decisions: explicit + verbatim-anchored. Carry a confidence (high/medium/low).
+   C. Actions: concrete follow-ups with one owner each. Confidence + due if stated.
+   D. Quotes: 3-8 load-bearing verbatim quotes.
+   E. Research seeds + memory updates. Cross-check existing entities BEFORE adding.
+3. Produce: a recap doc with the schema below, plus a separate transcript file.
+
+Rules:
+- Verbatim where possible. No paraphrasing of decisions.
+- Every decision + action carries a confidence field.
+- Ambiguous owner = "Both" + confidence: low, surface it for the operator.
+- Never invent dates. Relative ("by Thursday") -> absolute, anchored to the meeting date.
+- If an entity is already documented, LINK it. Do not re-introduce.
+
+Doc structure:
+- Frontmatter (date, attendees, project, doc-type: meeting-recap)
+- TL;DR (3-5 bullets)
+- Decisions (table with id, text, owner, confidence)
+- Actions (table with title, owner, due, category, confidence)
+- Key quotes (3-8)
+- Transcript link (separate file)
+- Next Actions (link to trackers + PRs)`
+
+export const meetingCapture: PatternAdapter = {
+  name: 'meeting-capture',
+  defaultRunner: 'hermes',
+  costCap: 3.0,
+
+  matches(task: Task): boolean {
+    const t = task.text.toLowerCase()
+    if (KEYWORDS.some((k) => t.includes(k))) return true
+    if (FILE_EXTENSIONS.some((ext) => t.includes(ext))) return true
+    return false
+  },
+
+  prepare(task: Task, memory: MemoryHit[]): RunnerInput {
+    const fewshot =
+      memory.length > 0
+        ? `\n\nPast meeting recaps for similar contexts (most relevant first):\n${memory
+            .map((m, i) => `${i + 1}. ${m.body}`)
+            .join('\n')}`
+        : ''
+    return {
+      prompt: task.text,
+      systemPrompt: SYSTEM_PROMPT + fewshot,
+      allowedTools: ['Bash', 'Read', 'Write', 'Edit', 'Grep', 'Glob'],
+      maxCostUsd: 3.0,
+      metadata: { pattern: 'meeting-capture' },
+    }
+  },
+}

src/patterns/research-doc.ts

@@ -0,0 +1,54 @@
+import type { PatternAdapter, Task, MemoryHit, RunnerInput } from '../types.js'
+
+const KEYWORDS = [
+  'research',
+  'investigate',
+  'audit',
+  'look into',
+  'find out about',
+  'survey',
+  'compare',
+  'evaluate',
+]
+
+const SYSTEM_PROMPT = `You are a research agent. Your job is to produce one durable research doc, not a chat answer.
+
+Workflow:
+1. Search the existing research library first (grep over README.md files) to avoid duplicating work and to find related docs.
+2. Fetch the actual source - climb the ladder WebFetch -> exa web_fetch -> Playwright -> Wayback. Do not write off a search snippet.
+3. Classify every source FULL / PARTIAL / FAILED. Escalate PARTIAL/FAILED through the full ladder before writing.
+4. Write the doc with required frontmatter: topic, type, status, last-validated, related-docs, original-query, tier.
+5. End with a Next Actions table linking to concrete todos / PRs / calendar items.
+
+Rules:
+- Be specific. Include at least 3 numbers (versions, prices, dates, counts).
+- No vague language: never use "consider", "it might be worth", "you could explore". State the decision.
+- Recommendations FIRST in a Key Decisions table at the top.
+- Cite every source URL. Mark each FULL/PARTIAL/FAILED.`
+
+export const researchDoc: PatternAdapter = {
+  name: 'research-doc',
+  defaultRunner: 'hermes',
+  costCap: 5.0,
+
+  matches(task: Task): boolean {
+    const t = task.text.toLowerCase()
+    return KEYWORDS.some((k) => t.includes(k))
+  },
+
+  prepare(task: Task, memory: MemoryHit[]): RunnerInput {
+    const fewshot =
+      memory.length > 0
+        ? `\n\nPast research outputs for similar topics (most relevant first):\n${memory
+            .map((m, i) => `${i + 1}. ${m.body}`)
+            .join('\n')}`
+        : ''
+    return {
+      prompt: task.text,
+      systemPrompt: SYSTEM_PROMPT + fewshot,
+      allowedTools: ['WebFetch', 'WebSearch', 'Read', 'Write', 'Edit', 'Grep', 'Glob', 'Bash'],
+      maxCostUsd: 5.0,
+      metadata: { pattern: 'research-doc' },
+    }
+  },
+}

tests/patterns.test.ts

@@ -0,0 +1,102 @@
+import { describe, it, expect } from 'vitest'
+import { researchDoc } from '../src/patterns/research-doc.js'
+import { meetingCapture } from '../src/patterns/meeting-capture.js'
+import { hermesBugFix } from '../src/patterns/hermes-bug-fix.js'
+import { classify } from '../src/router.js'
+import type { Task, MemoryHit } from '../src/types.js'
+
+function task(text: string): Task {
+  return { id: 't-1', text, createdAt: new Date().toISOString() }
+}
+
+describe('researchDoc pattern', () => {
+  it('matches on "research", "investigate", "audit"', () => {
+    expect(researchDoc.matches(task('research vector dbs for ZAO'))).toBe(true)
+    expect(researchDoc.matches(task('Investigate the Bonfire API limits'))).toBe(true)
+    expect(researchDoc.matches(task('audit the agent stack'))).toBe(true)
+  })
+
+  it('does NOT match on bug-fix or meeting language', () => {
+    expect(researchDoc.matches(task('fix the type error in foo.ts'))).toBe(false)
+    expect(researchDoc.matches(task('process this meeting recording'))).toBe(false)
+  })
+
+  it('prepare returns a research-shaped RunnerInput with WebFetch allowed', () => {
+    const input = researchDoc.prepare(task('research vector dbs'), [])
+    expect(input.allowedTools).toContain('WebFetch')
+    expect(input.allowedTools).toContain('WebSearch')
+    expect(input.systemPrompt).toContain('research agent')
+    expect(input.systemPrompt).toContain('Next Actions')
+    expect(input.maxCostUsd).toBe(5.0)
+    expect(input.metadata?.pattern).toBe('research-doc')
+  })
+
+  it('few-shot injects past memory hits into the systemPrompt', () => {
+    const past: MemoryHit[] = [
+      {
+        name: 'research-doc:t-prev',
+        body: 'Past task: "research vector dbs", Outcome: doc 200 recommends pgvector',
+        sourceTag: 'hermes:research-doc:completed',
+      },
+    ]
+    const input = researchDoc.prepare(task('research vector dbs again'), past)
+    expect(input.systemPrompt).toContain('Past research outputs')
+    expect(input.systemPrompt).toContain('doc 200 recommends pgvector')
+  })
+})
+
+describe('meetingCapture pattern', () => {
+  it('matches on meeting / transcript / recap keywords', () => {
+    expect(meetingCapture.matches(task('process this meeting recording'))).toBe(true)
+    expect(meetingCapture.matches(task('Recap that call'))).toBe(true)
+    expect(meetingCapture.matches(task('extract action items from the standup'))).toBe(true)
+  })
+
+  it('matches on media file extensions', () => {
+    expect(meetingCapture.matches(task('/Users/me/Downloads/call.mp4'))).toBe(true)
+    expect(meetingCapture.matches(task('process /tmp/voice-memo.m4a'))).toBe(true)
+    expect(meetingCapture.matches(task('here is the recording.wav'))).toBe(true)
+  })
+
+  it('does NOT match on plain research or bug-fix', () => {
+    expect(meetingCapture.matches(task('research vector dbs'))).toBe(false)
+    expect(meetingCapture.matches(task('fix the type error in foo.ts'))).toBe(false)
+  })
+
+  it('prepare returns a capture-shaped RunnerInput', () => {
+    const input = meetingCapture.prepare(task('process the call'), [])
+    expect(input.allowedTools).toContain('Bash')
+    expect(input.allowedTools).toContain('Write')
+    expect(input.systemPrompt).toContain('meeting-capture agent')
+    expect(input.systemPrompt).toContain('Decisions')
+    expect(input.systemPrompt).toContain('Actions')
+    expect(input.maxCostUsd).toBe(3.0)
+    expect(input.metadata?.pattern).toBe('meeting-capture')
+  })
+})
+
+describe('router picks the right pattern when multiple are registered', () => {
+  const patterns = [hermesBugFix, researchDoc, meetingCapture]
+
+  it('picks hermes-bug-fix for fix/bug language', async () => {
+    const d = await classify(task('fix the type error in src/foo.ts'), { patterns })
+    expect(d.pattern).toBe('hermes-bug-fix')
+  })
+
+  it('picks research-doc for research language', async () => {
+    const d = await classify(task('research the best vector db for our agents'), {
+      patterns,
+    })
+    expect(d.pattern).toBe('research-doc')
+  })
+
+  it('picks meeting-capture for media file paths', async () => {
+    const d = await classify(task('process /tmp/Arthur-x-Zaal.mp4'), { patterns })
+    expect(d.pattern).toBe('meeting-capture')
+  })
+
+  it('returns unknown when no pattern matches', async () => {
+    const d = await classify(task('write a haiku about ducks'), { patterns })
+    expect(d.pattern).toBe('unknown')
+  })
+})