Virtual Persona

Virtual Persona (VP) is an autonomous AI assistant that reads your Teams messages, decides whether to respond, and replies in your authentic communication style — so colleagues get timely answers even when you’re unavailable.

How It Works

Monitor

VP polls your Teams DMs on a timer (default every 2 minutes). It reads new messages, aggregates multiple messages per chat into one query, and classifies each: does this need a response? Is it something the AI can handle?

Decide

Each message gets a confidence score. High confidence → auto-respond. Medium → respond with disclaimer. Low → silently escalate to you. Messages that don’t need a response are skipped entirely.

Respond (or not)

For auto-responses, VP delegates to Octo’s full agent stack (subagents, tools, skills) for the actual answer, then rewrites it in your personal communication style. For escalations, VP stays completely silent — the sender never knows.

Decision Types

VP classifies every incoming message into one of five decisions:

Decision	Confidence	What Happens	Sender Sees
skip	—	No response needed (acknowledgments, social chatter)	Nothing
respond	≥80%	Auto-respond in your voice	Your response + 🤖
disclaim	60-79%	Respond with a “verify with real me” caveat	Response + disclaimer + 🤖
escalate	below 60%	Thread locked, private notification to you	Nothing
monitor	—	Suggested answer sent privately to you	Nothing

Monitor is the default. Until you explicitly /vp allow someone, VP only observes — it reads messages and sends you private suggestions but never responds to the sender.

Message Aggregation

People often send 3-7 consecutive messages. VP collects all unprocessed messages per chat, merges them into one query, and runs the pipeline once per chat — not once per message. This means:

No duplicate responses to a burst of messages
Better context for confidence scoring (sees the full request)
If you already replied in a thread, VP skips messages before your last reply

Two-Stage Pipeline

VP doesn’t generate knowledge itself. It uses a multi-stage architecture:

┌─────────────────────────────────────────────────────────┐
│ Stage 1: VP Graph (decision engine)                     │
│   access_check → classify → gather_context → route      │
└───────────────────────┬─────────────────────────────────┘
                        │ respond / disclaim / escalate / monitor
                        ▼
┌─────────────────────────────────────────────────────────┐
│ Stage 2: Octo Supervisor (knowledge engine)             │
│   context-gatherer → subagents + MCP tools → raw answer │
└───────────────────────┬─────────────────────────────────┘
                        │
                        ▼
┌─────────────────────────────────────────────────────────┐
│ Stage 3: VP Graph (persona formatting)                  │
│   raw answer → rewrite in your style → send             │
└─────────────────────────────────────────────────────────┘

This means VP responses benefit from every agent, tool, and skill Octo has — without reimplementing anything.

Cross-Conversation Context

Before generating an answer, VP gathers context from multiple sources: Layer 1 — Local knowledge (instant, no API calls):

Searches the indexed conversation knowledge base for related threads by topic keywords
Finds all recent threads involving the same person
Feeds topic summaries and prior discussion context into the answer generation

Layer 2 — Deep search (via context-gatherer agent):

For respond/disclaim decisions, Octo’s context-gatherer agent searches:
- Email threads with the same person on similar topics (via Outlook MCP)
- Other Teams conversations involving the same person (via Teams MCP)
- External context if relevant (via web search)
This ensures VP answers reflect the full history you’d have in mind when replying

Invisible Escalation

When VP can’t handle a message confidently, escalation is completely invisible to the sender:

No “routing to real person” message
No typing indicator or acknowledgment
The thread is silently locked — VP won’t touch it again until you release it
You get a private notification (via Telegram or console) with the full context and a suggested answer
Reply to the Telegram notification to respond directly — Octo routes your reply to the Teams chat and auto-releases the delegation lock
Reply “ignore” to mute the chat entirely

VP never reveals its existence to the sender during escalation. From the sender’s perspective, they’re just waiting for you to reply.

Persona Generation

VP needs to know how you communicate to sound like you. The /vp persona generate pipeline analyzes your actual Teams conversations:

Fetch Conversations

Pulls 30-50 recent Teams chats with messages and participant info via MCP.

Per-Conversation Analysis

Each conversation is analyzed separately for behavioral patterns: tone, language, humor, response style, social dynamics. This captures how you behave differently in different contexts.

Web Research

Searches for your public communication presence — blog posts, conference talks, GitHub activity, social media. Uses Octo’s full agent stack with web search tools.

Synthesis

Merges all per-conversation analyses and web research into a unified communication profile: consistent patterns, situational variations, personality traits, anti-patterns.

Prompt Generation

Generates a system prompt that captures your authentic voice, with behavioral rules, escalation guidelines, and example Q&A pairs.

More conversations = better persona. Aim for 20+ conversations with a good mix of 1:1 and group chats for the most accurate results.

Access Control

VP uses a YAML-backed access control system with two lists:

List	Behavior
allow_ai	VP responds automatically (subject to confidence scoring)
always_user	VP observes and sends you suggestions, but never responds

Users not on either list are treated as always_user by default (monitor mode).

Per-User Confidence Modifiers

Each allowed user can have a confidence modifier that adjusts scoring:

/vp allow alice@company.com +10    # More confident when answering Alice
/vp allow bob@company.com -15      # More conservative with Bob

1-on-1 Chat Boost

VP automatically gives a +15% confidence boost to 1-on-1 (DM) conversations. People expect replies when they message you directly, so the threshold is lower. Group chats use standard scoring.

Priority Users (Never-Ignore)

Some contacts should never be ignored, even if their chat is on the ignore list. Add them to the priority list:

/vp priority add boss@company.com       # Always process their messages
/vp priority remove boss@company.com    # Remove from priority list
/vp priority                            # List all priority users

Priority users bypass the chat ignore list — their messages are always processed regardless of which chat they’re in.

Ignoring Chats

Mute noisy chats (bot channels, automated notifications) to reduce VP noise:

/vp ignore <chat_id> [label]     # Mute a chat
/vp unignore <chat_id|all>       # Unmute
/vp ignored                       # List ignored chats

Delegation Locks

When VP escalates a thread, it automatically locks that conversation:

VP won’t respond to any messages in a locked thread
You handle it manually
Release when done: /vp release <chat_id> or /vp release all
View locked threads: /vp delegated

People Profiles

VP maintains a contact map with interaction history:

Auto-enrichment: New contacts are enriched from Teams member info (name, title, department)
Tone presets: casual_technical (default for engineers), casual, semi_formal
Topic tracking: Tracks what each person usually discusses with you
Interaction count: How often VP has processed messages from them

Set a tone override: /vp profile alice@company.com tone semi_formal

Conversation Knowledge

VP maintains a topic cache for each thread it processes:

Incremental sync: Thread summaries updated on each poll cycle
Topic classification: LLM-based topic labeling for grounding
Context feeding: Thread context is passed to the confidence scorer and Octo supervisor

This prevents VP from confusing topics across threads or giving responses without context.

Configuration

VP_ENABLED=true                       # Enable VP subsystem (default: true)
VP_SELF_EMAILS=you@company.com        # Your email(s), comma-separated
VP_POLL_INTERVAL=2m                   # Polling frequency (2m, 30s, 1h, or bare seconds like 120)
VP_ACTIVE_HOURS_START=08:00           # Local time
VP_ACTIVE_HOURS_END=22:00             # Local time

VP_SELF_EMAILS is required for VP to work correctly. Without it, VP can’t distinguish your own messages from incoming ones.

Commands

Command	Description
`/vp status`	Show VP state, allow/block counts, messages today
`/vp enable`	Enable VP and start poller
`/vp disable`	Kill switch — stop all VP activity
`/vp allow <email> [+mod]`	Add user to auto-respond list
`/vp block <email>`	Add user to monitor-only list
`/vp remove <email>`	Remove from all lists
`/vp list`	Show both lists with modifiers
`/vp test <email> [query]`	Dry-run access check + confidence scoring
`/vp stats [days]`	Aggregated stats (default 7 days)
`/vp audit [N]`	Last N audit entries
`/vp confidence <email> <mod>`	Update per-user confidence modifier
`/vp sync`	Force full conversation knowledge sync
`/vp profile <email>`	Show person’s profile
`/vp profile <email> tone <preset>`	Set tone override
`/vp threads [N]`	Show recent classified thread topics
`/vp delegated`	List locked (delegated) threads
`/vp release <chat_id\|all>`	Release locked thread(s)
`/vp ignore <chat_id> [label]`	Mute a chat (skip during polling)
`/vp unignore <chat_id\|all>`	Unmute a chat
`/vp ignored`	List ignored chats
`/vp priority [add\|remove] <email>`	Manage never-ignore user list
`/vp persona generate [name]`	Run full persona generation pipeline
`/vp persona show`	Preview current persona prompt
`/vp persona analysis`	Show communication analysis summary
`/vp persona path`	Show file paths

Data Files

All VP data lives in .octo/virtual-persona/:

File	Purpose
`access-control.yaml`	Allow/block lists, enabled flag
`system-prompt.md`	Persona prompt (generated or hand-edited)
`profiles.json`	People profile map
`knowledge/threads.json`	Thread topic index
`audit.jsonl`	Append-only decision log
`stats.json`	Counter aggregates
`message-cache.json`	Processed message dedup
`delegated.json`	Locked thread map
`communication-analysis.json`	Synthesized profile from persona gen
`per-conversation-analyses.json`	Raw per-chat analyses
`web-research.md`	Public presence research results

Setup Guide

Step-by-step VP setup and first test

Teams Integration

Setting up the Teams MCP connection

Get Started

Features

Resources

Virtual Persona

How It Works

Decision Types

Message Aggregation

Two-Stage Pipeline

Cross-Conversation Context

Invisible Escalation

Persona Generation

Access Control

Per-User Confidence Modifiers

1-on-1 Chat Boost

Priority Users (Never-Ignore)

Ignoring Chats

Delegation Locks

People Profiles

Conversation Knowledge

Configuration

Commands

Data Files

Setup Guide

Teams Integration

Get Started

Features

Resources

​How It Works

​Decision Types

​Message Aggregation

​Two-Stage Pipeline

​Cross-Conversation Context

​Invisible Escalation

​Persona Generation

​Access Control

​Per-User Confidence Modifiers

​1-on-1 Chat Boost

​Priority Users (Never-Ignore)

​Ignoring Chats

​Delegation Locks

​People Profiles

​Conversation Knowledge

​Configuration

​Commands

​Data Files

Setup Guide

Teams Integration

How It Works

Decision Types

Message Aggregation

Two-Stage Pipeline

Cross-Conversation Context

Invisible Escalation

Persona Generation

Access Control

Per-User Confidence Modifiers

1-on-1 Chat Boost

Priority Users (Never-Ignore)

Ignoring Chats

Delegation Locks

People Profiles

Conversation Knowledge

Configuration

Commands

Data Files