Email Pipeline + Edita PA — Runbook

Email Pipeline + Edita PA — Runbook

MC: #8521 | Related: #8466 (OWN classifier fix) | Date: 2026-04-20

Overview

The email pipeline classifies incoming emails and routes them to Mission Control, HiveMind, or archive. Edita PA is the autonomous email assistant operating in phased rollout (currently Phase 1).

Architecture

• Daemon: ~/system/daemons/email-agent.js
• LaunchAgent: com.john.email-agent (via wrapper email-agent-wrapper.sh)
• Vault: Bitwarden session (/tmp/bw-session) required for IMAP credentials
• Triage LLM: llama3.1:8b (Ollama ANVIL, preloaded via ollama-triage-preload.sh)

OWN Classifier Logic

The OWN classifier identifies machine-generated emails from ALAI's own systems to prevent task spam.

Constants (email-agent.js lines 118-123)

const OWN_SYSTEM_PREFIXES = [
  'noreply@', 'no-reply@', 'sentinel@', 'alerts@', 'auto@', 'daemon@',
  'mailer@', 'notification@', 'notifications@', 'bounces@', 'bounce@',
  'donotreply@', 'do-not-reply@', 'system@'
];
const OWN_SYSTEM_DOMAINS = ['@alai.no', '@basicconsulting.no'];

isOwnSystemEmail() Function (lines 446-456)

Two-tier check:

1. Exact match: OWN_ADDRESSES array (hardcoded machine addresses)
2. Prefix + domain: Any prefix in OWN_SYSTEM_PREFIXES on domains in OWN_SYSTEM_DOMAINS

Critical: alem@alai.no is NEVER in this list. VIP check runs FIRST (line 464), bypassing OWN classifier entirely.

TLDR_SKIP Routing

Newsletters from dan@tldrnewsletter.com do NOT create MC tasks. They are handled exclusively by tldr-briefing.js daemon.

// line 126
const TLDR_SENDER = 'dan@tldrnewsletter.com';

// line 474
if (lowerFrom.includes(TLDR_SENDER)) {
  return { category: 'TLDR_SKIP', priority: 'low', summary: 'TLDR newsletter — handled by tldr-briefing.js', action: '' };
}

VIP Ordering

Classification priority (lines 464-481):

1. VIP: CEO/family → bypass ALL filters, force ACTION/high, skip Ollama
2. TLDR_SKIP: Newsletter → skip MC INTAKE, route to tldr-briefing.js
3. OWN: System emails → archive, no task
4. Other: Spam allowlist check → Ollama classification

Edita PA Phases

Phase 0: --dry-run (Log-Only)

Classification + logging only. No archive, no escalate, no respond.

node ~/system/daemons/email-agent.js --dry-run

Phase 1: --allow-archive (CURRENT)

Archive low-priority emails only. Escalate and respond actions are held (logged but not executed).

node ~/system/daemons/email-agent.js --allow-archive

Plist config: com.john.email-agent calls email-agent-wrapper.sh, which passes no flags → defaults to Phase 1 (archive-only mode is internal default in daemon code).

Phase 2: Full Live (NOT YET APPROVED)

Archive + escalate + respond. Requires CEO explicit approval.

node ~/system/daemons/email-agent.js --allow-all

Unit Testing

Test classifier without IMAP/Vault dependencies:

node ~/system/daemons/test-email-classifier.js

Scenarios (16 total):

• VIP bypass (alem@alai.no, CEO family)
• TLDR_SKIP routing
• OWN system emails (noreply@alai.no, sentinel@basicconsulting.no)
• Spam patterns with allowlist exceptions (GitHub, Cloudflare, Anthropic)

Rollback

Revert to dry-run mode:

launchctl unload ~/Library/LaunchAgents/com.john.email-agent.plist

# Edit wrapper to add --dry-run flag
echo 'exec /opt/homebrew/bin/node "$HOME/system/daemons/email-agent.js" --dry-run' >> ~/system/tools/email-agent-wrapper.sh

launchctl load ~/Library/LaunchAgents/com.john.email-agent.plist

Monitoring

• Logs: ~/system/logs/email-agent-launchd.log
• Errors: ~/system/logs/email-agent-launchd-error.log
• MC tasks: node ~/system/tools/mc.js list --owner edita
• DLQ: Failed vault sessions stored in email-agent.js in-memory DLQ (logged only, no persistence)

Generated by Skillforge | ALAI, 2026