Email Triage — Discovery Brief

Project: Forge Email Triage  |  Date: 2026-03-25  |  Discovery — Future Planning
Prior Art: Vault/projects/email-system/DISCOVERY.md (Ralph, 2026-02-25) — covers inbound task extraction. This brief widens scope to full triage, competitive positioning, and second-order effects.

1. Problem Statement

The Pain

Jason operates across six active projects (Forge, MasteryOS, HC Protocol, Brad Himel Beta, Credential Vault, Voice Agent), coordinates with a distributed team (Will, Derek, Sumit, Lee, Ashwini), and manages JV relationships (Brian Muka, Samuel, Brad). Email is the default communication channel for:

What Falls Through the Cracks

  1. Commitment tracking — Jason agrees to "send that by Friday" in an email thread. No system extracts that as a task. It lives only in the thread.
  2. Follow-up gaps — Emails sent with no reply within 3–5 days. No automated follow-up reminder. Relationships cool.
  3. Priority blindness — Newsletter noise buries an urgent email from a JV partner. No automatic surfacing.
  4. Response latency — Emails that need a simple acknowledgment ("Got it, working on it") sit for hours because Jason is in a Claude Code session.
  5. Context loss — Email threads contain decisions, deadlines, and commitments that never reach the Task Board, Knowledge graph, or JASON-DEPS.
Time cost estimate: 45–90 minutes per day on email triage, response drafting, and manual follow-up tracking. At $200/hr operator time, that is $1,500–3,000/month of drag.

What "Solved" Looks Like

2. Competitive Landscape

Tool Triage Draft Responses Commitment Extraction CRM / Relationship Self-Hosted Forge Integration Price
Apex.host Auto-triage + prioritize Auto-draft Tracks commitments Context memory across tools Yes None Unknown (waitlist)
Superhuman Auto-labels, summaries Auto-draft per email Reminders, follow-ups Basic No None (closed API) $30/mo
Shortwave AI splits, filters Ghostwriter (voice-matched) Tasklets (Slack/Notion) No No None (Gmail-only) $7–24/mo
SaneBox Folder-based sort (98%) No SaneRemind No No None (IMAP) $7–36/mo
Forge (proposed) Hybrid rules + LLM Draft via Comms service Task Board + Ralph queue Knowledge graph Yes (VPS) Native (full stack) $2–5/mo

Key Observations

Apex.host is the closest competitor to Forge's vision. Both are self-hosted, context-aware, always-on AI assistants. Apex treats email as one channel among many (Slack, calendar, KPIs). Forge does the same with its modular service architecture. The difference: Forge is purpose-built for Jason's workflow, and Jason owns every line of code.

Superhuman/Shortwave solve the "email client" problem, not the "operating system" problem. They make email faster but do not connect email to a task queue, an autonomous agent (Ralph), or a knowledge graph. They are polished UIs on top of Gmail — not intelligence infrastructure.

SaneBox is the "good enough" option. Rule-based filtering at $7/mo, 98% accuracy, works with any email provider. If Jason only wanted noise reduction, SaneBox is the answer. But Jason wants commitment extraction, relationship scoring, and Forge integration — SaneBox does none of that.

No existing tool provides:

3. Build vs. Buy Analysis

Option A: Buy SaneBox + Manually Bridge

What you get: Noise filtering (newsletters, receipts, automated junk) at 98% accuracy. SaneLater folder. Daily digest. Follow-up reminders.

What you do not get: Task extraction, draft responses, Knowledge graph, Ralph integration, Telegram alerts.

Cost: $7–12/month. Verdict: Solves 30% of the problem.

Option B: Buy Shortwave as Primary Email Client

What you get: AI summaries, auto-categorization, Ghostwriter (writes in Jason's voice), Tasklets.

What you do not get: Forge Task Board integration, Ralph queue routing, Knowledge graph, PII-safe routing, Telegram native alerts.

Cost: $14–24/month. Verdict: Best standalone email client with AI. But replaces Gmail rather than augmenting Forge. Lock-in risk. No API.

Option C: Build Natively in Forge

What you get: Full triage pipeline, commitment extraction, draft responses via Comms, Ralph queue routing, Knowledge graph enrichment, Telegram notifications, PII-safe LLM routing, full control.

What you do not get: Pre-built UI polish (Superhuman-level design). However, Telegram + Dashboard is sufficient.

Cost: $2–5/month operational. 2–3 weeks dev time across phases.

RECOMMENDED: Build (Option C), with SaneBox as optional complementary noise-reduction layer.

Rationale:
  1. 80% of infrastructure already exists (Gmail OAuth, Comms, ClawdRouter, Knowledge, Commander bot)
  2. The value is in the connections, not the classification. Classification is commoditized. Routing into a personal AI OS with an autonomous agent, knowledge graph, and task queue — that is the moat.
  3. $2–5/mo vs $30/mo for Superhuman (which cannot integrate with Forge)
  4. Second-order effects (commitment extraction, relationship intelligence, CRM) are only possible with native build
  5. Aligns with Forge's "every capability is a service" architecture

4. Technical Approach

4.1 Gmail API (Already Available)

The forge4j@gmail.com account already has OAuth2 credentials, Gmail API enabled (used by the Knowledge service poller), and 6 Google Calendar connections (same OAuth scope extends to Gmail watch).

Inbound mechanism: Gmail API users.watch() with Google Cloud Pub/Sub for push notifications. Watch expires every 7 days — a timer renews it. Alternative for Phase 1: extend the existing historyId-based poll every 2–5 minutes.

4.2 Classification Pipeline

Inbound Email
     |
     v
[Rule Engine] -- newsletters, receipts, noreply@ --> ARCHIVE (no LLM cost)
     |
     | (ambiguous emails only, ~30% of volume)
     v
[ClawdRouter: pii_sensitive cascade]
     |  Claude Max CLI --> Qwen 7B Local --> Haiku (scrubbed) --> Sonnet (scrubbed)
     v
[Structured Classification JSON]
     |
     +---> type: task --> Extract commitments --> Task Board / Ralph queue
     +---> type: urgent --> Telegram notification with action buttons
     +---> type: info --> Store in Knowledge graph, no notification
     +---> type: reply_needed --> Draft via Comms --> Telegram approval

PII safety: All email content routes through the pii_sensitive ClawdRouter cascade. Claude Max CLI first (local, free), then Presidio-scrubbed fallbacks to trusted-jurisdiction models only (Anthropic). No Chinese-origin models ever see email content. Local Qwen is acceptable because data stays on VPS.

4.3 Comms Service Integration

The Comms service on port 5013 already handles draft queue, anti-bot send scheduling, and Telegram approval. Extensions needed:

4.4 Telegram Notifications

URGENT EMAIL from Brian Muka
Subject: "Need revised proposal by Thursday"

[Reply: "On it"]  [Snooze 2h]  [Open in Gmail]  [Delegate to Ralph]

Inline keyboard buttons map to Comms compose, snooze timer, Gmail deep link, and Ralph queue insertion.

5. Architecture Options

Option Description Pros Cons
A: New Service
forge-email-triage		 Standalone service on port 5019. Own systemd unit, health endpoint. Gmail watcher + classifier + router. Clean separation. Independent deploy/restart. Follows Forge convention. Another service to monitor (15+ already). Duplicates some Gmail auth.
B: Extend Comms Add inbound processing to existing Comms service. Shared Gmail auth and thread tracking. Fewer services. Conceptually "Comms owns email." Comms becomes multi-concern. Outbound anti-bot logic is delicate. Harder to restart independently.
C: Gmail + External Automation Gmail watch via Cloud Console. n8n/Make processes webhooks. Minimal custom code. Quick prototype. n8n deprecated in Forge. No Forge context access. No PII-safe routing. Breaks service architecture.
Recommended: Option A (New Service). Email triage has its own lifecycle (watch renewal, classification tuning, follow-up scheduling). It deserves its own service. The Knowledge service already has enough responsibility. The Comms service should stay focused on outbound safety.

6. Effort Estimate

Phase Description Days Owner Human Required?
1 Foundation (watch + rules + schema) 2–3 Ralph No
2 LLM classification via ClawdRouter 1–2 Ralph No
3 Task extraction + routing 2–3 Ralph + Jason review Light review
4 Response drafting via Comms 1–2 Ralph No
5 Follow-up engine 2–3 Ralph + design Design decisions
6 Intelligence layer (relationship scoring, "Ask Forge") 3–5 Ralph + design Design decisions
Total 11–18

Cost: Phases 1–4 are free (Ralph time + $2–5/mo LLM). No new infrastructure. No new subscriptions.

7. Second-Order Effects (The Real Value)

First-Order: Email Gets Triaged

Emails classified and routed automatically. Response time drops from hours to minutes. Nothing falls through the cracks.

↓

Second-Order: Commitment Extraction Feeds the Factory

Every "I'll do X by Y" extracted from email becomes a Task Board item. Automatable tasks go to Ralph queue — factory works on email-sourced tasks overnight. Follow-up reminders ensure commitments are tracked to completion.

Cascade: Email → commitment → task → Ralph → done → auto-follow-up

↓

Third-Order: Relationship Intelligence Emerges

Email patterns reveal: who Jason communicates with most, response times, sentiment trends. Knowledge graph learns relationship context. Relationship scoring feeds into JV qualification (Athio pipeline), client health, team responsiveness.

Cascade: Email patterns → relationship scores → JV qualification → revenue

↓

Fourth-Order: CRM Without a CRM

Contact enrichment from email signatures, domains, LinkedIn links. Interaction history across channels (email + Telegram + meeting notes). Deal stage tracking from email content.

Cascade: Email intelligence → contact graph → deal tracking → pipeline visibility

↓

Fifth-Order: The Inbox Becomes a Sensor

Email volume/urgency trends signal when Jason is overcommitted. Missed follow-ups signal relationship decay. Response time trends signal operational health.

Cascade: Email metrics → workload awareness → proactive capacity management

Compound Effect: Email triage is not about reading email faster. It is about turning the inbox into an input stream for an autonomous operating system. Every email becomes structured data. Every commitment becomes a tracked task. Every relationship becomes a scored node in a graph. The inbox stops being a liability and becomes a sensor.

8. Build-or-Buy Recommendation

RECOMMENDED: Build natively as forge-email-triage service (Option A), with optional SaneBox ($7/mo) as a noise-reduction pre-filter.

Why Build

  1. 80% of infrastructure already exists (Gmail OAuth, Comms, ClawdRouter, Knowledge, Commander bot)
  2. The value is in the Forge-native integrations (Task Board, Ralph queue, Knowledge graph), not in the classification itself
  3. $2–5/mo operational cost vs $30/mo for Superhuman (which cannot integrate with Forge)
  4. Second-order effects (commitment extraction, relationship intelligence, CRM) are only possible with a native build
  5. Aligns with Forge's "every capability is a service" architecture

Why Not Now

When to Build

Quick-Win Alternative

If email pain increases before then: SaneBox at $7/mo for immediate noise reduction. Takes 15 minutes to set up, no engineering required. Can run alongside future Forge Email Triage service.

References

Forge Email Triage Discovery Brief — 2026-03-25 — Generated by Claude Code