email_agent/AGENTS.md

Email Agent

Build spec for Paperclip. Follow the phases in section 11 in order. Do not skip a phase's Definition of Done. When something is unclear, ask Sam — do not guess.

---

1. Purpose and outcome

A web app that triages the family inbox. All mail lands in one Gmail account (samuelrolfe@gmail.com); PurelyMail forwards every domain address there.

The app sorts mail, recommends actions, performs them in bulk, drafts replies, and answers questions from the Hub.

Outcome (the single success statement):

Sam opens one page, clears the inbox in a few minutes, and knows nothing important was missed.

2. Scope

In scope

• Read mail from Gmail.
• Classify each message: category, importance, suggested action.
• Recommend: delete, unsubscribe, reply, file, keep.
• Bulk actions: delete, move, mark.
• AI reply drafts from templates and rules.
• A rules editor (text) that the classifier uses.
• A daily summary.
• Answer the Hub.

Out of scope

• Sending marketing or bulk mail.
• DNS, MX, SPF, DKIM. That is infrastructure, not this app.
• The other domains' websites.
• Any auto-send without review unless a rule explicitly allows it.

3. Users and roles

• One user: Sam. No multi-user. A session login protects the app.
• Adding a second user is out of scope.

4. Functional requirements

Each item is testable. Numbering is stable; do not renumber.

ID	Requirement
FR-1	Fetch new mail from Gmail with the Gmail API, on an n8n schedule. Store metadata and body.
FR-2	Record which address the mail came through. Read X-Pm-Original-To (added by PurelyMail). Fallbacks, in order: To, the PurelyMail Received ... for <address> clause, Return-Path.
FR-3	Classify each message with JEV: category, importance, suggested action, and a confidence.
FR-14	Edit the category list in the UI: add, rename, remove. JEV uses the current list.
FR-4	Apply user-written rules from the rules editor to override or refine the classification.
FR-5	Show a triage list grouped by category, with the recommended action on each row.
FR-6	Bulk actions: select all in a group, then delete, move, or mark.
FR-7	Detect unsubscribe links. Keep an "ignore unsubscribe" list. Record what was unsubscribed.
FR-8	Draft a reply from a template or rule with the LLM. Let Sam edit, then send.
FR-9	Move messages to folders (Gmail labels): replied-to by address, a delete folder, and a saved folder. Deletes are recoverable.
FR-10	Produce a daily summary and deliver it by email.
FR-11	Rules editor: a plain-text area, saved, versioned. JEV and the LLM read it.
FR-12	Expose the Hub API: GET /capabilities, POST /ask (read-only), POST /notify.
FR-13	Log every action: what, when, which message, which rule.
FR-15	Reply identity: send as the address the mail arrived at. Otherwise send as Sam Rolfe.

5. Screens and UX flow

Screens

1. Login — session form. One dark indigo band (the design system's login moment).
2. Inbox (main) — three areas: left nav (categories and folders), centre list (grouped, with select-all and bulk actions), right detail pane.
3. Message detail — body, recommended action, reply draft, and history.
4. Rules editor — plain text, save, and a version list.
5. Reports — the daily summary, with delete-report.

Main flow

Login → Inbox → scan categories → bulk-delete junk → review "important"
      → approve or edit reply drafts → send → close

Rules

• Works on a phone browser. Touch targets ≥44px.
• No spinners on decorative elements. Subtle opacity pulse only.

6. Look and feel

Reuse the Family Home Lab design system. Do not invent a new one.

• Source: /home/sam/Docker/Containers/family-home-lab/DESIGN.md
• Tokens: /home/sam/Docker/Containers/family-home-lab/portal/static/tokens.css
• Copy tokens.css into this project. Do not import across projects.

Key rules: canvas #f6f5f4, one action blue #0075de, Inter, 8px grid, pill CTAs, 4px inputs, hairline borders, layered micro-shadows. Include dark mode.

7. Data model

SQLite, inside the service. Suggested tables:

Table	Holds
messages	Gmail id, thread id, from, to, subject, date, body text, source address, received-at
classifications	message id, category, importance, action, confidence, model, created-at
rules	id, text, version, created-at
actions_log	message id, action, actor, rule id, created-at
unsubscribe	domain, status (done, ignored), created-at
drafts	message id, body, template id, status
summaries	date, text, delivered-to

Retention: keep message bodies for one month. Logs are append-only and size-capped.

8. Integrations

System	Use
Gmail IMAP + SMTP	Read mail, move labels, send replies. App password (GMAIL_USER, GMAIL_APP_PASSWORD). Force IPv4 — see section 9.
PurelyMail	Source of the forwarding. Read-only reference. Not required for v1.
Amazon SES	App/website mail only. Not for personal replies in v1.
JEV	Classification (typed category + confidence). HTTP API.
OmniRoute	All other models.
n8n	Triggers the scan on a schedule.
Apprise	Delivery of the daily summary.
Hub	POST /ask (read-only), POST /notify, GET /capabilities.

The Hub interface is frozen in central_ai_llm_master/docs/ask-contract.md. Build to it. /ask never changes state.

9. Technology (locked)

Item	Value
Backend	Go
UI	HTMX + server-rendered HTML
Database	SQLite
Mail read	IMAP imap.gmail.com:993 (SSL), IPv4 only
Mail send	SMTP smtp.gmail.com:587 (STARTTLS), IPv4 only
Mail auth	Gmail app password from environment.d. No OAuth, no console.
Design	Family Home Lab tokens.css
Host	.13, Docker
Container	This project's own image and compose file. Not shared with any other project.
Folder	/home/sam/paseo/projects/email_agent/
Port	8790 (chosen; container listens on 8790, published as ${EMAIL_AGENT_PORT:-8790}). Confirm against infra-facts AUD-60.
URL	mail.lab.audasmedia.com.au via Caddy on .35
Auth	Separate session login. Not the console SSO.

10. Rules and constraints

1. All models go through OmniRoute, except JEV.
2. Secrets come from ~/.config/environment.d/10-secrets.conf. Never commit them.
3. /ask is read-only. It never changes state.
4. Sending a reply or deleting in bulk needs Sam's action in the UI. No auto-send unless a rule explicitly allows it.
5. Spam is recommended, never auto-deleted. Important mail sometimes looks like spam. Confirm every spam action.
6. Deletes are recoverable. Move to a delete folder or trash, never hard-delete.
7. Replies return to the source and stay short.
8. Every action is logged.
9. Do not merge this service into another. Own folder, own container, own port.

11. Build order

Each phase ends with a Definition of Done (DoD). Do not start the next phase until the DoD passes.

Phase 1 — Skeleton Go service, SQLite, config, session login, Dockerfile, compose, Caddy route. DoD: the app runs on .13, login works, an empty dashboard renders, /health returns ok.

Phase 2 — Read mail Connect to Gmail over IMAP with the app password. Force IPv4. Fetch, store, and list real messages. DoD: the inbox shows real mail. Confirm the source address from FR-2 against a real forwarded message (see section 13).

Phase 3 — Classify JEV classification plus the rules table. DoD: every message has a category, importance, and suggested action.

Phase 4 — Triage actions Delete, move, mark; bulk select. DoD: select-all then delete works, and the message is recoverable.

Phase 5 — Reply drafts Templates, LLM draft, edit, send. DoD: a reply is drafted, edited, and sent, and the thread updates.

Phase 6 — Unsubscribe and rules editor DoD: a rule written in the editor changes the classification of a real message.

Phase 7 — Summary and Hub Daily summary; /capabilities, /ask, /notify. DoD: the Hub asks "anything important in my inbox?" and gets a correct short answer.

Phase 8 — Documentation DoD: project-docs run; Gitea, Obsidian, Outline, Vikunja, and the map are updated.

12. Acceptance criteria

The project is done when:

• Sam can log in from a phone browser.
• Real mail appears, grouped, with the source address shown.
• Bulk delete works and is recoverable.
• A reply can be drafted, edited, and sent without leaving the app.
• The rules editor changes real classifications.
• The Hub gets a correct short answer to "anything important in my inbox?".
• Every action is in the log.
• The app runs in its own container on its own port, behind Caddy.

13. Decisions and remaining questions

Resolved:

1. Reply identity: the address the mail arrived at. Otherwise Sam Rolfe.
2. Categories: action, important, info, spam, newsletter, auto-reply. Editable in the UI (FR-14).
3. Spam: recommend and confirm. Never auto-delete.
4. Daily summary: by email.
5. Retention: one month of message bodies.
6. Mail access is IMAP + app password — no console, no OAuth. Working on .13 and .51.
7. Source address is X-Pm-Original-To. Confirmed on a real forwarded message (X-Pm-Original-To: Hello@where-woof.com). PurelyMail adds this header, so Gmail's overwritten Delivered-To does not matter. See FR-2.
8. PurelyMail forwarding is live, confirmed end to end.

Remaining:

1. Send-as: sending as a domain address needs Gmail "Send mail as" set up for that address.

Infrastructure notes (found during setup)

• IPv6 to Gmail is broken from .13. The first IMAP attempt failed with a socket EOF over IPv6. Forcing IPv4 works. The agent must force IPv4 for IMAP and SMTP.
• Credentials live in ~/.config/environment.d/10-secrets.conf on .13 and .51, as GMAIL_USER and GMAIL_APP_PASSWORD.
• Header fields seen on a real forwarded message (id 619, from favrolfe@gmail.com):

  X-Pm-Original-To: Hello@where-woof.com   <- use this
  To: Hello@where-woof.com
  Return-Path: <Hello@where-woof.com>
  Delivered-To: samuelrolfe@gmail.com       <- Gmail overwrites this; ignore