email

name	email
description	Search, read, triage, and organize email via Maildir and notmuch
user_invocable	true
argument_hint	[inbox [account]|search <query>|read <id|query>|triage [account]|sync]

/email — Maildir Email Workflows

Interactive and agentic email workflows backed by notmuch search/tag and Maildir access.

Infrastructure

• Maildir: ~/Mail/<account>/ — one directory per account, synced by mbsync
• mbsync: IMAP sync on a 5-minute systemd timer. Manual sync: mbsync -a && notmuch new
• notmuch: Search, index, and tag engine. Database at ~/.notmuch/
• aerc: TUI mail client for composing and replying — direct the user there for sending
• emma: Background LLM agent that classifies mail and generates digests via notmuch tags. Claude handles on-demand interactive analysis — complementing, not replacing, emma
• Protonmail: Routes through protonmail-bridge on localhost

Constraint: notmuch flag sync is disabled. Tags applied in notmuch do NOT propagate back to IMAP. This is by design — tags are local metadata.

Maildir Discovery

Accounts are discovered at runtime — do not hardcode them.

ls ~/Mail/                                    # List accounts
ls ~/Mail/<account>/                          # List folders for an account
notmuch search --output=tags '*'              # Discover tag vocabulary (including emma's tags)

Each folder contains cur/, new/, tmp/ subdirectories. Filename flags: :2,S (seen), :2,R (replied), :2,F (flagged).

Subcommands

/email inbox [account]

Review unread inbox. Always delegate to subagent (batch email reads).

Workflow:

1. Sync if requested or stale: mbsync -a && notmuch new
2. Count unread per account: notmuch count tag:unread AND folder:<account>/INBOX for each discovered account
3. Subagent reads unread messages (all accounts or filtered to [account]):

   notmuch show --format=json tag:unread AND folder:<account>/INBOX
   

4. Subagent returns: sender, subject, date, priority assessment, action needed — grouped by account
5. Present summary to user

Subagent prompt should request a synthesized summary (1,000–2,000 tokens), not raw message relay.

/email search <query>

Search email with notmuch. Runs directly (notmuch output is compact).

notmuch search <query>                        # Thread summaries
notmuch search --format=json <query>          # JSON for parsing
notmuch count <query>                         # Count matches

For natural language queries, translate to notmuch syntax (see Query Reference below). Show results directly in compact format. If result count exceeds ~30, show count first and ask the user to narrow down.

/email read <id|query>

Read a specific message or thread.

• Single message / short thread (≤3 messages): Read directly.
• Long thread (>3 messages): Delegate to subagent for synthesis.

notmuch show --format=json id:<message-id>    # Single message
notmuch show --format=json thread:<thread-id> # Full thread

Extract from JSON: .[0][0][0].headers for headers, .[0][0][0].body[0].content for body.

/email triage [account]

Analyze unread mail, classify, tag, and summarize. Always delegate to subagent.

Workflow:

1. Subagent discovers existing tags: notmuch search --output=tags '*'
2. Reads all unread: notmuch show --format=json tag:unread (filtered by account if specified)
3. For each message, classifies:
   • Category: personal, work, newsletter, promotional, transactional, notification, spam
   • Priority: low, normal, high, urgent
   • Action needed: yes/no with brief description
4. Applies tags directly (no approval step): notmuch tag +<category> +priority-<level> -- id:<msg-id>
   • Before tagging, check existing tags on each message to avoid conflicting with emma's classifications
   • If emma already tagged a message, preserve emma's tags and only add non-conflicting ones
5. Returns structured summary grouped by priority, then category

/email sync

Force mail sync. Runs directly.

mbsync -a && notmuch new

Report sync result (new message count from notmuch output).

/email (no args)

Show brief help listing subcommands with one-line descriptions, then ask the user what they want to do.

notmuch Query Reference

from:[e-mail]              to:[e-mail]
subject:invoice                    body:"keyword"
tag:inbox                          tag:unread
folder:account/INBOX               path:**/INBOX/**
date:today                         date:2026-01..
date:yesterday                     date:2026-01-01..2026-01-31

Operators: AND (default), OR, NOT, parentheses for grouping.

Tagging

notmuch tag +important -- id:<message-id>
notmuch tag +archived -inbox -- <query>
notmuch tag -unread -- tag:spam

Context Rules

Subcommand	Delegation	Reason
inbox	Always subagent	Batch email reads fill context
triage	Always subagent	Reads + classifies many messages
search	Direct	notmuch output is compact
read	Direct if ≤3 msgs, subagent if >3	Thread length determines context cost
sync	Direct	Single command

Subagent config: subagent_type: general-purpose. Target return: 1,000–2,000 tokens, synthesized not relayed.

Agentic Integration

Email workflows can be scheduled via /schedule for autonomous execution. Example patterns:

• Morning inbox digest: Scheduled task runs /email inbox, summarizes to a file or desktop notification
• Priority alert: Scheduled task checks notmuch count tag:unread AND tag:priority-urgent, sends desktop notification if >0
• Periodic triage: Scheduled task runs /email triage to keep classifications current between emma runs

The email skill provides the capability; the schedule skill provides the execution framework.

Notes

• mbsync runs every 5 minutes automatically — manual sync is rarely needed
• notmuch tags are local to ~/.notmuch/ and do not affect IMAP state
• Maildir is safe for concurrent access (mbsync, notmuch, aerc operate simultaneously)
• For composing or replying, direct the user to aerc
• When triaging, always check notmuch search --output=tags '*' first to respect emma's existing tag vocabulary