psd-email-triage

name	psd-email-triage
summary	Smart email triage — opt-in, per-user rules, Gmail labels, daily digest, Chat escalations. Configure entirely from chat.
description	Set up or adjust smart email triage — per-user rules, Gmail labels, daily digest, and Chat escalations. Use when the user wants to start or change triaging their inbox.
allowed-tools	Bash(node:*)

psd-email-triage

Smart email triage controlled entirely through chat. The user says things like "start triaging my email", "ignore vendor-x", "always page me when CEO emails", "show recent classifications", "stop triaging" — and this skill writes the rules into DynamoDB where a background Lambda (psd-agent-triage-poll-<env>) picks them up on its next 5-minute tick.

The skill itself does not classify mail. It manages CONFIG. The classifier Lambda does the per-email work.

When to call this skill

• User explicitly asks to enable / disable / configure email triage
• User says "ignore X", "always page me from Y", "auto-archive newsletters" → translate to a rules.* or escalation.* subcommand
• User asks "show recent triage", "what did you classify?" → training recent
• User says "I keep getting Important on noisy senders" → training correct + rules mute
• User wants a daily summary → digest enable / digest time HH:MM
• User says "only ping me when you're sure" / "stop pinging me for everything" / "quiet down the notifications" → escalation mode <high-confidence|rules-only|none> (+ escalation threshold)
• User says "triage my existing inbox" / "sweep my backlog" / "catch up on my old mail" → sweep
• User says "show me your suggestions" / "apply that" / "yes do the first one" / "ignore that suggestion" → suggestions list|apply <id>|dismiss <id>

Don't call this skill for:

• Generic Gmail operations (use psd-workspace / gws-*)
• The classifier's runtime behaviour (that's the Lambda — see docs/operations/email-triage.md)

Quick reference

node /opt/psd-skills/psd-email-triage/run.js <subcommand> --user <email> [args]

--user <email> is required on every call (matches the existing skill convention; comes from the [caller: …] header).

Lifecycle

Subcommand	Effect
enable	Create Gmail labels (default @psd/Important, @psd/Later, @psd/News), seed default rules, anchor Gmail history cursor, optionally schedule daily digest, and kick an initial-inbox sweep (last 30 days, ≤1000 messages, no Chat pings). Idempotent — running enable on an already-enabled triage is a no-op that returns current status.
disable	Pause classification. Keeps rules + labels intact for re-enable.
disable --forget	Hard reset: clears rules, learned patterns, decision history; deletes Gmail labels; removes digest schedule.
status	Returns enabled state, escalation mode/threshold, label set, rule counts, sweep progress, pending-suggestion count, last poll, recent decision counts.
sweep	(Re)run the initial-inbox backfill for an already-enabled user — classifies the last 30 days of INBOX (≤1000 messages) through the normal pipeline with no Chat escalations. Resumable + escalation-silent. No-op if a sweep is already pending/running.

Rules

Subcommand	Effect
rules list	Show all configured rules grouped by type.
rules add-vip <email>	Always classify as Important. Beats every other rule.
rules mute <pattern>	Auto-archive sender. Supports * wildcards (e.g. noreply@*, *.vendor.com).
rules add-keyword <kw> --label <important|later|news> [--subject|--snippet|--from <domain>|--external]	Add a keyword rule. Defaults to subject-match.
rules remove <type> <value>	Remove a rule. type is vip, mute, or keyword.

Escalation (Chat pings)

Subcommand	Effect
escalation list	Show current escalation mode, threshold, senders + keywords + label triggers.
escalation add-sender <email>	Always Chat-ping when this sender's mail classifies as Important.
escalation add-keyword <kw>	Always Chat-ping when an Important message's subject/snippet contains this keyword.
escalation remove <type> <value>	Remove an escalation rule. type is sender or keyword.
escalation labels <label1,label2>	Override which labels trigger Chat. Default: important only.
escalation mode <all|high-confidence|rules-only|none>	Per-user policy for which Important classifications ping Chat. all (default) = every Important pings. high-confidence = only rule matches + LLM decisions ≥ threshold. rules-only = only your VIP/escalation-rule matches; plain LLM Important never pings. none = label only, digest is the sole surface. Explicit escalation sender/keyword rules always ping (except none).
escalation threshold <0-1>	LLM-confidence bar for high-confidence mode. Default 0.85.

Training feedback

Subcommand	Effect
training recent [--limit N]	Last N classifications (default 20) with sender, subject, label, source (rule vs llm), confidence, reason.
training correct <messageId> <newLabel>	Re-label one message in Gmail and record a training signal.

Learned suggestions

A nightly job mines the user's corrections into weighted hints (fed to the classifier automatically) and suggested rule changes ("you archived 3 Important emails from X → mute X"). Hard rules are suggest-only — a suggestion becomes a real rule only when the user approves it here. New suggestions also arrive as a Chat card.

Subcommand	Effect
suggestions list	Show pending rule suggestions (each has an id, kind = mute/vip, target, and reason).
suggestions apply <id>	Approve a suggestion — writes the real rule (vip → VIP sender, mute → mute pattern) and clears it from pending.
suggestions dismiss <id>	Reject a suggestion — cleared from pending and never raised again.

Simulation

Subcommand	Effect
simulate --from <email> [--subject "..."] [--snippet "..."] [--external] [--has-user-reply]	Dry-run the rule engine against a synthetic email. Useful for "would my mute rule catch this?" without waiting for real mail.

Labels

Subcommand	Effect
labels list	Show current label names + Gmail label IDs.
labels rename <key> <new-name>	Rename one of important / later / news (renames the Gmail label too).

Digest

Subcommand	Effect
digest enable	Schedule a daily summary card at the user's configured time.
digest disable	Remove the daily schedule (rules/labels stay).
digest time <HH:MM>	Set the time (24-hour, user's timezone from the agent users table). Default 08:00.

Tasks (user-gesture, Phase 1.5)

When the user applies the @psd/Task label to a Gmail message, the classifier Lambda detects the gesture. What happens next depends on tasksMode:

Subcommand	Effect
tasks mode none	(default) Lambda ignores the label — the message just sits with @psd/Task applied. No automation.
tasks mode invoke-agent	Lambda invokes AgentCore with the email metadata. The user's agent (per their MEMORY.md instructions + skills) creates a task in their preferred task system. On success the email is archived (INBOX + @psd/Task removed). On failure the email is left alone and a Chat card surfaces the reason.
tasks notify-success on|off	(default off) When on, every successful task creation posts a one-line confirmation card to Chat. Failures always notify regardless.
tasks status	Returns current mode, notify setting, recent task-creation entries.

For invoke-agent mode to actually create tasks, the user's MEMORY.md must include instructions for what to do when invoked with a [psd-email-triage task request] prompt. The reply must be exactly one line in either Created <system> <type> <id>: <title> or FAILED: <reason> shape — the Lambda parses this deterministically.

Output

stdout is JSON — one object per call. Shape:

{
  "ok": true,
  "subcommand": "enable",
  "summary": "Watching <email>. 3 Gmail labels created. Default rules seeded.",
  "data": { /* subcommand-specific */ }
}

On error: { "ok": false, "subcommand": "...", "error": "...", "code": "..." }.

The agent should:

• Show summary to the user as the headline
• Use data to render any follow-on cards (rule lists, training tables, etc.) via chat-card

Examples

Onboarding

node /opt/psd-skills/psd-email-triage/run.js enable --user [e-mail]

"Ignore noisy vendor X"

node /opt/psd-skills/psd-email-triage/run.js rules mute "[e-mail]" --user [e-mail]

"Always page me when the CEO emails"

node /opt/psd-skills/psd-email-triage/run.js escalation add-sender [e-mail] --user [e-mail]

"I'm getting Important on github noreplies"

# 1. Move recent github noreply messages to Later
node /opt/psd-skills/psd-email-triage/run.js training correct <messageId> later --user [e-mail]
# 2. Add a mute rule so future ones auto-go
node /opt/psd-skills/psd-email-triage/run.js rules mute "[e-mail]" --user [e-mail]

Rule 13 reminder

If the user says "set this up for the whole organisation" or anything that touches OTHER USERS' triage rows, refuse. This skill is strictly per-caller. Multi-user setup is admin work and lives at /admin/agents/[userEmail]/triage.