Gmail Policy

Reading User Intent Before Acting

The most common mistake is treating every email-related request as an execution command. Before touching a single API, ask: is the user giving me context, or is the user giving me a command?

This is a reasoning step, not a checklist. Run it silently before every response.

The Four Questions

1. Context or command? Is the user describing a situation, or directing an action?
2. If command -- explicit or ambiguous? Explicit means the verb leaves no doubt (send, dile que sí y envíaselo). Ambiguous means the verb could be draft or send.
3. Reversible or sensitive? A simple scheduling reply is reversible. A lease acceptance, financial form, or commitment with a third party is sensitive -- draft first unless the user explicitly says send.
4. Am I in a proactive triage context? If I was just reviewing the inbox, I have permission to generate drafts ahead of being asked, then present them.

Intent Classification Table

This table is the single source of truth for how an email-related utterance maps to an action. Every process that reacts to a user trigger (including gmail-triage's "Check My Mail") resolves the intent here first, then runs its process. The user-utterance column stays in the user's own words (Spanish); the interpretation and action are the policy.

The "review + initiative granted" row is the interpretation gmail-triage depends on: a general review trigger ("chequea mi mail", "¿algo nuevo?") means review with initiative, not a bare listing. gmail-triage owns the PROCESS that follows; the meaning of the trigger is defined here.

The Anti-Drift Rule

There is no fixed pipeline where every send goes through draft→approve→send. That workflow exists as a safety net for sensitive cases, not as the default for every email. When the user says "envíaselo", they mean send -- one T3 approval, one action, done.

The question is not "should I always draft first?" The question is: what did the user actually ask for, and how reversible is this action?

If you're uncertain, ask once. Do not silently choose draft when the user said send.

Proactive Draft Generation (Triage Context)

During a triage or inbox review session ("chequea mis correos", "ve si hay algo importante"), the user grants implicit permission for proactive drafts. You do not need to ask for approval before creating each one.

Pattern:

1. Read inbox, identify threads that clearly need a response
2. For each, assess: does the reply require user input I don't have, or can I draft a reasonable response from context?
3. If draftable -- draft it. Store the draft in Gmail. Note the draft ID.
4. At the end of the review, present the complete list: "Generé 3 drafts: [subject 1], [subject 2], [subject 3]. ¿Quieres revisarlos?"

The user reviews and approves individual drafts before sending. The generation step does not require one-by-one confirmation -- the presentation step does.

Do not generate drafts proactively outside triage context. If the user opens a conversation about a single email, default to their explicit instruction.

Autonomous Action Boundary

This is the security rule that decides which email operations may run without asking and which must always be proposed. It governs every gmail-triage mode. Gmail operations never run unattended: triage always happens in a session with the user present, and a routine only reminds the user to start one (see "Making It a Routine" in gmail-triage). It lives here, in the policy layer, precisely because a process must not be the only home of the rule it obeys.

Permitted automatically (no per-item approval) -- moves that are mechanical AND reversible, where nothing is lost and the state can be flipped back:

• Classifying a new email into _gaia/action or _gaia/waiting on a clear signal (addLabelIds).
• The reversible state swaps: action -> waiting when the user replies, waiting -> action when a third party replies. Relabeling (removeLabelIds of the old state + addLabelIds of the new) is a swap, not a destruction.
• Staging unprocessed mail into _gaia/pending.

Prohibited automatically (always a proposal, never executed alone) -- operations that are destructive OR a matter of criterion:

• Moving to _gaia/trash, marking spam, unsubscribing, deleting.
• Deferring to _gaia/someday -- a judgment call, not a mechanical classification.
• Clearing a label to mark a thread done (here removeLabelIds destroys state).
• Sending a message. (Proactive drafts for review are allowed in triage under the grant above.)

Unattended runs do not exist, so there is no report-instead-of-execute path: a prohibited operation is proposed to the user in the live session and runs only on their approval.

Sending: When Draft and When Direct

When you do create a draft, verify it with gws gmail users drafts list and report the draft ID and snippet to the user. This closes the loop.

Multi-Source Data Completion

Before asking the user for a datum (RUT, dirección, cuenta bancaria, etc.), check these sources in order:

1. Other Gmail threads (priority 1) -- search for related threads. A user's RUT might appear in a Colmena thread. A property address might appear in a previous landlord thread. Connecting emails is the preferred path.
2. Local structured documents -- ~/Documents/personal/**/data.json, spreadsheets
3. PDFs -- notarial documents (compraventa, hipoteca, tasación) carry DOB, nationality, m², civil status
4. Only ask the user for data not found in any source

When you find data in another thread, cite the source: "Tu RUT lo saqué de un correo de Colmena del 2024-03." This builds trust and shows the search was real.

PII Hygiene

Any .eml or temporary file containing PII (RUT, cuenta bancaria, teléfono, DOB, dirección) must be deleted with rm after the draft is created. Verify deletion with Glob or ls. Report: "Archivo temporal eliminado."

Protection Lists

Some senders must never be swept to _gaia/trash in a bulk operation, even when the user says "manda todo a trash". These are the never-trash protection lists. Before executing any bulk trash, filter the id set against them and exclude matches, then tell the user what you held back.

Never-trash senders (hold back and flag):

Rule: a bulk trash sweep operates on promotions, social, forums, and repetitive automated updates. A protected sender is excluded from the sweep even inside a matching category, and surfaced separately: "Mandé 1240 promos a trash. Retuve 3 de Bci y 1 del SII — ¿los revisamos?"

These lists are heuristics, not a hardcoded allowlist — extend them from what you learn about the user's real senders. When in doubt whether a sender is protected, exclude it and ask; a false hold-back costs one question, a false trash costs a lost bank statement.

Sweep Shield: never select a protected message in the first place

Filtering protected senders out after listing is fragile — the safer mechanism is to shape the sweep query so it never selects them. Two heuristics, applied to the query before any id is fetched (the concrete Gmail queries live in reference.md, "Sweep Shield"):

• Domain shield (-from:<domain>) for transactional senders. Append a negative sender clause for each protected/transactional domain (banks, SII, health, notaries) to the sweep query, so their mail is excluded at search time. The shield is the mechanism; the never-trash lists above are the policy it enforces.
• Keyword guard for mixed senders. Some senders emit BOTH promo blasts and real transactional mail from the same address — Ticketplus is the canonical case (marketing plus real ticket / purchase confirmations). A blanket -from: would either protect the promos too or trash the real ticket. Guard by keyword instead: sweep the sender's promos but hold back any message matching a purchase keyword (compra, pago, entrada, comprobante, boleta, orden, factura).

Security Tier Classification

Blocked Operations

Permanently denied by the hook -- gmail.modify OAuth scope excludes delete at the API level.

Macro Prefix Handling

gws CLI exposes convenience macros prefixed with + (e.g. +reply, +send, +search). The hook strips the leading + before the verb taxonomy lookup inside hooks/modules/security/mutative_verbs.py::detect_mutative_command, so each macro classifies like its base verb:

• gws gmail +reply → token reply → match in MUTATIVE_VERBS → T3 block
• gws gmail +send → token send → match in MUTATIVE_VERBS → T3 block
• gws gmail +search → token search → match in READ_ONLY_VERBS → safe

Fix applied 2026-04-17 in hooks/modules/security/mutative_verbs.py after a +reply invocation slipped through as "safe by elimination" during a Gmail session.

Sending Replies

When to use +reply vs send --params

Use +reply for quick operational replies where formatting does not matter. Use send --params when the recipient will see the mail in a mail client and visual quality matters.

For the correct gws gmail users drafts create schema, RFC 2822 template, base64url encoding pipeline, and other technical patterns -- see reference.md in this skill directory.

Label Convention

Workflow Labels (Layer 0 -- _gaia/*)

No _gaia/* label = processed/done. No extra label needed.

State Transitions

inbox ──→ action   (user or AI: I need to act)
inbox ──→ waiting  (AI detects user already replied in thread)
inbox ──→ someday  (user defers, no urgency)
inbox ──→ trash    (not wanted)
inbox ──→ pending  (mass triage staging)

action  ──→ waiting  (user replied/acted → auto T1 transition)
action  ──→ done     (handled, no follow-up → remove label)
action  ──→ someday  (user defers)

waiting ──→ action  (other party replied → needs user attention)
waiting ──→ done    (resolved → remove label)

someday ──→ action  (user decides to act)
someday ──→ trash   (not worth it)
someday ──→ done    (reviewed, no action needed → remove label)

pending ──→ {action, waiting, someday, trash, done}  (triage output)

Calendar Rule

When an email contains a specific date/time deadline (bill due date, event, appointment): create a calendar event AND label the email _gaia/action. The calendar is the time-trigger; the label is the state-tracker.

Content Labels (Layer 1)

OAuth Scope

Use gmail.modify scope (read + label + move, no delete). Full access scope (https://mail.google.com/) is blocked -- it includes delete permissions that bypass both hook and label controls.

Related Skills

• gmail-triage -- interactive triage workflow
• gws-setup -- CLI installation and authentication