Profile Miner (The Mirror)

What it is

Profile Miner — shipped as "The Mirror" — is a background engine that reads the local copy of your Gmail that Omniscio already keeps for its email features, and quietly looks for short, factual statements about you. It is interested in things that stay true: a preference, a constraint, the way you like to work, a relationship, a goal, a problem, or a need. Each finding is stored on your own machine as a suggestion, together with a short supporting quote from the message it came from and how confident the engine is about it.

It is deliberately careful. It reads the mail copy but never writes to it, never sends mail and never syncs Gmail directly, and in this first version it writes nothing at all into your AI Coaching profile — every result stays a proposal, waiting. The whole feature ships off by default, and there is no screen for it yet: this is the engine stage, with a review surface planned for a later version.

Where to find it

There is no panel, sidebar row or settings screen for Profile Miner yet. It is a background engine, and it is off by default.

• When it is switched on, it simply runs in the background; you do not have to open anything for it to work.
• The three things you can inspect today — whether it is enabled and connected, the list of suggestions it has found, and a cost estimate for a full historical scan — are reachable through Omniscio's local control server, the same interface Omniscio's own agents use.
• Everything it finds lives on this computer only. The results are not uploaded anywhere, and signing out of Google erases them.

How it behaves

• It runs in two ways. An incremental pass drains new mail every fifteen minutes, working forward from the last message it saw. A one-off historical backfill works backwards through your whole mailbox, but only after you approve it and set a maximum dollar amount it is allowed to spend.
• Every message goes through a three-stage funnel that goes from cheap to expensive. A free first stage drops obvious bulk and machine-generated mail on its labels and sender. A cheap second stage looks only at the sender, the subject and the snippet — never the body — and drops anything below a salience floor. Only the survivors that have a readable body reach the expensive third stage, which extracts the actual signals.
• Each finding is one of six kinds: a fact, a preference, a relationship, a goal, a problem, or a need. It carries a confidence level and, at most, a short quoted line as evidence.
• Nothing is invented. If a model call fails, is cancelled, or returns something malformed, the message proposes nothing rather than a guess, and no partial or default signal is saved.
• Spend is capped at every stage. Each stage checks how much has already been spent today before it makes a call and stops at the cap, and the historical backfill can never spend past the amount you approved.
• It is idempotent. Re-reading the same thread never produces a duplicate suggestion, so a resumed or repeated run is safe.
• It is polite to your computer. It paces its work in small batches, checks how busy the machine is before each one, and only runs when the engine is switched on, its email companion is switched on, the Gmail path is in use, and a Google account is signed in.
• Privacy is enforced, not just promised. A short quote is the only email text stored; rejecting a suggestion erases that quote outright rather than hiding it; and signing out of Gmail purges every mined suggestion for that account, because the app's own database is not wiped the way the mail copy is.

For agents

The engine lives under src/main/services/profile-miner/: service.ts (the incremental tick, the backfill and the gating), prefilter.ts (the free stage), triage.ts and prompts.ts (the cheap metadata-only stage), extract.ts (the expensive stage) and miner-config.ts (the caps and tuning). The shared types and constants — the six SIGNAL_CATEGORIES, the SALIENCE_FLOOR of 2, and the 280-character EVIDENCE_QUOTE_MAX_CHARS cap — are in src/shared/profile-miner-types.ts. The store is src/main/db/queries-profile-miner/mined-signals.ts, backed by the mined_profile_signals and profile_miner_meta tables. The feature flag is profileMinerEnabled (default off) in src/shared/types/settings/ai-features-settings.ts. Signal statuses are proposed → confirmed / rejected / muted, but v1a only ever writes proposed; nothing in the renderer or the IPC layer reads, confirms or rejects a signal yet, so the review surface is still to come.

The CLI control server exposes four routes: GET /profile-miner/status, GET /profile-miner/signals (limit 1–200, default 50), GET /profile-miner/backfill/estimate, and POST /profile-miner/backfill/run with a body of { approvedCapUsd }. The three GETs are read-only; the POST is global-token-only and spends real money (triage plus extraction model calls), so it must never be called without a cap the user has agreed to.

Related

• AI Coaching sidebar — the guided-interview project whose saved artifacts inject as a "things to know about the user" profile block; the profile the miner is built to feed.
• Coaching Tips — the other half of the coaching family: contextual hint cards that nudge you.
• Teach Omniscio About Me — the guided mission that interviews you and saves an "About me" block, the manual cousin of what the miner proposes.
• Global memory — the agent-written memory library, another store of long-lived facts about you.
• Gmail integration — the Gmail inbox inside Omniscio, the account the miner reads through.
• Supermail — the email client whose local Gmail mirror the miner reads, strictly read-only.