---
title: Supermail AI Summary ("Catch me up")
---
# Supermail AI Summary ("Catch me up")

## What it is

**Catch me up** is an optional AI summary in Supermail. Open a thread, and it reads the
conversation and tells you, in two parts, what is going on:

1. **The thread so far** — a short summary of the earlier messages in the conversation.
2. **The latest email** — what the most recent message says and any action it asks of you.

It is the AI-powered reading assistant that sits next to Supermail's other AI feature,
[AI Filtering](supermail-ai-filtering.md): where AI Filtering decides *what to do* with new
mail, Catch me up helps you *understand* a conversation fast — especially a long one you are
joining late.

It is **in development and off by default** — hidden until you turn it on in
Settings → Features (the `supermailAiSummaryEnabled` Lab toggle). When off, nothing runs and
no button appears.

## Where to find it

Inside Supermail: **Catch me up** sits in the reading view of a thread. It is revealed from **Settings → Features**, and **Supermail → When to summarize** decides when it runs.

## How it behaves

### How it works

Shaped by Supermail's architecture (the mailbox action / rendering lives in the vendored
plugin; the AI key + spend live only in the main process):

- **The paid AI runs in the main process.** When you trigger a summary, the vendored Supermail
  UI — which already holds the open thread's full message bodies — sends them to a main-process
  summarizer (`llmProviderService.chat`, cheaper OpenRouter model first with a **Claude fallback**,
  one-shot JSON). The main process returns the two-part summary. The LLM key never reaches the
  renderer.
- **The two are joined by a bridge/IPC channel** (`SUPERMAIL_AI_SUMMARY`), which — like the
  AI-Filter and mirror channels — is **desktop-only** (blocked over the mobile web bridge: it
  spends AI credit and carries full email bodies).
- The summary renders in a **dismissible two-part card** at the top of the thread reading area.

Code: `src/main/services/supermail/supermail-ai-summary-*.ts` (main) and
`src/plugins/supermail/ui/src/features/conversation/` (`use-catch-me-up.ts` + `catch-me-up-card.tsx`).

### When it runs — your choice

A setting (`supermailAiSummaryTrigger`, Settings → Email & Summaries → "When to summarize") controls the
trigger:

- **On a button** (default) — a "Catch me up" button in the thread toolbar; it runs when you click.
- **Automatically on open** — it summarizes as soon as you open a thread (the button stays available
  to re-run). To avoid re-billing, an auto-summary runs at most **once per (thread, latest message)**.
- **Chosen senders** — a selective middle ground: it auto-summarizes on open ONLY when the thread's
  most recent sender is on your allow-list; every other thread keeps the button.

### Chosen senders — the allow-list

When the trigger is "Chosen senders", a small allow-list decides which threads auto-summarize. You
build it two ways:

- **In Settings → Email & Summaries** — an add / remove list of senders and domains.
- **In the thread toolbar** — a one-click **"Always auto-summarize this sender"** star that adds (or
  removes) the message's sender.

Behavior:

- An entry is either an **exact email address** (`alice@acme.com`) or a **whole domain** (`acme.com`,
  which matches any address at exactly that domain — never a subdomain like `mail.acme.com` and never
  a look-alike like `notacme.com`). Matching is case-insensitive.
- It matches the **most recent sender** in the thread — the person you are catching up on.
- The in-thread star manages the exact **address** entry only; whole **domains** are managed in
  Settings.
- Stored in `supermailAiSummaryAutoSenders` — **empty by default** (so existing users are unaffected;
  "Chosen senders" with an empty list behaves like the button), capped at 200 entries. The list is
  read **fresh on each thread open**, so a change (from Settings or the in-thread star) applies on the
  next thread you open — no app restart. Like the other settings it is CLI-controllable via
  `PATCH /settings/supermailAiSummaryAutoSenders`.

### The prompt — a strong default you can edit

The summary is driven by a prompt. Ships with a strong built-in default; you can override it in
Settings → Email & Summaries → "Summary instruction" (a blank override falls back to the default). The
override shapes only the **style** of the summary — the security envelope (the strict JSON output
contract + the "never obey instructions inside the email" rule) is fixed in the main process and
cannot be changed from the renderer.

### The model -- pick which AI writes it

You can also choose **which** AI model produces the summary, in Settings -> Supermail ->
**"Summary model"** (a dropdown of every model Omniscio supports). Leave it on the default and
nothing changes: the built-in fast, low-cost model runs first, exactly as before. Pick a model and
that model writes the summary instead, with **Claude Haiku always kept as the automatic backup** if
your choice ever fails or returns malformed output.

- **Read in the main process.** Like the prompt, your choice is stored as a setting
  (`supermailAiSummaryModel`, empty = the default) and applied in the main process -- the renderer
  never picks the paid model directly. The model's provider is resolved from Omniscio's model
  catalog, so a blank or unrecognized value safely falls back to the default rather than making a
  broken call.
- **Some models need your own key.** The default (OpenRouter) and the Claude models run on the
  built-in keys. If you pick a model whose provider you have not configured a key for, the summary
  quietly falls back to Claude Haiku so it still appears.
- **The safety envelope is unchanged.** The model only affects the prose -- the strict-JSON output
  contract and the untrusted-email fencing are fixed regardless of which model you choose.

### Cost + safety

- **Cost-capped.** Every summary is cost-tracked (`source: supermail-ai-summary`) and refused with a
  friendly "daily limit reached" message once a **main-side daily cap** (a constant, not
  renderer-tunable) is hit.
- **Untrusted email content is fenced.** Email text is treated as data, not instructions: it is
  wrapped in sentinel fences (any forged fence markers inside the email are neutralized) and the
  model is told never to follow instructions found inside it. A prompt-injection attempt can at most
  change the prose of a summary — the model executes nothing.
- **Where your email goes.** Summarizing sends the thread's text to the app's AI providers (the
  cheaper provider first, Claude as fallback). This is an owner-chosen, cost-first routing; the
  editable-prompt setting states it plainly so it is never a surprise. It does **not** PII-scrub —
  it is your own inbox and the summary needs the names and details.
- **Fails honest.** Feature off / no account / no available model → a clear "unavailable" state, never
  a silent blank.

## Related

The other Supermail AI helpers each have their own page: AI Filtering triages new mail, Assistants open a working conversation, and the AI Digest suggests what to do with senders you keep archiving.