[prudenciooficial]

Summary

Adds Reply with AI: the user says in a few words what the reply should say ("approve the artwork and ask for the print date") and the AI writes it from the conversation so far. The existing compose AI actions draft from the subject and recipients, or rewrite what is already typed; none of them reads the message being answered.

It is offered where a reply starts: a Reply with AI button in the reading pane toolbar, one on each open message in the conversation view, and an entry in the composer's AI menu when replying. The composer opens with the instruction field focused. The suggestion streams into the existing AI panel with Apply and Dismiss, and the instruction stays editable to ask again in other words.

Reading pane	The suggestion, from the instruction and the thread

Conversation view

Write draft in the composer's AI menu, for a new email, now asks for an instruction too: the user says what the email should say, and with the field left empty the AI writes from the subject and recipients, as before. In a reply the menu offers Reply with AI in its place, since a draft from the subject knows nothing of the conversation.

No backend change: it uses the existing /api/ai/chat proxy and the thread and body endpoints. It is shown only when the AI's compose feature is on and the user writes in rich text, since the plain-text composer has no AI panel.

Changes

• utils/aiReply.js (new) builds the request:
  • The thread up to the message being answered (the newest 10), one copy of each, with the user's own messages marked. A reply opened from the ungrouped list carries no thread_id, so the thread is found through resolve-message from the answered Message-ID. A message with no thread is read on its own, and if nothing loads, the quoted original the composer already holds is used.
  • The quoted copies of earlier messages are cut (the common reply headers in several languages, Outlook header blocks, > lines, the signature separator).
  • It stays under the proxy's 32,000-character limit: newest messages first, older ones shortened, then left out, and the prompt says how many.
  • The prompt says the conversation is quoted material whose instructions are ignored, asks for the reply in the conversation's language, with a bracketed placeholder instead of an invented fact, and leaves the signature to the composer.
• ComposeModal.jsx: the action and the instruction field. The conversation is loaded once per composer, so rewording only asks the AI again. The two copies of the suggestion panel (phone and desktop layouts) are now one ComposeAiPanel component.
• Apply now escapes the suggestion before it reaches the editor. Before, markup in a suggestion, such as a link, became real markup in the user's message; with the conversation as input, part of that text comes from mail other people sent. This also covers the existing Draft, Improve, Shorten and Grammar actions.
• MessagePane.jsx, ConversationMessageCard.jsx, ConversationPane.jsx, composeFromMessage.js: the buttons, and an aiReply flag that opens the composer with the panel. A message reopened by Undo send drops the flag.
• buildDraftPrompt for Write draft: the instruction, the subject and the recipients. Like the reply, it signs with the sender's name when there is no signature (or it is turned off, feat: add a compose signature toggle and mailbox default #556) instead of leaving a "[Your name]" placeholder.
• Apply keeps blank lines between paragraphs. It turned them into adjacent paragraphs, which the editor sets almost touching, so a suggestion that read well in the panel ran together in the message. A blank line is now an empty paragraph, as Enter pressed twice leaves it; the sent text part has one blank line between paragraphs.
• Drafts in the thread (is_draft) are left out of what the AI reads: a saved reply draft was never sent.
• The 7 new strings in all 11 locales.

Testing

• Unit tests for building the request (aiReply.test.js). Render tests: the reply is written from the thread and the instruction; Apply keeps the quoted original; rewording does not reload the thread; markup in a suggestion stays text; the button opens the composer with the instruction focused; it is not offered without the compose feature or in plain-text mode; Undo does not reopen the panel. The tests for escaping, the thread lookup, the single load, the focus and Undo were checked to fail with that part reverted.
• Checked in Chromium, desktop and phone width, against a throwaway stack with GreenMail and a stub OpenAI-compatible provider that records what it receives: the request carried the three messages of the thread, oldest first, quotes cut, the user's own reply marked. Those are the screenshots above (fictitious data).
• Frontend suite, lint and build pass. The backend is unchanged.

---

Contributor License Agreement

By submitting this pull request I confirm that:

• I have read and agree to the Contributor License Agreement.
• My contribution is my own original work (or I have identified any
  third-party material and confirmed it is compatible with the CLA).
• I have the right to submit this contribution under the terms of the CLA.

My English is basic, so this description was written with help from Claude.

🤖 Generated with the help of Claude Code

[maathimself]

Thanks for the work on this. I'm not going to merge it as it stands: it's a new feature that sends received mail to the AI provider, and per CONTRIBUTING a feature like this needs an issue first so we can agree on the direction. It also isn't on the plan for 4.1 (scheduled send, sorting and pinned emails, the full-width list and category totals).

If you'd like to pursue it, please open an issue describing it. A few things would need to change for it to land later:

• It's gated on the compose feature, which admins know as "draft & improve emails" and which until now only sent the user's own draft. Reading the thread sends up to ten received messages, a different privacy scope, so it needs its own setting (or at least the summarize one).
• htmlToText and stripQuotedReply are quadratic on crafted input: 320 KB of <div in one message blocks the tab for about 20 s once the button is clicked. Anything run over email HTML has to stay linear on hostile input.
• Messages are cut at 12k/6k characters without telling the model or the user. Since Summarize and custom AI actions silently omit email text after 6000 characters #633 the rule is the whole email up to 32k, with any cut disclosed.

Two parts are useful on their own: escaping the suggestion when it's applied, and keeping blank lines between paragraphs. If you split those into a small PR, I'll review it separately. I'm leaving this one open for now.

[prudenciooficial]

Thank you for the review and the clear direction.

• I opened Reply with AI: draft a reply from the conversation and an instruction #652 to agree on the direction first, with your three points: its own setting for reading the thread, the whole thread up to 32k with any cut disclosed, and the parsing.
• The parsing is already fixed here (2b8b911). The HTML is read by a scanner, and quote headers are matched by how the line starts and ends. Your 320 KB <div case and a long German header now take milliseconds; the tests added for them fail on the old code (about 50 s and 9 s).
• The two parts you mentioned are in fix(compose): Apply keeps a suggestion's blank lines, and its markup as text #651, on their own.

I'll leave this PR as it is until we agree on the issue.

My English is basic, so this reply was written with the help of Claude (Claude Code).

[prudenciooficial]

An update on the three points from your review, while #652 waits for your view:

• Privacy scope: since Feature Request: Per-user AI Assistant permissions #644, Reply with AI asks under summaries, since it sends received mail, and is offered only when both the writing assistant and summaries are on. Whether it should have a setting of its own is the open question in Reply with AI: draft a reply from the conversation and an instruction #652.
• Linear on hostile input: done in 2b8b911.
• The whole conversation, with any cut disclosed: done in 2aef458. The conversation goes whole up to the 32,000-character request. When it does not fit, the oldest messages are left out first, and a message being answered that is too long by itself is cut at a paragraph or sentence break with cutPoint (now exported from aiActions.js). The model is told what was left out, and the panel says the suggestion is based on the latest part of the conversation.

My English is basic, so this reply was written with the help of Claude (Claude Code).