LiteressSeptember 27, 2026

Literess works on her own, and every change can be checked

From the Transept team

An illustration on a dark background, headed “She’ll take it from here.” On the right, Literess, drawn as a blonde chibi girl with a gold hair clip, frowns in concentration with her tongue out, writing in an open book. On the left, a stack of app cards: a goal strip reading “Goal · polish chapter three, 37 / 42” with a nearly full progress bar and an “Allow All” pill on its corner; two cards marked “Applied automatically”, one changing “Not without effort,” to “With some effort,” in block 36 and one restoring the italics on “expensive” in block 37, with more cards piled behind them; and at the bottom a “Changes you can check” card reading “Literess · 37 changed” with a Review button. — With Allow All on, Literess works through a goal on her own, and everything she changed waits under “Changes you can check”.

This note covers everything that shipped between mid-August and the end of September, so it runs longer than usual. Most of it is about two questions: how much Literess can do without you pressing a button for every step, and how you check the result once she, a workflow, or a proofread has changed your text.

Underneath, runs got faster and no longer wait behind other people’s, the proofread scan finds more, and scans are now billed for the reading they do rather than as if they were translations.

Added

Allow All. With it on, the changes Literess proposes apply the moment she makes them, and she keeps working through the task instead of stopping for your click after each one. Every chip still appears in the chat, marked “Applied automatically”, as the record of what she did. Inviting a teammate and deleting a workflow always ask first. Switch it under Settings → Notifications or with the sparkle button in the Literess panel header; it is off by default. See Allow All & goal runs.
Goal runs and a Stop button. Give Literess a long job (“translate this whole document”, “check every heading against the glossary”) and she keeps going turn after turn until it is done, then rereads the result before she calls it finished. Each run has an allowance of turns and compute, and Pause and Resume sit on the goal strip. While she writes, the send button becomes Stop: work already applied stays applied, and only what she had not started is dropped.
Literess runs workflows on any document, and tells you when they end. Name a document in the chat, from the dashboard too, and she can start a workflow on it; the chip names the target. When a run she started completes, fails, is cancelled or pauses for review, she says so in the conversation she launched it from. She can also stop a run in flight and delete a saved workflow, both on a chip. Customize on a run chip opens the usual run dialog, so you can change the scope or turn on parallel mode before confirming.
She takes the chat with her. Asked to open a document or an app page (workflows, a specific glossary or style guide, settings, billing, teams, CRM), Literess moves the current tab there on a chip, the conversation comes along, and she carries on with the task on arrival. A confirmed move stays clickable in the chat history, so you can follow it later. See What Literess can do.
Her replies reach you when you are elsewhere. A reply that lands while you are on another document shows a toast with Open; in another tab or app you get a browser notification and a count in the tab title; with no Transept tab open, a push notification, which opens her chat when clicked.
Literess searches your translation memory for how a phrase was translated before, and she can see how far along each document is, how many chats it has, and its recent runs. In the editor, her panel now opens with Translate this doc on an untranslated document and Polish this doc on a finished one.
Changes you can check. Any change to text you already had that you did not type yourself (Smart or Vision Proofread, a workflow’s proofread or rewrite step, Regenerate, a rewrite with an instruction, Literess under Allow All) can now be reviewed afterwards. Each run adds a row under Changes you can check above the first block, such as “Smart Proofread · 3 changed”, and the run’s finished toast has Review changes. Each change shows the note it was made for, with Keep or Revert to previous. Reverting replies on that note, so the next proofread treats the point as settled instead of making the same change again.
One review screen for every review. Workflow findings, a Smart Proofread, the edits Literess proposes and a document’s open comments all open on the same screen. You choose a list or one card at a time, one suggested fix or three options, and Fast, Standard or Pro, with the credit cost shown before anything is generated. Fix all writes fixes onto the cards and leaves them for you to check. Your picks, edits and generated suggestions are kept, so you can leave and pick up where you stopped. See Reviews waiting on a document.
Get a fix on any comment. A comment’s card, in the margin or in the Comments panel, has Get a fix, with the cost shown first. The suggestion appears on the block, and Apply writes it as a new version and resolves the comment. A workflow flag that came with a suggested fix carries it on its comment the same way. Opening a thread in the panel now also selects it in the margin, with its card level with its block.
Everything waiting on a document, in one list. The Review button in the editor is replaced by a list above the first block that appears only when something waits: a run’s findings, a glossary review you saved, a generated style guide you never applied. Closing a glossary auto-build review now asks what to do with it, and Save for later keeps your ticks and reworded hints so you can continue from that list. See Build from a whole document.
Project defaults for glossaries, style guides and memory. A project can now carry default glossaries, style guides and translation memory settings, and sub-projects inherit whatever they leave unset. Glossaries and style guides are copied into a document when it is created in the project; memory defaults are read live on every run, so changing them re-tunes every document in the project that has no memory settings of its own. See Projects.
Joining hard-wrapped lines. Text pasted from a PDF or another paginated source often arrives with one block per printed line, and the model then translates line fragments. Transept now checks for this on every import route (upload, paste, Notion, the API) and offers to join the lines back into paragraphs. It is a suggestion you accept with a click, and it is not offered once a document has translations or comments.
Sort the dashboard by last updated, name or date added, in both the personal and team views. The choice is remembered.

Changed

The Free plan runs batch actions and workflows. Free accounts spend their 1,500 credits however they like: a selection, a page, the whole document, or a workflow. Only the model is limited: Fast and Standard, with Pro on paid plans. A document or workflow step set to a model your plan does not include now runs on Standard instead of refusing, and the templates and preset workflows that asked for Pro now use Standard. See Plans, credits and the free tier.
Scans are billed as scans. A proofread charges a read fee of 0.2 credits per source word it reads (0.5 with vision) plus the normal rate on the blocks it actually fixes; a find run charges a small read fee by mode and tier. Before, a flag-only scan could cost as much as translating the document. Vision proofreading now runs only where there is layout to look at: on a prose-only Word file or plain text it runs as a text proofread at the text rate, and the run dialog says so. Lines in your credit history now say what they were for, such as “Scanned 4,200 words”. See Smart Proofread and Vision Proofread.
Blocks Literess translates herself bill the premium rate, 3 credits per source word, since she writes with the premium model. They were free before. Rejected proposals cost nothing. She can also propose up to 50 blocks at once, each with its own row on the chip.
Chinese and Japanese are counted per character. Each Han or kana character counts as one word, the convention CAT tools use. A Chinese sentence used to count as a single word.
Standard mode runs on a newer, faster model, and every run dialog now starts on Standard unless you pick something else.
The proofread scan finds more. It runs on the Pro-tier model again. On the same document it flagged about twice as many blocks as the faster model, mostly smaller issues such as calques and register, and blocks the fix step judges fine are still not billed.
Runs start sooner. Up to five of your runs can now work at once, and one person’s large run no longer holds up anyone else’s; before, three runs were shared across all users. Translation memory lookups on large memories are several times faster, and a run in progress carries on through a Transept update.
Memory context for passages with no close match. When no memory entry is at least 30% similar to a block, the two closest passages still go into the prompt as related context, without a match percentage. In fiction this roughly doubled the number of blocks that get any memory at all. See Translation memory.
Invitations have a daily limit. Team invites, review invites and a workflow’s emails to a client share one daily limit per person, and a name, team name or document title containing a link cannot be used to send one. Adding someone you share no team or document with now sends them an invitation instead of adding them directly. See Teams and invites.
Notifications stack. Two or more toasts collapse into a pile and open on hover into a scrolling column, and Clear all dismisses the lot while leaving a running job’s progress in place. The ✕ on a single toast now closes it; before, it opened the document behind it.
Read in context is gone. Every flag is already a comment on its block, with its suggested fix, so the separate read-only view was removed.
The editor works on a phone. Translation memory, the selection bar, the review screens, the glossary candidate table and the formatting toolbar were all reworked for a narrow screen.
Glossary terms are listed alphabetically by source term everywhere they appear.

Fixed

Smart Proofread came back empty. From 2 to 5 September most proofread runs finished with no fixes, because the scan’s findings were discarded when they arrived without a note. A finding without a note now still gets its fix.
Accepting a review fix creates a version. Accepting used to overwrite the active translation in place, with no earlier version to switch back to. A block with several flags could also lose its fix and leave a comment open.
Literess keeps formatting when she edits a block. Bold, italics, links and line breaks were lost on blocks she rewrote by hand.
Glossary auto-build gives every term a hint again. Candidates were arriving without context hints, and builds into an existing glossary could fail outright; a reply cut short now keeps the terms it completed. Save for later no longer reports success when the save failed.
A block refused by the AI provider’s content filter no longer fails the run. Refused blocks are retried one at a time, the rest of the document translates normally, and any block that still fails says why.
Style guide notes no longer show raw markdown such as ### or asterisks when Literess or a generation wrote them.
A saved review no longer disappears when you open it straight after its run finishes.
Compact works on short conversations that carry a large document; the button used to do nothing there.
Project defaults apply on every route. Notion imports and documents created from a template ignored the project’s default glossary, style guide and language pair.
Security. An internal audit found and closed gaps in profile updates, Google account linking and HTML previews. Checks on production found no sign that any of them was used.

Have a request or hit a snag? Ask Literess in the app, or write to [email protected].