Inline actions
Rewrite, shorten, fix or translate the lines you choose with one AI request per line, from the Grid's or the Line editor's menu.
Inline actions run one AI task on the lines you choose, such as a rewrite or a translation, without an Agent panel thread. Each line goes to the model in a request of its own, apart from Translate, and every action changes only the line's text. Inline actions don't show in the Agent panel.
Where they are
- AI in the Grid's right-click menu, after Paste special, runs an action on the selected lines.
- AI in the right-click menu of the Line editor's text, and the AI button at the end of its strip of tag buttons, run one on the active line. With text selected in the box, only that text changes, and the rest of the line goes as context.
Two commands have no key at first. Give them one in File › Settings › Shortcuts:
- AI actions… opens the AI menu at the active line, or at the Line editor's AI button when you're typing in it.
- Repeat last AI action runs the last action again on the lines you have chosen now.
With Turn off all AI features ticked in File › Settings › AI, the menus and commands are hidden. An action is on when its role has a model in File › Settings › AI › Models, or when Chat has one, since a role without its own model uses the Chat model. An agent such as Claude Code doesn't count. With no model at all, the menu's last item is Add a model in Settings › AI…; with one, it is Edit actions….
The actions
| Action | Model | Asks |
|---|---|---|
| Rewrite… | Chat | How to rewrite it, such as less formal, fewer words. Your last five instructions are there to pick again |
| Shorten to 15 CPS | Chat | Nothing. It sends only the lines over 15 characters per second, the limit the Grid and Line editor show |
| Fix spelling and grammar | Reviewer | Nothing |
| Translate to English | Translator, then Reviewer | Nothing. The language is Translate to in Settings |
| Translate to… | Translator, then Reviewer | A language |
| Your own actions | The one you chose | Nothing, or their instruction when Ask each time is on |
Each model is the one you chose for that role in File › Settings › AI › Models. A role without a model uses the Chat model.
Comment lines, empty lines, lines of tags only, drawings and karaoke lines aren't sent, and HikariSub says how many it left out. Override tags stay where they are: each tag block goes as a numbered marker, and the answer must give every marker back once, in order. If it doesn't, the line is left as it is, with Not changed, because the answer dropped a tag. An answer the same as the line isn't a change.
Translating
Translate runs a translator and a reviewer over your lines, and keeps names and terms from Honya's GLOSSARY.md and CHARACTERS.md. Both Translate to actions work this way, on one line or many.
- In chunks. The lines go in Grid order, in chunks of about 1,000 tokens, and a line is never split. Each chunk goes with the glossary terms and characters it mentions, the style notes, and the last five lines translated.
- Checked. HikariSub checks each answer first: every line answered, every tag marker back once and in order, and the glossary's fixed terms kept. Then the reviewer approves the chunk, or sends it back with what to fix, up to three times. When the tries run out, the best one is kept, and its lines need review.
- Across chunks. With more than one chunk, the reviewer then reads the whole run for names, honorifics or terms written two ways.
- The glossary grows. An approved chunk adds the new terms and characters it used. An entry that is already there is never changed.
The Translator model translates, and the Reviewer model reviews, or the Translator's if Reviewer has none. Translate sends no lines of context; it sends each line's actor instead. A Project's source language goes with the lines; without one, the models detect the language. Its estimate counts three passes over the lines, and retries cost more.
The result
- In Translation mode it fills the Translated text, and the Original text stays as it is. A line the reviewer didn't approve is marked Unconfirmed, and the bar says so: The reviewer didn't approve line 4, so it's marked Unconfirmed.
- Otherwise it replaces the text, as the other actions do, and the bar lists the lines the reviewer didn't approve: Check it before you use it.
- In Manual, each result the reviewer didn't approve says Not approved by the reviewer: and why.
The glossary files change only after the lines are written: when the run ends, or on Apply in Manual.
The glossary
The glossary, the characters and the style are Honya's project files, PROJECT.md, GLOSSARY.md, CHARACTERS.md and STYLE.md, so Honya and HikariSub share them. HikariSub keeps everything in them it doesn't use. They are in:
- a Project's folder, or the folder inside it that
glossarynames in the[translation]table of.hikari/project.toml, which can't lead outside the Project; - outside a Project, the folder that the file's
Honya Projectline in[Script Info]names, from the subtitle file's folder, such asHonya Project: ../honya.
Without either, Translate works with no glossary, and the terms it finds aren't saved.
A Project's models
A Project can name its translator and reviewer models in .hikari/project.toml, as "<provider id>/<model>", but never a key:
version = 1
[translation]
glossary = "honya"
translator = "openrouter/google/gemini-3-flash"Here glossary puts the glossary files in the Project's honya folder, and reviewer takes the same form as translator. The provider id is the Id under Name in the provider's form in File › Settings › AI: openrouter, openai, anthropic, gemini or custom, then -2, -3 and so on for more of one kind, such as openrouter-2. A new provider's form shows the id it will get.
- The provider must be one you added in File › Settings › AI; otherwise your own model runs.
- They apply only once you trust the Project's folder. Until then, your own models run, and the bar says This project's translator and reviewer models are used once you trust its folder; your own ran instead.
- An entry whose name looks like a key, token, password, URL or endpoint, or a value that holds an address or a key, is never read, and the Log warns about it. Keys stay in your keychain.
Applying the results
Inline actions follow File › Settings › AI › Agent › Mode for new threads:
- Auto and Full access: when the run ends, all its changes go into the file together, as one turn in the review and one step in the undo history, such as AI: Rewrite (10 lines). Each changed line waits for Keep or Undo like the Agent panel's edits.
- Manual: the results wait in a list over the Grid, one row per line, each ticked. Click a row to select its line in the Grid. Apply 3 (the number of ticked lines) makes those changes, the same way. Discard makes none. A line you edit after it was sent is marked Out of date and can't be applied.
Running
One action runs at a time. The bar over the Grid shows how far it has got, such as 4 of 10 done, and the cost so far. Stop, or Esc, keeps the lines that are done and drops the rest. Editing a line while it is out drops it from the run.
An action on more than 50 lines asks first. It says how many requests it will send, to which provider and model, about what they will cost, and how much is left of today's cost cap. Inline actions count toward the daily cost cap, but not a thread's. Reaching the daily cap stops the run as Stop does.
A line that fails shows under the run with the reason. Retry runs the failed lines again.
Settings
File › Settings › AI › Inline actions has:
| Setting | At first | Does |
|---|---|---|
| Your actions | None | Add an action gives it a Name, an Instruction, a Model (Chat, Translator or Reviewer) and Ask each time, starting from this instruction. Move up, Move down, Edit and Delete manage the list. Your actions show in the AI menus |
| Translate to | English | The language of the Translate to action. A Project's target language comes first in its windows |
| Lines of context | 2 | How many lines before and after each line go with it, 0 to 5. 0 sends none. Translate sends none |
Your actions save at once. The other two settings save with OK or Apply.
What is sent lists everything a request carries.
Skills
Give the agent instructions for kinds of work, from HikariSub, your user folder or a Project, and see in Settings which skills a window's threads get.
Connect Claude Code or Codex
Let Claude Code, Codex or another MCP agent read your open subtitle files and edit them through HikariSub, with every edit waiting for you to keep or undo it.