Chat, edit and agent
Sirius is the default participant in the editor’s chat view, so there is nothing to invoke: open the chat (Ctrl+Shift+I), type into Describe what to build, press Enter. The mode selector in the input offers Ask, Edit and Agent.
What happens to a message
Section titled “What happens to a message”Every message runs the same loop against the model you picked (or the chat view’s own choice):
- Sirius sends the model its system prompt, your project rules,
the file you have open with the selection or the lines around the cursor, the earlier
turns of the conversation, and what you attached: the text of attached files and
selections (up to 100 KB each, and about half the model’s context window in all; what
does not fit, and binary files, go by path), and attached or pasted images for a
model that can see. Instruction files the editor attaches —
AGENTS.md,CLAUDE.md(including your personal~/.claude/CLAUDE.md, under the editor’schat.useClaudeMdFilesetting),*.instructions.md— join the instructions once, not the message. - The model answers, or asks for a tool. A tool call shows as Running read file… in the transcript; the result goes back to the model, and the loop continues — for up to 25 tool rounds per message. After that Sirius stops with Stopping here — this task took more tool rounds than expected. Say “continue” to keep going.
- Tool output is cut at 24,000 characters so one large file cannot swamp the context.
Small local models that write a tool call as text instead of using the tool API are caught: Sirius recognises up to six such calls per round and runs them as real tools.
The tools
Section titled “The tools”Sirius’s own, always offered:
| Tool | What it does |
|---|---|
readFile |
Reads a workspace file, optionally a line range (large files: ask for a range) |
searchFiles |
Searches workspace files for a literal string, like grep; an optional glob narrows the files |
listDirectory |
Lists a directory, folders first; skips node_modules, .git, dist and the like |
problems |
The current errors and warnings, optionally filtered by path |
edit_file |
Replaces one exactly-matching span in a file |
create_file |
Creates a file that does not exist yet |
edit_file insists the search text occurs exactly once; if it does not, the model is told
to read the file again or include more context, and nothing is changed.
Every one of these stays inside the workspace: a path that resolves outside the open
folders — ../../etc/passwd, /home/you/project/../.ssh/…, or a symlink in the project
that points elsewhere — is refused with … is outside the workspace, for reading,
searching and writing alike. For the web, the agent uses the
integrated browser tools.
The editor’s own tools, by tier. The workbench provides terminal, task, test and browser tools; Sirius offers them in two sets so that a model is never handed more schemas than it can hold:
- Core — every model: run a command in the terminal, read terminal output, keep a todo list.
- Extended — adds the rest of the terminal and task tools, test running, sub-agents, plan review and questions back to you, and the integrated browser tools.
The tier is chosen from what Sirius knows about the model: a model without a tool API gets core; a local model of 4 GB or more gets extended; a model whose size is unknown gets extended if its context window is at least 32,000 tokens. Hosted models get extended.
Commands the agent wants to run in the terminal go through the editor’s own approval flow — the prompt you see is the workbench’s, with its allow-once and allow-always choices.
Edits come back as diffs
Section titled “Edits come back as diffs”edit_file and create_file stream their changes into the editor’s chat-editing
session, which is what gives you:
- the change shown as a diff in the file, hunk by hunk;
- Keep / Undo per hunk and per file, and Keep All Edits / Undo All Edits in the chat;
- Open Changes in Diff Editor;
- checkpoints: Restore Checkpoint, Restore to Last Checkpoint and Start Over from the conversation, which roll the files back to how they were before a turn.
Ask, Edit, Agent
Section titled “Ask, Edit, Agent”The three modes are the editor’s, and each gets a different set of tools:
| Mode | The model can |
|---|---|
| Ask | read and search the workspace and see the problems list — it explains and answers, and cannot change a file or run a command |
| Edit | all of Ask, plus edit_file and create_file |
| Agent | all of Edit, plus the editor’s terminal, task, test, sub-agent and browser tools, by tier (above) |
Pick per message; it sticks until you change it. The conversation carries over when you switch. A tool the current mode does not offer is refused even if a model asks for it by name.
Thinking
Section titled “Thinking”Models that reason (Claude, Gemini, and local models whose server declares thinking —
🧠 in the picker) are asked to, at the effort you set (Sirius: Set Thinking Effort
Level); Sirius: Toggle Thinking turns it off. The reasoning shows in the transcript as
a collapsible Thinking section above the answer.