EditVolt Chat
Open the Assistant from the activity bar, or run EditVolt Assistant: Focus on Chat View from the Command Palette. Every conversation runs in one of six modes — pick how much autonomy you want from the mode chip in the composer, or cycle them with ⇧Tab.
| Mode | What it does | Touches your files? |
|---|---|---|
| Ask | Questions and explanations, with grounded retrieval | Never |
| Plan | Designs an approach step by step | Only after you approve |
| Debug | Evidence-first loop — hypothesize, instrument, reproduce, fix, clean up | Staged diffs you approve |
| Agent | Executes tasks end to end with tools (the default) | Staged diffs you approve |
| Design | Agent plus a live in-editor browser it can see and drive | Staged diffs you approve |
| Multitask | Decomposes the request into 2–6 subtasks and runs them as parallel workers | One consolidated review |
Ask and Plan are read-only by design — nothing changes until you explicitly approve.
Set the mode a new conversation starts in with product.modes.default (default agent).
@-mention context
Type @ in the composer to open the context picker. The categories are:
| Category | What it pins |
|---|---|
| Files & Folders | Search and pin files or folders from the workspace |
| Workspace | Hybrid semantic + keyword search over the repo before sending |
| Symbol | A workspace symbol's definition |
| Commit (Diff of Working State) | Your current Git working-tree diff vs HEAD |
| Branch (Diff with Main) | Commits on your branch since diverging from main |
| Git Ref (Diff vs) | Any branch or tag as the diff base |
| Commit (Diff vs) | A commit SHA to diff against HEAD |
| Pull Request (Diff) | An open GitHub PR, via the gh CLI |
| Problems | Editor diagnostics — errors and warnings |
| Docs | Your indexed documentation library, or a pinned URL |
| Past Chats | A previous conversation's transcript |
| Terminals | An open terminal's recent output |
| Browser (Fetch) | Lets the turn fetch live web pages and URLs |
| Browser | Browser tools — inspect elements, console logs, network traffic (needs product.browser.enabled) |
| Memory | Search persistent EditVolt memory |
| Skills | Run a workspace skill from .editvolt/skills |
You can also paste, drag, or attach images — vision works across all connected providers. Attachment size limits are configurable (product.composer.attach.imageMaxMb, …ipynbMaxMb, …referencedFileMaxMb).
The base branch used by Branch (Diff with Main) is product.context.changesBaseBranch (default main).
Reasoning effort and output style
Two composer-level controls shape the answer:
- Effort (
product.reasoning.effort, defaultmedium) — how hard the model thinks.autoalso serves medium;lowandhighpin a level across turns. - Output style (
product.assistant.outputStyle, defaultdefault) —concise,explanatory, orstructured.
Dictation
Press ⌘⇧Space to dictate into the composer. Speech-to-text runs through Chromium's engine or a local Whisper binary (product.voice.sttEngine: auto | chromium | whisper). Set product.voice.autoSubmit to send automatically when you say one of your product.voice.submitKeywords. Turn the whole feature off with product.voice.enabled.
Reviewing what the AI did
Every agent edit is staged as a diff, never auto-applied. Use Review My Changes for a unified multi-file diff, and checkpoints to roll the workspace back to any earlier point in the conversation.
Related settings: product.edits.inlinePreview (inline diff preview, on), product.edits.formatOnApply (on), product.chat.editResend.confirm (confirm before re-sending an edit, on).
Slash commands
Type / in the composer, or run EditVolt: Slash Command from the palette:
| Command | What it does |
|---|---|
/explain | Explain the current selection or file (reasoning-heavy) |
/fix | Fix the diagnostics on the current selection; returns a unified diff |
/test | Generate a test using the detected test framework |
/review | Best-practice scan of the current file |
/architect | Architecture-level reasoning, ADR-lite (Context / Decision / Consequences) |
/why | Workspace-grounded investigation of "why is X?" |
/debug | Switch to Debug mode and run the evidence-first loop |
/delegate | Kick off a parallel subagent — presets: explore, agent, shell, review, debug-repro, ci-investigator |
Each command is also its own palette entry, so you can bind it to a key. You can add your own — see Customization.