Agent mode
Describe an outcome; the agent plans the work, edits across files, and runs commands — with guardrails at every step. Agent is the default mode.
Everything happens in the editor window. There is no separate agents window.
The agent's tools
Reading and searching files (including glob and semantic search), proposing edits, creating, moving and deleting files, running shell commands and reading their output, fetching URLs and searching the web, notebook read/edit, notepads, todos, delegating to subagents, scheduling runs, invoking skills, and image generation. Edits are always staged as diffs for your approval.
Two settings bound a single turn:
| Setting | Default | Description |
|---|---|---|
product.agent.maxToolRounds | 6 | Tool round-trips allowed within one turn (1–10) |
product.agent.maxTurnTokens | 120000 | Per-turn token ceiling; a runaway loop stops here |
product.tools.enabled | true | Master switch for the tool surface |
product.chat.parallelToolCalls | true | Let the model issue independent tool calls concurrently |
Command consent
You control how much the agent can run on its own via product.commands.runMode, exposed as the Run Mode chip in the composer and under EditVolt Settings → Preferences:
- Auto-review (default) — every command is shown to you first
- Allowlist — commands matching your allowlist run automatically; everything else asks. The allowlist lives in
.editvolt/permissions.json(workspace or global) - Run everything — no prompts (use with care)
Regardless of mode, sudo and non-deleting destructive commands always prompt, and config-file writes are always gated. File-deletion protection (product.commands.protectFileDeletion) is on by default; rm-class commands only run unattended if you both switch to Run Everything and turn that protection off.
Commands can also be confined by network policy with product.commands.sandboxNetworkMode, and routed to a dev container instead of the host with product.container.runCommands (auto | host | container).
Honest reporting
The agent's claims about what it did are checked against what actually ran. A per-turn ledger of tool results grounds statements like "I updated the file" or "the tests passed"; if a claim isn't supported, EditVolt runs one observation round and then says so plainly rather than asserting success. Explicit user restrictions ("don't run any commands") are enforced before the consent prompt, not after.
Checkpoints
Before the agent changes anything, EditVolt records a checkpoint. Roll back to any point in the conversation if a direction doesn't work out — your workspace is restored exactly.
For riskier runs, product.chat.shadowWorktree stages the agent's work in a separate git worktree; EditVolt: Discard shadow worktree throws it away.
Multitask
Multitask mode decomposes a request into 2–6 independent subtasks and runs them as parallel workers, which defer into one consolidated review at the end. Concurrency is auto-sized and capped by product.multitask.maxConcurrency (default 5).
Design mode
Design mode is Agent plus a live browser: it opens a Browser tab in your editor group, and the agent can launch a URL, click, type, scroll, read console and network output, and see screenshots of what it built. It needs the in-editor browser enabled (product.browser.enabled, off by default). Under strict privacy, only localhost and loopback URLs are allowed.
Working without a folder open
With no folder open, the agent's first write lazily creates a managed project under ~/.editvolt/Projects/<name>/files/ and opens it in place — no modal, no reload, same conversation. Managed workspaces are deliberately not granted project scope: an agent-writable root cannot author its own rules, MCP servers, or command allowlist. EditVolt: Save Workspace as Project copies a managed workspace to a location you choose.