Chat modes

The chat input has a mode picker: Agent, Ask, Orchestrator, and Plan. A new chat defaults to Agent. Changing mode on an existing chat keeps the model you already picked. If the model field is empty, it fills the role default for that mode.

Each mode inlines a built-in skill into the system prompt and exposes only the tools on that mode's allowlist. Tools that are not on the list do not exist for that turn. Permissions still gate writes, shell, git, web, and MCP on top of the allowlist.

Capabilities

Nested sub-agents

spawn_subagent takes capabilities: read-only (the default) or write. Ask and Plan reject write-capable spawns. Agent and Orchestrator can spawn either.

A read-only nested agent can use read_file, list_dir, grep, glob_files, codebase_explore, codebase_search, codebase_impact, codebase_status, git inspect (git_status, git_diff, git_log, git_branch), lsp, diagnostics, load_skill, web_fetch, the shell suite (run_terminal, terminal_output, stop_terminal), and MCP (get_mcp_tool, get_mcp_tools, call_mcp_tool, list_mcp_resources, read_mcp_resource, get_mcp_prompt).

A write-capable nested agent also gets write_file, edit_file, apply_patch, delete_file, move_file, git_commit, git_checkout, and git_branch_create.

Nested agents cannot spawn further sub-agents (no spawn_subagent or steer_subagent). They also do not get plan tools (create_plan, update_plan, update_plan_todo), in-chat update_todos, ask_user, move_workspace, or resolve_models.

Agent

Agent implements changes in this chat. Prefer it when you want the work done here rather than handed to a durable plan. The skill tells it to prefer write and edit tools over shell redirects, and not to commit unless you ask. Use create_plan when the work should become a plan with Build or Orchestrate, rather than an in-chat task list.

Ask

Ask is read-only exploration. Shell is allowed for investigation; approvals still apply. It cannot mutate source or git, and it cannot create_plan. Responses cite files and symbols, and point at Agent or Plan when a change is needed.

Plan

Plan is Ask plus plan tools. It still cannot mutate source. Research first (reads, graph, MCP, web, sub-agents), then write PLAN.md. After create_plan succeeds, writes, shell, git mutations, MCP, spawn, steer, and a second create_plan are blocked until you choose Build or Orchestrate on the plan tab. update_plan and update_plan_todo stay available. How those buttons start the next chat is on Work with plans. Why plans stay on disk is on Philosophy.

Orchestrator

Orchestrator coordinates through sub-agents. The parent has no file-write tools and no git mutations, so it cannot edit source itself. It does have the shell suite. The built-in skill says to use that shell only for validation (CI, tests, lint, typecheck, gh), not for mutating commands. After a folder or worktree exists, the parent may move_workspace before spawning implementers.

Background spawns are preferred. The parent leaves a short status and ends the turn. The harness resumes the parent when a sub-agent finishes. terminal_output is not how that wait works. Network goes through web_fetch, user MCP, or validation shell commands such as gh.

See Orchestrate sub-agents for spawn, steer, and how to run them.

Git neutrality

Vixl has git tools. Built-in prompts never tell the agent to commit, branch, or follow a git flow. Agent can commit if you ask. git_commit stages the paths it is given, then runs git commit -m. It will not stage everything with git add -A. The commit message is the one from the tool call. Vixl does not add itself or the model as a co-author.