Skip to content
Docs0.9GitHubInstall

Using itManaged agents

Managed agents

Agents started by Vendetta run in native panes inside one pinned Agents tab (window 0) per project, on your normal tmux server. Agents you start by hand stay where they are. Launches made through the chooser, vendetta agent create, schedules, Telegram, and the Master all use the same launcher, unless you ask for a terminal window instead.

  • n opens a form seeded with the selected project and directory, and the [harnesses] default provider. You can pick an omp or Claude capability profile, and a model and effort: ←/→ cycles the provider’s suggested models, or type any. Codex, omp, Claude Code, and OpenCode can be launched. Gemini and Aider are only discovered when you start them yourself. An agent you start yourself is listed within 2 seconds.
  • N creates an ordinary tmux session (a name and an existing directory) without an agent.
  • Enter on an agent selects and zooms its pane. tmux scrollback and copy-mode work normally.
  • Alt+Z shows all of the tab’s agents side by side, and again zooms the one in focus. When the agent you are looking at exits, tmux shows the rest side by side; Alt+Z zooms the one it moved to.
  • Background launches never steal your focus. A missing project session is created.
  • Closing the Agents tab or its session ends its agents. The tab closes itself when its last agent exits.

Claude Code agents that Vendetta launches skip the workspace trust prompt for their git root, or for the directory itself outside a repository. Vendetta records Claude’s own hasTrustDialogAccepted setting before launch. Claude started by hand still asks.

Managed launches use YOLO mode by default: tools are auto-approved, and Codex also runs without its sandbox. Use this only in trusted projects. Configure each harness:

[harnesses]
default = "claude"   # preselected in the TUI, `vendetta agent create`, and Telegram
[harnesses.codex]
args = ["--dangerously-bypass-approvals-and-sandbox"]
[harnesses.omp]
args = ["--approval-mode=yolo"]
[harnesses.claude]
args = ["--dangerously-skip-permissions"]
# models = ["opus", "sonnet"]   # offered at creation; any other may be typed
[harnesses.opencode]
args = []
env = { OPENCODE_PERMISSION = '"allow"' }

These are the defaults. A model and effort chosen at creation are part of the launch: a restore starts the conversation on them again. Efforts are Claude’s low to max, Codex’s minimal to xhigh, and omp’s off to max; OpenCode takes none.

  • models replaces the models offered at creation. Unset, Claude offers what its /model picker does for your account (from Claude Code’s cached model catalog and extra models), else opus, sonnet, haiku; Codex offers what its /model lists (models_cache.json); omp and OpenCode offer none. Any model may be typed.
  • args are passed literally, and env applies only to the launched process.
  • args = [] or env = {} removes an injection; an omitted field keeps its default.
  • Each launch saves its arguments and profile, and restoring that conversation reuses them.
  • Apart from the workspace trust above, provider safety checks and approval dialogs still need you; Vendetta never answers them.

The notification worker stops managed omp, Codex, and Claude conversations that are bound to Telegram or the app after six hours without activity:

[plugins.telegram.idle_timeout]   # and the same in [plugins.app]
enabled = true
timeout_secs = 21600

It only stops agents that are managed, have a saved transcript, and are idle with no pending question or approval. It never stops:

  • the Master;
  • agents you started by hand;
  • agents that are running or in an unknown state.

Typing in the preview or in a terminal showing the agent counts as activity, and so do new transcript records and clients attaching to its pane. Status polling does not. Worker restarts keep the elapsed time. /agents in Telegram shows each agent’s eligibility.

Stopping is silent, and the conversation and topic stay. Send a message in the topic to restore the same conversation and deliver the message. While a restored agent waits for approval or readiness, up to 16 prompts are held in memory. If the worker restarts, they are lost, and you are told how many.