Skip to content

Core Concepts

A session is a PTY shell process. Each session has its own terminal stream, working directory, and execution state. You can have multiple sessions open as tabs. Sessions are ephemeral — the PTY is respawned on app restart, but your history and workflows persist.

Execution states: bootingreadyrunningready (or interruptingready, or desynced).

Command mode: your text is sent directly to the shell as a raw command. No AI involvement.

Ask mode: your text is treated as natural language intent. The planner generates a command plan for your review.

The mode toggle is explicit. CommandUI never guesses which mode you meant.

A plan is a structured proposal: the AI’s translation of your intent into a shell command, annotated with risk level, explanation, assumptions, and safety flags. Plans are proposals — they never execute automatically.

Plan lifecycle: generated → reviewed → approved (executed) or rejected.

Every interaction is recorded: raw commands, semantic requests, generated plans, approvals, rejections, and execution results. History items include the original input, the generated command (if any), the actually-executed command (if different from generated), exit code, duration, and working directory.

A workflow is a saved command (or multi-step sequence) promoted from history. Workflows can be:

  • Created from a single successful command (“Save Workflow”)
  • Promoted from detected patterns (the memory system suggests sequences you repeat)
  • Scoped to a project directory

Workflows are reusable — run them with one click.

Memory items are learned preferences derived from your behavior:

  • Preferred CWD: directories you work in frequently
  • Recurring commands: commands you run often
  • Workflow patterns: command sequences you repeat

Memory items have confidence scores (0–1) that increase with evidence. They are visible, editable, and deletable. The planner uses accepted memory to generate better plans.

Every generated plan has a risk assessment:

  • Low: read-only commands (ls, git status, pwd). Visible review, no extra confirmation.
  • Medium: commands that modify state (file moves, package install, git checkout). Configurable confirmation.
  • High: destructive commands (recursive delete, sudo, broad writes). Required confirmation.

The command planner has two backends:

  • Ollama: local LLM generates real command plans with explanations
  • Mock: fallback stub that echoes intent. Used when Ollama is unavailable or in browser preview mode.

The plan panel shows which source generated the current plan.