agenter-terminal

name: agenter-terminal description: Create and drive durable terminal sessions. Use this when work needs a long-lived or interactive process; killed-terminal recovery is an explicit forensic path.

Use this skill when work needs a durable process or an interactive shell. Killed terminal instances are dead history evidence; recover one only when you explicitly need that exact durable id.

Quick start:

Run terminal list first to inspect processPhase, currentPath, currentTitle, and prior stop facts before guessing lifecycle from stale output.
Run terminal create when no suitable terminal exists yet. Public terminal create auto-bootstraps by default, so a fresh terminal may briefly show lifecycleTransition = bootstrapping before it settles into processPhase = running.
If you need the durable launch command, launch cwd, geometry, or metadata, run terminal get-config instead of inferring config from observed runtime output.
If terminal list shows processPhase as not_started, run terminal bootstrap before expecting read/write to work. If the terminal was already killed, switch to terminal history; prefer terminal create for normal follow-up work and bootstrap the killed id only for explicit forensic recovery.
If terminal list or terminal get-config shows lifecycleTransition = bootstrapping or killing, wait and reread instead of stacking another lifecycle or config mutation.
Decide whether the next payload is raw or mixed.
If the exact payload shape is unclear, run terminal write --help or terminal input --help first.
Run terminal write for literal raw bytes, or terminal input for mixed DSL.
Run terminal await when you need to wait for bounded evidence such as output change, idle state, or a deterministic text match.
Run terminal read only for immediate inspection or recovery of current terminal state.
Run terminal set-config when the durable terminal identity is correct but launch truth needs to change for the next bootstrap.
Run terminal stop when you want to halt the PTY, remove it from the live list, and keep its durable history evidence for later bootstrap.

Key laws:

A runtime does not start with a terminal by default.
Long-lived and interactive work belongs in terminals, not one-shot bash.
terminal is a collaborative process surface, not a root-workspace shell.
Shared terminals keep real-home semantics and do not inherit runtime-local env/CLI by default, even when cwd starts inside the avatar root workspace.
terminal list is the lifecycle and observed-identity inspection surface. Read processPhase, currentPath, currentTitle, and stop facts there before inferring state from raw output.
terminal create auto-bootstraps by default. A newly created terminal may briefly expose lifecycleTransition = bootstrapping; wait and reread instead of firing a redundant second bootstrap.
Provisioned terminals do not auto-start when you read or write. Use terminal bootstrap explicitly.
Killed terminals leave terminal list. Use terminal history to inspect dead-instance facts. Prefer terminal create for normal follow-up work; bootstrap a killed id only when you intentionally need forensic recovery of that same durable terminal.
terminal read consumes this actor's read cursor. Other actors keep independent cursors on the same shared terminal output.
Use terminal read deliberately: remark:false inspects without advancing your cursor, while normal reads advance only your actor's cursor.
terminal await is the bounded observation primitive. Use it instead of reconstructing waits with shell sleep, repeated terminal read, and grep.
terminal await returns clean bounded snapshot lines and match context from the terminal canvas. Those lines are evidence of the stable screen state, not raw ANSI bytes or PTY transition chunks.
Prefer terminal await's command-level wait.timeoutMs for post-mortem evidence. Shell-level timeout may still cancel the command, but it can prevent the JSON result from reaching you.
lifecycleTransition is a coordination lock, not a work item. If it is bootstrapping or killing, wait and reread terminal list or terminal get-config before sending another lifecycle or config mutation.
terminal get-config is the durable launch/config inspection surface. Use it for command, launchCwd, geometry, metadata, and other next-bootstrap truth.
terminal set-config patches durable launch/config truth without changing the terminal id.
For running PTYs, cols and rows may apply live immediately. Launch-affecting fields such as command, launchCwd, env, processKind, gitLog, and logStyle update durable truth first and take effect on the next bootstrap.
If work needs a port listener, local web server, watch mode, REPL, or retryable boot sequence, start it in terminal.
terminal write is raw mode. It sends literal bytes and never invents Enter, waits, or special keys for you.
terminal input is mixed mode. Use it for <key .../>, <wait .../>, or literal <...> text wrapped in <raw>...</raw>.
In mixed mode, literal tag-like lines must stay inside <raw>...</raw>, and Ctrl combos use ctrl="true" such as <key data="d" ctrl="true"/>.
Interactive stdin programs usually split into two phases: start the program with terminal write, then feed content and special keys with terminal input.
terminal write and terminal input only prove that input delivery succeeded; they do not prove the process succeeded.
terminal stop halts the PTY, removes the terminal from the live list, and preserves durable history evidence. terminal archive hides dead history from the default work queue, and terminal delete is the final destructive removal.
After starting a listener in terminal, inspect its real state and verify the exact promised URL or path before you tell a room or user that it is ready.
terminal read snapshots and "the process is still running" only describe terminal state; they do not prove the promised URL or API path actually responds.
When the task already names the workspace and delivery target, the normal next move is to create or select a live terminal, not to browse unrelated room or attention detail first.
If a one-shot shell hits binding or sandbox errors while you are trying to make a service reachable, stop and switch to terminal.
When more than one workspace is mounted, choose an explicit absolute cwd.
terminal create, terminal write, terminal await, and terminal read are JSON-first commands. Through root_bash, default to command=<bare terminal action> plus JSON stdin.
Only use a single argv JSON payload when it is trivially short and clearly cheaper in tokens than a separate stdin field.
If terminal create --help, terminal write --help, terminal input --help, terminal await --help, or terminal read --help marks compact as Suggested or Available, the matching command also accepts --compact positional arrays. If the positional array becomes unclear, go straight back to standard object JSON.
If raw vs mixed choice is unclear, read references/input-modes.md before guessing.
If terminal write --help or terminal input --help still is not enough, run skill info agenter-terminal, derive the real skill directory, and read only the reference file you need.
If the durable launch truth is unclear, read references/terminal-config.md before guessing.
For multi-line writes, nested JSON, or heredoc-heavy payloads, the next file to open is references/file-writing.md.

References:

references/terminal-lifecycle.md: create/recover/bootstrap/await/read/write/stop strategy and recovery patterns
references/terminal-config.md: durable launch/config inspection and mutation rules
references/input-modes.md: when to use raw vs mixed, and how <raw>...</raw> works
references/file-writing.md: safe patterns for sending multi-line file writes through terminal raw/mixed input