Ways to start
Bare
stampede opens the TUI when stdin and stdout are both a terminal, and the line chat when either is a pipe or a file, so scripts keep working. Set console = "chat" at the top level of ~/.stampede/config.toml to make the line chat the default on a terminal too.
Start the console in the directory you want the model to work in:
chat, tui and agent accept --skill-lane stable|canary to choose which skill catalog the run loads (stable by default). See Skills.
Choosing the person
Every conversation runs as one person. Without a choice it is the default person,main. To start as another person:
stampede person use NAME sets the default person for later conversations. The current conversation keeps the person it started with.
The terminal interface
The TUI prints the conversation into your terminal’s own scrollback, so scrolling, selecting and copying work as usual. The input box, a separator and a status line stay pinned at the bottom.
A multi-line paste is shown as
[Pasted text #N +M lines] and expanded to the full text when you send it. Type exit or quit to leave.
With require_approval = true under [tools], the console asks before every tool call. Press y to run the call; any other key denies it.
The line chat
The line chat is the same conversation, one line at a time, for pipes and plain terminals. Typeexit or quit, or press Ctrl-D, to leave. An empty line is ignored. Ctrl-C during a turn requests cancellation and waits for the run to stop safely.
The full history is sent to the model on every turn, which is what gives it memory of the conversation.
Addressing a person
In the TUI, a line that starts with@name is a request to that person. It runs in the background while you keep typing, and the console prints its progress:
A request routes only when the mention is the first word, names an existing person, and has text after the name on the same line. A mention later in the line is ordinary text for the current model. The request is capped at 16 KiB.
Person mail does not carry the current conversation, selected file or earlier tool results. Make each request self-contained: name an exact path or paste the snippet.
Progress appears as
[people] NAME is working, a live activity row with a task label and elapsed time, and finally one line such as [people] editor finished with Completed in 25m 0s. Leaving the TUI cancels its background work; unanswered mail stays queued for later. Persons, policies and mail are covered in Persons and Mail and task rooms.
Person routing is a TUI feature. In the line chat, use /person send.
Slash commands
A line that starts with/ is a command. An unknown command is reported and never sent to the model. /help lists the commands and /help QUERY shows one in full:
/model takes a profile name or list number, a model name (claude-* implies Anthropic, gpt-* implies OpenAI, other names stay on the current provider), or an explicit provider, model and base URL:
Questions from the model
On a terminal the model may stop and ask you up to four questions with theask_user tool. Each question shows a short header, the question, and numbered options when it has any. Answer with a number, several numbers separated by commas when more than one may be chosen, or your own words.
An empty answer, or Esc in the TUI, leaves the question unanswered and the model reads {"status":"no_answer"}. A run with nobody to ask (a piped or scheduled run, a person working through its mailbox, a sub-agent) is not offered the tool, and a call that arrives anyway reads {"status":"no_operator"} at once.
A conversation keeps the tool set it started with. One started from a pipe, which has no ask_user, refuses to continue in a console that offers it, and says so.
Saved conversations
stampede resume continues where you left off. Starting a fresh conversation renames the previous last to last-previous, and only one such generation is kept. Name anything you want to keep with /save NAME, then reopen it with stampede resume NAME.
One-shot questions
Headless runs
stampede agent gives the model the task and the tool catalog, runs tool calls and feeds results back until the model answers in plain text. Progress and narration go to stderr. Stdout carries only the final answer when it is piped, so the run fits in scripts:
--ephemeral turns that off explicitly.
When the run ends, stderr shows
[stopped: WORD], where the word is one of completed, budget_exhausted, wall_clock, no_progress, cancelled, provider_failed, interrupted or failed. Budgets are described in Limits.
Inspecting and stopping runs
Console turns and agent runs are recorded the same way.runs show prints a redacted record of the run: its events, checkpoints, budget use and the stop: word. It never prints exact model or tool content. --tokens adds a line per model call with its tokens and cost.
To stop a run:
- In the TUI, press Ctrl-C or Esc. In the line chat or a foreground
stampede agent, press Ctrl-C. - From another terminal, run
stampede runs cancel RUN_ID. The cancellation is journaled and reaches the process that owns the run.
runs resume continues the run exactly as it was recorded and refuses when the configuration, tools or workspace have drifted or an effect’s outcome is unknown. runs conclude asks the model for a final answer with no tools, stating what is uncertain.