Работа с LetronnaТерминал
Scripts and one-shot runs
letronna -p: asking once from a script, its options, JSON output and running with nothing set up.
Эта страница ещё не переведена, поэтому показана на английском. Вы можете перевести её на GitHub. GitHub ↗
letronna -p asks one question, prints the answer and exits. Nobody watches such a run, so it never waits on a question: it decides from the task and the project folder.
letronna -p "list the TODOs in src and say which are urgent"In the default mode, a -p run cannot write files or run commands, and warns you of it. Choose a mode that allows what it needs:
letronna -p --mode accept-edits "fix the failing test"Options
| Option | Does |
|---|---|
--output <form> | text (the answer), json, or stream-json |
--max-rounds <n> | tool rounds a message gets before it stops (100) |
--model <name> | use this model instead of the one set up |
--tools <a,b> | only these tools, by name |
--budget <tokens> | stop once this many tokens are spent, over every message the run is sent |
--input-format <form> | text (the question), or stream-json: messages, interrupts and answers as JSON lines on stdin |
--permission-prompt-tool stdio | with stream-json in, ask stdin about each call that needs a yes |
--json-schema <file> | answer in JSON this schema checks; the schema itself can stand in for the file |
--dump-config | print the settings and limits a -p run with the same flags gets, as JSON, without running |
--json-schema needs a new session, so it does not go with -c, -r or stream-json in.
JSON output
--output json prints one JSON object with the answer and what it took: the session, the model, the cost, and how many calls were refused. --output stream-json prints a JSON line each time the run moves on, then that object as the last line.
letronna -p --output json "how many files are in this folder?"A time limit
LETRONNA_DEADLINE gives the run a number of seconds. At 80% of it, the run is told to write its outputs; at the end it stops, with a result whose status is deadline.
LETRONNA_DEADLINE=600 letronna -p --mode bypass "build the report"Running with nothing set up
In a container or on a fresh machine, a -p run can take its server from the environment, saving none of it:
| Variable | Does |
|---|---|
AGENT_BASE_URL | ask this OpenAI-compatible server instead of the gateway set up |
AGENT_API_KEY | the key to send it |
AGENT_MODEL | the model to ask, unless --model names one |
AGENT_BASE_URL=http://localhost:11434/v1 AGENT_MODEL=llama3.2 letronna -p "hello"