Expand description
Execution of seed = { command = "..." } region seeds.
§A command seed runs a shell command in the run’s workdir at spawn and puts its combined stdout/stderr into the region - but only when the command succeeds; a non-zero exit is reported as an error so a diagnostic never masquerades as data. It is the only seed source that executes anything, and it does so before the first inference - therefore before any tool-approval prompt - so it is deliberately hemmed in:
it is skipped entirely unless SeedCommandPolicy::allowed (the
[security] allow_seed_commands config switch and the --no-seed-commands
launch flag); -
it must be covered by [safe_commands], since a seed is precisely the case
where there is nobody to prompt - see SeedCommandPolicy::run; -
it runs inside the entry stage’s sandbox when the agent declares one,
using the same ShellExecutor::build_command routing as the built-in
shell tool, so a seed can’t escape the isolation the stage asked for; -
it is capped in wall-clock time ([limits] script_shell_timeout_secs) and
in output size (cap_script_io); -
it never runs on restart - crate::daemon::spawn only resolves seeds on
a fresh spawn.
Structs§
- Seed
Command Policy - How command seeds are executed for one spawn.
Type Aliases§
- Seed
Command Runner - Runs one seed command:
(command, workdir, timeout) -> combined output.