Expand description
Policy and process wiring for agent-provided shell commands.
Every agent bash process reaches resolve_sandbox_spawn and carries the
resulting SpawnPlan into one of the two process-creation primitives:
detached pipes or PTY. Foreground orchestration uses the detached registry
too, so it does not create a third process-creation path.
AFT also starts processes for its own implementation. Those are outside this
seam because they do not execute an agent command: external formatters,
linters, and type checkers in format; LSP servers and Windows LSP cleanup in
lsp::client and lsp::child_registry; git probes in search_index,
readonly_artifacts, commands::configure, and commands::conflicts; login
shell PATH discovery in effective_path; and Windows process liveness or
termination helpers in artifact_owner, fs_lock, and
bash_background::process. Image and PDF handling in commands::read is
in-process and creates no child. Keeping this inventory here makes the
agent-command boundary explicit without accidentally applying agent policy to
AFT’s internal tooling.
Structs§
Enums§
- Authenticated
Principal - Principal data supplied by the server-side transport, never by a bash body.
- Principal
Trust - Server-authenticated trust classification for a route bind.
- Requested
Sandbox Tier - Sandbox tier requested by the caller.
- Sandbox
Task Kind - Agent-command path that is about to create a process.
- Spawn
Plan - Complete process-launch decision consumed by a spawn primitive.
Constants§
Functions§
- resolve_
sandbox_ spawn - Resolve policy for an agent-command process.