You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Added
SSH execution backend for remote tool runs. New backend: ssh option for security.sandbox runs shell, python, and script tools on a remote host over OpenSSH (ControlMaster-multiplexed, auth via ~/.ssh/config + ssh-agent). Slots into the existing SandboxBackend Protocol next to bwrap, docker, and none. SSH is remote execution, not isolation, so the schema rejects bind_mounts, allowed_read_paths / allowed_write_paths, and network: bridge at load time with explicit migration messages; python_exec reports the unsupported extra_mounts case to the model instead of crashing the run. See docs/security/ssh-sandbox.md.
Per-run token budget for one-shot CLI mode. New guardrails.run_token_budget schema field and --token-budget N CLI flag cap cumulative token usage across a single initrunner run invocation. The tracker is shared via ContextVar so it propagates through the timeout thread-pool to delegate tools fired during the parent's model call, meaning inline-delegated sub-agents count against the parent's budget. See docs/configuration/guardrails.md and docs/getting-started/cli.md.
Fixed
initrunner new dispatch missing token_budget kwarg. The token-budget feature added a required keyword to _run_agent but the call site in cli/new_cmd.py (used after initrunner new ... --run) wasn't updated. Now passes token_budget=None.