Bash

​

Authentication and enablement

​

Inputs

• command (required): the shell command to execute. Runs under /bin/bash -c.
• inputs (optional): array of task variable names (e.g. ["VAR_27", "VAR_5"]) to write as JSON files before the command runs. Each name becomes /workspace/vars/<NAME>.json.
• timeout_seconds (optional): execution timeout. Default 120, max 600.

​

Workspace contract

• /workspace/inputs/ — task files staged in (e.g. PDFs, CSVs, images uploaded or produced by prior steps).
• /workspace/vars/ — JSON files corresponding to inputs array (e.g. /workspace/vars/VAR_27.json contains the variable’s value).
• /workspace/out/ — write downloadable files here. Each file becomes a new task file and a task variable reference (e.g. /workspace/out/results.csv → new variable VAR_N with mime type inferred from extension).
• /workspace/.out/<step>.log — fuller stdout/stderr written here if the in-result output is truncated (see Limits below for its own cap).

​

Output

• exit_code (integer).
• stdout (string, capped ~16 KiB in the returned result; full output in the log file if truncated).
• stderr (string, same cap and truncation behavior).
• truncated (boolean): true if stdout or stderr was capped.
• log_path (string, optional): sandbox-internal path to the full output log, present only when truncated is true.
• files (array of objects): one per file written to /workspace/out/. Each has name (the file’s name), var (the new VAR_N reference other steps can use), and url (download URL).

​

Limits and side effects

• Timeout: 120s default, 600s max. Exceeding it terminates the command and returns a timeout error.
• Output truncation: stdout and stderr are each capped at ~16 KiB in the returned result. When truncated, the fuller output is written to /workspace/.out/<step>.log inside the sandbox workspace itself — it is not exported to any UI (there is no task-detail “Logs” tab view of it), so a later step in the same task must read it back via the bash tool. That log can itself be incomplete: the sandbox shim that runs your command caps stdout at ~1 MiB and stderr at ~256 KiB before Go ever sees the bytes, so a command producing more output than that loses the excess before it reaches /workspace/.out/<step>.log too.
• File export: up to 20 files per call, max 10 MiB per file. Files exceeding either cap are named in the result’s “not exported (too large, over the per-command file limit, or unreadable): …” line; they are never silently dropped.
• File persistence: files in /workspace/out/ are staged as task files for the next step, but do not persist across independent task runs. Each new task run re-stages /workspace/inputs/ from the latest versions.
• Network: during the POC the sandbox has open outbound egress and carries no platform credentials — it is enabled on internal workspaces with test data only. Egress lockdown is planned hardening after the POC.
• Billing: each call is a billed step visible in the task feed.

​

Expected errors

• "sandbox runner not configured" — feature not available in this environment (dark-launched feature off).
• Variable name not found in task context (e.g. VAR_999 doesn’t exist).
• "output truncated" — stdout or stderr exceeded the cap. /workspace/.out/<step>.log inside the sandbox has the fuller output — unless the sandbox shim itself already truncated upstream, in which case the note says so and the log is capped too.
• Exit code non-zero — command failed; inspect stdout/stderr and the log file for details.
• File export error — file too large or over the 20-file limit.