workbench

Use Workbench

Workbench is an AWorld-owned candidate-management capability that is bundled but disabled by default. Use it only when it was explicitly selected and multiple candidates, rollback, checkpoint preservation, or measured local comparison provide real value. Direct and low-risk tasks should usually proceed without Workbench ceremony.

Its checks and receipts are agent self-check evidence only. They do not inspect hidden checks, compute task reward, or represent caller/canonical acceptance. Final external evaluation is outside Workbench.

WORKBENCH_* values are optional ordinary environment variables, not a security boundary. Changing them cannot grant access beyond the terminal's existing operating-system permissions and cannot produce canonical verifier or reward evidence. When they are absent, the CLI uses the real current directory as the workspace, puts state in AWorld's external control/state directory, and derives a bounded scope from AWorld task/session identity or a workspace fingerprint. Explicit values remain supported.

Choose the configured AWorld interpreter when present; otherwise python3 launches the CLI and the CLI uses sys.executable:

WORKBENCH_PYTHON="${AWORLD_PYTHON_EXECUTABLE:-python3}"
"$WORKBENCH_PYTHON" /skills/workbench/scripts/workbench.py --help

The CLI exposes no command-line shortcuts for replacing workspace, state, or scope. Workbench never infers outputs, checks, policy, or scope from task text, dataset names, or benchmark identity.

Initialize an explicit self-check contract

Create a JSON contract inside the current task workspace. Every contract must declare authority: "agent_self_check", at least one output, at least one executable check, and an explicit selection policy. Use paths relative to $WORKBENCH_WORKSPACE_ROOT; output and input paths must remain inside that configured directory.

{
  "schema_version": "workbench.contract/v1",
  "authority": "agent_self_check",
  "outputs": [
    {
      "id": "report",
      "path": "report.json",
      "checks": [
        {"id": "report-json", "kind": "json", "path": "report.json", "root_type": "object"}
      ]
    }
  ],
  "inputs": [],
  "checks": [],
  "policy": {
    "mandatory_checks": ["report-json"],
    "hard_constraints": []
  }
}