hitl-gatelisted

# hitl-gate Implements a **durable human-in-the-loop interrupt**. The agent requests approval, the pause is written to disk, and the run can be resumed by a *different process* — approve to proceed, deny to re-enter the loop with feedback. The point: the pause survives a crash or restart, so an unattended agent never executes a risky action on its own. ## When to use - Before irreversible or high-blast-radius actions (sending mail, deleting data, spending money, pushing code). - When a human reviewer is asynchronous — the agent may exit and the decision arrives later. - Encoding an approval policy where deny should *teach* the next attempt (feedback loop), not just stop. ## Run it (three separate processes — that's the point) ``` python -m harness.hitl request --payload "send the invoice email" # prints a run_id, status=pending python -m harness.hitl approve --run <run_id> --feedback "looks good" python -m harness.hitl resume --run <run_id> # -> proceed # deny path: python -m harness.hitl deny --run <run_id> --feedback "wrong recipient" python -m harness.hitl resume --run <run_id> # -> loop-with-feedback ``` `resume` on a still-pending run exits non-zero (status 3) — so a supervisor can tell "not yet decided" apart from "approved". ## Statuses - `pending` → awaiting a decision (persisted at the run's `interrupt.json`). - `approved` → `resume` returns *proceed*. - `denied` → `resume` returns *loop-with-feedback* carr