| name | unattended-operation |
| description | Run the backlog unattended — drain it continuously, park on escalation instead of waiting, break the circuit on repeated failure, and stay inside a run budget. Use this whenever the loop runs without a human watching: overnight runs, CI-driven drains, "leave it working while I'm away", or any `aegis run` without `--once`. Trigger it before starting an unattended session and when deciding how the loop should behave on failure, escalation, or a stuck requirement. |
unattended-operation
The loop is safe to run unattended on a fixed, well-specified backlog — that is what
resumability (LOOP-003), atomic claim/release (RTMX-002), and the audit log are for.
But "unattended" changes how failure and escalation must behave: nobody is there to
unblock a stuck requirement or notice the run going sideways. Four behaviors make it safe.
The four behaviors
- Continuous drain.
aegis run (without --once) pulls rtmx next and works
requirements until the backlog is empty or a stop condition fires. One requirement at a
time, same discipline as a single iteration.
- Park on escalation, don't wait. Unattended,
LOOP-002's escalation can't mean
"block until a human answers." It means: mark the requirement blocked, log why, release
the claim, and move to the next. One hard requirement must never stall the whole night.
- Circuit breaker. Stop the run after M consecutive failures. Consecutive failures
mean something broke — the endpoint died, the model regressed, a dependency moved — and
grinding through the rest just produces garbage you'll have to unwind. Stop and surface.
- Run budget. Cap the session: max requirements and/or wall-clock. A bounded run is a
reviewable run; an unbounded one is how a calibration drift quietly burns a night.
All four are config (internal/config), read by internal/loop. Defaults are
conservative: small consecutive-failure threshold, a finite budget, park-not-wait on.
Autonomy is bounded by your tests
The loop's safety ceiling is the acceptance tests behind each requirement. A weak or
lenient test lets the model "close" a requirement the wrong way, and rtmx verify won't
catch what the test doesn't check. So unattended autonomy is only as trustworthy as the
tests — before a long run, trust the backlog's tests, not the model. When in doubt, shrink
the run and review the audit log before extending it.
What unattended operation does NOT do
It executes an existing backlog. It never authors or decomposes requirements on its own —
that is decomposition, and it is human-gated by design. An unattended run that invented
its own work, then graded itself against criteria it also invented, is exactly the closed
self-validating loop the architecture exists to prevent.
On return
Everything is in the audit log and the metrics. Review: what closed, what parked (and
why), whether the breaker tripped, and the dashboard trend. The run reports a summary;
the log is the record. Do not infer success from "the backlog is empty" — infer it from
the verify results and the parked list.
Example
aegis run --max 40 --break-after 3 overnight: drains up to 40 requirements, parks any
that exhaust retries (logged, claim released, next pulled), and halts immediately if three
in a row fail. Morning review: 31 closed, 6 parked on under-specified tests, breaker never
tripped — a clean, bounded, reviewable night.
