Recovery, retry and replan

When a step’s runtime fails — a missing key, an expired login, a crash, a timeout — the Run does not guess. It stops in needs a decision and waits for you in Attention. Aqorin never switches runtime on its own, and nothing is retried on its own.

The failed turn’s own words, in the Run’s transcript, say what the runtime answered.

The three ways on

r, on the Run in Attention, offers:

Choice	When to use it	What happens
Retry same assignment	The cause was outside Aqorin — a missing key, a login — and you fixed it.	Aqorin checks again, as things stand now: the installation, connection, credentials, model, eligibility and resources. If the step still cannot run, the retry is refused. Otherwise a new attempt of the same step runs.
Replan to another runtime/model	The runtime or model is the problem.	The step’s role moves to another runtime, connection and model, with a reason. The new plan is admitted like any other.
Cancel Run	You are done with it.	The Run ends as cancelled.

After a decision, R resumes the Run.

A retry is not offered for:

• validation failures;
• review rejections;
• policy or restricted-role refusals;
• integrity failures;
• workflow or schema errors;
• cancellation.

Those are outcomes, not accidents.

At least once

A step that is attempted again starts over. It is not the runtime resuming its old conversation, so work the failed attempt already did may be done twice.

Resume

aqorin resume <run> — or R — continues a Run that was interrupted, or whose decision is recorded, from where it stopped, on the runtime of its current assignment — the one a replan chose, if there was one.

From the CLI

aqorin replan <run> --role developer --same --reason "Key fixed"
aqorin replan <run> --role developer --runtime codex --model <model> --reason "Pi is down"
aqorin resume <run>
aqorin cancel <run>