Recover a run without blind retries

When a run fails or a response is lost, first establish what already happened. A repeated paid call can create more work; an error message does not prove that nothing was submitted or charged.

Observed state	Read next	Recovery
Revision conflict during an edit	Fresh read_project_graph	Rebuild the edit against the current graph and revision.
busy	Returned active run_id and retry_after_s	Wait for the active run or the stated interval; do not launch copies.
need_accept_warnings	Current quote and warnings	Resolve or deliberately accept the specific overwrite warning.
Estimate above ceiling	New quote and intended scope	Reduce the work or obtain an accepted ceiling; do not silently raise it.
error, stopped or interrupted	get_run, related jobs and billing	Determine which nodes completed and which submitted jobs remain.
not_saved	The returned error and current graph	Preserve the completed work; follow the cache recovery route before considering new generation.
Completed job but missing node result	Job status, graph and material list	Locate the material and attachment outcome; generation and graph saving are separate.
Expired download link	get_artifact for the same material	Obtain a fresh link, without rendering again.

Recover reusable work

For a completed request whose result was not attached, an unchanged up_to run can recover a cached result. First verify that the submitted job has finished, the relevant inputs are unchanged and the returned instructions support that route. Quote again. Do not switch to regen as a recovery shortcut: that explicitly requests new work.

A job may outlive a stopped graph run. The server can settle already submitted work after the run ends, and the next unchanged run can attach its cached result. Until the job and billing records settle, describe the cost as uncertain instead of promising a refund or a free retry.

Montage and server Decoration jobs retain their phases, progress and saved results across browser reloads. Follow the returned job with get_job before submitting another operation. A full decoration run reserves its existing quote until the result succeeds; a failed job releases the pending reservation and keeps any saved plan. A successfully saved plan-only design is already the paid result, and rendering that unchanged plan later remains $0. An interrupted unfinished authoring call is reported rather than automatically repeated. See assembly and decoration jobs.

Report the actual outcome

State which nodes completed, what was saved, what remains pending and what cost is recorded. If the critic failed, report the failed check separately from successful text generation. If content in a report is shortened or omitted, say so and use the full node view.

If a source file is inaccessible, a model rejects an input or the run route does not support a node, follow the specific refusal. Repeating the same request does not repair ownership, file type or unsupported execution.

See results and cache and runs and billing.

Last checked: 21 October 2026. This recovery guide does not promise that a new generation is free, that submitted work can be cancelled, or that a job finishes within a set time.