Recover the signal before retrying.
When a coding task stalls, repeated retries can obscure the first useful error. Preserve the failing command or request, its output, and the repository state. Then identify whether the problem belongs to code, configuration, access, or the selected provider.
Classify the failure
A missing executable is an environment problem; an assertion failure may be a behavior problem. Authentication errors and rate limits need different responses. Avoid asking the assistant to rewrite application code to fix an unavailable service. Redact tokens and private response data before sharing logs.
Resume with a smaller checkpoint
Re-establish the active directory and current changes. Restate the last verified result and the next narrow action. If the assistant lost context, provide a compact handoff with relevant paths. If a provider is unavailable, check its current documentation before changing endpoint settings or retry behavior.
Checkpoint
- First useful error preserved
- Failure assigned to the right layer
- Last verified state identified
Read-only transmission
Read only. Diagnose this failure using [redacted output]. Separate code defects, environment issues, access problems, and provider responses. State the strongest evidence for each plausible cause and suggest the smallest check that distinguishes them.
Recovery should increase what you know. Stop repeating a check when it produces the same result and no new evidence.
Next: Prepare launch