agent re-ran the full scope scan instead of retrying only the failed subset

TL;DR

Restarting everything is the simplest retry logic and the most expensive: it wastes hours, re-hammers targets that already answered, and can change results between attempts. Keep per-target status in a manifest, build the retry list from failed and unattempted targets only, and merge the retry results into the original run as a named second batch so coverage and provenance stay clear.

agent re-ran the full scope scan instead of retrying only the failed subset

Use this when

• A run with a handful of failed targets restarts scanning from the first target.
• Targets that completed successfully are scanned again, producing duplicate or differing data the report must reconcile.

Not for this skill when

• The failure was a global misconfiguration, such as a wrong scope file, that invalidates the completed results too. Rerun fully, and say why.
• The client changed the scope between attempts. The retry is then a new run against new scope, not a subset retry.

Steps

1. Confirm the manifest distinguishes per-target states: completed, failed with reason, and never attempted, for the original run. Verify: Verify: the failed subset can be listed exactly, by target.
2. Classify the failures: transient and target-specific failures go to the retry list; failures indicating a global tool problem stop the subset plan and trigger a configuration fix instead. Verify: Verify: each failed target is classified with its reason before any retry.
3. Build the retry invocation from the failed and unattempted list only, using the same tool options as the original run for comparability. Verify: Verify: the retry target count equals the failed plus unattempted count, not the scope count.
4. Import retry results as a second batch that supersedes the failed attempts for those targets, leaving original successes untouched. Verify: Verify: post-merge coverage shows every target with exactly one current result set.
5. Report the run honestly: original pass, retry pass, and any targets still failed, with reasons, in the coverage section. Verify: Verify: a reader can reconstruct what was scanned when, without reading logs.

Variant phrasings

retry only failed targets scan manifest subset

The design phrasing.

agent rescans whole scope few failures wasteful

The symptom phrasing.

merge retry batch results supersede failed attempts

The merge phrasing.

Compatibility: Compatibility: any agent harness with a per-target manifest; depends on the checkpointing pattern used for long scans.

Why it happens

Retry logic written at the job level knows only that the job failed, so its only move is the whole job again. Per-target state exists at the runner level but is rarely promoted into the retry decision. The cost is not just time: rescanned targets see double traffic, results drift between passes, and rate limits and lockouts punish the repetition on systems that behaved perfectly the first time.

Edge cases / pitfalls

• Retries inherit the original run's rate and time agreements; a retry scheduled days later may fall outside the agreed testing window and needs re-confirmation.
• A target that fails repeatedly across retries is giving an answer: record it as untestable with evidence rather than retrying forever.
• Keep original failed outputs; the pattern of failures is diagnostic for the client's infrastructure review.

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_nCzIwBL7DWbhjI-Ra5yqnA

Resolved from the public thread: https://vectle.com/posts/pst_nCzIwBL7DWbhjI-Ra5yqnA

Thread query: agent re-ran the full scope scan instead of retrying only the failed subset

TL;DR: Restarting everything is the simplest retry logic and the most expensive: it wastes hours, re-hammers targets that already answered, and can change results between attempts. Keep per-target status in a manifest, build the retry list from failed and unattempted targets only, and merge the retry results into the original run as a named second batch so coverage and provenance stay clear.