Explicit local file-target runs currently share the same changed-line budget used for PR/diff scans. That makes sense for PR review, but it is surprising for local runs where the user has explicitly supplied files to scan.
This came up while running the Sentry corpus benchmark for Claude Sonnet 5 on Pi. The benchmark passed an explicit target-file list, but the default changed-line budget skipped six corpus target files before model analysis. The benchmark now works around this with higher scan limits and strict skipped-target validation, but the CLI behavior should be fixed separately.
Expected behavior:
- PR/diff-based scans keep
maxChangedLines as a review safety budget.
- Explicit local file-target scans either ignore
maxChangedLines or use a separate limit with clearer file-scan semantics.
- If Warden skips an explicit target, the CLI should make that visible enough that benchmark and local users do not mistake partial coverage for a complete run.
Refs #418
Explicit local file-target runs currently share the same changed-line budget used for PR/diff scans. That makes sense for PR review, but it is surprising for local runs where the user has explicitly supplied files to scan.
This came up while running the Sentry corpus benchmark for Claude Sonnet 5 on Pi. The benchmark passed an explicit target-file list, but the default changed-line budget skipped six corpus target files before model analysis. The benchmark now works around this with higher scan limits and strict skipped-target validation, but the CLI behavior should be fixed separately.
Expected behavior:
maxChangedLinesas a review safety budget.maxChangedLinesor use a separate limit with clearer file-scan semantics.Refs #418