fix: remove double-counted write_actions from partially_reducible evidence; rename Minutes → Action Minutes#23075
Conversation
…dence; rename Minutes header to Action Minutes Agent-Logs-Url: https://github.com/github/gh-aw/sessions/ee2348c1-4f1f-4285-9cec-2e8c60c10108 Co-authored-by: mnkiefer <8320933+mnkiefer@users.noreply.github.com>
There was a problem hiding this comment.
Pull request overview
Fixes agentic analysis reporting and logs console output clarity by removing an unreliable/double-counted write-actions signal from one assessment and renaming a confusing column header.
Changes:
- Remove
write_actionsfrom thepartially_reducibleassessment evidence string inaudit_agentic_analysis.go. - Rename the logs console column header from
MinutestoAction Minutesforaction_minutesinlogs_report.go.
Reviewed changes
Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.
| File | Description |
|---|---|
| pkg/cli/logs_report.go | Renames the console table header for action_minutes to reduce ambiguity. |
| pkg/cli/audit_agentic_analysis.go | Drops write_actions from partially_reducible evidence output. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| Severity: severity, | ||
| Summary: fmt.Sprintf("About %d%% of this run's turns appear to be data-gathering that could move to deterministic steps.", deterministicPct), | ||
| Evidence: fmt.Sprintf("agentic_fraction=%.2f turns=%d write_actions=%d", fingerprint.AgenticFraction, metrics.Turns, writeCount), | ||
| Evidence: fmt.Sprintf("agentic_fraction=%.2f turns=%d", fingerprint.AgenticFraction, metrics.Turns), |
There was a problem hiding this comment.
write_actions was removed from this evidence string due to the write count being double-counted, but writeCount is still computed as len(createdItems) + processedRun.Run.SafeItemsCount earlier in buildAgenticAssessments (and also in buildBehaviorFingerprint). Since SafeItemsCount is set from len(extractCreatedItemsFromManifest(...)) and createdItems is derived from the same manifest, this sums the same signal twice and can misclassify ActuationStyle / ResourceProfile and inflate other evidence (e.g. resource_heavy_for_domain). Suggest deriving writeCount from a single source (e.g. SafeItemsCount with fallback to len(createdItems) if needed) and reusing it consistently.
Two fixes based on PR review feedback on the agentic fraction / action minutes tracking work.
Double-counted
write_actionsin evidence (audit_agentic_analysis.go):writeCount = len(createdItems) + SafeItemsCountinflates the count because both terms derive from the same source (extractCreatedItemsFromManifest). Removedwrite_actions=%dfrom thepartially_reducibleevidence string — the reliable signals (agentic_fraction,turns) are sufficient for that assessment.Ambiguous console column header (
logs_report.go): Renamed theActionMinutesstruct tag header fromMinutestoAction Minutesto match the JSON field name and avoid confusion with wall-clock duration.⌨️ Start Copilot coding agent tasks without leaving your editor — available in VS Code, Visual Studio, JetBrains IDEs and Eclipse.