Auto Run summary only reports last loop iteration, not cumulative session stats

[kringle25]

Problem

When an Auto Run session spans multiple loop iterations over days/weeks, the "Auto Run stopped" summary entry only reports stats from the last loop iteration, not the cumulative session totals. This significantly undercounts user activity.

Observed Data

An 8-day Auto Run session ("Arnold") with 741 history entries:

Metric	Summary Reported	Actual (all entries)	Under-reported by
Tasks	61	741 (647 success, 22 failed)	12x
Duration	5h 55m	145.9 hours	25x
Tokens	1.87M	16.09M	8.6x
Cost	$1,249.60	$6,280.57	5x

Root Cause

The summary entry (entries[0], type AUTO, success: false) is generated when the auto-run stops. It appears to only aggregate stats from the current/last loop iteration (loop iteration 00001 in the prompt), not across all entries in the session history file.

Expected Behavior

The "Auto Run stopped" summary should aggregate across all entries in the session:

• Total tasks = count of all entries (excluding the summary entry itself)
• Total duration = sum of all elapsedTimeMs values
• Total tokens = sum of all inputTokens + outputTokens across all entries
• Total cost = sum of all totalCostUsd across all entries

Impact

• Achievement/level progress is calculated on undercounted stats
• User dashboard shows incorrect session totals
• Cost tracking is inaccurate (showed $1.2K instead of $6.3K)

Related

• Auto-run stop/kill unreliable — hung processes survive termination #733 — Auto-run stop/kill unreliable (same session exhibited both bugs)

Environment

• macOS Darwin 25.4.0
• Maestro desktop app (Electron)
• Session ran March 29 – April 5, 2026

[pedramamini]

Thanks for the detailed report and analysis, @kringle25! The data table really helped pinpoint this.

We believe we have a fix in #735. The root cause is that the in-memory cumulative counters (totalCompletedTasks, totalInputTokens, etc.) reset each time startBatchRun is invoked — e.g. after an app restart during a long-running session. Individual task history entries persist on disk across restarts, but the final summary was only using the in-memory counters from the last invocation.

The fix reconciles the summary with all persisted history entries for the session before generating the final stats, using Math.max(inMemory, historyDerived) so sessions spanning restarts now report accurate cumulative totals.

Would love your thoughts on the approach — does this match what you were seeing?

[kringle25]

no problem, I found this in my attempt to understand the other issue with Claude and Auto-Run hanging all the time in the middle of sessions and unable to stop the session and restart it