Long-loop run history truncation drops container + early-iteration step logs (compactStepsForHistory)

[os-zhuang]

Follow-up to #1505 (region step logs) — now that per-iteration / per-region body steps are recorded (engine) and surfaced (objectui#2667), a durable-history edge case remains.

The gap

AutomationEngine.compactStepsForHistory() bounds the persisted step log to the last MAX_PERSISTED_HISTORY_STEPS entries:

// packages/services/service-automation/src/engine.ts
private compactStepsForHistory(steps: StepLogEntry[]): StepLogEntry[] {
    return steps.slice(-MAX_PERSISTED_HISTORY_STEPS).map(...);
}

This keeps steps_json cheap under retention (#2585), but with the #1505 structured-region work a single loop can now emit many steps (iterations × body-steps). For a long loop the tail-slice can drop:

• the loop/parallel/try_catch container step itself (it precedes all its body steps), and
• the early iterations' body steps.

Impact

listRuns / getRun serve the durable row after a process restart or ring-buffer eviction (the in-memory log has full detail; durable history is the fallback). So post-restart, the Runs surface shows only the loop's last few iterations, and — because the container step was sliced off — the objectui FlowRunsPanel (#2667) can no longer attach those body steps to their container. Its buildStepTree degrades them safely to top-level rows (visible, so nothing vanishes), but the per-iteration grouping and early-iteration detail are lost exactly for the big loops where they matter most for debugging/audit.

Possible directions (to decide)

• Retain region boundaries when compacting: always keep each container step + a per-iteration summary (e.g. first/last iteration, failed iterations), downsampling the middle.
• Or keep the newest N per region rather than N globally, so early iterations of a still-relevant region survive.
• Or raise/relax the cap for region-bearing runs while keeping a hard byte ceiling on steps_json (Automation run observability — follow-ups (retention, discoverable Runs surface, durable run detail) #2585).
• Whichever: keep buildStepTree's orphan-degradation as the safety net.

Files

• packages/services/service-automation/src/engine.ts — compactStepsForHistory, MAX_PERSISTED_HISTORY_STEPS, recordTerminal persistence path.
• Consumer: objectui FlowRunsPanel.tsx (buildStepTree).