[deep-report] Fix safe-output aggregator undercount (SafeItemsCount=0 / read_only since 05-31)

[github-actions]

Description

The Safe Output Health Report (#42907) and API Consumption Report (#42942) both note a known bash_safeoutputs aggregator undercount since 2026-05-31: every run reports SafeItemsCount=0 and actuation_style=read_only because the Process Safe Outputs step logs are not pre-bundled, so message-level (per-handler) safe-output counts are unavailable. Job-level safe_outputs conclusions remain authoritative (100% success, 6 clean days), but all message-level metrics are blind.

Fix: ensure the Process Safe Outputs step log is bundled into the run artifact before the aggregator reads it, so SafeItemsCount and per-handler counts (create_issue/create_discussion/add_comment/...) are populated again.

Expected Impact

Restores fleet-wide safe-output message-level observability (currently a lower bound of 0 for 30+ days), unblocking accurate cost/actuation reporting.

Suggested Agent

Copilot SWE Agent / Agentic Maintenance.

Estimated Effort

Medium (1-4 hours) — log-bundling ordering fix in the safe-outputs pipeline.

Data Source

DeepReport Intelligence Briefing 2026-07-02; Safe Output Health #42907, API Consumption #42942.

Generated by 🔬 Deep Report · 266.6 AIC · ⌖ 17.9 AIC · ⊞ 10.2K · ◷

• expires on Jul 4, 2026, 7:36 AM UTC-08:00

[pelikhan]

@copilot compute safe outputs aggregates data in the conclusion job and include the usage artifact so that the command finds that data without having to download extra artifact.