[deep-report] Expose bundled safe_outputs step logs (from #47855) via audit/logs MCP tooling #54756

Description

PR #47855 (merged 2026-07-25) added stdout/stderr capture for the Process Safe Outputs step, teed to /tmp/gh-aw/process-safe-outputs.stdout.log / .stderr.log and bundled into the safe-output-items artifact specifically so failures could be root-caused. However, today's Safe Output Health Report (discussion #54725) could not retrieve this log text for two safe_outputs job failures — the agenticworkflows/logs MCP tool's supported artifact sets (activation, agent, all, detection, evals, experiment, firewall, github-api, mcp, usage) don't include a set that surfaces these bundled log files, so the fix from #47855 is effectively unreachable by the tooling meant to consume it.

Expected Impact

Unblocks root-causing of safe_outputs job failures (like the PR Sous Chef issue above) without needing raw GitHub Actions console access — the data already exists on disk/in the artifact, it just isn't wired into the audit tool's read path.

Suggested Agent

New Agent / general-purpose — extend the agenticworkflows MCP bridge to fetch/expose the existing process-safe-outputs.stdout.log/.stderr.log artifact contents (truncated) as a safe_outputs artifact set or via steps[].error_excerpt.

Estimated Effort

Medium (1-4 hours)

Data Source

DeepReport analysis of Safe Output Health Report (discussion #54725, 2026-08-22), Work Item 2, cross-referenced with merged PR #47855.

Generated by 🔬 Deep Report · agent · 109.8 AIC · ⌖ 8.86 AIC · ⊞ 12.4K · ◷

• expires on Aug 23, 2026, 10:29 PM UTC-08:00