[integrity] DIFC Integrity-Filtered Events Report — 2026-03-19

[github-actions[bot]]

Executive Summary

In the last 7 days, 1,543 DIFC integrity-filtered events were detected across 52 workflow runs. The most frequently filtered tool was list_issues (1,132 events — 73% of all events), and filtering was heavily concentrated on a single day: 2026-03-19 accounted for 1,473 events (95%), compared to just 70 on 2026-03-18. This sharp concentration suggests the surge is driven by a high volume of issue-triggering activity on March 19 rather than a systemic change.

The two dominant workflows — Auto-Triage Issues (616 events) and AI Moderator (615 events) — together account for 80% of all filtered events. Both workflows routinely call list_issues and search_issues on content contributed by external users, which the DIFC system correctly flags as having lower integrity (none:all tagged) than the agent's required level. This is expected behavior for community-facing workflows processing untrusted issue/PR bodies. The Sub-Issue Closer and Daily Documentation Updater show notable spikes (100 and 73 events respectively) that may warrant review.

Key Metrics

Metric	Value
Total filtered events	1,543
Unique tools filtered	7
Unique workflows affected	12
Most common integrity tag	none:all (1,543 events)
Events with unapproved:all tag	210
Events with private secrecy tag	8
Busiest day	2026-03-19 (1,473 events)
MCP server(s) involved	github (100%)

📈 Events Over Time

The timeline reveals a sharp spike on 2026-03-19, with virtually no activity on 2026-03-18 (70 events). This is consistent with batch processing runs scheduled or triggered on March 19. The near-total concentration in a single day means the 7-day trend is essentially a point-in-time snapshot — additional days of data will be needed to assess whether this level is typical or elevated.

🔧 Top Filtered Tools

list_issues dominates at 1,132 events (73%), followed by search_issues (244, 16%) and search_pull_requests (97, 6%). All seven affected tools are GitHub MCP tools that return content authored by external users — the filtering is triggered because those resources carry none:all integrity tags, which are insufficient for the agent's trust requirements. get_me and list_discussions appear as minor contributors (5 and 3 events respectively), suggesting they are edge cases rather than systemic issues.

🏷️ Workflow & Tag Breakdown

All 1,543 filtered events carry the none:all integrity tag, meaning every event involves resources that have not been assigned any positive integrity classification. 210 events (14%) additionally carry unapproved:all, indicating resources authored by contributors who have not been approved. Only 8 events involve private secrecy-tagged resources, meaning secret leakage risk is minimal.

📋 Per-Workflow Breakdown

Workflow	Filtered Events
Auto-Triage Issues	616
AI Moderator	615
Sub-Issue Closer	100
Daily Documentation Updater	73
Smoke Codex	44
Dev	43
Daily Team Evolution Insights	21
Security Alert Burndown	15
Daily Documentation Healer	5
GitHub MCP Structural Analysis	3
Smoke Update Cross-Repo PR	3
Smoke Gemini	2
Workflow Health Manager - Meta-Orchestrator	2

📋 Per-Server Breakdown

MCP Server	Filtered Events
github	1,543

All filtering originates from the github MCP server. No other MCP servers (playwright, custom tools, etc.) triggered integrity filtering in this period.

📋 Top Filtered Issues (by specific resource)

The most frequently blocked individual resources by reason string:

Issue Resource	Events	Tag Required
issue:#18480	22	unapproved:all + approved:all
issue:#21709	14	unapproved:all + approved:all
issue:#21708	14	unapproved:all + approved:all
issue:#21722	13	unapproved:all + approved:all

These specific issues appear repeatedly across multiple workflow runs (e.g., Auto-Triage Issues, AI Moderator), suggesting they are high-activity community issues that are continuously being processed by automated workflows.

💡 Tuning Recommendations

1. Auto-Triage Issues & AI Moderator — consider bulk list_issues integrity elevation: These workflows call list_issues 20–25 times per run in rapid succession. If the intended behavior is to process all open issues regardless of contributor status, consider configuring the activation job to apply a baseline read-only:issues integrity grant at the workflow level, allowing the agent to consume issue lists without triggering per-resource DIFC blocks.

2. unapproved:all resources (210 events): Issues authored by contributors who are not on the approved list are generating a large share of filtered events. Review whether the Auto-Triage Issues and AI Moderator workflows should be permitted to act on unapproved contributor issues, and if so, add an explicit integrity policy that covers this case.

3. Sub-Issue Closer spike (100 events in one run): This is a high count for a single run. Investigate whether the workflow is fetching the full open issue list (which would produce many none:all tagged resources in one batch). If the run was processing a large backlog, the spike may be a one-time event; if it's recurring, consider paginating or filtering the issue list before passing resources to the agent.

4. Daily Documentation Updater (73 events, one run): Similarly, 73 events in a single run of this workflow suggests it is calling search_pull_requests or list_issues at scale. Verify whether the workflow's prompt or tool usage has changed recently and whether integrity tags could be elevated for internal PRs.

5. Private secrecy-tagged events (8 events): A small but non-zero set of events involves private secrecy-tagged resources (synthetic issue:#0 entries). Confirm these are expected (placeholder resources from activation context) and not real private data leaking through the pipeline.

6. get_me filtering (5 events across AI Moderator): The get_me tool should not normally return integrity-tagged resources. These events likely indicate the GitHub MCP server is applying conservative defaults. Consider whitelisting get_me as a zero-integrity-risk tool if confirmed safe.

7. Monitor daily trend as data accumulates: With only 2 days of data in this window, it is not possible to establish a baseline. Run this report daily for 1–2 weeks to determine whether 1,000–1,500 events/day is normal or an anomaly.

---

Generated by the Daily Integrity Analysis workflow
Analysis window: Last 7 days | Repository: github/gh-aw
Run: §23318632802

AI generated by Daily DIFC Integrity-Filtered Events Analyzer · history

• expires on Mar 22, 2026, 10:16 PM UTC

[github-actions[bot]]

This discussion has been marked as outdated by Daily DIFC Integrity-Filtered Events Analyzer.

A newer discussion is available at Discussion #21933.