stale-counts findings in SKILL.md files (Home-Lab, Steam) need investigation

[TMHSDigital]

Background

Phase 2 Session D-0 audit ran the drift checker against all 9 tool repos. stale-counts produced 62 warnings total.

61 of 62 warnings live in repo-root agent files (AGENTS.md, CLAUDE.md) where aggregate narratives like 177 skills, 71 rules are clearly stale ecosystem-summary boilerplate.

5 warnings are different: they are inside skills/<name>/SKILL.md files, where the aggregate count may be legitimate technical content rather than stale narrative.

Specific findings

Repo	File	Count
Home-Lab-Developer-Tools	skills/secrets-management/SKILL.md	1
Steam-Cursor-Plugin	skills/steam-api-reference/SKILL.md	4

Hypothesis

These may be matches against patterns like:

• "3 firewall rules" (Home-Lab secrets-management referencing security configurations)
• "5 API tools" or "12 commands" (Steam API reference describing Steam Web API surface)

If so, these are technical specifications, not stale ecosystem narratives. The drift checker's regex (\b\d+\s+(skills?|rules?|MCP\s+tools?|tools?|commands?|hooks?)\b) cannot distinguish between "3 firewall rules" (legitimate domain content) and "71 rules" (stale ecosystem summary).

What needs investigating

1. Read each of the 5 flagged contexts and classify as legitimate-domain vs stale-narrative.
2. If all are legitimate-domain: design a way for stale-counts to skip them. Options:
   • Add drift-ignore pragma in the SKILL.md frontmatter (workable today, but burdens tool authors with manual annotation).
   • Tighten the regex to exclude common technical-noun pairings (fragile).
   • Add a config knob like stale_counts_skip_patterns (most flexible).
3. If any are actually stale: leave the warning, fix the content.

Related

• Linked to: validate-counts vs stale-counts reconciliation (will be cross-linked from the other Session D-0 issue).
• Phase 2 Session B introduced the check (PR feat: drift checker additional checks and policy data (Phase 2 Session B) #8).
• Phase 2 Session B Q3: warn-only, suggest removal.

[TMHSDigital]

Cross-link: see #12 for the broader validate-counts vs stale-counts reconciliation.

[TMHSDigital]

Investigation complete. 5 findings categorized:

• 4 category A (genuine count claims): 3 currently accurate, 1 stale
• 1 category C (false positive from example dialogue)
• 0 category B (narrative aggregates)

Resolution:

• Steam content-fix PR fix: correct stale tool count in steam-api-reference SKILL.md Steam-Cursor-Plugin#13 (merged 2026-04-25 23:04 UTC, commit c9cea61a): corrects the off-by-one drift on steam-api-reference/SKILL.md line 123 (the one stale finding).
• Drift-check enhancement drift-check: stale-counts should skip example dialogue sections in SKILL.md #37: addresses the category C false positive systematically.
• Comment on Reconcile validate-counts vs drift checker stale-counts pattern across ecosystem #12 documents the "do not suppress SKILL.md" finding for the eventual policy session.

Full investigation written up in dtd13-investigation.md (workspace root).

The drift-check stale-counts pass did its job: caught a real off-by-one bug that had drifted unnoticed. Validates the infrastructure investment from Phase 2.

[TMHSDigital]

Correction to my closing comment above.

The framing "drift-check did its job. Validates the infrastructure investment from Phase 2." was incorrect.

DTD#12 investigation (resolved in #38, shipped as v1.9.0) revealed that the cursor-plugin and mcp-server type-level skip_checks entries in standards/drift-checker.config.json were suppressing stale-counts entirely (whole-check skip, not file-scoped). The Steam off-by-one bug was found by manually running stale_counts.py during DTD#13 investigation — not by the production drift-check workflow, which had stale-counts suppressed for Steam-Cursor-Plugin.

The bug fix in TMHSDigital/Steam-Cursor-Plugin#13 is still correct (the SKILL.md content matches reality now). But the verification narrative was wrong: production drift-check did not catch this; manual investigation did.

DTD#12's resolution removes the over-broad suppression and encodes the AGENTS.md/CLAUDE.md narrative-aggregate policy in the check itself (plus DTD#37's example-dialogue skip). Empirical post-merge verification on Steam now surfaces 4 stale-counts findings — all accurate (because Steam#13 fixed the off-by-one), but the check is actually running against SKILL.md files for the first time since v1.7.5.

Net: production drift-check now actually catches this class of bug across all tool repos. Closing my earlier overstatement of the verification chain.