fix: exclude zero-spend (skipped) runs from token-optimizer AIC ranking

[mrfelton]

Problem

The Aggregate top workflows by AIC usage step in workflows/agentic-token-optimizer.md keeps every completed run, including runs whose agent skipped (activation/preconditions) and therefore consumed 0 AIC. Those zero-spend runs:

• pad run_count for skip-heavy workflows,
• dilute avg_ai_credits (the 0s drag the average down), and
• can surface never-spent workflows into the top-10 that drives target selection.

High-frequency, event-driven workflows (e.g. label/PR-triggered) are exactly this case — many completed runs, agent skipped on most.

Full write-up in #120. Distinct from (but compounding) the self-targeting guard in #119 / #121.

Fix

 [.runs[]
   | select(.status == "completed")
+  | select((.aic // 0) > 0)
   | { workflow_name, ai_credits: (.aic // 0), ... }
 ]

A "top workflows by AIC spend" ranking should only count runs that actually spent AIC. Applied to both workflows/ (template) and .github/workflows/ (active) copies.

⚠️ Compiler-version dependency — please read before merging

This is only safe once the compiler reliably computes AIC. On gh-aw < v0.79.4 — including v0.72.1, which this repo currently pins — the Copilot-engine 0-AIC bug is present (provider github_models not normalized; fixed in v0.79.4 via github/gh-aw#38314, with #38364/#38327/#38353/#38412). There, genuinely-expensive Copilot runs can report aic == 0, so this filter would drop real spend.

Recommendation: pair this with / gate it on bumping these workflows' compiler to ≥ v0.79.4. Once there, aic == 0 reliably means "agent skipped / no consumption" and the filter is correct. We run it in production on v0.79.4.

Notes

• Source .md only — regenerate .lock.yml in-repo (fork-built locks carry repo-scoped cron jitter / differing action-pin resolution).
• Validated gh aw compile --validate --no-emit (v0.72.1): 0 errors, 0 warnings.
• Touches a different region of the file than fix: stop token-optimizer self-targeting the AIC monitoring family #121, so the two are independent; if fix: stop token-optimizer self-targeting the AIC monitoring family #121 merges first this still applies cleanly.

Fixes #120

Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com