[copilot-token-optimizer] Optimize Copilot cloud agent: reduce PR re-trigger cascades and setup overhead #27

Target Workflow

Copilot cloud agent — the highest-token agentic workflow in this repository (150,233 tokens across 30 runs per last audit), selected as the primary optimization target. No "Token"-excluded and not recently optimized with complete data.

Analysis period: 2026-05-05 – 2026-05-14 | Runs analyzed: 20

Token Profile

Ranked Recommendations

1. 🔴 Suppress rapid re-triggers on the same PR (Est. savings: ~500–2,500 tokens/day on active review cycles)

Evidence: On 2026-05-05, PR #9 alone triggered 6 consecutive agent runs between 07:22–08:01 UTC, with 2 cancellations (runs §25364643611 and §25362630400 cancelled). PR #10 triggered 2 additional runs. This pattern – multiple comments or pushes on a PR firing sequential agent runs before previous ones complete – is the primary source of both wasted tokens and cancellations.

Action: Add a concurrency group to the Copilot cloud agent invocation configuration to cancel in-progress runs for the same PR when a new trigger arrives, or add a debounce delay so that rapid successive triggers (within 60–90 s of each other) consolidate into a single agent run.

concurrency:
  group: copilot-agent-${{ github.event.issue.number || github.event.pull_request.number }}
  cancel-in-progress: true

This would have eliminated ~4 redundant runs on May 5 alone, saving an estimated 20,000 tokens that day.

2. 🟡 Cache or skip smoke-test compilation on every agent invocation (Est. savings: ~7–9 s setup overhead per run, negligible tokens)

Evidence: copilot-setup-steps.yml runs gh aw add ... && gh aw compile --validate on every Copilot agent invocation (step "Smoke test published workflows via gh aw add": 7 s on recent runs). This compiles the two token-audit/optimizer workflows into a temp directory every time, regardless of whether those workflows have changed.

Action: Gate the smoke-test step with a condition, or move it to only run on changes to copilot-setup-steps.yml (which already triggers via push: paths). Consider caching the gh aw compile output using actions/cache keyed on the workflow source hash to skip redundant compilation.

References: §25846852022, §25813362645

3. 🟡 Investigate high-variance processing times (Est. savings: 500–2,000 tokens on long-tail runs)

Evidence: Processing Request times ranged from 113 s (1.9 min) to 563 s (9.4 min) across observed runs — a 5× spread. Long runs correlate with complex PR contexts (many comments, large diffs).

Action: Review whether the agent is loading unnecessary context (full repository history, all PR comments) on every invocation. Add instructions to limit context to the last N comments (e.g., last 5) and to only read directly changed files. This is the highest-leverage lever for token reduction since it acts on every turn in long sessions.

Caveats

• Token counts are estimated from a 30-run rolling audit (150,233 total); per-run breakdowns are not available in the current pre-downloaded data.
• Cancellation savings are conservative — cancelled runs may have consumed 0–100% of a full run's tokens depending on when they were cancelled.
• Recommendation 1 (concurrency) requires verifying that the Copilot cloud agent's invocation workflow supports concurrency groups.
• Only 20 runs were analyzed; the May 5 PR review spike may not be representative of typical usage.

Generated by Copilot Token Usage Optimizer · ● 9.4M · ◷

• expires on May 21, 2026, 8:47 AM UTC