[copilot-token-optimizer] Token Optimization: Workflow Health Manager - Meta-Orchestrator #28967

Target workflow: workflow-health-manager.md — selected as the highest-token workflow not optimized in the past 14 days.

Analysis Period

• Window: 2026-04-21 → 2026-04-28 (7 days)
• Runs analyzed: 1 (single scheduled run)
• Conclusion: success (0 errors, 0 warnings)

Token Profile

Cache efficiency of 48.2% is well below the typical 70–80% range. Since the system prompt (~4,077 tokens estimated) is repeated every turn in a long-context conversation, uncached turns dominate cost.

Ranked Recommendations

1. 🔴 Trim the workflow prompt — remove verbose restatements (~450K tokens/run estimated)

Estimated savings: ~450,000 tokens/run
Evidence: The markdown source (456 lines, ~4,077 prompt tokens) contains redundant elaboration. Large sections like "Important Guidelines", "Success Metrics", and the full 60-line dashboard template are re-injected into every one of the 28 turns. At ~65K avg input/turn, even a 25% reduction in per-turn context saves ~18K tokens × 28 turns ≈ 500K tokens.

Actions:

• Remove the "Important Guidelines" section (~30 lines) — it restates the phase descriptions.
• Remove the "Success Metrics" section (~15 lines) — qualitative framing the agent doesn't use at runtime.
• Replace the full dashboard template with a concise skeleton (~15 lines instead of 60). The agent already infers format from heading structure.
• Keep the 5-phase execution plan and the tool/permission declarations — those are essential.

2. 🟠 Move discovery + compilation check to a deterministic pre-agent bash step (~150–200K tokens/run estimated)

Estimated savings: ~175,000 tokens/run (≈3–4 turns eliminated)
Evidence: Phase 1 (Discovery) instructs the agent to list 120+ .md files, parse frontmatter, and run gh aw compile --validate. These are fully deterministic — they don't need inference. Running them in a frontmatter steps: block before the agent starts eliminates 3–4 early turns and hands the agent a pre-built workflow inventory.

Actions:

steps:
  - name: build-inventory
    run: |
      gh aw compile --validate 2>&1 > /tmp/gh-aw/agent/compile-validate.txt
      ls .github/workflows/*.md | grep -v shared/ > /tmp/gh-aw/agent/workflow-list.txt

Then reference /tmp/gh-aw/agent/workflow-list.txt and /tmp/gh-aw/agent/compile-validate.txt in the prompt.

3. 🟠 Pre-load shared metrics in a deterministic step (~100K tokens/run estimated)

Estimated savings: ~100,000 tokens/run (≈1–2 turns eliminated)
Evidence: The prompt instructs the agent to read metrics/latest.json and metrics/daily/*.json from repo-memory. These file reads cost 1–2 full turns where context is established before the agent can interpret the data. Pre-loading the metrics in a frontmatter step and summarizing them (jq select top 10 failing workflows) would replace 1–2 agent turns with a single pre-computed summary injected into the initial prompt.

Actions:

steps:
  - name: load-metrics
    run: |
      cat /tmp/gh-aw/repo-memory/default/metrics/latest.json \
        | jq '[.workflow_runs | to_entries[] | select(.value.success_rate < 0.8)] | sort_by(.value.success_rate) | .[0:20]' \
        > /tmp/gh-aw/agent/failing-workflows.json 2>/dev/null || echo '[]' > /tmp/gh-aw/agent/failing-workflows.json

4. 🟡 Batch sequential GitHub API calls (~50K tokens/run estimated)

Estimated savings: ~50,000 tokens/run
Evidence: The run made 4 GitHub API calls (from github_api_calls: 4). The github toolset (with default + actions) is configured, and the behavior fingerprint reports tool_breadth: narrow with only 28 turns — the calls are likely sequential single-workflow queries. Batching workflow run queries (e.g., fetch recent runs for multiple workflows in one list_workflow_runs call filtered by date) reduces round-trips.

Summary of Expected Impact

Expected reduction from 1.84M → ~1.07M tokens/run (~42%).

References:

• §25052372422 — analyzed run (2026-04-28)

Generated by Copilot Token Usage Optimizer · ● 1.2M · ◷

• expires on May 5, 2026, 7:09 PM UTC