Summary
The Copilot Token Usage Optimizer (run 24673974785) selected CI Optimization Coach (ci-coach.md) as its optimization target. The agent completed its analysis successfully but the detection step failed, preventing the issue from being published. This issue captures the optimizer's findings.
Target Selection
CI Optimization Coach was the #3 highest-token workflow in the 7-day window:
| Rank |
Workflow |
Runs |
Total Tokens |
Avg Tokens/Run |
| 1 |
Q |
2 |
5,995,014 |
2,997,507 |
| 2 |
Test Quality Sentinel |
13 |
3,863,221 |
297,171 |
| 3 |
CI Optimization Coach |
1 |
2,792,855 |
2,792,855 |
| 4 |
Documentation Noob Tester |
1 |
2,521,089 |
2,521,089 |
| 5 |
Contribution Check |
1 |
1,544,888 |
1,544,888 |
Token Usage Profile
| Metric |
Value |
Target |
| Total tokens |
2,792,855 |
300K-600K |
| Turns |
38 |
— |
| Runtime |
16 min |
— |
| Cache efficiency |
49% |
— |
| Input:output ratio |
181:1 |
— |
| Over target |
4.7x |
— |
| Over alert threshold (1M) |
2.8x |
— |
Root Causes
1. Massive prompt overhead (~26KB combined)
The workflow prompt is built from multiple large files:
ci-coach.md: 14,526 bytes — includes a verbose PR template with full example comparisons embedded directly in the promptshared/ci-data-analysis.md: 5,618 bytesshared/ci-optimization-strategies.md: 6,325 bytesshared/reporting.md: additional shared import
This combined ~26KB prompt is loaded into context on every turn. Across 38 turns, even a 4KB reduction saves ~152K tokens.
2. Half of turns spent on data gathering
About 50% of the 38 turns are iterative data-gathering — the agent explores analysis and refinement cycles rather than executing a single optimized path. Much of this could be moved to deterministic pre-steps.
3. Unnecessary/scattered tool calls
search_repositories called only once despite being available (wasted tool schema tokens)issue_read invoked 5 times for reading CI-related issues not directly relevant to optimization- Full
[default] GitHub toolset loads ~22 tools when only a few are needed
4. Expensive pre-steps generating large context
The ci-data-analysis.md shared component runs a full build cycle (make deps-dev, make lint, make build, make recompile, make test-unit) and downloads 100 CI runs + artifacts. The resulting context files are large and the agent reads them across multiple turns.
5. Poor cache efficiency (49%)
Each run generates new artifacts, and the prompt structure doesn't align well with how the agent processes information, leading to poor context reuse.
6. Excessive turn count for task complexity
38 turns is excessive for what should be: read CI config → identify problems → propose fixes → create PR. The agent explores deeply instead of exiting early when there's nothing actionable to optimize.
Recommendations
R1: Trim the embedded PR template — Est. savings: ~152K tokens/run
The ci-coach.md prompt contains a verbose PR template with full example before/after comparisons. Move this to a separate reference file or condense it to a skeleton that the agent fills in, saving ~4KB from the system prompt × 38 turns.
Action: Refactor the PR template out of the inline prompt in ci-coach.md.
R2: Narrow GitHub toolset — Est. savings: ~50-60K tokens/turn overhead
Replace toolsets: [default] with a minimal set. The workflow primarily needs issues (for reading CI-related issues) and pull_requests (for creating PRs). Remove repos, code_search, and other unused toolsets.
Action: Change toolsets: [default] → toolsets: [issues, pull_requests] in ci-coach.md frontmatter.
R3: Pre-compute analysis data in bash steps — Est. savings: ~10-15 turns
Move the data-gathering phase into deterministic bash pre-steps. Parse CI run data, identify failure patterns, and compute statistics before the agent starts, then pass structured summaries rather than raw data.
Action: Add bash pre-steps to ci-coach.md or shared/ci-data-analysis.md that produce a structured JSON summary.
R4: Add early exit criteria
If the CI is healthy (no significant failures, no regressions), the agent should exit early instead of exploring for 38 turns. Add explicit instructions: "If no actionable optimizations found after Phase 1 analysis, publish a no-op issue and exit."
Action: Add an early-exit clause to the prompt after the initial analysis phase.
R5: Condense shared components
The two shared components (ci-data-analysis.md at 5.6KB and ci-optimization-strategies.md at 6.3KB) may contain redundant content. Review and consolidate where possible.
Action: Audit both shared components for overlap and condense.
R6: Enable cli-proxy
Add cli-proxy: true to features to enable direct gh CLI access, reducing dependence on the full GitHub MCP toolset for simple operations like reading files.
Action: Add cli-proxy: true to the features section.
Projected Impact
| Scenario |
Tokens/run |
| Current |
2,792,855 |
| R1 (trim PR template) |
~2,640K |
| R1 + R2 (narrow toolset) |
~2,100K |
| R1 + R2 + R3 + R4 (pre-compute + early exit) |
~600K-1.2M |
| All recommendations |
~300K-600K (target range) |
Source
Analysis by Copilot Token Usage Optimizer, run 24673974785 (April 20, 2026). The optimizer completed its analysis but the detection step failed, so this issue was manually extracted from the agent's output.
Summary
The Copilot Token Usage Optimizer (run 24673974785) selected CI Optimization Coach (
ci-coach.md) as its optimization target. The agent completed its analysis successfully but the detection step failed, preventing the issue from being published. This issue captures the optimizer's findings.Target Selection
CI Optimization Coach was the #3 highest-token workflow in the 7-day window:
Token Usage Profile
Root Causes
1. Massive prompt overhead (~26KB combined)
The workflow prompt is built from multiple large files:
ci-coach.md: 14,526 bytes — includes a verbose PR template with full example comparisons embedded directly in the promptshared/ci-data-analysis.md: 5,618 bytesshared/ci-optimization-strategies.md: 6,325 bytesshared/reporting.md: additional shared importThis combined ~26KB prompt is loaded into context on every turn. Across 38 turns, even a 4KB reduction saves ~152K tokens.
2. Half of turns spent on data gathering
About 50% of the 38 turns are iterative data-gathering — the agent explores analysis and refinement cycles rather than executing a single optimized path. Much of this could be moved to deterministic pre-steps.
3. Unnecessary/scattered tool calls
search_repositoriescalled only once despite being available (wasted tool schema tokens)issue_readinvoked 5 times for reading CI-related issues not directly relevant to optimization[default]GitHub toolset loads ~22 tools when only a few are needed4. Expensive pre-steps generating large context
The
ci-data-analysis.mdshared component runs a full build cycle (make deps-dev,make lint,make build,make recompile,make test-unit) and downloads 100 CI runs + artifacts. The resulting context files are large and the agent reads them across multiple turns.5. Poor cache efficiency (49%)
Each run generates new artifacts, and the prompt structure doesn't align well with how the agent processes information, leading to poor context reuse.
6. Excessive turn count for task complexity
38 turns is excessive for what should be: read CI config → identify problems → propose fixes → create PR. The agent explores deeply instead of exiting early when there's nothing actionable to optimize.
Recommendations
R1: Trim the embedded PR template — Est. savings: ~152K tokens/run
The
ci-coach.mdprompt contains a verbose PR template with full example before/after comparisons. Move this to a separate reference file or condense it to a skeleton that the agent fills in, saving ~4KB from the system prompt × 38 turns.Action: Refactor the PR template out of the inline prompt in
ci-coach.md.R2: Narrow GitHub toolset — Est. savings: ~50-60K tokens/turn overhead
Replace
toolsets: [default]with a minimal set. The workflow primarily needsissues(for reading CI-related issues) andpull_requests(for creating PRs). Removerepos,code_search, and other unused toolsets.Action: Change
toolsets: [default]→toolsets: [issues, pull_requests]inci-coach.mdfrontmatter.R3: Pre-compute analysis data in bash steps — Est. savings: ~10-15 turns
Move the data-gathering phase into deterministic bash pre-steps. Parse CI run data, identify failure patterns, and compute statistics before the agent starts, then pass structured summaries rather than raw data.
Action: Add bash pre-steps to
ci-coach.mdorshared/ci-data-analysis.mdthat produce a structured JSON summary.R4: Add early exit criteria
If the CI is healthy (no significant failures, no regressions), the agent should exit early instead of exploring for 38 turns. Add explicit instructions: "If no actionable optimizations found after Phase 1 analysis, publish a no-op issue and exit."
Action: Add an early-exit clause to the prompt after the initial analysis phase.
R5: Condense shared components
The two shared components (
ci-data-analysis.mdat 5.6KB andci-optimization-strategies.mdat 6.3KB) may contain redundant content. Review and consolidate where possible.Action: Audit both shared components for overlap and condense.
R6: Enable
cli-proxyAdd
cli-proxy: trueto features to enable directghCLI access, reducing dependence on the full GitHub MCP toolset for simple operations like reading files.Action: Add
cli-proxy: trueto the features section.Projected Impact
Source
Analysis by Copilot Token Usage Optimizer, run 24673974785 (April 20, 2026). The optimizer completed its analysis but the detection step failed, so this issue was manually extracted from the agent's output.