⚡ Copilot Token Optimization2026-06-24 — Build Test Suite

[github-actions]

Target Workflow: Build Test Suite

Source report: #5476
AIC per run: ~120.2 (highest in the fleet; token data unavailable — see instrumentation note below)
Total tokens per run: unavailable (0 reported — api-proxy token accounting not active)
Estimated cost per run: $0.00 reported (no token accounting)
LLM turns per run: unknown (not instrumented)
Runs analyzed: 3 (all successful)

⚠️ Instrumentation gap: All 50 Copilot runs in the latest report show 0 tokens/$0.00. Recommendations are based on static workflow analysis and AIC (Actions Intelligence Credits) as a proxy for model usage. AIC avg 120.2/run is the highest of all 15 workflows in the fleet.

Current Configuration

Setting	Value
Tools loaded	2 — bash: ["*"], github: (no toolsets restriction ≈ 22 tools)
Tools actually used	bash, add_comment, add_label (3 effective tools)
Network groups	12 — defaults, github, node, go, rust, crates.io, java, dotnet, bun.sh, deno.land, jsr.io, dl.deno.land
Pre-agent steps	❌ None
Post-agent steps	❌ None
Prompt size	~8,806 chars (~2,200 tokens)
Tasks	8 sequential ecosystems: Bun, C++, Deno, .NET, Go, Java, Node.js, Rust
Test repos cloned	8 (all inside agent via bash tool calls)

Recommendations

1. Restrict GitHub MCP Tool Toolsets

Estimated savings: ~10,200 tokens per LLM turn (~80% of tool schema overhead)

Currently github: loads the full GitHub MCP server with no toolsets: restriction. This loads ~22 tool schemas into the context window on every LLM turn. The workflow only uses three GitHub operations: post a PR comment, add a label, and optionally read the PR. Restricting to [pull_requests, issues] reduces loaded tools to ~5.

Change in .github/workflows/build-test.md frontmatter:

# BEFORE
tools:
  bash:
    - "*"
  github:
    github-token: "${{ secrets.GH_AW_GITHUB_MCP_SERVER_TOKEN }}"

# AFTER
tools:
  bash:
    - "*"
  github:
    github-token: "${{ secrets.GH_AW_GITHUB_MCP_SERVER_TOKEN }}"
    toolsets: [pull_requests, issues]

At ~600 tokens per tool schema, removing ~17 unused tools saves ~10,200 tokens per LLM turn. With an estimated 15–20 turns per run, this is ~153K–204K tokens saved per run.

---

2. Move Repo Clones and Runtime Setup to Pre-Agent Steps

Estimated savings: 8–16 LLM turns per run (eliminates clone bash-call/response cycles)

Currently the agent clones 8 repos, installs Bun via curl, installs Deno via curl, and creates ~/.m2/settings.xml — all as bash tool calls that consume LLM turns. These are deterministic setup operations with no branching logic; they do not require LLM judgment. Moving them to steps: pre-computes them before the agent starts.

Add to .github/workflows/build-test.md frontmatter:

steps:
  - name: Clone test repositories
    run: |
      git clone https://github.com/Mossaka/gh-aw-firewall-test-bun.git /tmp/test-bun &
      git clone https://github.com/Mossaka/gh-aw-firewall-test-cpp.git /tmp/test-cpp &
      git clone https://github.com/Mossaka/gh-aw-firewall-test-deno.git /tmp/test-deno &
      git clone https://github.com/Mossaka/gh-aw-firewall-test-dotnet.git /tmp/test-dotnet &
      git clone https://github.com/Mossaka/gh-aw-firewall-test-go.git /tmp/test-go &
      git clone https://github.com/Mossaka/gh-aw-firewall-test-java.git /tmp/test-java &
      git clone https://github.com/Mossaka/gh-aw-firewall-test-node.git /tmp/test-node &
      git clone https://github.com/Mossaka/gh-aw-firewall-test-rust.git /tmp/test-rust &
      wait
  - name: Install Bun
    run: |
      curl -fsSL (bun.sh/redacted) | bash
      echo "$HOME/.bun/bin" >> $GITHUB_PATH
  - name: Install Deno
    run: |
      curl -fsSL (deno.land/redacted) | sh
      echo "$HOME/.deno/bin" >> $GITHUB_PATH
  - name: Configure Maven proxy
    run: |
      mkdir -p ~/.m2
      cat > ~/.m2/settings.xml << 'SETTINGS'
      <settings>
        <proxies>
          <proxy>
            <id>awf-http</id><active>true</active><protocol>http</protocol>
            <host>squid-proxy</host><port>3128</port>
          </proxy>
          <proxy>
            <id>awf-https</id><active>true</active><protocol>https</protocol>
            <host>squid-proxy</host><port>3128</port>
          </proxy>
        </proxies>
      </settings>
      SETTINGS

Update the prompt body to remove the clone, Bun install, Deno install, and Maven proxy setup instructions from Tasks 1, 3, 6. Replace each with: Repos are pre-cloned to /tmp/test-<ecosystem>. Runtimes are pre-installed. Maven ~/.m2/settings.xml is pre-configured.

This eliminates ~22 LLM turns (8 clones × 2 turns each + 4 install turns + 2 config turns) and removes ~1,800 chars (~450 tokens) from the prompt body.

---

3. Compress Prompt Body

Estimated savings: ~600–800 tokens per run (constant per-turn overhead)

The 8,806-char prompt has three areas to cut:

a. Remove redundant bash command detail — Tasks 1–8 each include exact bash commands the agent already knows (e.g., bun install && bun test, go mod download && go test ./...). Once repos are pre-cloned, instructions can be shortened to:

Test each project: run install then test using the ecosystem's standard tool.

b. Collapse the combined output table template (~50 lines, ~1,200 chars). Replace the 18-row table template with a short prose description of the desired format:

Produce a single markdown table with columns: Ecosystem, Project, Build/Install (✅/❌), Tests (N/M passed), Status (✅ PASS / ❌ FAIL). Add build-test label if all pass.

c. Shorten the Error Handling section (currently 15 lines). Replace with:

On clone failure: record CLONE_FAILED and continue. If ALL clones fail, call safeoutputs missing_tool with reason "ALL_CLONES_FAILED".

Estimated prompt reduction: ~2,500 chars → ~1,600 chars (save ~225 tokens × N turns ≈ 3,375 tokens/run at 15 turns).

---

4. Add Post-Agent Validation Step

Estimated savings: 1–3 LLM turns per run (agent stops second-guessing its output)

Currently the agent has no post-validation hook, so it tends to re-read its output and self-check before posting the PR comment. Adding a post-steps: assertion that verifies the PR comment was posted reduces the agent's uncertainty-driven extra turns.

post-steps:
  - name: Verify PR comment posted
    uses: actions/github-script@v9
    with:
      script: |
        const comments = await github.rest.issues.listComments({
          owner: context.repo.owner,
          repo: context.repo.repo,
          issue_number: context.issue.number,
        });
        const hasResult = comments.data.some(c => c.body.includes('Build Test Suite Results'));
        if (!hasResult) core.setFailed('Agent did not post results comment');

---

Expected Impact

Metric	Current	Projected	Savings
GitHub tools in context	~22	~5	−77% tool schemas
Tokens/turn (tool overhead)	~13,200	~3,000	~−10,200/turn
Total tokens/run (est.)	~200K–300K	~60K–100K	~−55–70%
LLM turns/run (est.)	~20–25	~8–12	~−10 turns
AIC/run	~120	~45–60	~−50–60%
Prompt tokens	~2,200	~1,400	−800 tokens
Pre-agent repo clone time	In-agent (serial)	Parallel pre-step	−40–60% setup time

Implementation Checklist

• Add toolsets: [pull_requests, issues] under github: tool in frontmatter
• Add steps: block with parallel repo clones, Bun install, Deno install, Maven config
• Remove clone/install/Maven instructions from Tasks 1, 3, 6 body text
• Collapse combined output table template to prose description
• Shorten error handling section
• (Optional) Add post-steps: verification block
• Recompile: gh aw compile .github/workflows/build-test.md
• Post-process: npx tsx scripts/ci/postprocess-smoke-workflows.ts (if applicable)
• Trigger a test run via workflow_dispatch and compare AIC before/after
• Enable token accounting (api-proxy) to get real token data for future reports

Generated by Daily Copilot Token Optimization Advisor · 39.1 AIC · ⊞ 6.7K · ◷