⚡ Copilot Token Optimization2026-06-30 — Test Coverage Improver #5715

Target Workflow: test-coverage-improver

Source report: #5714
Estimated cost per run: N/A (token telemetry absent — AIC proxy used)
AIC per run: 304.1 (highest in report)
Actions minutes/run: 20.0 min
GitHub API calls/run: 10
Model: claude-haiku-4-5
Runs analyzed: 1 (2026-06-30)

⚠️ Token telemetry is absent for all runs in the report. AIC (Actions Intelligence Credits) is used as the proxy signal. Recommendations focus on reducing LLM input token volume based on static prompt analysis.

Current Configuration

Recommendations

1. Truncate the existing test file injection to the first 100 lines

Estimated savings: ~2,800–5,500 tokens/run (~15–25% of estimated total input)

The Select target file and inject content step injects the entire existing test file into the prompt for style reference. When a well-covered target is selected, this can be enormous:

The agent only needs the first 100 lines (import setup, mock patterns, first few describe blocks) to infer style — not the full file.

Change in .github/workflows/test-coverage-improver.md, step Select target file and inject content:

-      TEST_FILE="${TARGET%.ts}.test.ts"
-      echo "TARGET_TEST_FILE=$TEST_FILE" >> "$GITHUB_OUTPUT"
-      {
-        echo "TEST_CONTENT<<EOF"
-        cat "$TEST_FILE" 2>/dev/null || echo "(test file does not exist yet)"
-        echo "EOF"
-      } >> "$GITHUB_OUTPUT"
+      TEST_FILE="${TARGET%.ts}.test.ts"
+      echo "TARGET_TEST_FILE=$TEST_FILE" >> "$GITHUB_OUTPUT"
+      {
+        echo "TEST_CONTENT<<EOF"
+        if [ -f "$TEST_FILE" ]; then
+          LINE_COUNT=$(wc -l < "$TEST_FILE")
+          head -100 "$TEST_FILE"
+          if [ "$LINE_COUNT" -gt 100 ]; then
+            echo ""
+            echo "// ... ($((LINE_COUNT - 100)) more lines truncated for brevity — file exists at $TEST_FILE)"
+          fi
+        else
+          echo "(test file does not exist yet)"
+        fi
+        echo "EOF"
+      } >> "$GITHUB_OUTPUT"

2. Replace COVERAGE_SUMMARY.md injection with a minimal fresh table

Estimated savings: ~850 tokens/run (~5–8% of estimated total input)

The step Read COVERAGE_SUMMARY.md reads the static COVERAGE_SUMMARY.md file (4,356 bytes). This file is a stale artifact from a previous PR containing irrelevant sections: "Coverage Improvements in This PR", "Before/After comparisons", "Integration with CI/CD", "How to View Coverage" docs, etc. The agent only needs current per-file coverage percentages.

The run already generates coverage/coverage-summary.json in the Run coverage step. Use it directly for a concise table:

Change in .github/workflows/test-coverage-improver.md, replace the Read COVERAGE_SUMMARY.md step:

-  - name: Read COVERAGE_SUMMARY.md
-    id: coverage-md
-    run: |
-      {
-        echo "COVERAGE_MD<<EOF"
-        cat COVERAGE_SUMMARY.md 2>/dev/null || echo "(COVERAGE_SUMMARY.md not found)"
-        echo "EOF"
-      } >> "$GITHUB_OUTPUT"
+  - name: Generate compact coverage table
+    id: coverage-md
+    run: |
+      {
+        echo "COVERAGE_MD<<EOF"
+        node -e "
+          const d = JSON.parse(require('fs').readFileSync('coverage/coverage-summary.json','utf8'));
+          const t = d.total;
+          console.log('Overall: stmts=' + t.statements.pct + '%, branches=' + t.branches.pct + '%, fns=' + t.functions.pct + '%, lines=' + t.lines.pct + '%');
+          console.log('');
+          console.log('| File | Stmts | Branches | Fns | Lines |');
+          console.log('|------|-------|----------|-----|-------|');
+          Object.entries(d).filter(([k]) => k !== 'total').forEach(([k,v]) => {
+            const f = k.split('/').pop();
+            console.log('| ' + f + ' | ' + v.statements.pct + '% | ' + v.branches.pct + '% | ' + v.functions.pct + '% | ' + v.lines.pct + '% |');
+          });
+        " 2>/dev/null || echo "(coverage summary not available)"
+        echo "EOF"
+      } >> "$GITHUB_OUTPUT"

This produces ~400 bytes instead of 4,356 bytes — a ~90% reduction for this injection alone.

3. Update the priority target list to actual implementation files

Estimated savings: Qualitative — improves work relevance, reduces wasted agent turns

The priority list points to re-export stubs:

Writing tests for 577-byte re-export files provides minimal coverage value. The agent spends turns discovering this and pivoting. The actual uncovered implementations are:

Change in .github/workflows/test-coverage-improver.md, step Select target file and inject content:

-        const priority = ['src/docker-manager.ts','src/cli.ts','src/host-iptables.ts','src/squid-config.ts','src/domain-patterns.ts'];
+        const priority = [
+          'src/container-lifecycle.ts',
+          'src/config-writer.ts',
+          'src/compose-generator.ts',
+          'src/artifact-preservation.ts',
+          'src/github-env.ts',
+        ];

⚠️ Note: The source files for these targets are 12–17 KB each. When injecting the full file (via SOURCE_CONTENT), this adds ~3,000–4,300 tokens per run. Consider truncating source injection too (to 300 lines) for very large files, with a note that the agent can cat for the rest.

4. Trim verbose "Repository Context" section from prompt body

Estimated savings: ~200 tokens/run (~1–2%)

The "Repository Context" section in the prompt body restates general project knowledge the model already has from the system context. It can be reduced to one sentence:

-## Repository Context
-
-This is **gh-aw-firewall**, a network firewall for GitHub Copilot CLI that provides L7 (HTTP/HTTPS) egress control using Squid proxy and Docker containers. As a security-critical tool, comprehensive test coverage is essential for:
-
-- **iptables manipulation** - NET_ADMIN capability usage
-- **Squid ACL rules** - Domain pattern validation and filtering
-- **Container security** - Capability dropping, seccomp profiles
-- **Domain validation** - Pattern matching and injection prevention
+## Repository Context
+
+This is **gh-aw-firewall** — an L7 firewall for Copilot CLI using Squid + Docker. Security-critical paths: iptables rules, Squid ACL, container capabilities, domain validation.

Expected Impact

i️ Actual savings depend on which target file is selected each run. Recommendation #1 (test file truncation) has the highest variance but also highest ceiling.

Implementation Checklist

• Update Select target file and inject content step to truncate TEST_CONTENT to first 100 lines (Rec Improve links in readme to AW project #1)
• Replace Read COVERAGE_SUMMARY.md step with compact JSON-generated table (Rec Secret proxying #2)
• Update priority list to point at actual implementation files (Rec feat: add integration test for rostbuness #3)
• Trim "Repository Context" section in prompt body (Rec fix: add missing Docker image pulls to robustness test workflow #4)
• Recompile: gh aw compile .github/workflows/test-coverage-improver.md
• Verify a manual workflow_dispatch run completes and produces a valid PR
• Compare AIC on new run vs 304.1 baseline

Generated by Daily Copilot Token Optimization Advisor · Source report: #5714

Generated by Daily Copilot Token Optimization Advisor · 58 AIC · ⊞ 6.7K · ◷