[copilot-token-optimizer] Architecture Guardian: reduce 47-turn analysis via bash pre-step consolidation

[github-actions]

Target Workflow

Architecture Guardian (architecture-guardian.md) — selected as highest-token eligible workflow (1.97M tokens, not optimized in the last 14 days).

Analysis Period

• Window: 2026-04-23 (7-day window, 1 run available)
• Run analyzed: §24842109769

Token Profile

Metric	Value
Total tokens	1,969,440
Effective tokens	2,197,150
Avg tokens/run	1,969,440
Total turns	47
Avg turns/run	47
Cache efficiency	49.3%
Cache write tokens	0
Action minutes	9
Conclusion	✅ success

Agentic Assessments (from run classifier)

• resource_heavy_for_domain (severity: high) — 47 turns for a local static-analysis task is excessive
• partially_reducible (severity: low) — agentic_fraction=0.50 suggests ~50% of turns are data-gathering that could be deterministic

---

Ranked Recommendations

1. 🔴 Add a bash pre-step to collect all file metrics up front

Estimated savings: ~800K–1M tokens/run (40–50%)

The prompt instructs the agent to iterate file-by-file: run wc -l, grep -n, awk, and repeat for each changed file. With 47 turns at ~42K input tokens/turn (accumulated context), this is the dominant cost.

A single pre-step bash script can:

1. Identify all changed .go/.js/.cjs files in the last 24 hours
2. Run wc -l, grep -n "^func ", and export-count greps across all files at once
3. Run go list ./... 2>&1 | grep -i "import cycle" once
4. Write a structured JSON summary to /tmp/gh-aw/agent/arch-metrics.json

The agent then reads one file instead of running 20–30 individual shell commands, reducing turns from ~47 to ~15–20.

Concrete action: Add a steps: block in the workflow frontmatter with a bash pre-step that runs the full metric collection script.

Evidence: 47 turns, actuation_style=read_only, all outputs are analysis (no write actions), agentic_fraction=0.50.

---

2. 🟡 Remove the repos GitHub toolset

Estimated savings: ~30K–50K tokens/run (reduced tool-list context)

The workflow is configured with github: toolsets: [repos]. The entire analysis is performed with local bash commands (git log, wc, grep, awk, go list). No GitHub API reads are needed for local static analysis. With mount-as-clis: true, the gh CLI is available if needed.

Removing the toolset eliminates the tool-definition tokens injected into every request.

Concrete action: Remove the github: block from the tools: section (or replace with toolsets: []).

Evidence: 3 github_api_calls recorded for the run — these are likely the safe-output create_issue call and repo metadata lookups, not repos-toolset calls. The prompt never references GitHub API operations for analysis.

---

3. 🟡 Remove the edit tool from the tools list

Estimated savings: ~10K–20K tokens/run (tool schema injection)

The edit: tool is listed under tools:. The workflow's primary task (code analysis + issue creation) doesn't require editing files — only bash and safe-outputs are needed. edit is used only during workflow recompile (footer-workflow-recompile message), not in normal runs.

Concrete action: Remove edit: from the tools: section.

Evidence: No edit tool calls observed in run data. actuation_style=read_only.

---

4. 🟢 Simplify the per-violation AI suggestion step

Estimated savings: ~50K–100K tokens/run (reduced output + fewer turns)

Step 5 asks the agent to generate a "concise refactoring suggestion" for every BLOCKER and WARNING violation, including (1) what the violation is, (2) why it's a problem, and (3) a concrete plan. This produces verbose output and may require additional read turns to gather context for suggestions.

Concrete action: Replace per-violation narrative suggestions with a templated 2-line note: violation description + a single-sentence fix hint. Move the detailed architectural guidance to the issue body as boilerplate.

Evidence: 12,444 output tokens generated (vs typical ~5K for issue-creation workflows), 47 turns suggest multiple reasoning cycles per violation.

---

Supporting Data

Run token breakdown

Token type	Count
Input tokens	1,956,996
Output tokens	12,444
Cache read tokens	1,903,775
Cache write tokens	0
Requests (turns)	47
Avg input/turn	~41,638

Behavior fingerprint

Dimension	Value
Execution style	exploratory
Tool breadth	narrow
Actuation style	read_only
Resource profile	heavy
Agentic fraction	0.50

Configured tools

• github: toolsets: [repos]
• bash: git log/diff/show, find, wc, grep, cat, head, awk, sed, sort
• edit:
• safe-outputs: create-issue, noop

Caveats

• Only 1 run available in the 7-day window (scheduled daily, weekdays only). Recommendations are based on structural prompt analysis supplemented by the run fingerprint.
• The ~800K token savings estimate for recommendation rejig docs #1 is conservative — actual savings depend on the number of changed files per run. Runs with few changes (triggering early noop) will save proportionally less.
• Cache write tokens = 0 in this run, which means the KV cache wasn't populated. This may be normal for the first run of the day or after a workflow change.

References:

• §24842109769

Generated by Copilot Token Usage Optimizer · ● 989.3K · ◷

• expires on Apr 30, 2026, 3:34 PM UTC