⚡ Claude Token Optimization2026-06-12 — security-guard

[github-actions]

Target Workflow: security-guard

Source report: #4810
Est. cost/active run: ~$0.94 (avg 194K tokens, claude-sonnet-4-5)
Est. cost/all runs (amortized): ~$0.33/run (65% are zero-cost early-exit noops)
Cache read/write rate: N/A — per-turn breakdown unavailable (single aggregate in logs)
LLM turns: avg 1.2 across all runs; bimodal: 2 (simple) vs 4–5 (complex)

Current Configuration

Setting	Value
Model	claude-sonnet-4-5
Tools loaded	github: with toolsets: [pull_requests, repos]
GitHub MCP tools actually called	None — agent uses bash gh pr diff/view/api instead
Pre-agent steps	Yes (3: fetch diff, fetch PR metadata, pass relevance count)
Max turns	6
Prompt size	~5.5 KB instructions + up to 100 KB pre-fetched diff
Early-exit noop rate	65% (11/17 runs; 0 tokens)

Bimodal Run Pattern

Pattern	Runs	Tokens	Turns	Est. Cost
2-turn simple review	3	~122K avg	2	~$0.59/run
4–5-turn complex review	3	~263K avg	4–5	~$1.26/run
Zero-token noop	11	0	0	$0.00

Root cause of multi-turn explosion: Each turn re-sends the full growing context as input (Anthropic charges per-turn input). A 5-turn run sends ~375K cumulative input tokens vs ~110K for 2-turn — 3.4×. The tool_usage log confirms agents make prohibited bash fetches in complex-diff runs despite the existing instruction:

bash_gh pr view 4797  → run #4000 (274K, 5-turn failure)
bash_gh pr diff 4797  → run #4000 (274K, 5-turn failure)
bash_gh pr view 4783  → run #3985 (264K, 5-turn success)
bash_gh pr diff 4784  → run #3991 (262K, 4-turn success)

Recommendations

1. Strengthen the no-bash-fetch constraint (highest impact)

The existing prohibition is buried mid-prompt. Move it to the top as a hard constraint:

## ⚠️ Hard Constraints

1. **No bash GitHub fetches** — no `gh pr diff`, `gh pr view`, `gh api`, `git diff`,
   `git log`, `git show`, or `curl` to GitHub URLs. All PR data is pre-fetched below.
2. **One extra fetch maximum** — if `[DIFF TRUNCATED ...]` appears, call
   `mcp__github__get_pull_request_diff` **once**, then emit your safe-output immediately.
3. **No file reads** — analyze the diff only.

Estimated savings: Eliminating 2–3 bash-fetch turns saves ~130–195K tokens/affected run (~50–70% on complex runs).

---

2. Remove repos toolset — loaded but never called via MCP

The repos toolset (~10–12 tool schemas) is loaded into every system prompt but zero MCP calls to mcp__github__get_file_contents or similar were observed. The agent uses bash instead.

# Before
toolsets: [pull_requests, repos]

# After
toolsets: [pull_requests]

Keep pull_requests for mcp__github__get_pull_request_diff on truncation.

Estimated savings: ~10 schemas × 600 tokens × 3.3 avg turns ≈ ~20K tokens/active run (~10%)

---

3. Reduce DIFF_LIMIT from 100 KB to 30 KB

At 100 KB / 3.5 chars per token, a saturated diff injects ~28K tokens/turn into the growing context. At 30 KB → ~8.5K tokens, saving ~19.5K tokens per turn, compounded across multi-turn runs.

# Before
DIFF_LIMIT=100000
# After
DIFF_LIMIT=30000

Estimated savings: ~19.5K tokens × turns for large-diff PRs ≈ ~30–50K tokens/affected run (~15–20%)

---

4. Reduce max-turns from 6 to 3

After Recommendations 1–2, expected turn pattern: turn 1 (analyze), optional turn 2 (fetch full diff), turn 3 (safe-output). Max-turns 3 is a structural guard against regressions.

max-turns: 3

Risk: Test on a PR touching both setup-iptables.sh and squid-config.ts before deploying.

---

5. Move variable content to end of prompt (cache optimization)

The security file count (__GH_AW_EXPR_66EB691F__) appears in section 2 of the prompt, breaking Anthropic's prefix cache early. Move it inline with the diff block:

## Changed Files (up to 30 KB) | Security-critical files changed: __GH_AW_EXPR_66EB691F__

This lets ~3–4 KB of stable instructions cache on turns 2+.

Estimated savings: ~15–30% of per-turn input cost on turns 2+ (unquantified without cache hit rate data — see action item below).

Cache Analysis (Anthropic-Specific)

Per-turn breakdown estimated for a 5-turn run (~267K total):

Turn	Est. Input	Est. Cache Write	Est. Cache Read	Net New
1	~55K	~55K	~0	~55K
2	~63K	~8K	~55K	~8K
3	~68K	~5K	~63K	~5K
4–5	~75K avg	~5K	~70K	~5K

Cache writes from Turn 1 (~55K) are read back at $0.30/M vs $3/M (10×) in turns 2–5. This works within a run but not across runs (unique diff per PR breaks the cross-run prefix).

Action item: Instrument the workflow to log Anthropic usage.cache_read_input_tokens per turn to confirm actual hit rate.

Expected Impact

Metric	Current	Projected	Savings
Tokens/active run	~194K	~120K	~−38%
Cost/active run	~$0.94	~$0.58	~−38%
LLM turns (complex runs)	4–5	2–3	−2 turns
Session time (complex runs)	~10–16 min	~5–8 min	~−50%

Implementation Checklist

• Add hard-constraint block above "## Your Task" (Rec. 1)
• Change toolsets: [pull_requests, repos] → toolsets: [pull_requests] (Rec. 2)
• Change DIFF_LIMIT=100000 → DIFF_LIMIT=30000 (Rec. 3)
• Change max-turns: 6 → max-turns: 3 (Rec. 4)
• Move security file count to end of prompt near diff block (Rec. 5, optional)
• Recompile: gh aw compile .github/workflows/security-guard.md
• Post-process: npx tsx scripts/ci/postprocess-smoke-workflows.ts
• Test on a PR touching both setup-iptables.sh and squid-config.ts
• Compare next 5 active runs vs baseline (target: ≤130K tokens/active run)
• Add Anthropic cache_read_input_tokens logging to verify cache amortization

Generated by Daily Claude Token Optimization Advisor · ◷

[github-actions]

This issue is being closed as outdated. A newer issue has been created: #4927

View newer issue

---

This action was performed automatically by the Daily Claude Token Optimization Advisor workflow.