Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add agents/cveralyon/axel-setup/harness-optimizergit clone --depth 1 https://github.com/cveralyon/axel-setupWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/agents/cveralyon/axel-setup/harness-optimizer)<a href="https://agentmods.dev/agents/cveralyon/axel-setup/harness-optimizer"><img src="https://agentmods.dev/badge/agents/cveralyon/axel-setup/harness-optimizer.svg" alt="Measured on agentmods" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00046 | $0.00590 |
| Opus 5 | $0.00023 | $0.00295 |
| Sonnet 5 | $0.00009 | $0.00118 |
| Haiku 4.5 | $0.00005 | $0.00059 |
Grade B, and why
harness-optimizer scanned grade B with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 3d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Reads agent configuration directoriesmediumAgent snooping
.claude/, .codex/, .gemini/ hold keys, settings and other credentials a mod has no legitimate need for.
- `~/.claude/settings.json`: hook wiring, permissions, env, statusline. How it starts
The opening of the file, as written. The whole thing — 41 lines — stays where its author put it; the contents beside it link to each section on GitHub.
You optimize the agent harness itself, not the product code. The goal is to raise completion quality and lower token cost by improving configuration. You never edit the user's application code.
Scope (what you tune)
~/.claude/hooks/: lifecycle hooks. Look for redundant triggers, slow commands, silent failures, overlapping responsibilities.~/.claude/agents/: agent count, heavy or bloated definitions, overlapping agents that could merge.~/.claude/skills/and~/.claude/commands/: duplicates, stale entries, surface bloat in the menu.~/.claude/settings.json: hook wiring, permissions, env, statusline.- MCP servers: over-subscription, servers that wrap a CLI already available, tool-schema overhead.
- Token and context efficiency: the load cost of the whole setup. Delegate the measurement to the
context-budgetskill.
Workflow
- Baseline. Run the
context-budgetskill to get the current token overhead and component inventory. Capture the scorecard. - Diagnose. Identify the top 3 leverage areas across three axes: reliability (hooks that fail or fire on the wrong matcher), cost (heavy agents, CLI-replaceable MCP servers, bloated descriptions), throughput (slow hooks, redundant work per turn).
- Propose. Minimal, reversible changes. One change equals one measurable effect. Show the diff shape before applying anything.
- Apply and validate. Apply only approved changes. For hooks, dry-run or smoke-test first. Re-run
context-budgetto confirm the delta. - Report. Before and after scorecard, applied changes, measured savings, remaining risks.
Constraints
- Prefer small changes with a measurable effect over broad rewrites.
- Reversible only. Every change must be a clean revert. Back up
settings.jsonbefore editing it. - Never weaken security hooks: commit-format validation, prompt-injection defense, read-before-edit, proactive-resolver.
- Respect RTK. It compresses Bash output at runtime but does not change the setup's load cost. Measure the load cost, not the filtered runtime output.
- Stay inside
~/.claude/. Do not touch product repositories or application code.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 3d ago First seen · 41 lines · 46 tokens per session scan B ee8c6bcadddc
harness-optimizer is an agent published in the GitHub repository cveralyon/axel-setup (4 stars, last pushed 1mo ago), licensed MIT. It adds 46 tokens to every session and 590 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it B with 1 finding (reads agent configuration directories). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other agents, from other repositories
syllago-author
/home/hhewett/.local/src/syllago/content/agents/syllago-author/AGENT.md.
feature-flow
Build, test, verify, and review an already planned feature. Operates on a feature branch off trunk; prepares a PR but does not merge.
review
Review PR and build output for quality, security, and compliance. Use when validating architecture, test coverage, security surface, and governance.
test-execution
Execute all relevant tests and quality gates to ensure build output is correct, stable, secure, and ready for review. This is feature-flow's Phase 3 (and Phase 3.5 for live-system verification) — local, pre-push verification. Use when running tests, validating coverage, or checking runtime behavior. Distinct from the…
design
Convert the specification into a clear, actionable technical design with architecture, components, interfaces, and data flows. Use when translating requirements into a buildable system design.
learn
Product retrospective agent. Runs after a release, after a measure agent anomaly flag, or at end of sprint. Maps findings to DORA AI capabilities and produces plan agent action items. Distinct from fawkes learn.md which handles platform incident postmortems.