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 skills/hhamja/claude-code-harness/loop-runnpx skills add hhamja/claude-code-harness --skill loop-rungit clone --depth 1 https://github.com/hhamja/claude-code-harnessWhat 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 | $0.00036 | $0.00906 |
| Opus 5 | $0.00018 | $0.00453 |
| Sonnet 5 | $0.00007 | $0.00181 |
| Haiku 4.5 | $0.00004 | $0.00091 |
Grade A, and why
loop-run scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 44 lines — stays where its author put it; the contents beside it link to each section on GitHub.
loop-run — verified maker/checker cycles
Arguments ($ARGUMENTS): --once = exactly one cycle. --verify-only = grade once, report, write NOTHING (no marker, no state updates — strictly read-only).
Before anything
- No
.claude/loop/: tell the user to run/loopy:loop-initand stop. - Read, in order:
state.md(entry point — resume from its unresolved list),goal.md,rubric.md,loop.config.md, andmemory.md's## Distilled rules. - Preflight (skip for
--verify-only): choose the implementer (codex vs claude) and write the run marker — procedure inreferences/preflight.md.
--verify-only path
Spawn the verifier agent with "grade .claude/loop/rubric.md". Print its report verbatim plus a one-paragraph summary. Update no file, then stop.
Cycle (repeat until a stop condition)
- Implement by the implementer chosen at preflight:
- claude (or any fallback): pick the unresolved rubric criteria (from state.md), make the smallest change that could pass them — skip any approach under
## Approaches tried. Prefer the cheapexploreragent for reconnaissance over broad main-context reading. - codex: ONE fresh
codex execper cycle, neverresume, prompt rebuilt from disk each cycle. Full procedure (prompt construction, the exact command, retry/fallback) inreferences/codex-exec.md.
- claude (or any fallback): pick the unresolved rubric criteria (from state.md), make the smallest change that could pass them — skip any approach under
- Phase gate — end of cycle only: spawn the
verifieronce with "grade.claude/loop/rubric.md". Never invoke it per file edit; it only reports. - Apply the report — the main agent does ALL updates to
rubric.md,state.md,memory.md, andreview.md. Exact fields and rules:references/apply-report.md.
Stop conditions
- Success: every rubric criterion passes → do NOT stop yet. Run the green gate ONCE (full runs only) before declaring done — procedure in
references/green-gate.md. - Safety rails (mandatory, never skip):
- iteration count would exceed
max_iterations(loop.config.md) → stop, record the reason in state.md. - the same criterion failed 3 consecutive cycles → do NOT escalate yet. Replan first per the loop-engineering skill's
references/replan.md: up toreplan_max(default 2) genuinely different strategies (change approach / decompose / spike), never weakening the criterion. Only once replan is exhausted, escalate with 2-3 options (interactive: ask; headless: record them + the stall reason in state.md, sethuman_gate: stalled, end the turn). - An unbounded "repeat until pass" loop is forbidden.
- iteration count would exceed
--once: stop after one full cycle regardless of outcome (state/memory/review still updated).
What ships with it
4 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.
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.
- 2d ago First seen · 44 lines · 36 tokens per session scan A 33662ff92094
loop-run is a skill published in the GitHub repository hhamja/claude-code-harness (1 stars, last pushed 1mo ago), licensed MIT. It adds 36 tokens to every session and 906 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
git-commit-convention
写 git 提交信息、整理提交历史时使用——Conventional Commits 的格式、类型选择和拆分原则。.
delegate-to-deepseek
默认把中等及以下、批量、重复或机械任务作为完整逻辑单元派给 DeepSeek，并由主 Agent 独立验收。适用于批量改文件、扫日志、翻译、ETL、脚本、测试、文档、CRUD、单领域重构、单组件或单 endpoint。主 Agent 可基于上下文和失败代价调整派工策略；用户显式指令、安全、权限、隐私边界和派工后验证不可突破。DEEPSEEKMODE=off 时跳过。.
code-review-checklist
审查代码改动（review diff / PR）时使用——按固定清单过正确性、边界、错误处理和测试盲区，避免只看顺眼不顺眼。.
sruja-project
Procedural workflows for working with the Sruja codebase. Teaches AI editors how to add components, validate changes, and follow patterns.
grill-me
This skill ensures AI and human share a design concept before writing code.
sruja-harness
Verification harness for any AI agent skill. Use with any other skills; before marking a task done, run verify-task. On failure, record a correction learning. Ensures lint, tests, drift, and intent checks pass regardless of which skill generated the code.