菜单
Benchmark live CortexFS multi-agent system with Inspect AI ctx CLI. Use when asked test ctx agents, compare architect/coder/reviewer/worker performance, measure quality, latency, TTFT, throughput, tokens or errors, diagnose benchmark regressions, produce repeatable agent evaluation report.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/LIghtJUNction/cortexfs --skill ctx-agent-benchmark命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
在 CortexFS 仓库中继续去重、减少重复逻辑、合并等价代码、减少代码量或函数数量、恢复上次 DRY 重构现场，或在新增代码前避免写出重复 helper 时使用。
This skill should be used when the user asks to "build a CortexFS tool", "create a CortexFS agent", "package a CortexFS extension", "write a cortexfs.object/v1 or cortexfs.object/v2 manifest", "install an external tool", or "install an external agent".
Use this skill when working in the CortexFS repository and the user asks to test CortexFS, test FUSE, test /ctx, test ctx, test agent.sh, run live tests, use Ollama or smollm2:135m, verify file/socket ABI behavior, check tests/mounts/cortexfs, inspect the systemd mount, update the AUR package, or diagnose mount/socket permission failures.
基于 SOC 职业分类
正在显示 SKILL.md
| name | ctx-agent-benchmark |
| description | Benchmark live CortexFS multi-agent system with Inspect AI ctx CLI. Use when asked test ctx agents, compare architect/coder/reviewer/worker performance, measure quality, latency, TTFT, throughput, tokens or errors, diagnose benchmark regressions, produce repeatable agent evaluation report. |
Use repository benchmark as a single implementation. Do not copy its JSONL parser or metric logic into this skill.
AGENTS.md, apply cortexfs-test, complete its live-agent proof every
selected role; abort if fails.rtk proxy ./inspect_benchmark/run_benchmark.sh system --output-dir results
Use Inspect logs focused quality run when needed:
rtk proxy ./inspect_benchmark/run_benchmark.sh ctx coder
Compare summary.json only across equivalent datasets, repeats, agents, and
provider routes. The benchmark requires pre-existing canonical listeners and
never starts or stops agents.
It creates unique sessions only after proving they did not exist, cancels exact
canonical run IDs on timeout, archives only an owned, inactive, non-current,
unreferenced session selected alone by exact GC preview.
Cleanup refusal is retained in receipts and makes the run nonzero.
Report runtime success rate and exact accuracy separately.
Include mean/p50/p95 latency, mean/p50/p95 TTFT, token totals, tokens/s or
chars/s, and errors grouped by code.
If start latency is unavailable, report it as n/a; report preflight timing
instead.
Include delegated live-proof outcome, configured models, cleanup outcomes, and
every skipped unavailable metric.
Do not add watchers, polling loops, provider-specific core branches, or second workflow ABI.