Getting it into your agent
It runs from inside its repository, so the clone comes first — what it calls does not travel with the file alone.
git clone --depth 1 https://github.com/Tencent-RTC/agent-skillsnpx agentmods add skills/tencent-rtc/agent-skills/trtc-evalWrote 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/skills/tencent-rtc/agent-skills/trtc-eval)<a href="https://agentmods.dev/skills/tencent-rtc/agent-skills/trtc-eval"><img src="https://agentmods.dev/badge/skills/tencent-rtc/agent-skills/trtc-eval/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/tencent-rtc/agent-skills/trtc-eval"><img src="https://agentmods.dev/badge/skills/tencent-rtc/agent-skills/trtc-eval.svg" alt="Reviewed on agentmods" width="80" 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 5.1 | $0.00079 | $0.05630 |
| Opus 5 | $0.00039 | $0.02815 |
| Sonnet 5 | $0.00016 | $0.01126 |
| Haiku 4.5 | $0.00008 | $0.00563 |
Grade C, and why
trtc-eval scanned grade C 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 9d 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.
Recursive force deletehighDestructive command
rm -rf with a variable or a broad path is one typo away from removing the wrong tree.
| 清理本次产物 | `rm -rf {run_dir}` | 用户确认不再需要 | How it starts
The opening of the file, as written. The whole thing — 345 lines — stays where its author put it; the contents beside it link to each section on GitHub.
TRTC 评测工具
触发条件
- 用户 prompt 中出现:跑分 / 评测 / eval / benchmark / 回归 / 看看效果 / 质量对比
- 如果含过滤条件(如 "只跑 iOS 的"、"只跑 smoke"),记录下来并在 Step 1 应用到用例筛选
执行步骤(你,主 Agent,严格按顺序执行)
工作目录:所有
python scripts/...命令都从本 skill 目录运行。开始前先cd .claude/skills/trtc-eval/(脚本通过__file__解析 skill_root,所以 cwd 实际不影响数据路径,但保持习惯让命令简短)。 eval-runs 路径:每次运行的产物落在仓库根的.claude/eval-runs/{ts}/,不在 skill 目录里。下面示例统一用相对 skill 目录的../../../.claude/eval-runs/{ts}表示。
Step 1:加载 eval set
- 执行
python scripts/selfcheck.py --phase=pre-run校验环境- 校验失败 → 停止,把 selfcheck.json 摘要给用户,让用户修
- 读取
tests/benchmark/cases.json - 按用户过滤条件筛出要跑的用例列表
- 创建本次运行目录
../../../.claude/eval-runs/{ISO8601}/,写 run.manifest.json
必须产出的证据(缺一不可):
- selfcheck pre-run 输出
passed=true(附 selfcheck.json 中 checks 全部 ok) - run.manifest.json 已写入且
case_count > 0 - eval-runs/{ts}/ 目录已创建
⛔ 门控规则:
- selfcheck 任何一项 check.ok=false → 立即停止,不得进入 Step 2
- case_count == 0(筛选后无用例)→ 告知用户筛选条件无匹配,不进入 Step 2
- "我先跑着,selfcheck 的 warning 之后再修" ← 这个想法是错的。修完再跑
Step 2:串行调用 orchestrator 跑每条用例
对筛出的每条用例(串行,不并发),用 execute_command 工具调用 orchestrator:
python scripts/case_runner_orchestrator.py \
--case-id={test_id} \
--run-dir=../../../.claude/eval-runs/{ts}
关键约束:
- 一次只跑一条用例,串行执行(首版避免并发导致真机资源冲突)
- orchestrator 内部串联 7 个步骤(run_ai → evaluator → demo_build → log_stream_start → demo_run → log_stream_stop → runtime_monitor),它是
trace.jsonl的唯一写入者 - orchestrator 的 stdout 只输出一行 JSON:
{"test_id":"...","exit_code":0,"summary_path":"<相对路径>"}。你只读这一行即可。 - 绝对不要
cat或读取ai_raw_output.md/runtime.log/compile.log—— 这是上下文污染源。 - 如果你需要看分数,只读
{run_dir}/cases/{test_id}/summary.json。
必须产出的证据(每条用例):
- orchestrator stdout JSON 已捕获(含 test_id + exit_code + summary_path)
- summary.json 已读取且包含 final_score 和 passed 字段
⛔ 门控规则:
- orchestrator exit_code != 0 → 记录失败但继续下一条(单条失败不阻塞整体)
- 失败升级触发(见下方「失败升级协议」):连续失败达到阈值 → 暂停或终止
Step 3:汇总 & 出报告
- 所有用例跑完后,执行
python scripts/report.py build --run-dir=../../../.claude/eval-runs/{ts} - (可选)如果用户要求 diff:
python scripts/report.py diff --baseline=<旧 run_dir> --current=<新 run_dir> - 执行
python scripts/selfcheck.py --phase=post-run --run-dir=../../../.claude/eval-runs/{ts}再次自查 - 把
report.md路径和selfcheck.json中的verdict字段一起给用户
What ships with it
60 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.
- bootstrap.sh 6.4 KB runs code
- config.example.json 149 B
- docs/DATA_FLOW_DIAGRAMS.md 37 KB
- docs/INVESTIGATION_SUMMARY.md 10 KB
- docs/quality-self-check.md 1.8 KB
- docs/QUICK_REFERENCE.md 5.0 KB
- docs/SCORING_ANALYSIS.md 18 KB
- docs/troubleshooting.md 1.6 KB
- EVAL_AUTORUN_ARCHITECTURE.md 24 KB
- EVAL_AUTORUN_FLOWS.md 32 KB
- prompts/ai_driver_system_prompt.md 7.1 KB
- README.md 21 KB
- scripts/__init__.py 28 B runs code
- scripts/case_runner_orchestrator.py 17 KB runs code
- scripts/demo_runner.py 7.9 KB runs code
- scripts/evaluator.py 5.2 KB runs code
- scripts/lib/__init__.py 22 B runs code
- scripts/lib/builder.py 398 B runs code
- scripts/lib/cli_driver.py 1.5 KB runs code
- scripts/lib/code_injector.py 39 KB runs code
- scripts/lib/creds_normalizer.py 9.9 KB runs code
- scripts/lib/dep_installer.py 6.7 KB runs code
- scripts/lib/device_picker.py 2.8 KB runs code
- scripts/lib/eval_config.py 6.1 KB runs code
- scripts/lib/file_naming.py 9.6 KB runs code
- scripts/lib/flow_codegen.py 22 KB runs code
- scripts/lib/launcher.py 612 B runs code
- scripts/lib/log_parsers/__init__.py 22 B runs code
- scripts/lib/log_parsers/logcat_parser.py 1.6 KB runs code
- scripts/lib/log_parsers/puppeteer_parser.py 7.0 KB runs code
- scripts/lib/log_parsers/syslog_parser.py 1.9 KB runs code
- scripts/lib/platforms/__init__.py 587 B runs code
- scripts/lib/platforms/android.py 3.1 KB runs code
- scripts/lib/platforms/base.py 2.2 KB runs code
- scripts/lib/platforms/ios.py 9.2 KB runs code
- scripts/lib/platforms/web.py 3.0 KB runs code
- scripts/lib/schemas.py 14 KB runs code
- scripts/lib/template_fetcher.py 1.6 KB runs code
- scripts/lib/web_profile.py 6.8 KB runs code
- scripts/log_streamer.py 4.4 KB runs code
- scripts/log-bridge.mjs 22 KB runs code
- scripts/package-lock.json 42 KB
- scripts/package.json 281 B
- scripts/report.py 8.5 KB runs code
- scripts/requirements.txt 49 B
- scripts/run_ai.py 6.6 KB runs code
- scripts/runtime_monitor.py 17 KB runs code
- scripts/selfcheck.py 22 KB runs code
- scripts/stats_trigger.py 2.0 KB runs code
- scripts/tools/__init__.py 0 B runs code
- scripts/tools/migrate_flows_to_dsl.py 23 KB runs code
- tests/benchmark/cases.json 51 KB
- tests/benchmark/schema.json 2.3 KB
- tests/unit/fixtures/sample_logcat.txt 426 B
- tests/unit/fixtures/sample_puppeteer_console.json 463 B
- tests/unit/fixtures/sample_syslog.txt 420 B
- tests/unit/test_creds_normalizer.py 6.0 KB runs code
- tests/unit/test_dynamic_web.py 5.3 KB runs code
- tests/unit/test_fingerprint.py 7.2 KB runs code
- tests/unit/test_flow_codegen.py 10 KB runs 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.
- 9d ago First seen · 345 lines · 79 tokens per session scan C 0f8d2f917a71
trtc-eval is a skill published in the GitHub repository Tencent-RTC/agent-skills (12 stars, last pushed 4d ago), licensed MIT. It adds 79 tokens to every session and 5,630 once invoked, about $0.0004 per session on Opus 5. A static security scan graded it C with 1 finding (recursive force delete). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.
Other skills, from other repositories
vs-chat
Conversational search runtime: send messages, keep sessions consistent, and verify retrieval behavior and responses.
self-testing
E2E self-testing workflow for vibing.nvim using a separate Neovim instance controlled over RPC. Use when writing or debugging E2E tests, running npm run test:e2e, or executing the 3-try auto-fix rule after implementing a feature. Covers the…
test-design
Automatically design comprehensive E2E test cases for newly implemented vibing.nvim features. Use immediately after completing feature implementation (Phase 5.4) and before running E2E tests. Generates test scenarios covering Happy paths, Error cases, Edge cases, and Integration points with priority ranking…
ci-gates
The CI gates of vibing.nvim (test:lua exit code, E2E timeout budget, check:doc, check) and the ways each one has silently stopped failing. Use when editing package.json scripts, .github/workflows/ci.yml, scripts/check-help.lua, doc/.txt, or any tests/.test.mjs that guards a gate — and whenever a gate passes but you…
Chat with PDF
Answer questions about PDF content, summarize, and extract information.
cometchat-angular-testing
Testing patterns for CometChat Angular UI Kit v5 (@cometchat/chat-uikit-angular@5) in Angular 17–21 projects. Covers Jasmine/Karma (default) and Jest, Angular TestBed for STANDALONE components (imports[] not declarations[]), mocking the static CometChatUIKit kit API…