An OpenAI Codex skill for COROS runners: daily readiness, sleep HRV recovery, training load, lap/FIT analysis, structured athlete context, validated training prescriptions, fueling, weather risk, injury risk, and race plans from 5K to marathon.
中文快速入口：这是一个面向高驰 COROS 跑者的 Codex skill，用于分析恢复、训练负荷、圈速分段、补给、天气热应激、伤病风险，并生成 5K/10K/半马/全马训练建议。
Use $running-coach to analyze my recent COROS running data and give today's training recommendation.
Most running dashboards show data. This skill turns COROS data into an agent-friendly coaching workflow:
- Check data quality before making a recommendation.
- Build personal 7/28/56 day baselines for sleep, HRV, resting heart rate, running volume, long runs, and quality sessions.
- Gate intensity with readiness before suggesting workouts.
- Analyze laps, segments, intervals, marathon-pace blocks, long-run drift, and late-run fade.
- Adjust for heat, humidity, wind, rain, fueling, fatigue, pain, and injury risk.
- Produce a plan-execute-review loop with stop rules and feedback questions.
- Build an
AthleteContextand validateTrainingPrescriptionoutputs before presenting training advice.
Install the skill into Codex:
mkdir -p ~/.codex/skills
git clone https://github.com/liumy-qd/running-coach-skill ~/.codex/skills/running-coachRun the synthetic examples:
cd ~/.codex/skills/running-coach
python3 scripts/analyze_laps.py --input examples/laps.json --json
python3 scripts/build_coros_summary.py --input examples/coros-summary.txt
python3 scripts/evaluate_context.py --input examples/context-summary.json --markdown
python3 scripts/build_athlete_context.py --input examples/athlete-bundle.json > /tmp/athlete-context.json
python3 scripts/validate_prescription.py --input examples/prescription-safe.json --context /tmp/athlete-context.jsonUse it in Codex:
Use $running-coach to review my last 14 days of COROS running data and tell me whether I should run today.
中文：
使用 $running-coach 分析我最近 14 天的高驰跑步数据，判断今天是否适合跑步，并给出训练安排。
Readiness: yellow
Action: downshift
Recommendation: Keep the habit, but reduce intensity and volume. Avoid intervals, threshold work, and long runs today.
Reasons:
- Short-term / long-term load ratio is above 1.3.
- Recovery percentage is low.
- The longest run takes too much of recent total mileage.
Next action:
- 20-40 min easy run, brisk walk, or full rest.
- Stop if pain changes your gait, heart rate spikes abnormally, or heat stress symptoms appear.
More examples:
| Area | What it does |
|---|---|
| COROS workflow | Uses COROS MCP first, with coros-mcp CLI as fallback. |
| Readiness | Combines sleep, HRV, resting heart rate, stress, recovery, load, pain, and recent workouts. |
| Personal baselines | Uses 7/28/56 day trends instead of generic thresholds only. |
| Lap analysis | Detects cardiac drift, pace fade, interval candidates, segment stability, and long-run risk. |
| Race planning | Supports 5K, 10K, half marathon, marathon, maintenance, recovery, and fat-loss phases. |
| Environment | Adjusts recommendations for heat, humidity, dew point, wind, rain, and time of day. |
| Fueling | Adds carb, hydration, electrolyte, and GI feedback checks for long runs and races. |
| Safety | Includes stop rules, downshift options, and injury-risk prompts. |
| V2 schemas | Produces compact AthleteContext and validates TrainingPrescription safety gates. |
- Classify the request: daily check, weekly review, activity review, race plan, or data inventory.
- Pull the minimum required COROS data.
- Build a compact AthleteContext with coverage, baselines, risk flags, and evidence.
- Gate training intensity deterministically before the model explains the plan.
- Deep-dive key activities when needed.
- Adjust for environment, fueling, fatigue, and injury risk.
- Validate the TrainingPrescription safety contract.
- Return a concrete action plan plus stop rules and feedback questions.
.
├── SKILL.md
├── agents/openai.yaml
├── references/
├── scripts/
├── examples/
│ ├── prompts/
│ └── outputs/
└── .github/
The skill expects a COROS MCP server when available. If direct MCP access is unavailable, it can use the local coros-mcp CLI fallback described in references/coros-tools.md.
Do not commit access tokens, OAuth callback parameters, passwords, or exported private health files.
python3 -m py_compile scripts/*.py
python3 scripts/analyze_laps.py --input examples/laps.json --json
python3 scripts/build_coros_summary.py --input examples/coros-summary.txt
python3 scripts/evaluate_context.py --input examples/context-summary.json --markdown
python3 scripts/build_athlete_context.py --input examples/athlete-bundle.json > /tmp/athlete-context.json
python3 scripts/validate_prescription.py --input examples/prescription-safe.json --context /tmp/athlete-context.json
python3 -m unittest discover -s tests -p 'test_*.py'If you have Codex's skill-creator validation script available:
python3 ~/.codex/skills/.system/skill-creator/scripts/quick_validate.py .- Add more synthetic examples for 5K, 10K, half marathon, marathon, recovery week, and heat-stress scenarios.
- Add more optional parser adapters for direct MCP JSON bundles from different clients.
- Add more route/weather/fueling examples without storing private user data.
- Add community-contributed coaching templates for different experience levels.
No. The repository examples are synthetic. Real COROS access is optional and should happen through your local MCP or CLI environment, not through committed files.
No. It is a coaching workflow for training decisions. Chest pain, fainting, abnormal heart rhythm, acute injury, or persistent pain should be handled by a qualified professional.
Yes, partially. The scripts can analyze synthetic or exported summaries. For live coaching, COROS MCP or coros-mcp gives the agent a better data source.
这个项目的核心不是“看数据”，而是帮助 Codex 按教练流程使用数据：
- 先判断数据是否足够可靠。
- 再用个人基线判断今天能不能上强度。
- 对关键跑步做圈速、分段、心率漂移和间歇稳定性分析。
- 把天气、热湿、补给、疲劳和疼痛纳入训练建议。
- 最后给出可执行训练、停止规则、降级选项和复盘问题。
安装：
mkdir -p ~/.codex/skills
git clone https://github.com/liumy-qd/running-coach-skill ~/.codex/skills/running-coach使用：
使用 $running-coach 分析我近期的高驰数据，判断今天适合休息、轻松跑还是质量课。
MIT