weather-search

Weather Area Query

InputCity or location + Activity radius (km) + Days, automatically performs:

1. Geocode to locate the center point
2. Pull forecast and air quality for 9 points: center + 8 directions
3. Aggregate extreme elements by local date, generate a Markdown report
4. Agent writes going-out protection suggestions

Dependencies

pip install esflow openmeteo-requests pandas

Usage

# 1. 首跑到 agent_advice 退出(exit 2)
python3 scripts/run.py --query 北京天安门
python3 scripts/run.py --query 北京大兴机场 --days 3 --radius-km 10 [--out ./out]
# stdout: envelope JSON(ok, data.job_dir / analyze_grid_artifact / advice_path)
# stderr: [to_agent] 续跑命令 + [样式] 表格约束

# 2. Agent 读 analyze_grid_artifact,写 data.advice_path

# 3. 续跑落盘
python3 scripts/run.py --resume <job_dir>
# stdout: envelope(ok, data.out_path / data.out_dir / data.job_dir)

The final report path is envelope data.out_path. When --out is not passed, write to export/weather_report.md; when --out <DIR> is passed, write to <DIR>/weather_report.md，--resume automatically inherits the first-run input parameters.

Parameters

Parameter	Default	Description
--query	-	City or specific location (required on first run)
--days	1	Forecast days
--radius-km	5	Activity radius km
--out <DIR>	-	Final report output directory
--resume <job_dir>	-	Resume-run TO_AGENT node
--schema	false	Output stdout JSON contract

Exit code:0 ok / 1 runtime / 2 to_agent / 3 validation. esflow events go to stderr and do not pollute the stdout envelope.

Nodes (8)

id	depth	Responsibility
parse_args	0	query + radius_km + days
geocode	1	Nominatim encoding
build_grid	2	List of center + 8-direction sampling points
fetch_bulk	3	Two-point coordinates bulk:forecast + air-quality, local-day aggregation
analyze_grid	4	Extreme value localization,llm_text / llm_json
format_table	5	Markdown report + independent LLM input
agent_advice	6	TO_AGENT, write advice.md
export	7	Merge and persist to disk

DAG

parse_args → geocode → build_grid → fetch_bulk → analyze_grid → format_table
                                                              ↓
                       export ← agent_advice(TO_AGENT)

TO_AGENT

1. On first run, pause at agent_advice with exit 2; parse the stdout one-line envelope,data contains job_dir、analyze_grid_artifact、advice_path
2. Read {job_dir}/.esflow/analyze_grid/artifact.json: prefer llm_text(alert segments take priority), or structured llm_json（alerts、points_merged、daily_reports , etc.); do not rely on the upstream markdown embedded in stderr
3. Write to agent_advice/advice.mdin the following style (i.e., data.advice_path), do not restate values already present in the report table
4. python3 scripts/run.py --resume <job_dir> Resume-run export

Going-out suggestion style (consistent with stderr [样式] )

• Title:## 外出建议
• Fixed table header:日期 | 风险点 | 外出建议
• Use short phrases for risk points, one sentence for actions; consecutive dates of the same type may be merged (e.g., 07-10 至 07-11）
• Do not write summaries outside the table

Technical choices

• Open-Meteo forecast + air-quality,openmeteo-requests Batch coordinates
• pandas Perform hourly→daily aggregation and typical times (8/12/16/20)
• No local response cache (personal use)
• Nominatim still uses standard-library preflight

Preflight

• Python >= 3.10
• Nominatim reachable