sf-civic-digest（科幻公民文摘）

SF Civic Digest — SKILL.md

Cold-start note: This file is designed to be self-sufficient. An agent reading only SKILL.md should be able to run reports, generate digests, and understand the full data stack. No other context required.

Quick Start

# Step 1: Find the absolute path to this skill's scripts/ directory
#         (it's always a sibling of this SKILL.md file)
SKILL_DIR="/absolute/path/to/sf-civic-digest"

# Step 2: Run scripts using that absolute path (--district N required)
python3 "$SKILL_DIR/scripts/sf_weekly_digest.py" --district 5 --json

# Step 3: Synthesize the JSON output into a narrative report (see STYLE.md)

Note: sf_weekly_digest.py makes ~22 network calls sequentially and can take 2-3 minutes. For a quick test, try a single script first: python3 "$SKILL_DIR/scripts/sf_journalism.py" --district 5 --days 7 --json

District: Pass --district N (1–11) to any script. User preferences and neighborhood context live in USER.md — read it before writing a report. There is no civic_config.json or civic_config.example.json; those files have been removed.

⚠️ Path rule — critical for any agent (any agent, IDE, or human):

• All scripts must be run with the absolute path to the scripts/ directory
• Find it: scripts/ is always a sibling of this SKILL.md file
• Example: if SKILL.md is at /home/alice/.openclaw/workspace/skills/sf-civic-digest/SKILL.md, then scripts are at /home/alice/.openclaw/workspace/skills/sf-civic-digest/scripts/
• Never use bare relative paths like scripts/foo.py — they break when the working directory is anything other than the skill root
• When in doubt: realpath on any script path gives you the absolute path

Architecture

Scripts are data pipelines. The agent writes the narrative.

The scripts fetch, parse, dedup, and output JSON. The agent (you) reads that JSON and synthesizes it into a report following the style guide. There is no auto-generated report — the editorial judgment is yours.

sf_weekly_digest.py --json
    └── calls all subscripts → returns combined JSON
            │
            ▼
    Agent reads JSON + STYLE.md
            │
            ▼
    Agent writes narrative report

For report generation, see Report Generation section below.

All Scripts

Canonical script location: skills/sf-civic-digest/scripts/

No browser required

All scripts now work without a browser. sf_board_of_appeals.py and sf_rent_board.py use the sf.gov Wagtail CMS API (api.sf.gov/api/v2/pages/). sf_planning_commission.py uses curl against server-rendered Drupal HTML (the sfplanning.org calendar pages are not JS-dependent despite initial assumptions).

District Configuration

All scripts accept --district N (1–11). That's all you need for a basic run.

User preferences and neighborhood context (streets, addresses, people to flag) live in USER.md — a human-readable profile file. There is no civic_config.json; that file has been removed. Read USER.md to understand what the user cares about before writing any report.

District→neighborhood mappings: references/sf-districts.md Build timeline benchmarks: references/sf-build-timelines.md — Read before writing about any housing project. Includes typical phase durations, financing tracking methods, and context for whether a project is on track or stalled.

Report Generation

Agent-driven (recommended for one-off reports)

# 1. Collect data (use absolute paths — see path rule above)
python3 /absolute/path/to/sf-civic-digest/scripts/sf_weekly_digest.py --json > /tmp/civic_data.json
python3 /absolute/path/to/sf-civic-digest/scripts/sf_housing_pipeline.py --district 5 --json >> /tmp/civic_data.json
python3 /absolute/path/to/sf-civic-digest/scripts/sf_volunteer_cleanups.py --district 5 --days 14 --json >> /tmp/civic_data.json

# 2. Agent synthesizes following STYLE.md
# Read STYLE.md, read the JSON, write narrative to reports/

Agent-driven reports (quarterly, annual, deep dives)

Give your agent the workspace path and let it orchestrate:

Read STYLE.md and USER.md. Run the scripts in scripts/ to collect data for District 5, then write a Q1 2026 civic digest. Save to reports/.

⚠️ Path gotcha: Agents default to wrong relative paths. Always resolve the absolute path to scripts/ from the location of this SKILL.md before running anything. Use full absolute paths — e.g. python3 /full/path/to/scripts/sf_housing_pipeline.py. Never just scripts/SCRIPTNAME.py.

Report style

Read STYLE.md before writing any report. Key rules:

• TLDR (5-7 bullets) first — last bullet is always one action
• Officials box: YOUR elected officials only (supervisor, mayor, state reps)
• Short sections (3-5 sentences), scan-and-skip
• Connect threads across sources (lobbying + permits + 311 + journalism = one story)
• End each thread with a specific action (date, email address, what to say)
• Rates over counts. "Up 15%" beats "95 reports."
• Tone: pro-building, pro-tech, car-free lens. See STYLE.md for full framing guidance.

Reference: see STYLE.md for rules and examples.

Key Data Gotchas

These will burn you if you don't know them:

Socrata column names:

• 311 (vw6y-z8j6): supervisor_district = '5.0' (float string), date field = requested_datetime
• Building Permits (i98e-djp9): supervisor_district = '5'. Address split across street_number/street_name/street_suffix — no combined field. estimated_cost is a string, cast in Python.
• Lobbyist Contacts (5f5n-tdbf): NO underscores — officialname, clientname, lobbyistname
• Lobbyist Contributions (e6py-fg8b): same — lobbyistname, candidatename, sourceoffunds
• Planning Records (qvu5-m3a2): pim_link returns {"url": "..."} dict, not string. No supervisor_district — cross-reference parcels (acdm-wktn) by block/lot.

Legistar: The API is broken for SF (returns "Agenda Draft Status" error). Use the RSS feed instead: https://sfgov.legistar.com/Feed.ashx?M=C&ID=17442&GUID=EEE85B7C-1A1C-4A56-873E-355A0A0DE5C3&Mode=All&Format=rss Returns ~100 meetings (~6 months). Parse meeting IDs from RSS, then fetch_meeting_detail() for agenda items.

Reddit: Direct curl blocked (403, datacenter IP). Use web_fetch tool — routes differently, works.

sfrecpark.org: Completely unreachable from this server. Granicus (view_id=91) is the only path.

311 trend pattern: Query two 7-day windows (current + prior week) and show week-over-week delta. "Encampments up 15%" is useful. "95 encampments" is not.

Journalism archive: sf_journalism.py persists articles to sf_journalism_archive.json (cap 2000). For quarterly history, use --archive --days 90.

Socrata Dataset Reference

Query pattern: https://data.sfgov.org/resource/{id}.json?$where=supervisor_district='5'&$order=date+DESC&$limit=50

Scraping Method Reference

Scheduling

Weekly digest (Monday morning)

{
  "schedule": { "kind": "cron", "expr": "0 14 * * 1", "tz": "America/Los_Angeles" },
  "payload": { "kind": "systemEvent", "text": "Run SF civic weekly digest and send summary" }
}

Daily delta (weekday noon check-in)

{
  "schedule": { "kind": "cron", "expr": "0 12 * * 1-5", "tz": "America/Los_Angeles" },
  "payload": { "kind": "systemEvent", "text": "Run SF civic daily delta and report any new agenda items" }
}

File Layout

skills/sf-civic-digest/
├── SKILL.md              ← you are here
├── STYLE.md              ← report style guide (read before writing any report)
├── USER.md               ← user profile: district, neighborhoods, streets, interests (fill this in)
├── .gitignore            ← excludes runtime state from version control
├── scripts/              ← all data pipeline scripts
│   ├── sf_weekly_digest.py      ← master aggregator
│   ├── sf_civic_digest.py       ← Board of Supervisors (Legistar)
│   ├── sfmta_hearings.py        ← SFMTA Engineering Hearings
│   ├── sf_planning_notices.py   ← Planning project notices
│   ├── sf_planning_commission.py← Planning Commission / HPC / ZA
│   ├── sf_board_of_appeals.py   ← Board of Appeals
│   ├── sf_building_permits.py   ← Building permits (Socrata)
│   ├── sfgov_events.py          ← Supervisor events
│   ├── sf_311.py                ← 311 service requests (Socrata)
│   ├── sf_ethics.py             ← Lobbyist contacts + contributions
│   ├── sf_rent_board.py         ← Rent Board
│   ├── sf_sfmta_board.py        ← SFMTA Board of Directors
│   ├── sf_rec_park.py           ← Rec & Park Commission
│   ├── sf_housing_pipeline.py   ← Housing pipeline (AB2011/SB35/density bonus)
│   ├── sf_journalism.py         ← News aggregator (5 outlets)
│   ├── sf_volunteer_cleanups.py ← Community cleanups (Refuse Refuse + DPW)
│   ├── sf_evictions.py          ← Eviction notices (Socrata)
│   ├── sf_bart_board.py         ← BART Board of Directors (Legistar)
│   ├── sf_sfusd_board.py        ← SFUSD Board of Education
│   ├── sf_sfcta.py              ← SF County Transportation Authority
│   └── config_loader.py         ← District config loader
├── reports/              ← generated reports
│   └── archive/          ← older versions
└── references/
    ├── sf-districts.md       ← district map, supervisor names, neighborhood lists
    └── sf-build-timelines.md ← build phase benchmarks, financing signals, project context

Participation Guide

How to comment at SFMTA Engineering Hearings

• Online: join at sfmta.com/EngHearing when your item is called, raise hand
• Phone: number + conference ID in hearing notice
• Written: email the contact listed per item with "Public Hearing" in subject
• Decisions posted the following Friday at sfmta.com/EngineeringResults

How to comment at Planning Commission / HPC

• Written comments accepted before each hearing via sfplanning.org
• HPC meets 1st and 3rd Wednesdays (low attendance, high impact)
• In-person: City Hall Room 400

How to reach your supervisor

• Mahmood (D5): bilal.mahmood@sfgov.org
• Constituent meetings: request via supervisor's website
• Board public comment: first Tuesday of each month, City Hall Room 250