Adu City Research

ADU City Research

Overview

Research city-level ADU regulations for any California city. California has 480+ cities, each with different websites, different municipal code platforms, and different formats. State ADU law (covered by the california-adu skill) addresses certain development standards, while cities have their own additional requirements — local building code amendments, grading ordinances, utility standards, and submittal expectations. This skill finds what a specific city requires at the local level.

Key principle: Every California city is required by law to publish their ADU regulations on their website. The information exists — this skill is the method for finding it.

Required tools: WebSearch and WebFetch at minimum. Chrome MCP (browser automation) only needed for Mode 3 (Browser Fallback).

Research Modes

This skill operates in three modes. When used standalone, run all three sequentially. When invoked by an orchestrator, run only the specified mode.

When to Use Each Mode

• Standalone (no orchestrator): Run Discovery → Extraction → Fallback (if needed). Single agent, sequential.
• Orchestrated (called by adu-corrections-flow): The orchestrator specifies which mode to run. Discovery runs in parallel with other research. Extraction runs after Discovery completes. Fallback runs only if Extraction has gaps.
• Fan-out ready: The Extraction mode takes a list of URLs + topics. An orchestrator can split this list across multiple agents for parallel extraction without changing the mode's logic.

Mode 1: Discovery

Tool: WebSearch only Input: City name + list of correction topics (optional — omit for comprehensive discovery) Output: city_discovery.json — key URLs organized by category Time: ~30 seconds

Process

Run 3-5 targeted web searches to find the city's key ADU-related pages. Do NOT fetch or read the pages — just find the URLs.

Search queries to run (in order of priority):

1. "City of [Name]" ADU accessory dwelling unit — finds the city's ADU page
2. "City of [Name]" municipal code — finds the code platform (ecode360, Municode, QCode, etc.)
3. "City of [Name]" standard detail construction OR "City of [Name]" standard plans engineering — finds standard detail PDFs
4. "City of [Name]" information bulletin building — finds IBs
5. "City of [Name]" building permit submittal requirements — finds checklists

If correction topics are provided, add targeted searches:

Platform detection: Identify the municipal code platform from search results. See references/municipal-code-platforms.md for the major platforms (ecode360, Municode, QCode, American Legal, Sterling Codifiers, city-hosted).

Building vs. Planning split: Larger cities often split ADU information across two departments. Flag both URLs if found.

Discovery Output Format

{
  "city": "City name",
  "discovery_timestamp": "ISO date",
  "urls": {
    "adu_page": "https://...",
    "adu_page_building": "https://... (if separate from planning)",
    "municipal_code": {
      "platform": "ecode360 | municode | qcode | amlegal | sterling | city-hosted",
      "base_url": "https://...",
      "building_code_title_url": "https://... (if found in search results)",
      "zoning_title_url": "https://... (if found in search results)"
    },
    "standard_details": [
      {"name": "Inferred name from search result", "url": "https://..."}
    ],
    "information_bulletins": [
      {"name": "Inferred name from search result", "url": "https://..."}
    ],
    "submittal_requirements": "https://...",
    "pre_approved_plans": "https://..."
  },
  "not_found": ["category names where no URLs were found"],
  "notes": "Any observations from search results (e.g., 'city appears to use Municode but ADU page links to a separate handout site')"
}

Important: Discovery only finds URLs — it does NOT read page content. Speed comes from not fetching anything. The Extraction mode does the reading.

Mode 2: Targeted Extraction

Tool: WebFetch primarily Input: Discovery results (city_discovery.json or equivalent URL list) + specific topics to extract Output: city_research_findings.json — extracted content organized by category Time: ~60-90 seconds (depends on number of URLs)

Process

For each URL from Discovery, fetch the page content and extract information relevant to the correction topics. Prioritize by impact.

Extraction priority (highest first):

1. Standard detail PDFs — If the corrections reference city standard details, fetching these PDFs is the single most valuable action. They show exactly what the city wants drawn.
2. Municipal code sections — Fetch the specific title/chapter relevant to corrections (building code amendments, ADU ordinance, grading chapter). Do NOT fetch the entire municipal code.
3. Information Bulletins — Fetch IBs relevant to correction topics (wind parameters, submittal guides).
4. ADU page content — Fetch the city's ADU page for any additional requirements or handouts.
5. Submittal requirements — Fetch if corrections reference missing sheets, notes, or plan check items.

What to extract from each page:

WebFetch tips:

• Most municipal code platforms (ecode360, Municode, QCode) serve clean HTML that WebFetch handles well.
• Standard detail PDFs may not be readable via WebFetch — note the URL and title for the professional to download.
• If WebFetch returns empty or error for a URL, add it to extraction_gaps for Browser Fallback.

Extraction Output Format

{
  "city": "City name",
  "extraction_timestamp": "ISO date",
  "municipal_code": {
    "platform": "ecode360 | municode | qcode | etc.",
    "url": "https://...",
    "code_edition": "2022 | 2025 California Building Standards Code",
    "current_through": "Ordinance # or date if available",
    "building_code_title": "Title XX",
    "relevant_chapters": [
      {
        "chapter": "XX.XX",
        "title": "Chapter title",
        "url": "direct URL",
        "key_content": "Summary of relevant provisions"
      }
    ],
    "local_amendments": [
      {
        "section": "XX.XX.XXX",
        "amends": "CBC/CRC/CPC section reference",
        "description": "What it changes from state code",
        "applies_to_items": ["correction item numbers"]
      }
    ]
  },
  "standard_details": [
    {
      "name": "Detail name",
      "url": "PDF URL",
      "description": "What the detail shows",
      "applies_to_items": ["correction item numbers"]
    }
  ],
  "information_bulletins": [
    {
      "number": "IB-XXX",
      "title": "Title",
      "url": "URL",
      "key_requirements": "Summary of requirements",
      "applies_to_items": ["correction item numbers"]
    }
  ],
  "specific_findings": [
    {
      "topic": "What was researched",
      "finding": "What was found",
      "source": "URL or code section",
      "applies_to_items": ["correction item numbers"]
    }
  ],
  "extraction_gaps": [
    {
      "category": "What we tried to find",
      "url_attempted": "URL that failed or returned empty",
      "reason": "Why extraction failed (timeout, empty response, PDF not readable, etc.)",
      "fallback_suggestion": "What to try in Browser Fallback mode"
    }
  ]
}

Fan-out design: The extraction input is a list of URLs + topics. To parallelize, split this list:

• Agent 1: Municipal code URLs + code-related topics
• Agent 2: Standard detail and IB URLs
• Agent 3: ADU page + submittal requirements

Each agent produces a partial city_research_findings.json. The orchestrator merges them.

Mode 3: Browser Fallback

Tool: Chrome MCP (browser automation) Input: City name + extraction_gaps from Mode 2 Output: Gap-filling additions to city_research_findings.json Time: ~2-3 minutes When to use: Only when Mode 2 has extraction gaps that matter for the corrections