integrate-tool

name	integrate-tool
description	Teaches an agent how to add a new search provider, scrape backend, or data source to Calixto Research Workspace. Use when the user asks to add support for Brave, Tavily, Google Scholar, Semantic Scholar, PubMed, GitHub, a custom API, or any new data source. Covers the SearchProvider and ScrapeProvider interfaces, registration in the script dispatchers, optional dependency groups, testing, and documentation. Also useful for adding new CLI scripts.
license	MIT
compatibility	Requires Python 3.11+, Calixto Research Workspace installed, familiarity with the providers/ and scripts/ directories in the Calixto repo.
metadata	{"category":"meta","mode":"developer","version":"0.1.0"}

Integrate a New Tool or Provider (Meta-Skill)

Teaches an agent how to add a new search provider, scrape backend, data source, or CLI script to Calixto.

When to Use

• The user asks to add support for a new search provider (Brave, Tavily, Google Scholar, Semantic Scholar, etc.)
• The user wants to integrate a new scrape backend
• The user wants to add a new data source (PubMed, GitHub, Stack Overflow, etc.)
• The user wants to write a new CLI script in scripts/

Do NOT use this skill for one-off scripts outside the toolkit, or for adding new skills (use create-skill instead).

Goal

Add a new tool or provider to Calixto by:

1. Implementing the appropriate interface (SearchProvider, ScrapeProvider, or a new one)
2. Registering the new tool in the relevant script dispatcher
3. Documenting the new tool so future agents can use it
4. Testing the new tool against the golden dataset

The goal is that an agent reading only the existing docs (AGENTS.md, the provider interface, this skill) can add a new provider without asking the user for clarification.

Provider Interface Contracts

SearchProvider

Located at providers/search/base.py.

from providers.search.base import SearchProvider, SearchResult, SearchError

class MyProvider(SearchProvider):
    name = "my_provider"

    def search(self, query: str, max_results: int = 10) -> list[SearchResult]:
        self.validate_query(query)
        # ... call the underlying API ...
        return [
            SearchResult(url=..., title=..., snippet=..., score=..., metadata={...})
            for r in raw_results
        ]

Required:

• name class attribute (lowercase, unique)
• search() method returning list[SearchResult]
• Rate limiting with backoff and retries
• Raise SearchError for recoverable failures
• Let unexpected exceptions propagate

ScrapeProvider

Located at providers/scrape/base.py.

from providers.scrape.base import ScrapeProvider, ScrapeResult, ScrapeError

class MyScraper(ScrapeProvider):
    name = "my_scraper"
    timeout_seconds = 30

    def scrape(self, url: str) -> ScrapeResult:
        self.validate_url(url)
        # ... fetch and convert ...
        return ScrapeResult(
            url=url, title=..., markdown=..., word_count=..., metadata={...}
        )

Failure handling (per requirements.md section 12.2):

• Paywall: set metadata["error"] = "paywalled", return partial content if any
• Timeout: raise ScrapeError("timeout")
• Non-HTML content: set metadata["error"] = "unsupported_content_type"
• Crash: let the exception propagate

Step-by-Step: Adding a New Search Provider

1. Create the provider file

providers/search/my_provider.py:

"""
my_provider.py: <Description>

Usage:
    from providers.search.my_provider import MyProvider
    p = MyProvider(api_key="...")
    results = p.search("query", max_results=10)

Rate limiting:
    - <Describe how rate limits are handled>

Error handling:
    - <Describe what raises SearchError>
"""

from __future__ import annotations

import logging
import time

from .base import SearchError, SearchProvider, SearchResult

log = logging.getLogger(__name__)


class MyProvider(SearchProvider):
    name = "my_provider"

    def __init__(self, api_key: str, delay_seconds: float = 1.0, max_retries: int = 3):
        if not api_key:
            raise ValueError("api_key is required")
        self.api_key = api_key
        self.delay_seconds = delay_seconds
        self.max_retries = max_retries
        self._last_request_time = 0.0

    def search(self, query: str, max_results: int = 10) -> list[SearchResult]:
        self.validate_query(query)
        if max_results <= 0:
            return []

        elapsed = time.monotonic() - self._last_request_time
        if self._last_request_time >   elapsed < .delay_seconds:
            time.sleep(.delay_seconds - elapsed)

         attempt  (, .max_retries + ):
            :
                response = call_my_api(.api_key, query, max_results)
                ._last_request_time = time.monotonic()
                 [
                    SearchResult(
                        url=r[],
                        title=r.get(, ),
                        snippet=r.get(, ),
                        score=r.get(, ),
                    )
                     r  response.get(, [])
                ]
             Exception  e:
                   (e).lower()  attempt < .max_retries:
                    backoff =  ** attempt
                    log.warning(, backoff, e)
                    time.sleep(backoff)
                    
                 SearchError()

         []


 () -> :
    
     {: []}

2. Register the provider in the script dispatcher

Edit scripts/search_web.py:

def get_search_provider(name: str, **kwargs: Any) -> Any:
    name = name.lower()
    if name == "duckduckgo":
        from providers.search.duckduckgo import DuckDuckGoProvider
        return DuckDuckGoProvider(**kwargs)
    if name == "brave":
        from providers.search.brave import BraveProvider
        return BraveProvider(**kwargs)
    if name == "my_provider":
        from providers.search.my_provider import MyProvider
        api_key = kwargs.pop("api_key", None) or os.environ.get("MY_PROVIDER_API_KEY")
        return MyProvider(api_key=api_key, **kwargs)
    raise ValueError(f"unknown search provider: {name}")

3. Add optional dependency

If the provider requires a Python package, add it to pyproject.toml:

[project.optional-dependencies]
my_provider = ["my-provider-sdk>=1.0"]

Users opt in with: pip install 'calixto-research-workspace[my_provider]' or uv sync --extra my_provider.

4. Document

• Add a module docstring with usage example
• Add a row to the "Optional Search Providers" table in requirements.md section 5.3
• Update AGENTS.md if the provider ships by default

5. Test

# Smoke test
python -c "from providers.search.my_provider import MyProvider; p = MyProvider(api_key='test'); print('importable')"

# End-to-end test
python scripts/search_web.py "test query" --workspace workspaces/test --max-results 3 --search-provider my_provider --no-scrape

Step-by-Step: Adding a New Scrape Provider

Same pattern as search, but the file goes in providers/scrape/, the interface is ScrapeProvider, and you register in scripts/search_web.py's get_scrape_provider function.

Step-by-Step: Adding a New Data Source

A "data source" is anything beyond web search and arXiv (e.g., Semantic Scholar, PubMed, GitHub). It is essentially a specialized search provider with a custom data model.

Approach:

1. Create providers/search/semantic_scholar.py (etc.)
2. Implement the SearchProvider interface, mapping the source's data model into SearchResult
3. Save results to sources/code/ or sources/papers/ depending on the type
4. Modify scripts/search_web.py (or add a new script) to dispatch to the new provider
5. Add a flag like --source-type to the script

Alternatively, create a new top-level script scripts/search_semantic_scholar.py if the data source has unique behavior.

Step-by-Step: Adding a New CLI Script

When you need a new script that is not just "another provider":

1. Create scripts/<script_name>.py
2. Add sys.path.insert(0, str(Path(__file__).resolve().parent)) at the top so _common imports work
3. Use _common.emit_ok(), _common.emit_error(), _common.emit_partial() for structured output
4. Use _common.load_workspace_config(), save_workspace_config(), load_source_index(), save_source_index() for workspace I/O
5. Register a console entry point in pyproject.toml under [project.scripts]
6. Document the script in AGENTS.md "Scripts Reference"
7. Add unit tests under tests/

Testing New Providers

Unit test (recommended)

# tests/test_my_provider.py
from providers.search.my_provider import MyProvider

def test_search_returns_results():
    p = MyProvider(api_key="test-key")
    results = p.search("python", max_results=3)
    assert len(results) <= 3
    assert all(r.url for r in results)

Integration test

# Fresh workspace
python scripts/init_workspace.py test-my-provider

# Run the search
python scripts/search_web.py "test query" --workspace workspaces/test-my-provider --max-results 3 --search-provider my_provider --no-scrape

# Verify
python scripts/workspace_info.py show test-my-provider
python scripts/workspace_info.py audit test-my-provider

Golden dataset test

Add a new search to tests/golden/config.json and verify the run is reproducible.

Updating Documentation

After adding a provider:

1. Update AGENTS.md "Scripts Reference" if applicable
2. Add a row to requirements.md section 5.3 (search providers) or section 5.4 (scrape providers)
3. Create or update providers/<type>/README.md with the new provider's usage
4. If the provider required a non-trivial design decision, create an ADR in docs/adr/
5. Add an entry to docs/initial-implementation-plan-decision-log.md for any non-obvious choice

Reference Implementations

Use the existing providers as references:

• providers/search/duckduckgo.py: free, no API key, simple rate limiting
• providers/search/brave.py: paid, requires API key, more sophisticated rate limit tracking
• providers/scrape/crawl4ai_provider.py: the only scrape provider, async wrapped in sync

Common Pitfalls

• Forgetting to register the provider: a new provider is invisible until it appears in the script's dispatcher
• Hardcoding the API key: read from env or pass as a constructor arg; never commit
• Not handling rate limits: every external API has them; check the API docs and bake the limits into the provider
• Returning raw dicts instead of SearchResult: the script layer depends on the dataclass
• Inconsistent error types: always raise SearchError / ScrapeError, not generic Exception
• Skipping tests: even a 5-line smoke test catches the obvious bugs

See Also

• providers/search/base.py: SearchProvider interface
• providers/scrape/base.py: ScrapeProvider interface
• skills/create-skill/SKILL.md: how to add a new skill
• requirements.md section 5.5: Adding New Providers
• docs/adr/001-choose-crawl4ai.md: example ADR