literature-search

Goal:

• Run the canonical literature search bridge for one evidence query and persist traceable evidence artifacts.

Inputs:

• active run directory
• a SearchRequestContract payload or enough context to build one
• optional existing literature/queries/<query_id>/REQUEST.json

Outputs:

• literature/queries/<query_id>/REQUEST.json
• literature/queries/<query_id>/PROVIDER_RECEIPTS.json
• literature/queries/<query_id>/CANDIDATE_PAPERS.json
• literature/queries/<query_id>/VERIFIED_PAPERS.json
• literature/queries/<query_id>/EVIDENCE_BUNDLE.json
• literature/queries/<query_id>/EVIDENCE_BUNDLE.md
• literature/queries/<query_id>/SEARCH_TRACE.jsonl
• literature/bundles/<bundle_id>.json

Context Loading:

• Open skills/shared-references/schema-index.md.
• Open skills/shared-references/literature-search-contract.md.
• Read packages/agent_contracts/literature.py before building or editing a search request or evidence bundle.
• Read tools/literature_search_client.py before invoking the bridge.

Execution Contract:

• Import and call from tools import search_literature.
• When the host exposes the configured co_scientist_search_bridge MCP server, its tools are an allowed transport for the same canonical bridge.
• If the MCP server is unavailable, fall back to the Python tools surface instead of performing informal web search.
• Do not perform informal web search as a substitute for tools.search_literature(...).
• Do not invent papers, DOIs, arXiv IDs, venues, citation counts, or abstracts.
• Preserve provider failure receipts. A failed provider is evidence about retrieval coverage and must not be hidden.
• If retrieval_metadata.status is blocked, return that blocked state to the caller instead of fabricating an evidence bundle.
• If retrieval_metadata.status is partial, downstream callers may continue only when they preserve the partial-source limitation.

Execution Steps:

1. Open skills/shared-references/schema-index.md, skills/shared-references/literature-search-contract.md, and packages/agent_contracts/literature.py.
2. Build a SearchRequestContract with a stable query_id, goal, query, query_type, providers, filters, and consumer.
3. Call tools.search_literature(run_dir, request) through the stable tools surface, or call the equivalent configured MCP search bridge tool when it is available.
4. Read the returned EvidenceBundleContract and confirm the run-local artifacts were written.
5. If the bundle status is blocked, stop and report the blocked retrieval state to the caller.
6. If the bundle status is partial or succeeded, return the bundle_id and query_id to the caller.
7. Run python -m tools.validation.contract_validation <run_dir> --skill literature-search.

Completion Rule:

• This skill is complete only when the search bridge has written a canonical evidence bundle or an auditable blocked retrieval state, and the literature artifacts validate.