academic-tools-mcp: Skill for Claude Code

Adding a new provider or OpenAlex entity

Adding a new OpenAlex entity

1. Add _normalize_* and _canonical_* functions in providers/openalex.py.
2. Add an async get_* function that checks cache, fetches, stores.
3. Add focused tool(s) in the matching tools/*.py module (OpenAlex metadata → tools/paper.py) that extract lean slices; shared param types live in app.py.
4. Add unit tests for normalization in tests/providers/test_openalex.py.

Adding a new API provider

Mirror providers/biorxiv.py — it is the fullest instance of the shape (both throttle wrappers, a cached_lookup getter, a PDF path). Read providers/crossref.py as a counter-example rather than a template: its rate constants are _resolve_policy() output, not literals. The shape (pooled client, _throttled_get + burst cap, _single_flight, cache → negative cache → fetch with re-checks inside the slot, 404 → negative cache) is documented in .claude/rules/providers.md and .claude/rules/net.md and .claude/rules/store.md. New clients live under providers/ and import shared infra one level up, by package (from ..net import clients, http, from ..store import cache, from ..util import config, doinorm, useragent). After mirroring it:

1. Nothing to register: stats.throttles() and the conftest reset fixture both discover the module's _throttle instance by scanning imported modules for a Throttle-typed attribute — the name is conventional, the discovery is by type. Declare NAMESPACE and LABEL at module level and pass both to the Throttle (namespace=NAMESPACE, label=LABEL) — tests/net/test_stats.py asserts each matches, because counters are filed under the namespace and the label is the name that reaches the agent. NAMESPACE is the on-disk .cache/ directory name, so it is a data migration, not a rename — it may differ from the module name and must not be changed to follow one (providers/acl.py keeps NAMESPACE = "acl_anthology"). Use LABEL at every other site naming the provider (http.error_dict, http.parse_error_dict, stream_to_file(provider_label=)); an AST scan in tests/test_politeness.py fails on a string literal at any of them.
2. Add env vars to .env.example and load via config.get().
3. If the provider serves PDFs, add a _Route row to manual._ROUTES (claims, NAMESPACE, canonical_key, pdf_path) — position it before the generic-DOI fallback — and, if it also serves metadata, an entry in manual._METADATA_SOURCE_BY_NAMESPACE. Without the row nothing routes to the new namespace. If its ids need a slash restored from a stem, corpus._restore_slashes needs a clause too, or corpus hits won't chain back.
4. Add tools in the matching tools/*.py module.
5. Tests in tests/providers/test_<name>.py (and tests/providers/test_<name>_properties.py for anything hypothesis is stronger at), covering normalization, parsing, backpressure, 404 negative-cache, and TTL eviction / force_refresh if relevant. tests/ mirrors src/; shared fakes live in tests/helpers/ and are imported absolutely.