[NielsRogge]

Do not merge yet. This depends on huggingface/paperswithcode#508 (GET /api/v1/papers/{identifier}/public-metadata), which is not live in production yet (curl https://paperswithcode.co/api/v1/papers/2610.18869/public-metadata → 404 as of 2026-10-14). If this ships first, uncatalogued papers fall back to the plain not-found error. This is the production chat's tool server, so merging or deploying needs explicit approval.

Why

PR #27 added client-side HF Papers / arXiv / Crossref lookups (external.py) for papers the catalog does not hold. The backend now serves the same data from one endpoint, which also returns labelled available_at links. This PR makes the MCP and the CLI use that endpoint. There is now one implementation, and both the agent and CLI users are told clearly that the paper is not indexed and where to read it.

MCP (pwc-mcp 0.2.7)

• get_paper_info, read_paper, get_related_papers: same in_catalog=false outputs as before (data=None, the abstract as Markdown for reads, a search fallback plus note for related papers). New: an available_at list (label, url, kind). The note now tells the agent to link to those sources, not a PwC page.
• If the endpoint maps the reference to a catalog paper (in_catalog=true), the tool retries with its arxiv_id, then catalog_id. If that still fails, the tool returns the original error rather than calling a catalogued paper "not indexed".
• Identifiers are sent in the path with slashes kept, as the backend route expects ({identifier:path}). 400 (titles) and 404 are cached for 1 h as misses. 429/503 are not cached and fall back to the original error.
• external.py and test_external.py are removed.
• ⚠️ Behaviour change: the endpoint does not accept titles, so a title the catalog does not hold now returns the plain not-found error. Before, the MCP looked it up on HF/arXiv. arXiv IDs/URLs, HF URLs and DOIs work as before.

CLI (pwc-cli 0.4.6)

pwc paper info / pwc paper read on a 404 (or "Paper title not found") now print this to stderr, before the unchanged error:

Not indexed on Papers With Code — available at:
- arXiv abstract: https://arxiv.org/abs/2610.18869
- arXiv PDF: https://arxiv.org/pdf/2610.18869

title: What Matters in Quality-Aware VLA Pretraining: A Benchmark and Empirical Study
published: 2026-10-13
authors: …
metadata_source: arxiv

## Abstract
…

pwc: API request failed (404): {"detail":"Paper not found","code":"paper_not_in_catalog"}

Stdout stays empty and the exit codes are unchanged (3 for a 404, 4 for an unknown title), so scripts behave exactly as before. An arXiv or HF URL that the API maps to a catalog paper is now read from the catalog. The fallback lives in main(), so embedders that call the handlers (the MCP, via queries.query) never trigger it (covered by a test).

Before / after for arXiv 2610.18869

The "after" run used the real response from #508's service code (f80bc013) against live arXiv, passed to this branch's MCP. Production catalog calls were live.

	before (0.2.6, client-side)	after (0.2.7, endpoint)
in_catalog	false	false
note	"This paper is not in the Papers With Code catalog; metadata comes from public arxiv records, so it has no catalog tasks, methods, evaluations, or curated code."	"This paper is not indexed on Papers With Code, so there is no Papers With Code page, catalog tasks, methods, evaluations, or curated code for it; metadata comes from public arXiv records. Link to where it is available instead: arXiv abstract: https://arxiv.org/abs/2610.18869; arXiv PDF: https://arxiv.org/pdf/2610.18869."
available_at	—	[{label: "arXiv abstract", kind: "abstract", …}, {label: "arXiv PDF", kind: "pdf", …}]
read_paper markdown	header + **Paper:**/**PDF:** lines + abstract	header + ## Available at link list + abstract
get_related_papers	same 2 search results + note	same 2 search results + note pointing at available_at + available_at

Tests

• uv run --project mcp_server pytest mcp_server/tests: 69 passed (endpoint client: path encoding and cached misses vs. transient errors; tool outputs with available_at; in-catalog retry; in-catalog-but-unreadable keeps the error)
• uv run --project standalone_cli --with pytest pytest standalone_cli/tests: 99 passed (stderr block for info/read × text/json, title miss exit 4, unknown everywhere unchanged, URL→catalog retry, embedded queries skip the endpoint)
• standalone_cli/scripts/check_contract.py ok; ruff shows no new findings vs. main

Before merging

• #508 deployed. curl https://paperswithcode.co/api/v1/papers/2610.18869/public-metadata returns 200 with in_catalog: false
• Re-run the live tool check against production
• Explicit approval to release pwc-mcp 0.2.7 / pwc-cli 0.4.6

🤖 Generated with Claude Code

[NielsRogge]

Live check against production (2026-10-14). huggingface/paperswithcode#508 is deployed (b39aebba).

Endpoint:

• /papers/2610.18869/public-metadata → 200, in_catalog: false, source arxiv, available_at = arXiv abstract + arXiv PDF
• /papers/https://arxiv.org/abs/1706.03762v7/public-metadata → 200, in_catalog: true, catalog_id 755
• 2610.99999 → 404 paper_not_found; a title → 400 invalid_paper_identifier

MCP (this branch, in-process, against production) for 2610.18869:

• get_paper_info → in_catalog: false, new note, available_at [arXiv abstract, arXiv PDF]
• read_paper → abstract Markdown with an "Available at" list
• get_related_papers → search fallback, with note + available_at

The output is identical to the "after" column in the description.

CLI (this branch):

• pwc paper info 2610.18869 and pwc paper read 2610.18869 --json: stdout empty, exit 3, "Not indexed on Papers With Code — available at: …" block, then the original 404 on stderr
• pwc paper info https://arxiv.org/abs/1706.03762v7: exit 0 with the catalog record. Before this PR it failed as "Paper title not found".
• pwc paper info 2610.99999: unchanged plain 404, exit 3

Still needed before merge: explicit approval to merge and release pwc-mcp 0.2.7 / pwc-cli 0.4.6.