[NielsRogge]

Production chat audit (7 days to 2026-10-14, ~5,500 turns): reliability is fine, but 5.7% of tool calls errored, mostly with lookups the agent could not recover from. This PR changes the MCP tools (and one CLI hint helper) so those calls return something useful.

Changes

• Catalog misses fall back to public metadata. get_paper_info, read_paper, and get_related_papers used to return a bare not_found for arXiv IDs, arXiv/HF URLs, DOIs, and titles the catalog does not hold (76 of 80 recent misses were real catalog gaps). They now fetch public metadata: title, authors, abstract, date, links, and linked repos.
  • Sources, in order: Hugging Face Papers, then the arXiv abstract page. Titles try the arXiv API after HF Papers search. DOIs go to Crossref.
  • get_paper_info and read_paper mark the result in_catalog: false; read_paper returns the abstract as Markdown.
  • get_related_papers returns a catalog search for the public title, with a note.
  • A DOI whose public title (or arXiv ID) is in the catalog returns the catalog paper.
  • Lookups are best effort (6 s timeout per request) and cached for an hour, misses included. The reference is never logged.
  • When nothing matches, the error now says the paper is not in the catalog and no public record matched it.
• Short names resolve. Title lookup accepts the one catalog paper whose title starts with the name before the colon: MuseTalk → MuseTalk: Real-Time High-Fidelity Video Dubbing … (2410.10122).
• read_paper placeholder cursors. cursor="next" (and start, 0, …) with nothing saved now reads from the beginning. This caused 28 failures a week, each followed by identical retries.
• Closest-result hints need real overlap. A hint now has to share at least half of the reference's words, contain or be contained in it, or be a near spelling (ratio ≥ 0.8). FaceNet no longer suggests WaveNet, and "person re-identification" no longer suggests Spoken Language Identification. No hint beats a misleading one.
• get_benchmark with an unknown metric (overall, pass@1, accuracy on a leaderboard without them) returns the unfiltered leaderboard. A metric_note names the metrics it does have, instead of raising an error.
• Bumps pwc-cli to 0.4.5 and pwc-mcp to 0.2.6.

Notes

• New outbound traffic: a catalog miss now sends the user's paper reference to huggingface.co, arxiv.org, or api.crossref.org.
• New output fields, all additive: PaperInfoResult.in_catalog/note, PaperReadResult.in_catalog, PaperPage.note, BenchmarkResult.metric_note.
• arXiv's export API is rate-limited from the production host, so it is only the last resort for titles. Lookups by arXiv ID use the abstract page instead.

Verification

• uv run pytest in mcp_server: 86 passed. Covers fallback, cursor, metric-note, short-name, caching, and parser tests.
• CLI tests: 90 passed.
• In-process live smoke run against the production API:
  • MuseTalk and Noise2Noise resolve in the catalog.
  • "Instant Neural Graphics Primitives with a Multiresolution Hash Encoding" returns arXiv metadata (2201.05989) with in_catalog: false.
  • https://doi.org/10.1088/1361-6471/ac5fb4 resolves to catalog paper 2108.05103 through its Crossref title.
  • read_paper(cursor="next") reads from the beginning.
  • get_method("FaceNet") and get_task("person re-identification") return no misleading hints.
  • get_benchmark("ImageNet", sort_metric="pass@1") returns the leaderboard with "available metrics: FID, PSNR, SSIM".

Chat-related: do not merge or deploy without explicit authorization.

🤖 Generated with Claude Code