feed.research named the tool and threw its papers away

[mgoldey]

What

feed.ask routed a research question correctly, called feed.research, and then dropped its results. The reader got a count and an apology.

Measured from a live Discord session, 2026-09-18. Asked to "research orthogonalized mixtures of experts", the agent answered that it had "only a summary count rather than the specific details or titles". Asked for "the explicit corpus" it apologised; asked to "retrieve the full assets" it said it would need an ID first. It was not being evasive — _compose really did hand it nothing else.

Two causes, both in mcp/ask.py

1. papers was not in _RESULT_KEYS. The tuple listed items, which is what feed.list returns, but feed.research returns papers. So _summarise fell through to the bare message. This is the exact failure that tuple's own comment warns about:

A route whose result key is absent here degrades to a bare count, which is silent.

Same bug already fixed once for kg.path and kg.neighbors; it recurred on the next route added.

2. refs was built only from rows carrying item_id. A research hit is a library row and has no item_id, so every paper was filtered out and the reader was told about work it had no way to open. Papers now ride on the doi/arXiv id cite.lookup takes.

Before / after

Same payload through _compose:

before  answer: "3 paper(s) from arxiv, pubmed; 3 new in the library"
        refs:   []

after   answer: "3 paper(s) from arxiv, pubmed: Towards a Statistical
                 Understanding of Mixture-of-Experts; COEC: Calibrated
                 Orthogonal-Equivalence Compensation"
        refs:   [{paper_id: 2609.03501, url: https://arxiv.org/abs/2609.03501},
                 {paper_id: 2608.21142, url: https://arxiv.org/abs/2608.21142}]

For contrast, feed.list was always fine — it returns items.

Gates

Both gates that pushed back were right, and neither was worked around:

• complexity refused the inline version (_compose at 14 against a limit of 10), so _refs is split out. The gate's call, not a preference.
• size ratchet 379 → 383, with the reason recorded beside the entry.

All eight pass: ruff format, ruff check, ty, uv.lock, complexity, bandit, xenon, pytest (1863 passed).

Regression test is test_feed_research_answer_names_the_papers_and_carries_their_links, pinned the same way as the kg.ask one it rhymes with.

Not fixed here

That session's library holds 3 PubMed rows and zero arXiv, though every paper shown was arXiv: arXiv intermittently answers 406 to the research client. Retries of the identical URL succeed and curl succeeds, so it reads as throttling rather than a malformed request. Worth watching — if it persists, ArxivSearch wants retry-on-406 — but not worth building machinery against one afternoon's evidence.

🤖 Generated with Claude Code