Offer a gene list however it is handed over, and remember it

[adamjohnwright]

Two conversations reported on 2026-09-29. In both, the model answered with website instructions, and those instructions were wrong: they sent a gene list to ReactomeGSA, which takes a matrix.

1. The chat said "include the genes in your message", and the reply "here is my gene list TP53, ERBB3 and JAX9" was not recognised.
2. "can you analyze the gene list that I gave you" was not recognised either; the list was in an earlier message.

Now

• Hand-over phrasing counts as a request: "here is / here's my gene list …", "my genes are …", "these are my genes …", "gene list: …".
• A reply to an invitation: straight after the GSA how-to asks for genes, the next message that lists them is offered, however it is phrased. It must look like genes: three or more items, or at least one gene-shaped token. Group labels such as "control, control, treated, treated" are not offered.
• Pointing back: the last list a reader sent is remembered, including one from a question the model answered. "analyse the gene list I gave you" offers that list, marked as coming from an earlier message.
  • If no list was sent yet, the chat asks for one; the model is not asked.
  • A message that points at a matrix, a file, a GSA or results is not treated as pointing back at a list.
• Declining an earlier-list offer names the genes to the model.
• One function decides. gene_list.read_message makes all of these decisions, as a pure function, and runs off the event loop.

Review

One round of adversarial review. It found two regressions in this flow, which are fixed and tested:

• the how-to's own group example was offered as a gene list;
• matrix requests that pointed back got "no gene list".

It also found chit-chat overwriting the remembered list, the word "are" opening lists everywhere, and 4× parse cost. All are fixed. Each guard has a test that fails when the guard is sabotaged. One sabotage survived at first, so I added cases that only that guard catches.

Measured

A fresh set for the new routes, written after review and measured once before its fixes: 13/16. The two misses fixed since are marked as tuned. The third, "BRCA1 and BRCA2" as a reply to an invitation, is pinned as a known limit. A case earlier moved into held-out set 4 after a fix aimed at it now sits in TUNED_LATER.

Verified

• ./checks.sh passes.
• In a browser against a local instance, both reported conversations were replayed verbatim:
  • the offer appears;
  • Run gives "Matched 2 of 3";
  • JAX9 is listed as not found.
• Also checked in the browser: a bare list after the how-to; an earlier list offered by name; "no list yet" followed by a pasted list; group labels not offered; a matrix request pointing back still gets the matrix how-to.
• The earlier 23-check suite still passes.

Not fixed here

The model's own answers about gene lists come from the website's user guide, and they tell people to paste a gene list into ReactomeGSA. That answer path is shared with the search page, so it needs a separate change.

🤖 Generated with Claude Code