Run a gene set analysis from a file in the chat

[adamjohnwright]

The last slice of spec 012. Attach an expression matrix, say which group each sample is in, get the significant pathways and the full table back as a download.

An attached file routes here instead of the graph

Not a tool the agent calls. The run takes minutes, longer than a chat turn, and the matrix is over a megabyte, which must never enter the model's context. A tool call would put the model in the middle of both problems.

Nothing reaches OpenAI

The result is described from the table and attached as a file — the user's own data going back to the user. Finished.for_model is computed and sent nowhere.

It is computed rather than skipped so the disclosure rules stay exercised by the tests. When a summary is added it must be handed the bounded allow-listed view, behind the existing warning — not the result.

Where the decisions live

gsa/chat.py and gsa/chainlit_flow.py hold them as functions over plain values: what counts as a usable reply, what the person is told, what happens when the service accepts a job and then dies. run_analysis takes its four chat operations as arguments — ask, say, revise, hand over a file — so it can be driven without a browser. The Chainlit handler is the only code that knows those are Chainlit, and what is left for a browser is whether the wiring matches.

Three things worth calling out

The label count is strict; the separators are not. Commas, semicolons and tabs all work, and case is folded for grouping while the user's own spelling is kept — the labels come back in the service's output and reading it against their own notes matters. The count is strict because three labels for four columns still analyses something, and returns a plausible answer to the wrong question.

Uploads are capped twice. .chainlit/config.toml goes from the template's 500 MB to 20 MB, and upload.py enforces 20 MB again. The first stops a browser sending it; only the second stops the server accepting it. The template also ships accept = ["*/*"], which invites every other kind of file into a chat that can do nothing with them — now TSV/CSV only, one file.

Results are written 0700 under a configurable directory, not a bare path in world-writable /tmp. They are a user's analysis on a shared host, and prune_results already bounds how long they stay.

Adversarial review

The progress line said "Done." on every path, including failure — set in a finally. A user whose analysis died was told it had finished, one line above the message explaining that it had not. It is removed now, and run_analysis reports the outcome itself on every path.

for_model being unused read as an oversight rather than a decision. It now says which it is, so the next person does not "fix" it by wiring the result into a prompt.

sabotage	test that failed
keep a rejected file	test_a_rejected_file_is_explained_and_deleted
never send the table	test_it_delivers_a_summary_and_a_file
swallow a post-submission failure	test_an_analysis_that_fails_after_starting_is_reported

27 new tests, ./checks.sh clean.

What this does not do

The public-dataset path (submit_public_dataset) is built and tested but has no chat route: reaching it needs the model to pick a dataset and a factor, which is conversation design rather than plumbing. That inverts the spec's priorities and the spec now records why — the upload path turned out to be the simpler one, because a file arrives with its own sample names and nothing needs disambiguating.

Wants a click-through on beta before it is called done.

🤖 Generated with Claude Code

---

Adversarial review of the finished branch

The upload survived the errors I had not thought of. Each branch deleted the file for itself, which covered every case I had in mind and none of the others.

• A file Chainlit has not finished writing raised FileNotFoundError straight out of validate, uncaught, past the handler. The user would have seen a raw error and the flow would have reported nothing. Reproduced before fixing, with an attachment pointing at a path that does not exist.
• An AskUserMessage that times out or is interrupted left the matrix on disk. Ten minutes is a long window, and an unhandled error is exactly when nobody is around to tidy up — on a host with 5 GB free.

The deletion is now a finally around the whole flow rather than a line in each branch. discard is safe to call twice, so the branches that already delete it still do; what changed is that the paths nobody enumerated are covered.

An unreadable file is also reported as something to retry rather than as the user's mistake — "that is not an expression matrix" is both wrong and unactionable when the real problem is that the upload did not land.

Also checked: a binary file with a .tsv name is refused cleanly rather than crashing. Somebody will attach a spreadsheet.

Verification

Replacing the finally fails four tests — including the two that passed before this commit. That is the point: they were passing because of their own branch's cleanup, not because the file was guaranteed gone.

30 tests in the flow, ./checks.sh clean.