[amrit110]

What

The tabctx-mcp agent demo (examples/agent_demo/) moves from telecom churn to 30-day hospital readmission, and its page loses most of its text in favour of visuals.

Data: Diabetes 130-US Hospitals (UCI; OpenML #4541)

• The canonical hospital readmission table: 101,766 stays of diabetic patients, 1999-2008. The task (rank today's discharges by risk, find drivers, test care changes, recommend a program) reads for hospitals, and as rank-by-risk-then-test-an-intervention for financial institutions (credit, claims, fraud).
• Prepared as in Strack et al. 2014: stays ending in death or hospice dropped, ICD-9 codes grouped into chapters, near-empty or constant columns dropped. Admission and discharge codes are named ("Home health service", not 6), and "None" lab results become "Not measured" (CSV readers took them for missing).
• 30,000 past stays and 2,000 current patients, one stay each, who appear nowhere in the history. Current patients are picked with a chance in proportion to their number of stays, as the history samples stays, so both sides have the same mix: 11.0% readmitted on each. (Picking patients uniformly gave a healthier current set, 9.6% against 11.6%; taking each patient's latest stay gave 4.7%.)
• task.py names the files and columns, the outcome labels (positive and negative class, window) and the page's step icons; prepare_data.py, score.py, run.sh (via python3 task.py files) and the page read it. Another dataset means changing task.py, prepare_data.py and PROMPT.md.

Page: visual over text

• The gateway-key paragraph is an ⓘ tooltip; the hero lost its paragraph and footer (the data card's lock already says the outcome is hidden).
• Agent cards: name, score badge and icon chips (MCP, no shell); model and tools on hover.
• Task: five step icons (Compare, Rank, Drivers, What-if, Report) with the prompt folded (open by default only locally, where it is editable).
• A failed tool call shows its first clause; the rest (e.g. a column list) on hover.
• The predict card shows the expected split (mean probability). For a rare outcome almost no row crosses 50%, so the old "No 100% / Yes 0%" said nothing.
• What-if rows: the new values, with the columns they go in beneath in small type.
• Tooltips escape untrusted values twice (tipAttr): data-tip is unescaped by the browser and then rendered as HTML.
• Score cards say "readmitted" from the dataset's labels; score.json keys are now positive_in_top_10/100 (fractions).

Fixes

• run.sh claude failed on macOS bash 3.2 without CLAUDE_MODEL (empty array "unbound" under set -u). The Docker image uses bash 5, so the hosted demo was not affected.
• make_showcase.py masks the login name (Codex's ls -la output showed it).

Results (prod gateway, real agents)

Agent (LLM)	tabctx calls	Wall time	ROC AUC on real outcomes	Readmitted in top 10 / top 100
Claude Code 2.1 (Opus 5.5)	10	160 s	0.674	80% / 36%
Codex 0.137 (gpt-5 via gateway)	9	188 s	0.674	80% / 36%
pi 0.73 (Sonnet 5.5 via gateway)	10	134 s	0.674	80% / 36%

• Base rate 11.0%: a random top 10 holds about 1; the top 100 is 3.3x the base rate. Readmission is hard to predict from administrative data (published scores reach AUC ~0.65-0.72), so these are honest, not churn-sized, numbers.
• evaluate on 7,500 held-out stays: tabpfn-3.5 0.697, tabicl-v2 0.690, gradient boosting 0.683. Fit 30,000 rows in 20-23 s; predict 2,000 in ~1.5 s; explain (50,000 rows in one predict) ~30 s.
• All three used every tool, dropped patient_id, chose tabpfn-3.5, and found the same drivers (prior inpatient stays, discharge destination, emergency visits, primary diagnosis). For the three riskiest patients going home, home health service lowered predicted risk by 2-10 points; medication and A1C changes by about 1 point or less. All three reports state these are learned associations, not proven effects, and recommend a risk-targeted transitional-care program measured against a comparison group.
• In an earlier run, Codex stopped after predicting with a progress message and no report (gpt-5 in codex exec sometimes ends its turn early); running it again finished.

Review

A review agent found one blocking issue and two should-fixes, all fixed in 483c70e:

• Blocking, XSS: data-tip values were escaped once, but the browser unescapes the attribute and the tooltip sets it as innerHTML, so an agent-supplied what-if value or column name could inject HTML on hover (older cards had the same pattern). Fixed with tipAttr(); checked in Chrome with an onerror payload in a column name, value, error and class label: no element created.
• Should-fix: the current set was healthier than the history (9.6% vs 11.6%); now stay-weighted (11.0% each), with a test. Showcase runs redone on the new split.
• Should-fix: a dataset swap still needed JS edits; the negative class, outcome window and step icons moved to task.py.
• Nits fixed: "Not scored" for a score.json without top-k keys, a clear error when task.py can't run in run.sh, stable sort in scoring, login masked only as an ls -l owner, dead counter and a churn comment removed.
• Verified with no change needed: the code-to-name maps against the dataset's IDs_mapping, the "None" → "Not measured" handling, run.sh in bash 3.2 and 5, the Docker image's imports, and every README/CHANGELOG number against the showcase files.

Tests

• New tests/unit/test_agent_demo_data.py: ICD-9 grouping, the patient-disjoint and stay-weighted split, scoring (top-k, base rate, row coverage), showcase runs matching the task, task.py files, and transcript sanitizing (paths, login name).
• ruff check, ruff format --check, and the unit suite (759 passed) pass locally.
• Checked in Chrome: local and hosted modes, live replays of the new runs, light and dark.

Deploy after merge: GCP_ACCOUNT=amritk@vectorinstitute.ai examples/agent_demo/deploy.sh.

🤖 Generated with Claude Code

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 84.78%. Comparing base (270b8f1) to head (9d2e3c7).
⚠️ Report is 1 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main      #87      +/-   ##
==========================================
+ Coverage   84.74%   84.78%   +0.03%     
==========================================
  Files          56       56              
  Lines        5021     5021              
==========================================
+ Hits         4255     4257       +2     
+ Misses        766      764       -2     

Flag	Coverage Δ
integration	37.22% <ø> (ø)
unit	81.87% <ø> (+0.03%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.
see 1 file with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.