Repository navigation
Conversation
A tabular foundation model learns from the past rows it is shown, so choosing them (the cohort) is the agent's main job, not just tidying. The brief gets a section after cleaning, "Which past rows will it learn from?": how many rows count (each once), how the export's outcome spellings were read (the question's own definition), and how many columns each row has. The data comes from the cleaning view, which now also reports these examples; the showcase's saved view is regenerated. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Codecov Report✅ All modified and coverable lines are covered by tests. ❌ Your changes status has failed because you have indirect coverage changes. Learn more about Unexpected Coverage Changes and reasons for indirect coverage changes. Additional details and impacted files@@ Coverage Diff @@
## main #145 +/- ##
==========================================
- Coverage 87.72% 87.60% -0.12%
==========================================
Files 67 67
Lines 6723 6723
==========================================
- Hits 5898 5890 -8
- Misses 825 833 +8
Flags with carried forward coverage won't be shown. Click here to find out more. 🚀 New features to boost your workflow:
|
The work panel listed raw bash and file writes (python3 - <<'EOF' ...), unreadable on a projector. Each command is now a tile with a plain title and a picture read from the command and its output: the files and their rows; what a check found (duplicates, a totals row, columns written many ways); the cleaning script's steps; the cleaning run as 6,127 -> 6,000 rows with each column's spellings collapsing into its values; edits as lines in and out. Consecutive checks fold into one tile. The raw command stays one click away, and a picture appears only when the output says it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
P (or ?present) gives the agent's work most of the screen at 1.3x, for a projector. A replay then stops before each of the agent's commands: →, space, Enter or a clicker's page-down shows the next one, and A plays the rest at a pace to talk over (2.4 s a command). A pill at the bottom shows the step and the keys. The work panel now follows the newest tile in every mode (a replay used to leave it at the top). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Presenter keys: Space and Enter on a focused button or link keep their meaning, and a step key only acts when a step is waiting. - A new run or replay stops the old one's pending timer, which could shift the new queue or start a second drain. - The work panel follows the newest tile only while the viewer is at the bottom; their own scrolling up is left alone. - actions.js: bounded regexes and output (an agent's long output could freeze the tab); "verify" no longer catches pandas .diff(); Gemini's replace/read_file/list_directory and pi's oldText/newText edits; python3 -u clean.py; the raw id comes from the scenario (app.py sends raw_id), and labels no longer say "policies". - cleaning.py: rows with no outcome are not examples (counted as "unlabeled", no "nan" class), and outcome spellings keep their spaces. - Tests: each agent's tool shapes, escaping of hostile output, a time bound on huge output, the Codex unwrap itself, unlabeled rows. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
For a group benefits audience: each row is an employer whose group health, dental, life and disability plan renews this year, and the question is which employers will leave at renewal. The drivers follow what group carriers describe: a renewal increase after a bad claims year (loss ratio), service issues, late payments, small plans, broker channel; tenure, broader coverage, ASO funding, online plan admin, a wellness program and member app use keep plans. An account review meeting halves the effect of an increase, the what-if lever. About one plan in eight leaves. The export keeps its mess, plus a loss ratio written as a percent in some rows and a fraction in others. The showcase is a new Claude Code run on this data: 6,127 -> 6,000 rows, held-out ROC AUC 0.765 vs 0.733 for gradient boosting, and against the withheld outcomes 0.776, 9 of the top 10 left, 48% of the top 100 (12% base rate). Also: - transcripts.py: a Claude Code system notice with a text message (permission_denied) no longer crashes the parser. - actions.js reads more of what agents print: pandas value counts as dicts, describe() extremes, "N rows in, M plans out", per-file prefixed lines, "(s)" plurals; a cleaning run that doesn't print its merges gets them from the page's own cleaning view. - Presenter keys work when the event target isn't an element. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Four changes to the agent demo, plus fixes from review, both so a room can follow the run on a big screen.
1. Which past rows will it learn from?
The brief now has a section after cleaning, Which past rows will it learn from? A tabular foundation model learns from the past rows it is shown, so choosing that cohort is the agent's main job, not only tidying the data.
It shows three things:
The data comes from
ui/cleaning.py:summarize()now returns anexamplesblock (rows, features, and the outcome's classes with their raw spellings). The showcase'scleaning.jsonis regenerated; its other fields are unchanged.2. The agent's commands, drawn as what they do to the data
"Show the work" used to list raw commands (
Bash python3 - <<'EOF' import pandas…), unreadable on a projector. A newui/static/actions.jsturns each command into a tile: a plain-English title, a small picture read from the command and its output, and the raw command one click away.ls; head; wc -lWrite clean.pypython3 clean.pyEditWrite report.mdA picture appears only when the command or its output says it; otherwise the tile shows just the title. Codex's
/bin/zsh -lcwrapper is unwrapped.3. Presenter mode, and a replay you step through
Press P (or open
/?present) to give the agent's work most of the screen at 1.3× for a projector; P again returns to the normal page.In presenter mode a replay stops before each of the agent's commands:
A pill at the bottom shows the step (
6 / 27) and the keys. The work panel now follows the newest tile in every mode; a replay used to leave it at the top. A live run is never held back.4. The insurance scenario becomes group benefits renewals
Each row is now an employer whose group health, dental, life and disability plan renews this year; the question is which employers will leave. Drivers follow what group carriers describe: a renewal increase after a bad claims year (loss ratio), service issues, late payments, small plans and the broker channel push plans out; tenure, broader coverage, ASO funding, online plan admin, a wellness program and member app use keep them. An account review meeting halves the effect of an increase, the what-if lever. About one plan in eight leaves. The export keeps its mess, plus a loss ratio written as a percent in some rows and a fraction in others.
The showcase is a new Claude Code run on this data (2.6 min, $0.92): 6,127 → 6,000 rows; held-out ROC AUC 0.765 vs 0.733 for gradient boosting; against the withheld outcomes 0.776, 9 of the top 10 left, 48% of the top 100 (12% base rate).
Fixes from review
actions.js: bounded regexes and output (long agent output could freeze the tab);verifyno longer catches pandas.diff(); Gemini'sreplace/read_file/list_directoryand pi'soldText/newTextedits;python3 -u clean.py; reads pandas value-count dicts,describe()extremes, "N rows in, M plans out" and per-file lines; a cleaning run that doesn't print its merges gets them from the page's own cleaning view.cleaning.py: rows with no outcome are not examples (counted asunlabeled), and outcome spellings keep their spaces.transcripts.py: a Claude Code system notice whosemessageis text (permission_denied) no longer crashes the parser.Tests:
tests/unit/test_agent_demo_cleaning.py(two new cases) and the newtests/unit/test_agent_demo_actions.py, which runsactions.jsunder node on the recorded Claude Code run (skipped without node). All agent demo tests pass (105), including new ones for each agent's tool shapes, escaping of hostile output, a time bound on huge output, unlabeled rows and the group benefits data's levers. Checked in Chrome on the recorded insurance run, in both modes (stepping, auto, and P on and off).🤖 Generated with Claude Code