release: 0.20.0

CHANGELOG.md

@@ -1,6 +1,8 @@
 # Changelog
 
 ## [Unreleased]
+
+## [0.20.0] - 2026-11-10
 ### Added
 - **An insurance scenario for the agent demo.** A coding agent gets two messy raw exports from a made-up insurer's policy system (synthetic data: dates in four formats, a province spelled five ways, rows exported twice, test policies, a totals row), writes and runs a cleaning script, then finds which policyholders will cancel at renewal, why, and which retention offers move their risk, and writes a report for the head of retention. `DEMO_SCENARIO=insurance` picks it; readmission stays the default and the hosted demo. Scenarios are now folders in `examples/agent_demo/scenarios/` (task, prompt, data), found through `scenario.py`; a task with a "clean" step gives each agent its shell (Claude Code's only inside its sandbox, which keeps it out of the answers). A Claude Code run against TabICL on a laptop took about five minutes.
 ### Fixed

deploy/helm/tabctx/Chart.yaml

@@ -5,8 +5,8 @@ type: application
 # Released with tabctx, at the same version: a release bumps both with
 # pyproject.toml (tests/unit/test_helm_chart.py checks). appVersion is
 # the image the chart deploys.
-version: 0.19.0
-appVersion: "0.19.0"
+version: 0.20.0
+appVersion: "0.20.0"
 home: https://vectorinstitute.github.io/tabctx/
 sources:
   - https://github.com/VectorInstitute/tabctx

pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "tabctx"
-version = "0.19.0"
+version = "0.20.0"
 description = "Serving engine for tabular foundation models (TabICL, TabPFN): fit a table once, keep its encoded context on the GPU, and serve fast repeat predictions to many tenants, with memory-aware admission and multi-GPU placement."
 readme = "README.md"
 license = "Apache-2.0"