0.3.0

openstudio-ai 0.3.0 adds a review-gated, user-local learning workflow for
personal modeling assistance in both Codex and Claude Code.

Added

Five MCP learning tools for capturing observations, creating and reviewing
candidates, listing candidates, and retrieving approved personal lessons.
openstudio-ai learning CLI commands to curate evidence, propose measure
candidates from repeated successful scripts, inspect candidates, and preview
or explicitly confirm pruning of stale unapproved candidates.
The curate-learning skill, which can run as a bounded background/subagent
task on hosts that support subagents without interrupting primary modeling.

Renamed the PNNL runtime Python package from openstudio_mcp to
openstudio_ai_mcp. Integrations importing the former package must update
their imports; the runtime command remains openstudio-ai-mcp.
The MCP interface contract is now 4. Marketplace plugins exported before
this release declare contract 3 and are reported as incompatible by
openstudio-ai doctor and runtime_plugin_compatibility. Re-export or
reinstall the plugin, then reconnect its MCP server.
Learning is explicit and local: evidence is stored only after a host invokes
a learning tool. Unreviewed candidates never change future assistant
behavior, and approving a candidate creates a personal lesson only; it does
not modify shared skills, knowledge, measures, or MCP tools.

Update any Python integration from openstudio_mcp imports to
openstudio_ai_mcp.
Refresh Codex and Claude Code plugin exports, run openstudio-ai doctor, and
restart or reconnect the OpenStudio AI MCP server before using learning
workflows.
Before publishing, complete the full release checks below, including the
platform-specific OpenStudio simulation and TestPyPI validation.