1.9 KiB
Changelog
0.2.0 — runtime composition repair
Root cause fixed
The 0.1.x runtime did not consume the example gallery. prompt.py emitted only a generic node/edge scaffold, the agent skill did not require reference inspection, VizSpec had no composition grammar, layout.py always used one topological layout, and the SVG renderer always drew the same rounded-card vocabulary with a visible title/question/footer. The linter also allowed several unconnected nodes.
Changing only examples/ therefore could not change generated output. A model could produce 25 differently labelled specs that all compiled to the same picture.
Added
- VizSpec 1.1
compositionblock. - Ten composition profiles and node/edge semantic roles.
- Context-sensitive reference selection with scores, matched terms, preview paths, and executable runtime-spec paths.
- Context/profile compatibility lint.
- Hard rejection of disconnected cards and excessive isolated nodes.
- Profile-specific structural gates.
- Profile-specific deterministic compositors.
- Diagram-only SVG output with hidden accessibility metadata only.
- Ten executable runtime profile fixtures.
techviz audit-batchlabel-independent topology-collapse detection.- Batch-collapse, context-profile, runtime-profile, and diagram-only tests.
Changed
techviz preparerecordsvisual_reference_candidatesin the context package.techviz promptembeds selected reference grammars and restricts the candidate profile set.techviz referencesprints score, matched terms, preview, and runtime-spec paths.- Manifests record composition profile, reference ids, and diagram-only status.
- Agent instructions require reference inspection and batch audit.
Compatibility
VizSpec 1.0 can still be parsed using the compatibility layout, but it emits a warning. New work should use VizSpec 1.1. Existing 1.0 specs must be upgraded to obtain profile-specific output and context-selection gates.