Limitations
Stated plainly, so that a gap is never mistaken for a bug.
Durable records
Development snapshots and exports written by pre-0.1 builds — before the unversioned durable-record cutover — are not migrated. Rebuild artifacts and discard those preview Eval and Playground records before upgrading. Current readers reject the superseded shapes instead of guessing at compatibility.
The development server
The Workbench binds to loopback only and is a foreground development session, not a hosted service.
Native harnesses
Native Claude and Codex harnesses require those CLIs to be installed and signed in. A missing, incompatible, or unauthenticated CLI is reported as a harness failure, which is distinct from a plugin failure. Their live smoke tests are opt-in and are not part of an ordinary test run.
Semantic grading requires a native Claude harness and a signed-in Claude Code session; the deterministic and Codex selections are refused when it is configured.
Evidence that cannot be observed
Codex exposes no authoritative Skill-activation event, so Codex activation evidence is inferred
and is never reported as observed.
Comparison facets that a run did not record — semantic grader identity, host CLI version, invocation — are labeled unverified rather than assumed aligned.
What the contract matrix does not prove
No matrix boundary proves browser App HTML or artifact-rebuild replay. In-memory runs and compiled
artifacts without event routes report runtime identity as honestly not-applicable.
The packed boundaries cannot load project route modules — source may be deleted and verified
absent — so serialized-round-trip, compat-probe, and version-skew checks are reported
not-applicable with a stated reason rather than silently skipped. host-install reports
module-backed checks the same way, because loading project modules would cross back into the
source and build tree.
When an advertised input schema declares additionalProperties: false, plain z.object tool
routes may still strip unknown keys without a protocol failure. The negative-inputs check records
that tolerance when other generated negatives still prove rejection paths.
Host validation
Neither Codex nor Cursor publishes a plugin-validation developer verb, so their emitted documents are checked against vendored pinned schemas, and host schema generators are treated as a drift signal rather than a substitute contract. See Artifact validation.
Rendering
Raw HTML, JSX/MDX, and Mermaid in Skill Markdown are inert in the Workbench renderer.