[rtiman-spotify]

Summary

Draft implementation of SLMD-1717: /confidence:onboard-data in the Confidence plugin.

• Add the CLI command, canonical skill and references for domain-grouped review, optional goals, 20-table paging, bounded revision-specific approval and partial completion.
• Define status/resume, selected retry, mapping resolution, inclusion of excluded tables, dependency-checked replacement planning and read-only refresh.
• Register confidence-data for supported plugin hosts and update routing/documentation. Authentication, API composition and validation polling remain in MCP.
• Add 14 synthetic decision-evaluation cases, paired with-skill/baseline runs, deterministic regression tests and a PR workflow for tests/typechecking.
• Fix a pre-existing missing URLSearchParams declaration in the Metric Explorer scorer that blocked repository typechecking; include a regression test.

Validation

• npm run typecheck
• npm run test:onboard-data — 63 tests passed
• git diff --check
• Independent technical/safety review; fixed installed-plugin path resolution and successful zero-evaluator exits, with regression tests
• Model-quality evaluation — not passing: with skill 2/14 (14.29%), baseline 5/14 (35.71%)
• Live MCP/local-progress integration and real CLI acceptance journey

The model runs are synthetic single-turn action-plan evaluations, not executed warehouse operations. Failures include action-vocabulary/state-convention mismatches and recovery/decision-preservation checks; they need investigation. Unit test success does not establish acceptable model behavior or end-to-end safety.

Why this is a draft / remaining work

Not ready to merge or release.

• Integrate the delivered data MCP contracts (SLMD-1921 and related discovery/import work) and durable local progress (SLMD-1931). The skill currently fails closed when required interfaces are missing; it does not implement an ad hoc journal or fabricate tool signatures.
• Keep context import gated on implemented ownership/storage/read-only agent access (SLMD-1858).
• Verify authoritative agent availability separately from registration/ACTIVE.
• Diagnose model/harness mismatches and genuine semantic failures; improve and rerun the paired evaluations without dropping safety assertions.
• Run the real CLI acceptance scenario: three available tables, one permission failure, one unresolved mapping and one exclusion, plus recovery and the 300-table review matrix.

No real warehouse imports, permission changes, analytical queries or remote data mutations were performed. No successful end-to-end recording exists; this CLI-only draft does not claim browser/video acceptance.

refresh is defined conservatively as read-only rediscovery and review, preserving previous decisions. Its V1 versus P1 launch priority remains undecided.