The workflow is grounded in the public usage tool reference and the deployed MCP catalog schema. Live read-only validation covered project defaults, explicit project filters, team scope, single-user filters, time buckets, and activity offsets. Disjoint project totals and per-user totals reconciled to team consumption and every model's CU, API-key CU, and job counts.
\nThe verified accounting contract uses consumption value and model point cost, both already after discounts. Raw consumption total includes discounts. Inclusive timestamp bounds replace the draft's incorrect exclusive-end assumption. Zero-discount live samples do not prove discount arithmetic; synthetic nonzero-discount and negative fixtures exercise that documented contract. Version 2 invalidates the draft's cache format.
The catalog fallback matters when a connected host retains an older direct tool signature after deployment. It stays within MCP. User-filtered usage replaces the draft's job-history workaround and supports both CU and jobs without inferring a joint distribution from marginal totals.
\nAll live customer snapshots and reports stay outside this public repository. Fixtures contain fictional identifiers and values. Host fallbacks are tested as document-only plans, not live certification of every CLI or desktop product.
\nThe committed validation evidence records the live application result and provides screenshots, a PDF, a dashboard, and CSV exports rendered from a reproducible fictional dataset. Public examples contain no live customer data.
\nThe baseline planner without the skill lacked executable discovery, cache, and joint-ranking procedures. Application testing gives a fresh agent only this skill, the connection skill, and a task, then grades the exact call plan against the fresh public tool schema. The deployed-contract plan passed scope, stale-schema routing, inclusive dates, discount accounting, complete user matrices, cache invalidation, and desktop fallback checks. Re-run after substantive contract changes.
\npython3 -B -m unittest discover -s tests/scenario-admin-analytics exercises the offline helper, accounting, cache isolation, scope/date validation, user coverage, exports, PDF creation, and dashboard filters. Install the suite's requirements to include PDF checks. Run pnpm test and pnpm validate before shipping. Render every PDF page for visual inspection separately from structural checks.