v0.3.0 - Grounded AI completion and cost analysis

Highlights

• Turn large Microsoft 365 Copilot usage reports into bounded activity counts and inactive-user pages, with report dates and explicit coverage.
• Calculate storage, backup, VM and run-rate estimates using deterministic decimal arithmetic, alongside the existing token-cost calculator.
• Get usable, like-for-like regional pricing rankings with explicit currency, volume-band and pagination evidence.
• Deliver real owner-bound CSV, XLSX, HTML and script downloads instead of internal filesystem links.
• Keep Azure changes reviewable: exact application approval is required, and destructive cleanup remains a user-run script.

Added

• GetCopilotUsage for licensed-user reports, activity selection, complete counts and bounded paging.
• CalculateCost for explicit quantities, rates, billing units, currency, tax, discounts and annualization.
• A maintainer /investigate-ai-sessions workflow for reviewing 100 distinct retained conversations without confusing HTTP success with task fulfillment.
• Durable execution outcomes, evidence-qualified scheduled runs and richer compute/Spot diagnostic evidence.

Fixed

• Chart calls using the legacy chart key and documented optional inputs that were incorrectly required.
• Unsupported Graph query options and incorrectly scoped reservation-utilization requests.
• Pricing output that ignored requested top-N rankings or omitted storage-volume thresholds.
• Missing SDK tool-failure accounting and silent genuinely empty model results.
• Incomplete final chat replies, Cost Management retry feedback and detailed-cost routing.
• File projection, conversation-bound upload continuity and generated-artifact delivery.

Changed

• Financial guidance preserves source currencies, units, dates and cohorts; estimates remain distinct from billed data.
• Follow-up guidance prioritizes finishing the requested task and avoids duplicate Crawl/public follow-up calls.
• Updated README, tool catalog and reliability documentation describe all 38 tools and their limitations.
• New deployments default to the configured Luna model; deployment availability and model-specific quota must still be verified.

Security

• Upgrade the separate telemetry collector to 0.161.0 to address the fixable AMQP vulnerabilities found by the production image preflight, without relaxing the vulnerability gate.
• Custom-tool-only execution, credential screening, owner-bound downloads and exact approved-write dispatch remain enforced.
• Generated HTML previews stay sandboxed. User consent and Azure RBAC remain the authorization boundary.

Validation and limitations

Credential-free backend, Python, frontend and desktop/mobile regressions cover the changed contracts. Deterministic tests do not certify every future model answer, external consent, current pricing or resource availability. Historical message-only audits are explicitly distinguished from full tool evidence.

Full changelog: v0.2.0...v0.3.0