Changelog

1.2.0 (2026-10-13)

• Personal-data and secrets guard in eval-models: scans the system prompt and every example before any call. Secrets are always blocked; personal data (Aadhaar with Verhoeff check, PAN, payment cards with Luhn check, email, UPI ID, Indian mobile) blocks the run unless --allow-personal-data is passed. Outputs are not saved when personal data is allowed unless --keep-outputs is passed. Decisions are recorded in results.json and summary.md; values are never printed or stored by the guard.
• Router pii_policy (block / warn / off, default warn) via make_routing.py --pii-policy.
• DATA-PROTECTION.md: data-flow and role analysis per component, DPDP Act obligation mapping, required Cloud Run design, open items for counsel.
• 4 new tests (20 total).

1.1.0 (2026-10-13)

• New skill eval-models: run your own examples against Claude, Gemini, OpenAI, or any OpenAI-compatible endpoint (including local Ollama/vLLM); reports pass rate, latency, tokens, and cost; cached and reasoning tokens handled; unknown costs never shown as zero; dry run and consent before paid calls.
• New skill route-models: cheapest-first routing from eval evidence; standalone router escalates on errors and failed guards; LiteLLM config with error-based fallbacks.
• New skill size-local-model: weights plus KV-cache estimate (MHA/GQA), concurrency, KV precision, headroom.
• find-model: host-neutral wording (works in IDEs and chat clients), coding-subtask guidance, read-only repo signals with privacy rules.
• Packaging for Claude Code, Codex, Cursor, Gemini CLI, and Copilot CLI from one repository.
• Test suite with a local mock of the three provider APIs.

1.0.3

• Preparation release with the find-model skill.