szhygulin/recon-crypto-mcp

Rolls up post-0.5.1 security and verification-UX changes.

Highlights

• Auto-run preview checks — agent-side ABI decode + pair-consistency pre-sign hash recomputation now run unprompted at preview_send and report in a CHECKS PERFORMED block (no more user-consent menu per check). PRs #57, #59.
• WalletConnect session-topic cross-check — get_ledger_status returns the WC topic; agent surfaces last 8 chars and asks the user to verify a matching session in Ledger Live → Settings → Connected Apps. Catches peer impersonation that the self-reported name/URL alone can't. PR #61.
• Clear-sign vs blind-sign UX — on-device check reminders now distinguish decoded-field checks (Aave / Lido / 1inch / LiFi / approve) from hash-match checks (everything else).
• Swiss-knife fallback — decoder URL surfaces as a Markdown hyperlink inside the CHECKS PERFORMED block when the agent's built-in ABI knowledge is low-confidence. PRs #59, #62.
• Plain-English trust pitch — new "Why trust VaultPilot?" section at top of SECURITY.md, mirrored briefly in README's Security model.
• Security-critical error hand-holding — SECURITY: prefix + explicit "do NOT retry this handle — re-prepare from scratch" guidance on payload-hash mismatch (EVM + TRON), previewToken mismatch, TRON device-address mismatch, and unexpected-signature-shape errors.
• Agent-prompt notation fix — {✓|✗|⚠} placeholders instead of […] so literal Markdown brackets and inline-code backticks survive the agent's paraphrase. PR #63.
• Integrator string — LiFi integrator set to "vaultpilot-mcp". PR #60.
• README demo gif — added near the top. PR #64.

Publishing

Published to npm with provenance via GitHub Actions (trusted publisher / OIDC) and synced to the MCP Registry.