[Moonbot-Tech]

Station can now serve an AI client independently of Telegram: the owner enables HTTPS from Terminal, approves one OAuth connection for selected cores and permissions, and the AI reads live facts or submits explicit, observable operations through ActiveLib.

Behavior

• Add 17 MCP tools for core state, both hedge positions, strategies and their schema, markets, settings, reports, bounded diagnostic log capture, MoonDoc knowledge, operations and cursor-based events.
• Mutations use a short-lived preview, revalidation on the connection owner, deduplicated apply and honest operation results. SDK order guards survive the runtime queue; strategy resolution carries immutable desired/actual snapshots. Settings reuse the existing full-settings writer, preserve untouched float bits/opaque fields and retain compact-write/manual-order priority.
• Add OAuth authorization-code flow with S256 PKCE, resource binding, rotating refresh tokens and persistent revocation. Owner consent is available only over the existing private administrative channel. Public client registration alone grants nothing.
• Add one-button Caddy setup with automatically renewed trusted certificates for a public IP or existing hostname; no MoonBot DNS service or user-managed certificate is required. Caddy has a separate service identity and no access to Station credentials.
• Add the AI access page in all seven locales and a revision-pinned, locally searched MoonDoc cache that refreshes every six hours.

Dependencies and boundaries

• SDK changes: feat(automation): expose typed settings and guarded action receipts MoonProto#18 (pinned public commit fd743049). No Delphi kernel or wire-protocol changes.
• Knowledge guides: docs(station): explain AI commands and bounded log diagnosis MoonDoc#1. Station reads public main, so the new guide sections become available after that PR is merged and the cache refreshes.
• A send receipt is not an exchange fill. There is no server-side settings CAS or multi-core transaction, and uncertain trades are never retried automatically. MCP events require a running/polling AI host; they do not wake a sleeping chat.

Validation

• Windows: 9,286 workspace tests passed, 211 ignored; workspace/all-target Clippy with warnings denied, rustfmt and Tour check passed using public locked dependencies.
• Linux: Station build, 89 unit tests, 9 integration contracts and Clippy passed using the public SDK pin; SDK suite passed 1,280 tests.
• Real test VPS: trusted public HTTPS, private owner consent, PKCE, the 17-tool catalog, read tools, scope/core denial, refresh/revocation, timed logs, MoonDoc reads and disabled-strategy create/edit/delete passed. The exact production setup call also recovered a stopped Station service.
• A real settings patch changed only the sound-alert flag, confirmed it from a fresh full core snapshot, then restored and confirmed the original value. The final HTTPS smoke passed after this fix. No live trading commands were sent; MCP grants and the temporary test administrator were removed after testing.
• Visual checks covered light/graphite at narrow width and dark/Obsidian at narrow and wide widths, loading/disabled state and real client/grant data. Full keyboard-only consent and onboarding in individual AI provider apps remain unverified.
• Two independent clean-context reviews covered security and command semantics; their findings were addressed.
• CI: https://github.com/Moonbot-Tech/MoonTerminal/actions/runs/42317383121. macOS diagnostic: https://github.com/Moonbot-Tech/MoonTerminal/actions/runs/42317383036.