4543ac3c1a
Motiviert durch die Mega-Debrid-Cooldown-Diagnose: der account-weite Cooldown musste aus Log-Arithmetik rekonstruiert werden (alle Fehlerzeilen zeigten ms-genau auf dieselbe Deadline). Genau diese Sicht-Luecke wird geschlossen. Neu: - GET /providers (debrid.ts getProviderRuntimeSnapshot): Live-Laufzeitzustand pro Mega-Account / Debrid-Link-Key — AKTIVER Cooldown (untilMs/remainingMs/Grund/ Kategorie/untilRestart), in-flight-Tiefe, Mega-Rotationscursor + Sticky-Count, Empty-Response-Streaks, Debrid-Link Key-/Host-Cooldowns + Runtime-Status. Account- Keys sind nicht-umkehrbare Hashes (mda_<fnv1a64>), keine Logins/Tokens. - GET /logs/conversion: der pro-Item Link-Aufloesungs-Lebenszyklus (Token, API getLink, Web, Account-Rotation, Abbrueche mit Zeiten) aus conversion.log. - Beide auch im /diagnostics-Aggregat (providers + logs.conversion) -> rd_diagnostics zeigt die Cooldowns jetzt direkt. - MCP-Bridge: neues Tool rd_providers + "conversion" in der rd_logs-Enum + README. Test: getProviderRuntimeSnapshot spiegelt einen geprimten Account-Cooldown (until/remaining/Grund) und ist vorher null. Suite 935 gruen, tsc 6.
3.0 KiB
rd-diagnostics-mcp
Standalone stdio MCP bridge to the Real-Debrid-Downloader debug-server. It runs on the machine where the
MCP client runs, takes a connection code for a downloader server, and exposes that server's
read-only HTTP diagnostics API (/diagnostics, /status, /errors, /logs/*, /accounts, …) as MCP tools.
One bridge serves all 5–6 servers; you pass a code (or a configured server name) per call.
This bridge is not bundled into the Electron app and adds no dependencies to it.
Setup
cd tools/rd-diagnostics-mcp
npm install
Register it with your MCP client as a stdio server that launches the bridge:
node "<repo>/tools/rd-diagnostics-mcp/src/bridge.mjs"
Provide servers via environment variables (codes contain a token — treat like passwords):
RDDIAG_CODE— a single default connection code (rddiag:v1:...)RDDIAG_SERVERS— JSON map of name → code, e.g.{"berlin":"rddiag:v1:...","fra":"rddiag:v1:..."}
Without env config, every tool simply takes a code argument.
Tools
rd_servers, rd_ping, rd_diagnostics, rd_status, rd_items, rd_packages, rd_errors, rd_logs
(main|audit|rename|trace|session|conversion|package|item), rd_history, rd_accounts, rd_providers
(live per-account/key cooldown + in-flight + rotation state), rd_host, rd_self_check,
rd_get (raw escape-hatch, any read-only path).
Each tool accepts code or server to pick the target.
Connection code
Format: rddiag:v1:<base64url(JSON)> with { v:1, h:host, p:port, t:token, n?:name, fp?:certFingerprint, s?:scheme }.
Generated by the app (Hilfe → Remote-Support → Ferndiagnose (MCP)) or via node src/gen-code.mjs --host H --port P --token T.
Security model (read before exposing a server)
- The debug surface is read-only for state/logs; the one control endpoint is
/trace/config(toggles the optional, time-bounded support trace). No persistent secrets are written into the logs it serves: passwords are redacted, debrid API keys/tokens and resolved download URLs are never logged;/settingsand/accountsare redacted. - Auth is a bearer token (24 random bytes). Over plain HTTP on a public network the token is sniffable, so:
- Preferred: keep the server bound to
127.0.0.1("Nur lokal") and reach it through a private tunnel (Tailscale / SSH / Cloudflare Tunnel). The tunnel encrypts and authenticates; no public exposure. - Direct network bind (
0.0.0.0) requires a non-empty IP allowlist (enforced fail-closed: with an empty allowlist only loopback is accepted). Use only inside a trusted LAN/VPN.
- Preferred: keep the server bound to
- Revoke instantly from the app ("Token neu" or "Deaktivieren") — the old code stops working immediately.
fppins a self-signed cert fingerprint and is verified onsecureConnect(before the token is sent). HTTPS is not the v1 default; plain HTTP behind a tunnel is the recommended transport.
Test
npm test # spins a fake debug-server, runs the bridge as a stdio child, asserts the full protocol path