Repository navigation
Russian authorization searches and small spelling mistakes now reach the owning workflows. Agents can discover the actual library before querying: list domains, choose a real topic key, select the requested operation and load the returned workflow ID.
- 45 domain keys and 162 topic keys are backed by all 241 existing workflow IDs. Exact topic lookup uses validated membership and never guesses an identifier.
- Subject tags rank owning capabilities ahead of incidental mentions. Registered-vocabulary typo repair accepts one unambiguous edit or adjacent transposition and protects exact IDs, short technical names and explicit filters.
- Vague reports such as a misspelled backend-glitch query route to Backend Probe, which carries the authorized repair and verification context into Patch Checking.
- The four-language console offers guided discovery and recovery from empty text results. Native/offline search shares the same vocabulary and ranking.
- The service adds no embedding model, new API key or inference-provider query transfer. Previous tool names, workflow IDs and legacy result shapes remain compatible.
Validation: 72 Python checks and 29 Node/search/real-HTTP checks pass. The frozen 71-query set still meets its accepted ranks; every catalog key and supported mode round-trips to actual workflow members in four languages. Browser checks cover four languages at 320 and 1280 pixels, guided selection, empty-result recovery, five positive and three negative protocol examples. Both Claude manifests validate.
The included 84-second video is an edited walkthrough of actual compiled Worker UI responses. It does not claim autonomous model selection or execution of a user project.
The Library ZIP is the thin MCP plugin; the separate Vibe Coding ZIP contains all offline skills.