Made on
Tilda
Querying the knowledge base from an agent session connected as described in Connecting a workstation, and how each call shows up on the MCP Live screen.
knowledge/note/<id>.The agent picks the tool. If you need a specific source, name the tool in your prompt.
| Task | Tool |
|---|---|
| Start working on an unfamiliar file or module | brief_me_on |
| A question about the project with no known place in the code | ask_question (returns matching locations, not a ready answer) |
| List entries on a topic | list_knowledge |
| Full entry | get_knowledge |
| Conventions before editing a file | check_conventions_before_editing; the full rulebook: list_conventions |
| Entity usages and dependencies | find_usages, get_dependency_graph |
| Search code by name or by meaning | grep_code, search_code |
| Tracker tickets | find_tickets_for, get_ticket |
| Quality status | get_quality_dashboard, get_riskiest_problems |
| Save a finding to the knowledge base | store_knowledge, create_convention |
The line AI Council: sent the project rulebook in full to the agent (N chars, off-screen) means the agent received the full rulebook. It is printed on the agent's first action that isn't a call to Council. Calls to Council during a response are collapsed into the line Called ai-council N times. The text of conventions and entries isn't printed to the terminal: it is meant for the model.
Until the session has called the knowledge base, the first code read is rejected. The agent only needs to call a knowledge base tool. Only the user can turn the check off, with the /council-without-kb command in their own session. The agent can't use this command.
Covers the whole installation. Tabs: Sessions shows the slowest tools and the list of sessions with an event feed; Effectiveness shows the share of sessions that reached each stage; Configuration shows the MCP server profile and tool list; Tool Stats shows calls by client and by tool. The text of user prompts isn't stored, only its length.