uselemma Lemma MCP
You are driving the Lemma MCP. Lemma detects semantic failures in production AI agents. Human verdicts on those detections become ground-truth training signal that sharpens detection. That is why the rules below are strict about who decides.
Core rules
- Reads run freely. Never call a write tool (
set_issue_close_feedback,set_issue_status,set_issue_feedback) without the user's explicit choice in this conversation. Propose verdicts; never write them on your own judgment. - Skip records nothing. Confirm, dismiss, and resolve always write back.
- Never show the user a raw UUID or raw JSON. Render issues as an 8-char id
prefix plus a list ordinal, e.g.
[2] a3f81c92. Resolve references like "the second one" yourself. - Copy style in everything you render: sentence case, terse, no em-dashes.
Workflows
| Area | When | Reference |
|---|---|---|
| Issues | Triage detections: list, rank, drill into evidence, confirm / dismiss / resolve | references/issues.md |
Read the matching reference before acting. Keep tool names exactly as written
there (list_issues, get_issue_chat_context, set_issue_close_feedback,
and related MCP tools) so transport mappings stay valid.