tdqs / flags
Flags
Flags mark categorical defects; smells mark dimensions scoring below 3. A flag is published beside the score, and only two of them touch the arithmetic.
- No Description — tool flag · Stage 2, hard gate
- The description is null or whitespace-only. There is nothing to evaluate: an agent selecting between tools sees only a name.
- Tautological Description — tool flag · Stage 2, evaluated up front and applied in stage 4
- The description, lowercased and trimmed, is exactly the tool name or title. A description that restates the name adds zero information, and models grading "looks reasonable" text would otherwise drift upward.
- Annotation Contradiction — tool flag · Stage 3, the LLM contradiction check
- The description contradicts the declared MCP annotations — a description that "creates a record" on a tool annotated readOnlyHint: true, for instance. An agent that trusts either signal is misled by the other.
- Shadowing Risk — server flag · Server coherence evaluation, from the invocation-cost prefilter
- A tool whose purpose is substantially covered by a sibling that is materially cheaper to invoke. Both can score well alone; an agent that can reach the answer either way tends toward the cheaper call, and when the dearer tool is the authoritative one the agent presents an answer it assembled itself. A deterministic prefilter proposes at most one candidate pair per tool and the coherence evaluation confirms whether the purposes genuinely overlap.