Create a scoring run for the current judge over a dataset's existing outputs (wraps runs_create with prompt_id omitted and output_column supplied). This only sets up the run; call runs_generate to actually re-judge the outputs so you can compare against human verdicts.
| Name | Required | Description | Default |
|---|---|---|---|
| name | Yes | ||
| metric_id | Yes | ||
| dataset_id | Yes | ||
| judge_model | Yes | ||
| output_column | No | Dataset column with the existing outputs to grade. Defaults to actual_output. |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description carries the full burden. It does disclose a key behavioral trait: 'This only sets up the run; call runs_generate to actually re-judge.' This is valuable. Yet it omits other behavioral details such as side effects, return values, or prerequisites (e.g., whether the judge must already exist). The transparency is adequate but not thorough.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is composed of two tight sentences. The first sentence delivers the core purpose and wrapper details; the second clarifies the workflow by referencing runs_generate. No unnecessary words or redundant content.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
The tool has 5 parameters, no annotations, and no output schema. The description outlines the setup-vs-generate flow and the intended comparison scenario, but it lacks details on what the run creation returns, how the judge is selected, or any error/where it fits in the broader workflow. It is serviceable but incomplete for full contextual understanding.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is only 20% (only output_column has a schema description). The description additionally explains that output_column is supplied and prompt_id is omitted, which adds some meaning. However, it does not clarify the semantics of name, metric_id, dataset_id, or judge_model, which are all required. Given the low coverage, the description should compensate more but does not.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool's function: 'Create a scoring run for the current judge over a dataset's existing outputs.' It also specifies that it wraps runs_create with a particular configuration, distinguishing it from the general runs_create sibling. This gives a specific verb, resource, and scope.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description explicitly instructs to 'call runs_generate to actually re-judge the outputs,' which guides the user on the next step and differentiates when to use this tool versus runs_generate. It also implies the use case of comparing against human verdicts. However, it does not explicitly exclude or contrast with other sibling tools like runs_regrade, so it's not a perfect 5.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Add one secure layer between your agents and this server.
Each tool targets a distinct resource and action, with clear separation across agreements, datasets, judges, metrics, prompts, runs, tags, and usage. Even similar tools like datasets_create vs datasets_create_from_url and runs_generate vs runs_rerun are explicitly differentiated in their descriptions.
The overwhelming majority of tools follow a consistent plural_resource_action snake_case pattern (e.g., datasets_create, metrics_update, runs_retry_failures). The only slight deviation is promptfoo_import, but it is still descriptive and does not break the overall predictability.
With 54 tools, the server far exceeds the 25+ threshold considered too many, and approaches the 50+ extreme mismatch level. Even for a broad LLM evaluation platform, this count is excessive and likely to overwhelm agents, making tool selection more error-prone.
The toolset provides full CRUD for core resources (datasets, metrics, prompts, runs, tags) plus lifecycle operations like publish, generate, regrade, and retry. It also includes cross-cutting utilities (usage, import, provider credentials). Minor gaps exist, such as no update/delete for agreements and no cross-run response search, but these are non-essential for the primary workflows.
