/**
* Centralized description strings for the Mandoline MCP integration.
*/
export const SCHEMA_DESCRIPTIONS = {
chatMessage: {
role:
"Role of the speaker: 'user', 'assistant', or 'system'. Must match the " +
'original conversation exactly.',
content:
'Verbatim text of the message. No summarization or truncation—this is ' +
'the ground-truth transcript line.',
},
contextAsset: {
name:
'Human-friendly handle or path of the resource the assistant READ ' +
"(e.g., 'utils/pager.py').",
kind: 'One of: code, doc, data, config, log, other.',
content:
'Full text the assistant consumed BEFORE generating its reply. ' +
'Include original versions of files that will later be modified so ' +
'reviewers can compare old vs new. You may truncate as long as a ' +
'reviewer can still reconstruct intent—show first/last sections and ' +
"insert an ellipsis marker like '[…]'.",
description: 'Single-line purpose summary (helps quick scans).',
metadata:
'Optional key-value info (language, framework, line-count, etc.). ' +
'If the file has a language, set metadata.language.',
},
outputAsset: {
name:
'Handle or path of the artifact the assistant WROTE or MODIFIED in ' +
"this turn (e.g., '.cursorignore', 'README.md').",
kind: 'Same taxonomy as contextAsset.',
content:
'Full text produced by the assistant. Include ALL generated artifacts ' +
'—code, docs, tests, configs—so evaluators can see the delta. Follow ' +
'the same truncation rules as contextAsset.',
description: 'One-line why-it-matters summary.',
metadata:
'Optional key-value info (language, purpose, etc.). Populate ' +
'metadata.language for syntax-highlighting when possible.',
},
toolInvocation: {
name: "Name of the tool invoked (e.g., 'web_fetch').",
arguments: 'Raw JSON string passed to the tool (exact, unformatted).',
result: 'Raw text or JSON returned by the tool (optional).',
},
prompt: {
root:
'Complete prompt context including conversation history, assets read, ' +
'and tool calls executed before the assistant reply being evaluated. ' +
'Must be provided as an object.',
messages:
'COMPLETE conversation history up to—but NOT including—the assistant ' +
'reply being evaluated. Include BOTH user and assistant turns in ' +
'chronological order.',
context_assets:
'Every ContextAsset the assistant read while producing the reply. ' +
'Omit only if genuinely none were consulted.',
invocations:
'All tool calls executed BEFORE the assistant reply (chronological).',
},
response: {
root:
"The assistant's complete response being evaluated, including the " +
'message content, generated assets, and any tool calls made. ' +
'Must be provided as an object.',
message:
"The assistant's reply for THIS turn (the thing being scored). " +
"Must have role 'assistant'.",
output_assets:
'Array of OutputAssets generated in this reply. Include updated docs ' +
'as well as code/config files.',
invocations:
'Tool calls executed DURING the assistant reply, if any. Leave empty ' +
'if the reply had no tool invocations.',
},
} as const;
export const TOOL_DESCRIPTIONS = {
create_metric: 'Creates a new evaluation metric in Mandoline.',
batch_create_metrics: 'Creates multiple metrics in one call.',
get_metric: 'Fetches a metric by its ID.',
get_metrics: 'Lists metrics with filtering and pagination.',
update_metric: "Modifies an existing metric's metadata.",
create_evaluation:
'Evaluates a single prompt/response pair and returns a score (-1 … +1).',
batch_create_evaluations:
'Evaluates the same pair against multiple metrics in one request.',
get_evaluation: 'Retrieves one evaluation by ID.',
get_evaluations: 'Lists evaluations with pagination and filtering.',
update_evaluation: 'Updates evaluation metadata (score is immutable).',
get_server_health: 'Check the server health status.',
} as const;
export const SHARED_PARAM_DESCRIPTIONS = {
skip: 'Number of results to skip (default 0).',
limit:
'Maximum results to return (default 100, max 1000). Large payloads can be expensive.',
include_content:
'When false, omits heavy prompt/response text (default true).',
filters: "Server-side field-value filters, e.g., { 'name': 'foo' }.",
properties:
"Optional evaluation metadata (e.g., { branch: 'feature/x', commit: '<sha>' }).",
} as const;
export const METRIC_PARAM_DESCRIPTIONS = {
name: 'Descriptive label for the metric.',
description: 'Detailed explanation of what the metric measures.',
tags: "Optional labels for organization (e.g., ['security']).",
metric_id: 'Unique ID of the metric to fetch or modify.',
metric_ids: 'Array of metric IDs for batch evaluation.',
metrics: 'Array of metric definitions for bulk creation.',
} as const;
export const EVALUATION_PARAM_DESCRIPTIONS = {
model_name:
'Identifier of the model that generated the response. Only provide if certain.',
evaluation_id: 'Unique ID of an existing evaluation.',
} as const;
export const PARAM_DESCRIPTIONS = {
...SHARED_PARAM_DESCRIPTIONS,
...METRIC_PARAM_DESCRIPTIONS,
...EVALUATION_PARAM_DESCRIPTIONS,
} as const;
