plugin.yaml

name: understudy
version: "1.0.0"
manifest_version: 2
api_version: 1
description: "Audition cheaper models on your real Hermes history before you switch: replays past decisions on candidate models, judges each against what your current model did, and reports takeover rate, misses and projected cost."
author: stffinfcti
license: MIT
homepage: https://github.com/stffinfcti/understudy
requires_hermes: ">=0.22.6"
tags: [models, cost, evaluation]
provides_tools: [understudy_audition, understudy_report]
provides_hooks: []
capabilities: [llm.model_override, llm.provider_override]
config_schema:
  sample: {type: int, default: 20, description: "Past decisions replayed per candidate"}
  days: {type: int, default: 14, description: "How many days back to sample conversations"}
  budget_usd: {type: float, default: 1.5, description: "Hard spending cap per audition, in USD"}
  concurrency: {type: int, default: 6, description: "Parallel model calls"}
  baseline_sample: {type: int, default: 8, description: "Decisions your current model re-answers, to measure its own consistency and cost"}
  judge_model: {type: str, default: "", description: "provider:model for the judge; empty uses your current model"}
  include_persona: {type: bool, default: true, description: "Send an excerpt of the agent's instructions with each replay"}
  exclude_sources: {type: list, default: [], description: "Session sources to skip, e.g. telegram"}
  prices: {type: dict, default: {}, description: "USD per 1M tokens: 'provider:model' or 'model' -> {input, output, cache_read?}"}