Configuration

The minimal setup is one key:

export OPENROUTER_API_KEY=sk-or-...

Everything else has a default. Override the row's config to change it — every field is documented in cordis.yml:

- id: jev
  name: 'dsh-jev'
  config:
    backend: typesafe            # openrouter (default) | typesafe
    apiKeyEnv: TYPESAFE_API_KEY  # credential reference (an environment variable name)
    model: jev-latest            # omit to use the per-backend default
    timeoutMs: 20000
    budgetMs: 45000
    maxRetries: 1

Field	Default	Meaning
`backend`	`openrouter`	`openrouter` \| `typesafe`
`apiKeyEnv`	per backend	Credential reference; `openrouter` → `OPENROUTER_API_KEY`, `typesafe` → `TYPESAFE_API_KEY`
`apiKey`	—	Literal key (`role('secret')`). Discouraged; prefer `apiKeyEnv`
`baseURL`	per backend	Override for a self-hosted gateway or proxy
`model`	per backend	`openrouter` → `jev-1.13`; `typesafe` → `jev-latest`
`timeoutMs` / `budgetMs`	20000 / 45000	Per-attempt timeout / whole-call budget (including retry backoff)
`maxRetries`	1	Retries after the first attempt
`maxQuestions`	64	Maximum questions per call
`maxStateChars`	96000	Character limit on the serialized `state`
`maxInstructionChars`	8000	Character limit on one instruction
`appName` / `appUrl`	`dsh-jev` / —	OpenRouter attribution headers `X-Title` / `HTTP-Referer`
`logRequests`	true	Structured log line per evaluation

A settings write is rejected when timeoutMs > budgetMs or baseURL is not a valid URL, so the settings document and the running plugin never disagree.

Model ids across the two backends

	OpenRouter	TypeSafe direct
Alias	`jev-latest` → `~typesafe/jev-latest`	`jev-latest`
Version	`jev-1.13` → `typesafe/jev-1.13`	`jev-1.13.0`
Pinned	`typesafe/jev-1.13-20260917`	same style
Extra fields	response carries `id`, `provider`, `usage.cost`	none

Usage guidance

Ask atomic questions. Instead of "rate this pitch", ask about market size, technical feasibility, and differentiation separately, then weight them with your own formula — changing a coefficient then means changing code, not a prompt.
Batch questions. Questions in one request are evaluated independently and in parallel, so asking many at once barely costs more latency.
Set your own thresholds. The probabilities exist so your code can branch; a choice answer being the highest-probability option does not make it certain — look at confidence.
Never present a Noul as a risk percentage. It is the probability of a yes/no statement.

dsh-jev

同名パッケージの別リポジトリ

インストール

ドキュメント

Configuration

Model ids across the two backends

Usage guidance