model-autofill

English | 中文

What it does

Auto-fills missing model metadata in the llm-pi-ai settings namespace ($DSH_HOME/settings.yaml, written by the web Models page). Whenever a model entry is missing any of these fields, they are filled in:

• display name (name) — from the pi-ai built-in model catalog
• context window (contextWindow)
• max output tokens (maxTokens)
• reasoning efforts (reasoningEfforts) — from a built-in rules table (new in v0.2)

Only missing fields are filled — values you set are never overwritten. It fires whenever the model configuration changes and also runs one idempotent pass at startup. Model ids with a provider/ prefix or an OpenRouter-style :suffix (e.g. openai/gpt-5.6-luna, xiaomi/mimo-v2.5-pro:free) are matched correctly.

The reasoningEfforts rules table

Reasoning capability is not discoverable the way catalog fields are — discoverModels returns only id/name/contextWindow/maxTokens. So the completion uses a declarative rules policy: model ids are matched against an ordered rule table, and the first matching rule supplies the value to write. The per-family effort sets mirror the locally installed pi-ai provider catalog (node_modules/@earendil-works/pi-ai/dist/providers/data/*.json — each model's own thinkingLevelMap), cross-checked against public docs:

Key semantics (from the pi-ai implementation):

• Valueless off: offers an Off level that sends nothing on the wire

(thinking: {type: "disabled"} in the anthropic dialect).

• minimal folds to wire low: the anthropic-messages

mapThinkingLevelToEffort collapses minimal→"low" (see pi-ai/dist/api/anthropic-messages.js), so every table writes minimal: "low" instead of emitting a wire value that does not exist.

• Undeclared levels are pinned unsupported, so each table IS the full offer

for its family; a level declared but not actually supported (e.g. kimi-k3's medium) fails at request time when selected.

• :free is not "non-reasoning": xiaomi/mimo-v2.5-pro:free is the same

model on a free tier, so the blanket :free → false rule is gone.

> ⚠️ Your gateway (knnns, anthropic-messages dialect) translates effort names > into the underlying model's own vocabulary; whether a level is accepted is > gateway/model-dependent. Declaring efforts tells the gateway "this model can > think". If a particular model does not actually accept the thinking parameter > (ignored or erroring), change its rule to efforts: false or delete that > entry's reasoningEfforts from settings.yaml. The completion writes into the > user settings layer, so everything it writes stays editable/removable — no > lock-in.

You can replace the whole table via the plugin config on the model-autofill row in your profile's cordis.patch.yml / cordis.yml:

- id: model-autofill
  name: "@birdman1992/model-autofill"
  config:
    reasoning:
      rules:
        - match: ["claude-", "claude/"]
          efforts: { off: null, minimal: low, low: low, medium: medium, high: high, xhigh: xhigh, max: max }
        - match: ["deepseek-", "deepseek/"]
          efforts: { off: null, high: high, max: max }

> Declaring efforts tells the gateway "this model can think". If a particular > model does not actually accept the thinking parameter (ignored or erroring), > change its rule to efforts: false or delete that entry's > reasoningEfforts from settings.yaml. The completion writes into the user > settings layer, so everything it writes stays editable/removable — no lock-in.

Install

Published to npm — install directly:

dsh plugin --profile web add @birdman1992/model-autofill

Or install directly from GitHub:

dsh plugin --profile web add github:birdman1992/dsh-birdman-plugins#path:plugins/model-autofill

Restart the profile afterwards:

dsh --profile web

> If you previously installed this plugin via a link: dependency, first run > dsh plugin --profile web remove @birdman1992/model-autofill and delete the > manually inserted model-autofill row from cordis.patch.yml before running > the install command above, to avoid inserting the same row id twice.