@felan-ai/ext-output-style

Portable output-style instructions for Felan sessions.

The extension appends one clearly bounded ## Output Style section to the system prompt before each model run. It provides the built-in concise, and explanatory styles plus a custom style for caller-provided instructions. concise is the default. It minimizes prose while preserving clarity, exact technical content, conditions, caveats, verification, and blockers, and it expands for ambiguity, safety-sensitive actions, errors, and complex plans. createOutputStyleExtension accepts instruction text only. Hosts may resolve extensionConfig.outputStyle.instructionsFile to that text before activation.

import { createOutputStyleExtension } from '@felan-ai/ext-output-style';

const extension = createOutputStyleExtension('explanatory');
const concise = createOutputStyleExtension('concise');
const experiment = createOutputStyleExtension('custom', 'Answer in one compact paragraph.');

Felan declares the style as extensionConfig.outputStyle.style. Custom instructions use exactly one of extensionConfig.outputStyle.instructions or extensionConfig.outputStyle.instructionsFile:

{
  "extensionConfig": {
    "outputStyle": {
      "style": "custom",
      "instructions": "Answer in one compact paragraph without omitting blockers."
    }
  }
}

{
  "extensionConfig": {
    "outputStyle": {
      "style": "custom",
      "instructionsFile": "output-style.md"
    }
  }
}

The local TUI, CLI, and /settings resolve these fields before activating the extension for root and child sessions. A custom style requires non-empty inline instructions or a host-loaded instruction file. This package does not read the file; the host passes the contents to createOutputStyleExtension.

Savings measurement

Only the built-in concise style contributes a savings measurement. For each successful assistant turn containing visible text, the extension estimates the actual output as UTF-8 bytes divided by four and rounded up. It estimates the disabled-style baseline as:

ceil(actual visible-output estimate * 4187 / 3500)

This uses the current repeated benchmark's pooled output-token result: 4,187 baseline tokens versus 3,500 concise tokens, a 16.4% reduction. The producer only sees UTF-8 visible text, so this is a documented byte-to-token proxy rather than an observed tokenizer count. The method identifier is concise-output-token-ratio-202609-v1.

This measurement isolates the visible-output boundary and is not a claim of whole-workflow savings; the benchmark's prompt tokens increased. Empty, errored, aborted, or unattributed turns do not report. explanatory and custom have no supported baseline and never contribute savings.

Package boundary and requirements

The package owns the supported style names, built-in instructions, custom-source validation, and prompt-section formatting. A host owns style selection, built-in enablement, instruction-file loading, and session lifecycle. The package requires a compatible @felan-ai/agent-core peer and does not read settings, prompt files, or ambient resources itself.

Development

Source: packages/ext-output-style in https://github.com/felan-ai/felan.

corepack enable
pnpm install --frozen-lockfile
pnpm --filter @felan-ai/ext-output-style build
pnpm --filter @felan-ai/ext-output-style type-check
pnpm --filter @felan-ai/ext-output-style test

Related documentation

• Configuration
• Extension catalog
• Architecture
• Efficient execution and savings

Attribution

The extension implementation is original Felan project code. The merged concise instructions adapt wording from Julius Brussee's MIT-licensed Caveman skill. See NOTICE and LICENSE.