@executioncontrolprotocol/harnesses-browser-coding

Browser Coding harness (@executioncontrolprotocol/harness-browser-coding) for coding models (Ollama qwen2.5-coder:1.5b small profile; Anthropic Sonnet medium / Opus frontier).

• TypeScript-first: Fluent workflows and typed EcpIntent / HarnessReply modules (no EQL).
• Chat + eval matrix — browser demo can bind this harness (e.g. with Ollama or Anthropic); eval matrix covers intent/authoring/assistant/chat.
• Profiles: small keeps compact plain inventory + heavier few-shots. medium / frontier use a Fluent capability catalog, thin grammar few-shots, generate messages[] for conversation history, and repair-as-turns (failed TypeScript as an assistant turn).

Generate contract

Providers implement @executioncontrolprotocol/model.generate with:

• prompt — current user turn (required)
• system — optional
• messages — optional prior { role: "user" | "assistant", content } turns
• files — attach to the current turn

The host (demo) passes conversationMessages on chat invokes; the harness forwards them on authoring/assistant shots. New chats omit messages.

Fixtures and evals

| Asset | Location | | ----- | -------- | | Prompt fixtures | fixtures/harness-prompts/*.prompt.json | | Eval cases | fixtures/eval-cases/*.cases.json | | Support fixtures | fixtures/workflows/, fixtures/runs/ | | Matrix tests | test/eval/ |

Run matrix from repo root: pnpm run test:eval:matrix:coding
Anthropic / Sonnet: pnpm run test:eval:matrix:coding:anthropic

For medium/frontier context flow (system prompts, messages[], Fluent catalog, repair turns), see docs/sonnet-coding-interaction-loop.md.

import { registerBrowserCodingHarnesses, BROWSER_CODING_HARNESS_CAPABILITY } from "@executioncontrolprotocol/harnesses-browser-coding"