Added

  • The gateway distills web pages in tool results (ADR-89). A raw HTML page that reaches the model in a tool result (curl output, an MCP fetch, a browser dump) becomes compact text first: title, headings, paragraphs, lists, table rows, code verbatim, [text](url) links and image alt text, without scripts, styles, navigation or hidden elements. Real pages shrink 84–97% in a few milliseconds. It applies on every harness and route (Anthropic tool_result, OpenAI/ChatGPT/Muse function_call_output, Gemini functionResponse), never to what you type or what the model wrote, and only to pages. Files the agent reads, commands that fetch no URL, and source such as JSX, templates, XML, SVG, diffs or code pass untouched. A page still over `gateway.compress.max…
  • context_expand reads long text in pages. Pass offset (and optionally limit, at most 100,000 chars) to read on from where a reply stopped. Each partial reply names the next offset.

Changed

  • The size cap now covers Codex-style tool outputs. Oversized OpenAI/ChatGPT/Muse function_call_output and custom_tool_call_output items are capped head and tail like Anthropic tool results, with the original kept for context_expand.

Update HarnessLink and every installed harness to its current release:

hnl update

New machine? Install with:

curl -fsSL https://harnesslink.sh/install.sh | bash

Newer: 17.15.1 · Older: 17.14.0 · All releases · Atom feed

Flag notifications