Skip to content

The cross-harness, self-improving memory layer for AI agents. Join our community: https://discord.com/invite/zdNChS2fBu

1.1 KB
Raw Blame

Claude Code payload fixtures

These are PostToolUse hook inputs captured from Claude Code 2.1.291, saved from the hook's stdin by a project-level hook that wrote its input to a file. The session ran in beacon-oracle's sealed container against its scripted mock model and fixture MCP server (scenario d19-presence-media-blocks), so the content is synthetic but the shape is exactly what Claude Code sends.

Covered payloads:

  • post_tool_use_mcp_media.json — an MCP tool whose server returned text, an image, audio, a text resource, a binary resource and a resource link. Claude Code rewrites the result before the hook sees it: the image arrives as an Anthropic image block with its base64 under source.data, and every other block arrives as text. Audio and the binary resource become notes naming the file Claude Code saved them to.
  • post_tool_use_mcp_empty.json — an MCP tool whose server returned no content blocks, which arrives as "tool_response": [].

The tokens shaped BCN-… are beacon-oracle canaries, not secrets. Replace a fixture with a fresh capture when a Claude Code release changes the shape.