Agent Skills: Trial Adapters

Build command-only adapters for `eval` run mode using strict stdin/stdout JSON contracts.

UncategorizedID: plaited/agent-eval-harness/trial-adapters

Install this agent skill to your local

pnpm dlx add-skill https://github.com/behavioral-sh/agent-eval-harness/tree/HEAD/skills/trial-adapters

Skill Files

Browse the full folder contents for trial-adapters.

Download Skill

Loading file tree…

skills/trial-adapters/SKILL.md

Skill Metadata

Name
trial-adapters
Description
Build command-only adapters for `eval` run mode using strict stdin/stdout JSON contracts.

Trial Adapters

This skill is for executable adapters used by:

agent-eval-harness eval '{"mode":"run",...}'

Required approach

  1. Use grill-me to pin down:
    • target coding-agent CLI and invocation
    • docs/MCP source for real command behavior
    • approval/automation flags
    • prompt/session mechanics
    • trajectory capture strategy
    • failure mapping and safe metadata
  2. If command semantics are unclear, ask the user for docs/MCP access or run --help/docs discovery first.
  3. Use tdd for adapter executable contract tests.

Adapter executable contract

  • Read one adapter-input JSON object from stdin.
  • Write exactly one adapter-output JSON object to stdout.
  • Write logs only to stderr.
  • Always include trajectory (can be []).
  • For completed status, include non-empty final message.

Use schema discovery as source of truth:

agent-eval-harness eval --schema adapter-input
agent-eval-harness eval --schema adapter-output

Contract tests to write first

  1. valid stdin -> valid stdout object
  2. exactly one stdout JSON object
  3. stderr-only logging
  4. schema-valid trajectory
  5. multi-prompt behavior (prompts array handling)
  6. explicit failure output behavior

Safe optional metadata examples

  • model/provider name
  • provider run id
  • adapter name/version
  • agent CLI version
  • token usage
  • cost estimate
  • safe config summary
  • suite tags