Trial Runner

Trial Runner

Use this skill to build or operate an eval suite around:

agent-eval-harness eval '<json>'

Required workflow

1. Use grill-me first to clarify:
   • evaluation goal and reliability target
   • target agent CLI(s)
   • task corpus source and metadata
   • grading evidence and pass policy
   • comparison strategy and calibration cadence
2. Use tdd to build in vertical slices:
   • one task JSONL row
   • one adapter contract test
   • one run smoke test
   • one grader
   • one grade smoke test
   • one compare check
   • one calibrate check
3. Use schema discovery for contracts (do this instead of copying field lists):
   • eval --schema task
   • eval --schema adapter-input
   • eval --schema adapter-output
   • eval --schema grader-input
   • eval --schema grader-output
   • eval --schema trial-row

Canonical pipeline

agent-eval-harness eval '{"mode":"run",...}' > raw.jsonl
agent-eval-harness eval '{"mode":"grade",...}' < raw.jsonl > graded.jsonl
agent-eval-harness eval '{"mode":"compare",...}'
agent-eval-harness eval '{"mode":"calibrate",...}'

run and grade stream compact trial_result JSONL rows to stdout.
compare and calibrate emit one bounded JSON object.

Suite-owned concerns (outside harness contract)

• workspace setup and cleanup
• resume/dedupe policy
• retry policy
• filtering and slicing strategy
• CI gating policy
• cost/token analysis scripts
• secrets via environment/wrappers (not JSON config)