nightshift/docs/devlog/phase18.md

792 B

Phase 18 Devlog: Prompt and Pipeline Experiments

Implemented

  • Added optional experiment.label and experiment.prompt_variant config fields.
  • Snapshotted agent prompt files into runs/<run-id>/prompts/.
  • Wrote run-metadata.md with project, experiment, agent backend, model, command, and prompt metadata.
  • Included experiment metadata in final task reports and run summaries.
  • Added tests for experiment config loading and prompt/metadata artifact creation.

Decisions Made

  • Experiment metadata is descriptive only and does not alter execution semantics.
  • Prompt snapshots are per-run, not per-task, because agent definitions are run-level configuration.

Notes

  • This creates enough metadata to compare prompt/backend runs from artifacts without adding a database.