Skip to contents

reproducibleai 2026.07.25.9000

Agentic-routing evaluation

  • Added versioned competency-question fixtures with private routing and answer rubrics.
  • Added deterministic competency derivation from maintained dev/ sections, mandatory human review, provenance hashes, and externally frozen benchmarks.
  • Added an external review bundle with an editable question CSV, explicit generator exclusions, generated QA instructions, and provenance-checked import.
  • Added canonical-answer token precision, recall, and F1 for generated retrieval and grounding questions.
  • Added a numbered review-run-interpret article workflow and the first durable self-evaluation health report.
  • Hardened live execution with codex exec interface preflight, a supported structured-output schema, canary guidance, and execution-aware recommendations.
  • Added explicitly approved, repeated, ephemeral, read-only codex exec evaluation with JSONL event capture and schema-constrained responses.
  • Added transparent routing, answer, completion, usage, and variability metrics.
  • Added aggregate Markdown health reports that omit raw prompts, rubrics, and traces from the evaluated repository.
  • Added contamination controls that keep gold fixtures and raw runs outside the repository under evaluation.
  • Added no-usage prerequisite diagnostics and robust discovery of the user-local standalone Codex CLI on Windows.
  • Clarified that deterministic package capabilities require neither Codex nor a cloud account, while live evaluation remains optional and subject to local application, identity, and network policy.

reproducibleai 2026.07.24.9000

Agentic-context standard 0.1

  • Replaced instruction modules, recipes, CHAT_INSTRUCTIONS.md, and session transcript tooling with a durable AGENTS.md-routed repository standard.
  • Added deterministic base and R-package scaffolds with a version manifest and seed hashes.
  • Added read-only profile detection and migration planning.
  • Added explicitly approved migration application that validates replacements before removing tracked, unchanged legacy files.
  • Added structured validation suitable for interactive use, tests, and CI.
  • Allowed use_sbom() to run non-interactively so its deterministic scaffold can be exercised in tests and automation.
  • Removed the superseded use_instructions(), instructions_available(), instructions_recipes(), extract_copilot_chat(), and cleanup_session_backups() interfaces without compatibility wrappers.