|
| 1 | +# Changelog |
| 2 | + |
| 3 | +All notable changes to SkillOpt are documented here. This project adheres to |
| 4 | +[Semantic Versioning](https://semver.org/) and the format is based on |
| 5 | +[Keep a Changelog](https://keepachangelog.com/). |
| 6 | + |
| 7 | +## [0.2.0] — 2026-07-02 |
| 8 | + |
| 9 | +The headline of this release is **SkillOpt-Sleep**: a nightly offline |
| 10 | +self-evolution engine that harvests a coding agent's real session |
| 11 | +transcripts, mines recurring tasks, replays them offline, and consolidates |
| 12 | +short-term experience into long-term memory and skills — all behind the same |
| 13 | +held-out validation gate that keeps SkillOpt training honest. It ships as a |
| 14 | +decoupled top-level package (`skillopt_sleep/`, zero dependency on the |
| 15 | +research code) and as the new `skillopt-sleep` CLI. |
| 16 | + |
| 17 | +### Added |
| 18 | +- **SkillOpt-Sleep engine** — nightly offline self-evolution cycle |
| 19 | + (harvest → mine → replay → consolidate) behind a validation gate, exposed |
| 20 | + as the `skillopt-sleep` console script and `python -m skillopt_sleep`. |
| 21 | + - Multi-objective reward (accuracy / tokens / latency) with user preferences. |
| 22 | + - Multi-rollout contrastive reflection under a token/time budget. |
| 23 | + - Experience replay + controllable dream rollouts (opt-in). |
| 24 | + - Slow-update long-term memory field (runs even with the gate off). |
| 25 | + - 3-way train/val/test split with `gate_mode on|off`. |
| 26 | + - Verifier-discipline validation gate, with a stress-test suite |
| 27 | + (thanks @Tanmay9223, #87). |
| 28 | +- **Cross-tool backends & plugin shells** for Claude Code, Codex, Copilot, |
| 29 | + Devin, and OpenClaw: |
| 30 | + - Codex Desktop transcript harvesting, skill-first Codex integration, and a |
| 31 | + reviewed task-file flow (thanks @Kirchberg, #48, #49, #60). |
| 32 | + - GitHub Copilot backend (`CopilotCliBackend`) + research-engine MCP plugin |
| 33 | + (thanks @Dongbumlee, #50). |
| 34 | + - Devin plugin: MCP server + ATIF-v1.7 harvest (thanks @xerxes-y, #88). |
| 35 | + - OpenClaw shell for SkillOpt-Sleep (thanks @Elzlxx, #59). |
| 36 | +- **SearchQA** split materialization helper and fail-fast on systemic rollout |
| 37 | + failures, with a `searchqa` install extra (thanks @summerview1997, |
| 38 | + #63, #64, #65). |
| 39 | +- WebUI environment loading and backend preflight (thanks @summerview1997, #63). |
| 40 | + |
| 41 | +### Changed |
| 42 | +- Decoupled the Sleep engine into a standalone top-level `skillopt_sleep/` |
| 43 | + package with zero dependency on the research code. |
| 44 | +- Made `EnvAdapter.reflect` a shared default so reflect kwargs are no longer |
| 45 | + dropped (thanks @imshunsuke, #44). |
| 46 | +- English-only pass across the engine, plugins, and docs. |
| 47 | + |
| 48 | +### Fixed |
| 49 | +- Windows robustness for the Claude/Codex backends, plus a hardened JSON |
| 50 | + fallback path (thanks @Yif-Yang, #79). |
| 51 | +- Reject prose pseudo-JSON wrapped in single quotes/backticks (#82). |
| 52 | +- Surface Codex auth/model/version failures instead of silently scoring 0 |
| 53 | + (thanks @dmmdea, #92). |
| 54 | +- Redact secrets before persisting cycle diagnostics. |
| 55 | +- Configure the `qwen_chat`/`minimax` backends so local LLM endpoints work |
| 56 | + (thanks @imrehg, #85). |
| 57 | +- Forward the Qwen target timeout and gate `enable_thinking` for vLLM targets |
| 58 | + (thanks @mvanhorn, #40). |
| 59 | +- Make `--bare` conditional on `ANTHROPIC_API_KEY` (#68), add a |
| 60 | + `SKILLOPT_SLEEP_PYTHON` override with a lookback-hours first-run fallback |
| 61 | + (#74), and fix ALFWorld gamefile paths relative to `ALFWORLD_DATA`. |
| 62 | + |
| 63 | +### Packaging |
| 64 | +- Bump `skillopt`, `skillopt.__version__`, and `skillopt_sleep.__version__` |
| 65 | + to `0.2.0`. |
| 66 | +- Restore `skillopt_webui` to the built wheel (it was dropped when the |
| 67 | + `packages.find` include list was made explicit). |
| 68 | +- Add the `searchqa` extra and include `json_repair` in the `claude`, `qwen`, |
| 69 | + and `all` extras. |
| 70 | + |
| 71 | +### Acknowledgements 🙏 |
| 72 | +v0.2.0 landed thanks to our community contributors — thank you! |
| 73 | + |
| 74 | +- @Kirchberg — Codex Desktop harvesting, skill-first Codex integration, |
| 75 | + reviewed task-file flow (#48, #49, #60) |
| 76 | +- @Dongbumlee — GitHub Copilot backend + research-engine MCP plugin (#50) |
| 77 | +- @summerview1997 — SearchQA materialization, rollout fail-fast, WebUI |
| 78 | + preflight (#63, #64, #65) |
| 79 | +- @xerxes-y — Devin plugin: MCP server + ATIF-v1.7 harvest (#88) |
| 80 | +- @Elzlxx — OpenClaw shell for SkillOpt-Sleep (#59) |
| 81 | +- @imshunsuke — shared `EnvAdapter.reflect` default + docs fixes (#43, #44) |
| 82 | +- @mvanhorn — Qwen timeout forwarding + `enable_thinking` gating (#40) |
| 83 | +- @dmmdea — surface Codex auth/model/version failures (#92) |
| 84 | +- @Tanmay9223 — verifier-discipline stress test (#87) |
| 85 | +- @imrehg — `configure_qwen_chat` for local LLM endpoints (#85) |
| 86 | +- @samuelgoofus-boop — community contributions |
| 87 | + |
| 88 | +Special thanks to @Yif-Yang for driving the SkillOpt-Sleep engine. |
| 89 | + |
| 90 | +**Full changelog:** https://github.com/microsoft/SkillOpt/compare/v0.1.0...v0.2.0 |
| 91 | + |
| 92 | +## [0.1.0] — 2026-06-02 |
| 93 | + |
| 94 | +Initial public release: the full training loop (rollout → reflect → |
| 95 | +aggregate → select → update → evaluate), multi-backend support |
| 96 | +(OpenAI / Azure / Claude / Qwen / MiniMax), six built-in benchmarks, and the |
| 97 | +WebUI dashboard. |
| 98 | + |
| 99 | +[0.2.0]: https://github.com/microsoft/SkillOpt/releases/tag/v0.2.0 |
| 100 | +[0.1.0]: https://github.com/microsoft/SkillOpt/releases/tag/v0.1.0 |
0 commit comments