Skip to content

CHANGELOG.md

Latest commit

 

History

History
66 lines (57 loc) · 3.24 KB

CHANGELOG.md

File metadata and controls

66 lines (57 loc) · 3.24 KB
Raw

Changelog

All notable changes to orbiter are documented here. Format follows Keep a Changelog; versioning follows Semantic Versioning.

Unreleased

0.1.0 — 2026-05-07

First public cut, prepared for the Xiaomi MiMo 100T Token Creator Incentive Program. Everything below ships with deterministic mock mode, so the entire pipeline is exercisable without a credential.

Added

  • 5-agent pipeline. Detective → Reproducer → Surgeon ⇄ Verifier (closed loop, max 3 iterations) → Reviewer.
  • MiMo V2.5 client (orbiter.llm.MiMoClient) over the OpenAI-compatible endpoint, covering the reasoner, multimodal, and TTS modalities; with tenacity retries, a per-run token-budget cap, and a one-shot JSON-repair pass on ValidationError.
  • Strict cross-agent contracts. 13 Pydantic v2 models in orbiter.schemas, all with extra="forbid", decoded via response_format=json_schema, strict=True.
  • Atomic RunState persistence under ~/.orbiter/<run_id>.json (tmp + rename); recovery via orbiter recover [<id>].
  • Sandbox verification. PytestSandbox snapshots the working tree, applies the Surgeon's hunks, runs the Reproducer's regression test, then re-runs the project's existing suite.
  • Patch hygiene. RepoView.apply_patch rejects any hunk whose original doesn't match the target file exactly once; refuses paths outside the repo root.
  • Typer CLI. orbiter run, recover, list, show, doctor.
  • Multimodal artefacts. Each accepted patch ships a markdown PR body, a 1280×720 PNG diagram (MiMo multimodal in live mode, PIL placeholder in mock mode), and a 60–90 s narrated WAV/MP3 (MiMo TTS in live mode, silent WAV in mock mode).
  • Demo fixture. examples/buggy_repo/ with two planted bugs (timing-attack in auth.py, bare-except in worker.py) plus four baseline tests that must remain green after patching.
  • Quality gate. 29 pytest cases (24 unit + 1 e2e + 4 CLI smoke), ruff clean, mypy --strict clean on Python 3.11 and 3.12.
  • CI. .github/workflows/ci.yml runs lint + types + tests + an end-to-end mock-mode smoke; .github/workflows/orbiter-nightly.yml self-scans this repo every night, automatically choosing live or mock mode based on whether ORBITER_MIMO_API_KEY is configured.
  • Demo recording. scripts/demo.sh + scripts/record-demo.sh produce docs/img/demo.{cast,gif,mp4} and docs/img/hero.png in one shot, used as the README hero animation and as the 02_demo.mp4 proof artefact for the grant form.
  • Docs. README.md, docs/architecture.md (deep-dive), docs/grant-submission.md (long-form rationale), docs/SUBMISSION_PASTE.md (1:1 form paste sheet), docs/demo-script.md (recording recipe).

Known limitations

  • Only Python projects are supported. TypeScript is on the roadmap.
  • Live mode requires an OpenAI-compatible MiMo endpoint; offline models behind a non-OpenAI shim aren't supported yet.
  • Reviewer runs sequentially per finding; for very large find lists this could be parallelised.