Skip to content

Changelog

0.1.2 (2026-09-11)

Added

  • Real LLM agent providers with function calling: OpenAIProvider (planner loop over injectable client, refusal-first, max-steps budget) and LangGraphProvider (StateGraph plan/execute pipeline with native fallback engine). Factory: create_agent("openai" | "langgraph"); POST /agent/run gains a provider field. Offline-testable via fake clients; llm extra installs openai + langgraph.
  • Real mAP evaluation: evaluate_model(mock=False) runs ultralytics val and reports mAP50/mAP50-95/precision/recall; CLI model evaluate --real and new model benchmark command (warmup/repeats FPS + latency stats) via vision/benchmark.py.
  • Quality pass: CLI in-process CliRunner tests (22), rule-based agent tests (20), trainer real-path tests, vision re-export tests — coverage 77% → 93% with CI gate at 85%; actions upgraded to checkout@v7 / setup-python@v7 / codeql-action@v4 / artifact actions latest (Node 20 deprecation fix); advisory pip-audit CI job with report artifact; docs/api.md endpoint reference with mkdocs nav entry.
  • Release engineering: docker workflow now pushes the API image to GHCR (master → latest + short SHA; tags → vX.Y.Z); release workflow builds desktop artifacts for linux, windows, and macos and attaches all available zips to the GitHub Release.

Changed

  • Replaced placeholder default API host with 127.0.0.1 so hokage-vision api starts without flags; container CMD uses 0.0.0.0.
  • Fixed dataset-manifest agent test leaking data/manifests/local.yaml into the repo root.
  • CLI generate_report tool now passes keyword arguments (was positional and raised TypeError).

0.1.1 (2026-09-10)

Added

  • Closed-loop training demo: scripts/closed_loop_demo.py generates a synthetic dataset, fine-tunes yolov8n on CPU, exports ONNX, and runs real inference through UltralyticsBackend (24/24 val accuracy; see models/model-card.synthetic-shapes.md).
  • Quickstart demo: python examples/quickstart.py walks detect → validate → smoke-train → agent → report in one command (see docs/quickstart.md).
  • Real ONNX inference integration tests, video pipeline tests (synthetic MP4), 10 API endpoint tests, GUI interaction tests, and 38 new unit tests (39 → 86+ total).
  • Demo video with real trained-model overlay: examples/videos/demo.mp4.
  • Coverage gate with XML artifact upload in CI.

Changed

  • Removed vendored legacy YOLOv5 toolchain (legacy/, 177 files); provenance preserved in git history and documented in docs/license-audit.md, docs/migration.md, THIRD_PARTY_NOTICES.md, and LICENSES/README.md.
  • Fixed pre-existing lint warnings in scripts/capture_real_shots.py and scripts/generate_evidence.py; ruff check is now clean across the repo.
  • Fixed environment-dependent assertion in tests/unit/test_core_paths.py that failed in Docker CI.

Added (previous)

  • Added Docker-first project structure and dependency-layer caching.
  • Added mock backend, shared inference service, rendering, CLI, API, GUI, Agent tools, dataset validation, annotation assistance, smoke training, model registry, evaluation, and comparison foundations.
  • Added package build, desktop build, CI, GUI tests, docs, release, CodeQL, Dependabot, and repository governance files.
  • Added license audit, third-party notices, and data/model distribution guardrails.