Project evolution: from vendored YOLO repo to agentic workbench

This page tells the engineering story behind the repo — useful if you are evaluating the project as a portfolio piece and want to know what was actually rebuilt, measured, and deleted along the way.

All numbers below come from the git history itself (75 commits, 2024-05 → 2026-09), so they can be re-verified with git log.

Phase 0 — the inherited starting point (2024)

The repo began life as a fork of a yolov5-PySide6 hobby project:

a full vendored copy of the Ultralytics YOLOv5 toolchain under legacy/old_project/ (177 tracked files, 13.8 MB) carrying GPL-3.0 obligations and 6 upstream TODOs,
no tests, no CI, no packaging, and a README with placeholder text.

Phase 1 — governance first (2026 early commits)

Before writing features, the project paid down license and hygiene debt:

vendored-code provenance audited and documented (license audit),
Apache-2.0 chosen for new code, GPL/AGPL boundaries documented in THIRD_PARTY_NOTICES.md,
hard CI gates, dependabot, docker health checks, bilingual READMEs.

Why first? A detection demo is worthless as a portfolio piece if its legal posture is "unknown GPL blob". Cleaning the license story made everything after it defensible.

Phase 2 — rebuild the skeleton (the 17 feature commits)

The new src/hokage_vision package was built one vertical slice at a time, each commit a working product increment:

config → detection models → mock backend → inference service → CLI
→ agent tool registry → data/validation → smoke training → model registry
→ ultralytics + legacy backends → PySide6 GUI → FastAPI → packaging → CI

Design rule enforced throughout: the default path needs no weights, no GPU, no network (deterministic MockBackend), so every layer — CLI, API, GUI, agent — is testable in CI from a fresh clone.

Phase 3 — the 13.8 MB deletion (commit `ddc2e3a`)

A read-only dependency audit proved nothing in src/ imported the vendored legacy/old_project tree: the only linkage was one explicit boundary class (YOLOv5LegacyBackend) loading it through torch.hub behind an opt-in backend name. The deletion removed 177 files / 24,111 lines in one commit, with docs rewritten to point at git history for provenance.

Deleting a vendored fork — with a documented audit trail instead of a silent purge — is the kind of decision portfolio reviewers rarely see, and it is fully reproducible: git show ddc2e3a --stat.

Phase 4 — closing the "does it actually work" gap

Early 2026 reviews rated the repo great skeleton, thin proof. The response was evidence, not arguments:

Gap	Fix	Commit
39 tests for 2548 LOC	160+ tests, 93% coverage, 85% CI gate	`dd41ed2`, `60ba0b8`
Zero real inference	bundled ONNX model + real mAP (0.995)	`f76769b`, `719da43`
No training loop proof	in-repo synthetic-data closed loop	`f76769b`
Agent was 4-line placeholder	OpenAI + LangGraph function-calling providers	`74d51f5`
Evaluation returned `None`	real ultralytics `val` integration + FPS benchmark	`a1721ad`

Phase 5 — shipping (v0.1.1 → v0.1.2)

Release engineering completed the story: GHCR image push, three-OS desktop packages, changelog-driven release notes, and two tagged releases with every CI workflow green.

What this history demonstrates

Audit before delete. 24k lines removed on the strength of a documented dependency audit, not vibes.
Mock-first is a testability strategy, not laziness — the same interfaces run mock in CI and real ONNX locally.
Every feature landed with its test; coverage went from ~40% to 93% with the gate raised behind the actual level.
Boring, verifiable claims — every metric in this page maps to a commit you can git show.