PatchPilot v0.4.0 — local-first preview and feedback thread

[harshitethic]

PatchPilot v0.4.0 is available as a local-first AI coding-agent preview.

It inspects a repository, produces a plan, applies guarded source edits to an isolated branch, runs permitted tests inside a fail-closed Docker container (when the approved offline image is available), and returns a Git diff for review.

What was independently checked: 116 backend regressions, Python 3.11/3.12 CI, a real Docker isolation workflow, Python CodeQL, and five controlled browser-state tests. The Git helper now rejects output above 4 MiB instead of silently displaying a partial diff, and cleans up its POSIX process group on timeout.

• Release and changelog
• Local setup guide
• Security boundaries

Feedback requested

If you try it, I'd especially like feedback on three things:

1. First-run friction: OS, Python version, Ollama model, Docker availability, and which setup step was confusing.
2. Agent accuracy: a small reproducible public repository and task where its plan or diff was unhelpful.
3. Isolation behavior: whether blocked test commands and unavailable Docker are explained clearly without falling back to the host.

Please leave an issue for reproducible bugs or reply here with an experience report. No stars or promotion needed—specific technical feedback is the goal.

Scope note: This is a preview for a trusted local workstation, not an audited multi-tenant sandbox. Repository cloning and trusted Git plumbing run on the host. Tests may fail closed when Docker or the approved image is unavailable. Browser scenarios use fixtures, not an independently evaluated AI model.