Project direction

oy gives coding agents repeatable audit → review → fix workflows through portable Agent Skills. Any agent that reads .agents/skills can run them; oy itself is not a second agent or model runtime.

Mission

Make repository review in coding agents visible and durable:

prepare known input
  → the agent analyzes it under the user's permissions
  → oy validates a report
  → fix one finding and rerun

Product boundary

Owner	Responsibilities
oy CLI	Repository/diff collection, ordering, limits, evidence identity, report validation, Markdown/SARIF output, skill installation
oy skills	Three workflow skills, one setup skill, and the oy persona
the agent and user	Models, providers, credentials, tools, sessions, UI, and project instructions

Principles

Do not add a parallel permission policy. The skills run under whatever permissions the user’s agent has; they never broaden them.
Inputs can be repeatable; conclusions cannot. Evidence and report normalization are deterministic, model reasoning is not.
Fail instead of silently sampling. Changed evidence, malformed reports, and explicit limits are visible errors.
Reports are handoff artifacts. Stable IDs and reruns matter more than chat-only output.
Keep host coupling narrow. The only agent-specific code is the optional OpenCode location refresh after setup.
Keep one useful persona. The oy persona emphasizes inspection, small changes, verification, and worktree safety without defining permissions.

Current product

Setup writes:

oy-audit, oy-review, and oy-enhance skills;
the oy-setup skill and the oy-persona.md it installs into the agent’s environment (improving the default agent or creating an oy agent).

The Rust CLI prepares evidence, verifies model-written candidates, normalizes finding metadata, writes Markdown/SARIF, installs the skills, and migrates legacy OpenCode plugin state.

Non-goals

becoming a second coding-agent or model runtime;
owning provider credentials or model routing;
adding permission overrides;
adding general shell, edit, web, clone, or search tools;
claiming deterministic security or quality conclusions;
running paid model evaluations in default CI.

See Architecture for implementation boundaries and ROADMAP.md for current work.