oad/review

Review the diff (staged, branch vs main, or a named PR). Every finding cites a file and line; a verdict without evidence is not a review.

1. Scope

git status --short
git diff --stat origin/main...HEAD    # or: git diff --cached --stat, or gh pr diff <n>

If the diff mixes unrelated concerns or runs past ~300 lines of mixed refactor and behavior change, say so and suggest a split before reviewing line by line.

2. Tests first

Tests show intent. For each behavior change: is there a test, does it assert behavior rather than internals, and would it fail without the change?

• Rust: #[cfg(test)] modules beside the change and appv3/crates/<crate>/tests/.
• Web: web/src/__tests__/ at the mirrored path (components/Foo.tsx → __tests__/components/Foo.test.tsx).
• Scripts and workflows: scripts/tests/.

Missing tests for non-trivial behavior is a required finding; point at oad/testing.

3. The five axes

Correctness — Does it do what was asked? Empty, null, boundary, and error paths? State that can drift between the web stores, SSE stream, and DB? Data as it really arrives (CRLF from multipart forms, rows written by v2 installs)?

Readability — Names and patterns match the nearest AGENTS.md? Dead code, leftover shims, unused imports, comments that restate the code?

Architecture — Route handlers stay thin (behavior in the owning crate)? TanStack Query for server state, Zustand for client state? Platform-specific behavior gated through the platform hooks? No second helper where a canonical one exists (safe_resolve / safe_join, validate_workspace, focusQuietly, utils/file-refs)?

Security — Treat as sensitive: auth and the desktop token, the Host/Origin network guard, workspace paths from external or model input, shell and file tools, MCP launch config, Tauri CSP and capabilities, keyring, updater signing. Secrets compare in constant time (auth::constant_time_eq); subprocesses take argument lists. When the diff touches any of these, run the security-review skill if it is available.

Performance — N+1 or unbounded DB queries, loops over whole session history, missing pagination, large SSE payloads, React re-renders from unselected store reads or inline object props, work on every streamed token.

Contracts — An SSE or wire change matches appv3/crates/agent/data/sse_events.json, updates the web client in the same change, and is recorded in appv3/REPORT.md. Release versions move only through scripts/bump_version.sh.

4. Classify findings

Lead with correctness and security, then architecture, then the rest.

5. Check the verification story

• Which gates ran: make verify-v3, make verify-web, make verify-docs, native targets (oad/testing)?
• UI changes: checked at a narrow (≤768 px) and a wide viewport, touch and pointer, against DESIGN.md?
• User-visible change: entry in documents/docs/features.md?

6. Output

## Review: <scope>

### Critical
### Required
### Consider / Nit / FYI
### Done well
### Verdict: Approve | Request changes

Approve when the change improves the code base overall, even if imperfect; do not block on taste when it follows project convention.