Review (toktrack override)

Follows the global /review multi-agent pattern, with the following overrides:

Override: UX Review Inactive

This project is a terminal TUI app — do not run the UX Review Agent. Run Code Review Agent only.

Override: Code Review Checklist Extension

In addition to the global checklist, append the following to the Code Review Agent prompt:

Critical (PLAN gate)

Category	Items
PLAN assumption refutation	Does the diff falsify the [agent-inferred]/[unverified-gate] assumptions from the PLAN (falsified → P1+, carry to wrap)

Author Review Gate (P0 — Identity / Matrix)

This repo is OSS and the reviewer is the author (receiving PRs). Generic code health is covered by Rust/TUI/clippy below; this section examines product-specific risks that only the author can validate. Perspective SSOT: session memory feedback_review_perspective.md.

Tier	Category	Items
reject	Identity ① ultra-fast	Adding latency·blocking IO·synchronous network to hot path/startup → reject
reject	Identity ② persistent cache / data preservation	Risk of cache·history loss or corruption → reject. backward-compat·graceful degradation are the means to uphold this
ruthless	Irreversible	breaking change · data migration · public behavior change
ruthless	Environment/state matrix (contributors test only their own machine)	OS(paths·home·file locks·Windows) / source schema version(forward new-schema+backward old-format, new source·parser PRs require old-version·unknown-schema fixture) / data shape(empty·large·corrupt·partial·permissions) / locale·TZ·DST / concurrency(read race while source file is being written) / terminal(width·color·TERM) / first-run vs upgrade migration
bias to YES	Other features·ideas	Stars are the goal — acceptance bias. But if it taxes either core(weighing down hot path·startup·cache paths), request location·approach adjustment only. additive/opt-in/lazy/behind a flag = welcome
lenient	Style·naming preferences	Linter takes precedence, bikeshed prohibited (uniform strictness = contributor friction → star loss)
signal	Direction/growth fit (not a gate)	Beyond code quality·conventions, every contribution(issue·PR) also gets a direction-fit verdict in the review output: strengthens wedge (persistent cache/data preservation) / neutral (additive) / dilutes focus (e.g. feature-parity chasing) + 1–2 lines of rationale. Weigh holistically: maintenance cost vs value, real value to target users(AI CLI heavy users), contribution to project direction. Never a reject reason — the two identities above remain the only gates, bias-to-YES stands

Rust-specific

Category	Items
Safety	Minimize unsafe usage, comment with justification
Ownership	Unnecessary .clone(), to_string(), to_owned()
Error	anyhow/thiserror pattern consistency, no unwrap() (tests excluded)
Performance	Unnecessary allocation, Vec vs iterator chain, Box<dyn> vs generic
SIMD	Check fallback branch in simd-json parsing path
Concurrency	Check shared mutable state in rayon parallel path

TUI-specific

Category	Items
Widget	ratatui Widget trait implementation consistency
Theme	Use theme.rs semantic colors (no hardcoded colors)
Layout	Handle terminal resize (Rect boundary check)
Input	Missing keyboard event handling (shortcuts registered in help vs actual handlers)

Clippy/Fmt Pre-check

Code Review Agent checks the following before review:

cargo fmt --check
cargo clippy --all-targets --all-features -- -D warnings

If there are clippy warnings, FAIL before starting review (should have been caught by verify).

Execution

1. Collect context (diff, conventions, architecture, Sprint Contract)
   • PLAN provenance tags/gates (if present) — [agent-inferred] / [unverified-gate]
2. Launch Code Review Agent only (feature-dev:code-reviewer)
   • Global agents/code-review.md prompt + above author gates + Rust/TUI checklist append
3. Parse verdict → PASS → /wrap, FAIL → fix → /verify → re-review

Rules

• Sub-agent model: launch the Code Review Agent via the Agent tool with model: "sonnet" (Sonnet 5).
• Do not run UX Review Agent (TUI project)
• PASS → run /wrap immediately
• FAIL → fix → /verify → re-review (max 3)