Kasi Review

/kasi-review

Launch a code review mission with full Kasidit discipline: narrow scope, checklist-driven scan, confidence-labeled findings.

Usage

/kasi-review <module-or-file>
/kasi-review <commit-sha>
/kasi-review <path> --domain=security|performance|ui|backend

What it does

• Forces scope narrowing before any scan — no "review the whole repo".
• Detects domain (backend / UI / security / performance) and loads the matching checklist from .kasidit/CHECKLISTS/.
• Auto-escalates Mode router → full for the duration; reverts on completion (v0.10).
• Dispatches audit-specialist --focus=quality (v0.10 — replaces standalone code-reviewer) per file (Haiku) or per module (Opus / Sonnet).
• Main synthesizes findings with [high | medium | low | unsure] confidence labels.
• Emits HIGH / MED / LOW severity buckets plus a Top-5 actionable list.

Flow

1. Ask user which module / file / commits — refuse vague scope.
2. Detect domain → load matching CHECKLISTS/*.md.
3. Auto-escalate Mode → full (v0.10).
4. Dispatch audit-specialist --focus=quality: 1 per file on Haiku, 1 per module on Opus / Sonnet.
5. Each agent scans mechanically against the checklist.
6. Main aggregates + deduplicates + labels confidence.
7. Output severity-bucketed findings + Top-5 priority.
8. Emit [kasidit-log] kind=review-<stack> mode=full turns=N outcome=... for router memory.
9. If counter hits 4 (Opus) / 2 (Haiku), escalate or hand back.
10. Revert to prior Mode.

When to use

• PR review before merge.
• Pre-release sanity pass on a changed module.
• Periodic checklist audit on a legacy area.

When NOT to use

• You want a fix, not findings — use Kasi-Fix.
• You need a security audit specifically — use Kasi-Security (different checklist set).
• Scope is "everything" — the command will refuse.

Tier behavior

• Opus — 1 subagent per module, counter cap 4.
• Sonnet — same as Opus, default workhorse.
• Haiku — 1 subagent per file (cheaper context), counter cap 2, mandatory verifier pass.

Anti-patterns

• ❌ Accepting "review all code" — always narrow.
• ❌ Declaring victory without runtime evidence (tests, traces, or user confirmation).
• ❌ Silently guessing [unsure] findings — list them separately.
• ❌ Skipping the checklist because "this file is small".

Since

Introduced in v0.1.0.

See also

• Commands (aggregate)
• Checklists
• Confidence-Labels
• Agent-Audit-Specialist — the agent dispatched (--focus=quality)
• Kasi-Mode — auto-escalation contract
• Kasi-Security