hoiung Ralph Review Trio
This skill triggers /ralph-review, which runs three sequential reviewer subagents at increasing depth. If any tier flags a failure, the loop restarts from Tier 1 after fixes, up to 3 restarts. The bound counts RESTARTS, not rounds; at restart 4 the loop escalates to a class sweep and then resumes with the count reset to zero for exactly ONE further loop - if that loop does not pass, the review stops and reports the outstanding findings with their classes rather than looping on.
When to trigger
- An implementation is code-complete on a feature / solo branch.
- All acceptance criteria for the underlying issue are believed satisfied.
- Pre-merge verification is needed before human review or merge-to-main.
- You want structured evidence that each tier's checklist was walked.
When NOT to trigger
- Work-in-progress branches mid-implementation — Ralph assumes the change is complete.
- Documentation-only diffs with no code — Tier 2/3 still run but most checks short-circuit to "doc-only PR" exemption; overkill for a single README edit.
- Hotfix branches where speed dominates verification — use the project's normal PR review.
How it works
/ralph-review
│
▼
Tier 1 — Haiku (surface checks) ─── fail ──> restart
│ pass
▼
Tier 2 — Sonnet (logic checks) ─── fail ──> restart
│ pass
▼
Tier 3 — Opus (deep analysis) ─── fail ──> restart
│ pass
▼
RALPH_PASS → merge OK
A <promise>HAIKU_PASS</promise> / <promise>SONNET_PASS</promise> / <promise>OPUS_PASS</promise> token is emitted by each tier on pass. All three required for overall pass.
Entry point
/ralph-review — defined in ../../commands/ralph-review.md.
Per-tier checklists
../../agents/haiku-reviewer.md— Tier 1 surface checklist../../agents/sonnet-reviewer.md— Tier 2 logic checklist../../agents/opus-reviewer.md— Tier 3 deep-analysis checklist
Extended reference content under references/ (same dir as this SKILL.md) is loaded on demand by each tier when a specific check requires more context.
Outputs
Each tier writes a fenced ## RESULT block with:
## RESULT
structural_search_available: yes|no # first line when discussing structural-search results
verdict: pass|fail|unknown
files_touched: [paths]
findings: [{path, line, claim, evidence}]
scope_gaps: [list or "none"]
The main agent reads the RESULT block and decides next action (restart, next tier, or PASS).