Ralph Review

Run a sequential three-tier code review on the current feature / solo branch. Each tier dispatches a reviewer subagent that runs its own checklist at increasing depth. If any tier fails, fix the findings and restart from Tier 1.

Agent dispatch — namespacing matters

Claude Code registers plugin-bundled agents under the <plugin-name>:<agent-name> namespace. Dispatching a bare haiku-reviewer will fail with Agent type 'haiku-reviewer' not found. Always dispatch with the full plugin prefix:

• ralph-review-trio:haiku-reviewer
• ralph-review-trio:sonnet-reviewer
• ralph-review-trio:opus-reviewer

Controller logic

1. Tier 1 — Haiku (surface). Dispatch the ralph-review-trio:haiku-reviewer subagent with the diff scope and pass criteria. Wait for <promise>HAIKU_PASS</promise> or a failure report.
2. On FAIL: main agent applies fixes, commits per-file, then restarts from Tier 1. Do NOT proceed to Tier 2 with a flag.
3. On PASS: proceed to Tier 2.
4. Tier 2 — Sonnet (logic). Dispatch ralph-review-trio:sonnet-reviewer. Same pass / fail logic.
5. Tier 3 — Opus (deep analysis). Dispatch ralph-review-trio:opus-reviewer. Same pass / fail logic.
6. On all three passes: emit <promise>RALPH_PASS</promise> and proceed to merge gate.

Why restart-on-fail (not continue-with-flag)

Bugs caught by a later tier often invalidate earlier-tier findings. The surface-check pass on commit hygiene is meaningless if the deeper review surfaces a removed-and-re-added file. A fail at any tier says "go back and re-verify everything" — not "note the problem and keep going".

Inputs

• Current git branch + diff against main (or the project's default branch).
• Issue number (if a solo branch named solo/issue-NNN-...).
• Expected Behavior / Acceptance Criteria from the issue body (if available).

RESULT block schema (required from every tier)

## RESULT
mcp_graph_available: yes|no      # first line when discussing graph queries
verdict: pass|fail|unknown
files_touched: [paths]
findings: [{path, line, claim, evidence}]
scope_gaps: [list or "none"]

The main agent parses the RESULT block. Prose body is informational only.

Bash output discipline

Any command a reviewer runs that produces > 200 lines (pytest, git diff, log tail): capture to a file, report only the path + verdict in the RESULT block. Do not paste full output back to the main agent.

Output

• On PASS: <promise>RALPH_PASS</promise> + summary of all three tiers' RESULT blocks.
• On FAIL (after all retries): the failing tier's RESULT block + main-agent-proposed fix list. No PASS promise emitted.