Ralph Review

STARTER_CHARACTER = 🔍

Ralph Review

Dispatch a subagent to review the ralph plan and bd graph for defects.

Dispatch

The launcher validates the bd graph, captures graph state, then invokes the subagent.

1. Validate the bd Graph

Derive the slug and locate the run epic:

PROJECT_ROOT=$(git rev-parse --show-toplevel)
SLUG=$(git branch --show-current | tr '/' '-')
EPIC=$(bd list -t epic -l "ralph:${SLUG}" --json | jq -r '[.[] | select(.status=="open")] | .[0].id // empty')

Run structural checks. Each failure is a HIGH finding:

• Exactly one open epic with label ralph:${SLUG}
• bd dep cycles reports no cycles
• Epic has at least one child (bd list --parent "$EPIC")
• At least one ready task (bd ready --parent "$EPIC")
• Every child has a description
• Parent set (bd list --parent "$EPIC") equals label set (bd list --label "ralph:${SLUG}" excluding the epic)

Check for isolated nodes (MEDIUM).

If any HIGH finding, report directly and stop. The graph must be sound before review.

2. Capture Graph State

bd list --parent "$EPIC" --json
bd ready --parent "$EPIC" --json
bd list --label "ralph:${SLUG}" --json

3. Invoke the Subagent

Dispatch a subagent with the review prompt below. Substitute <PROJECT_ROOT>, <SLUG>, <EPIC>, the graph validation findings, and the captured graph state.

Relay the response verbatim, prefixed with the graph validation findings.

Review Prompt

You are a plan reviewer. Read the plan, evaluate it against the graph state, and review for defects.

Source

Read <PROJECT_ROOT>/.llmtmp/ralph-plan.md. Then locate the ## Inputs section and parse the plan documents: line. Split on commas, trim whitespace, Read each path. These are plan documents: forward-looking intent the planner consulted (a PRD, a design note). The plan must faithfully render them. Flag the plan if it omits, contradicts, or stale-references a plan document requirement.

How Ralph Loops Work

A ralph loop is an autonomous, iterative execution mechanism. Two modes share the same artifacts:

• In-session (ralph-loop plugin): a Stop hook re-feeds the prompt to the same Claude session.
• External (scripts/ralph.sh): a bash while-loop spawns a fresh claude --print per iteration.

Mechanics:

• The bd graph is the state machine. Each run is organized under a per-run epic; concurrent worktrees stay isolated by --parent filtering.
• The run epic carries label ralph:<slug>. Every child task carries both --parent <epic> and the same label. The two signals are independent for orphan detection.
• The plan's "Run Identity" section records the literal slug and epic ID. The agent reads them from the plan each iteration.
• One task per iteration. The agent reads bd ready --parent <epic> --json, takes the first result, executes its description, closes the task, then stops.
• When the ready queue is empty, the agent closes the epic and emits the sentinel.
• Every action is a bd task (branch creation, BEGIN tag, code, docs, END tag), all parented to the run epic.
• Each bead carries full instructions in its description: files, verification command, commit message.
• Run-wide conventions live in ralph-plan.md, not in every description.
• Environment is a precondition. Credentials, toolchain, and runtime dependencies are the operator's responsibility.

Accepted patterns (do not flag): collapsing tasks into bundles, partial completion, the ralph:<slug> label as run-identity (not workflow dispatch).

What Counts as a Finding

A finding is a defect, gap, or risk the author should fix before execution. Conformance is the baseline, not a finding. Severity is HIGH or MEDIUM only. Omit anything less severe.

HIGH: zero or multiple open epics for ralph:<slug>; empty descriptions; graph cycles; no children under epic; children but no ready tasks; parent/label set mismatch; per-task workflow continues after bd close instead of stopping; per-task workflow omits epic closure before sentinel; missing Run Identity section; missing Inputs section with plan documents: line; plan contradicts a plan document; Approach prose declares ordering the dependency edges do not enforce.

MEDIUM: description omits verification command or commit message; isolated task in a multi-task graph; plan duplicates bead descriptions; bead touching README.md/CLAUDE.md/.llmdocs/ authors edits instead of running the docs skill (/docs, $docs).

Evaluation Areas

Review for these defect patterns only. Report defects found, nothing else.

1. Goal Clarity - vague objective; fresh-context iteration cannot determine success from the plan alone
2. Task Decomposition - wrong granularity; tasks not independently completable; missing or redundant tasks; technique-named titles; insufficient description detail for cold-start worker
3. Sequencing and Dependencies - incorrect edges; bd ready order produces wrong results; implicit dependencies missing explicit edges; cycles; isolated nodes
4. Plan Completeness - missing conventions/tools/commands; inaccurate build/test/lint commands; unclear per-task workflow; missing Run Identity; missing epic scoping in workflow; missing epic closure; missing STOP instruction; missing ralph/ branch and BEGIN/END tags; inline edits instead of docs skill delegation
5. Risk and Gaps - autonomous execution failures; tasks requiring human judgment; tasks that leave repo broken
6. Feasibility - tasks a single iteration cannot realistically complete; tasks exceeding fresh-context autonomy

Output Format

# Ralph Loop Review

## Goal Clarity
## Task Decomposition
## Sequencing and Dependencies
## Plan Completeness
## Risk and Gaps
## Feasibility
## Summary

Finding format: - **[high]** Title. Description. or - **[medium]** Title. Description.

Every section heading present. Defects only under each. 'No findings.' when a section has no defects.

Summary: overall assessment and top 3 recommendations.

Cite task IDs and section names. Produce the review, nothing else.