Skip to content

rfe-review

Scores work items against the type's rubric and auto-revises what fails. It accepts Jira keys (fetched if not already local) or reviews the local artifacts /rfe-create wrote.

It is an orchestrator: it never reads item bodies into its own context, only frontmatter. For each ID it launches an assess agent plus one agent per declared review dimension — technical feasibility for RFEs, feasibility plus strategic alignment for Initiatives — each with its own blocking and condition rules. A review agent then merges those results into {REVIEWS_DIR}/<ID>-review.md, and filter_for_revision.py selects which items a revise agent should rewrite.

Revised items are re-assessed from scratch for at most two cycles. Before each cycle the review and assess results are deleted so progress detection works, with preserve_review_state.py saving and restoring the cumulative before-scores and revision history across the teardown. An agent that dies mid-run leaves no output file, so the ID gets an error stub and drops out of the batch rather than stalling it.

Plugin: rfe-creator | User-invocable

Contract

Skill Contract canonical-skill-v1

Score work items of the resolved type against the type's rubric, run its review dimensions (technical feasibility; strategic alignment for Initiatives), and auto-revise failing items across up to two re-assessment cycles.

Identity

Functions
review
Success
  • Produces a review file per item with rubric scores, a verdict per review dimension, and a recommendation.
  • Auto-revises failing items and re-assesses up to 2 cycles; auto_revised stays true only when the content changed.

Optimization Targets

task_success judge SKILL.md @ c6eb99d
output_quality judge SKILL.md @ c6eb99d

Invariants

Must Preserve
  • Never read item bodies into orchestrator context — delegate to agents and read only frontmatter.
  • Do not exceed 2 re-assessment cycles.
  • Launch every agent with the type's launch block — the review prompts are the shared skeletons under .claude/skills/rfe-review/prompts/ filled from it.
Fixed Context
toolsGlob, Bash, Agent, AskUserQuestion
clipython3
knowledgerepository_contentpublic, task_inputtask_private, tool_outputtask_private

Traceability

Skill
.claude/skills/rfe-review/SKILL.md
Supporting

Diagram

rfe-review diagram

Arguments

/rfe-review [--type <t>] [--headless] [--caller <name>] <ID> [ID2 ...]
Argument Required Default Description
ID - One or more space-separated item IDs — the type's Jira key form or local id form. A key with no local task file is fetched from Jira first.
--type rfe Work-item type to review. An explicit --type always wins.
--headless - Suppress the end-of-run summary and emit a completion marker instead.
--caller none Identifies the calling skill so a headless run returns to it. /rfe-split passes --caller split.

Usage

/rfe-review RHAIRFE-1595
/rfe-review RHAIRFE-1595 RHAIRFE-1602 RFE-001
/rfe-review --type initiative INIT-001