Skip to content

rfe-speedrun

Runs the whole pipeline end to end — create, auto-fix, submit — from a single idea, a Jira key, or a YAML batch file, with as little interaction as possible.

Step 0 resolves the type, cleans tmp/, and pins every flag (notably batch_size, which must be a concrete integer) into tmp/{STATE_PREFIX}speedrun-config.yaml so the run is reproducible and survives compression. It then picks a mode: A a --input batch YAML, B an existing Jira key (skip create), or C a single free-text idea.

Mode A validates the batch with --strict before spending any agent budget, pre-allocates every ID, then fans out one /rfe-create agent per entry — passing each entry's clarifying_context verbatim inside a delimited block as data, never as instructions. Completion is a blocking barrier on the task files, not agent notifications: the same IDs pending across three exit-3 results means those agents are dead and get relaunched. Phase 2 hands off to /rfe-auto-fix via the Skill tool (Agent would background it and end the session) and verifies with check_autofix_complete.py, retrying the missing IDs up to three times. Phase 3 submits only the IDs collect_recommendations.py reports as passing.

Plugin: rfe-creator | User-invocable

Contract

Skill Contract canonical-skill-v1

Run the full pipeline for the resolved type end-to-end — create, auto-fix, and submit — from an idea, Jira key(s), or a batch file.

Identity

Functions
orchestrate
Success
  • Runs create, auto-fix, and submit in sequence, persisting IDs and config between phases.
  • Verifies completeness and retries missing IDs before submitting.

Optimization Targets

task_success judge SKILL.md @ c6eb99d

Invariants

Must Preserve
  • Always pass an explicit batch size to auto-fix for reproducibility.
  • Never emit a text-only response during pipeline execution, which terminates the CI process.
  • Forward the resolved type to every phase it launches.
Fixed Context
toolsRead, Write, Edit, Glob, Grep, Bash, AskUserQuestion, Skill
clipython3
knowledgerepository_contentpublic, task_inputtask_private, tool_outputtask_private

Traceability

Skill
.claude/skills/rfe-speedrun/SKILL.md

Diagram

rfe-speedrun diagram

Arguments

/rfe-speedrun [--type <t>] [--input <path>] [--headless] [--announce-complete] [--dry-run] [--batch-size N] <idea | Jira key>
Argument Required Default Description
idea or Jira key - A free-text idea or objective (mode C), or a single Jira key to process an existing item (mode B). Omit when using --input.
--type rfe Work-item type for the run. An explicit --type always wins.
--input - YAML file of batch entries (prompt, priority, labels, clarifying_context) — mode A. May also be a mapping with type and items keys.
--headless - Suppress questions and confirmations, for CI and eval runs.
--announce-complete - Print a completion marker when done, for CI and eval harnesses.
--dry-run - Skip the Jira writes in the submit phase.
--batch-size 5 Batch size forwarded to auto-fix. Always passed explicitly so runs stay reproducible.

Usage

/rfe-speedrun Users need to verify model signatures at serving time
/rfe-speedrun --input batch.yaml --headless --announce-complete
/rfe-speedrun --type initiative --dry-run RHAISTRAT-12345