rfe-speedrun

name	rfe.speedrun
description	End-to-end RFE pipeline. Accepts a single idea, Jira key(s), or a YAML batch file. Creates, reviews, auto-fixes (with splits), and submits. Supports --headless, --announce-complete, and --dry-run for CI.
user-invocable	true
allowed-tools	Read, Write, Edit, Glob, Grep, Bash, AskUserQuestion, Skill

You are running the full RFE pipeline in speedrun mode. Your goal is to go from problem statements to submitted Jira tickets with minimal interaction. You orchestrate by calling other skills — never duplicate their work.

Step 0: Parse Arguments and Persist Flags

Parse $ARGUMENTS for:

• --input <path>: Path to a YAML file with batch entries
• --headless: Suppress questions and confirmations (for CI / eval)
• --announce-complete: Print completion marker when done (for CI / eval harnesses)
• --dry-run: Skip Jira writes in submit
• --batch-size N: Override batch size (default 5), passed to auto-fix
• Remaining arguments: either a single Jira key (RHAIRFE-NNNN) or a free-text idea

Clean temp state and persist parsed flags. batch_size MUST always be a concrete integer — if the user did not pass --batch-size, substitute the speedrun default of 5. Do not write <N>, null, or omit the field.

python3 scripts/state.py clean
python3 scripts/prep_assess.py --clean-all
python3 scripts/state.py init tmp/speedrun-config.yaml headless=<true/false> announce_complete=<true/false> dry_run=<true/false> batch_size=<N or 5> input_file=<path or null>

Determine pipeline mode:

• Mode A (Batch YAML): --input flag present → batch create + auto-fix + submit
• Mode B (Existing RFE): argument is a Jira key (RHAIRFE-NNNN) → skip create, auto-fix + submit
• Mode C (Single idea): free-text argument, no --input → single create + auto-fix + submit

If no arguments provided, stop with usage instructions.

Step 0.5: Bootstrap Dependencies

Run bootstrap early so agent definitions (e.g. rfe-scorer) are installed in .claude/agents/ before they're needed in Phase 2. The CREATE phase gives the background agent rescan time to register them.

bash scripts/bootstrap-assess-rfe.sh

If bootstrap fails, retry once. If the retry also fails, continue — auto-fix will attempt bootstrap again in its own setup step.

Defaults

When the user doesn't specify, use these defaults:

• Priority: Normal
• Size: S or M (unless the input clearly describes a large initiative)
• RFE count: Single RFE per entry, unless an entry describes multiple distinct business needs
• Labels: None unless specified

Phase 1: Create

Mode A (Batch YAML): Read the YAML input file. Format:

- prompt: "Users need to verify model signatures at serving time"
  priority: Critical
  labels: [candidate-3.5]
- prompt: "TrustyAI operator crashes on large clusters"
  priority: Major

Validate the batch file before spending any agent budget on it. Use --strict so unknown fields and duplicate prompts (typically typos or copy-paste mistakes) block the run too, not just hard errors:

python3 scripts/validate_batch_input.py <input_file> --strict

If this exits nonzero, stop and report the printed ERROR:/WARNING: lines to the user instead of proceeding — do not fan out agents against a batch that's already known to be malformed.

Count entries and pre-allocate all IDs upfront:

python3 scripts/next_rfe_id.py --from-batch <input_file>   # input_file = the --input path; prints one RFE ID per entry

For each entry, launch an Agent to invoke /rfe.create. Pass the pre-assigned ID so each Agent knows which ID to use:

Agent for entry 1:  /rfe.create --headless --rfe-id RFE-001 [--priority <priority>] <prompt>
Agent for entry 2:  /rfe.create --headless --rfe-id RFE-002 [--priority <priority>] <prompt>
...
Agent for entry N:  /rfe.create --headless --rfe-id RFE-<N> [--priority <priority>] <prompt>

Each entry is a single business need — /rfe.create must produce exactly one RFE per invocation. Wait for all N agents to complete. You must have exactly N RFE IDs — if fewer were created, retry the missing entries. Never delete or re-create task files during Phase 1 — quality issues are addressed in Phase 2 (Auto-fix).

Mode B (Existing RFE): Skip Phase 1. The Jira key(s) from arguments become the processing list.

Mode C (Single idea): Invoke /rfe.create with the user's input:

/rfe.create [--headless] <idea_text>

If not headless, /rfe.create will ask clarifying questions. Collect created RFE IDs.

After Phase 1 (all modes), persist the ID list to disk:

python3 scripts/state.py write-ids tmp/speedrun-all-ids.txt <all_IDs>

Phase 2: Auto-fix

Re-read config and ID list from disk (in case context was compressed during Phase 1):

python3 scripts/state.py read tmp/speedrun-config.yaml
python3 scripts/state.py read-ids tmp/speedrun-all-ids.txt

Invoke auto-fix using the Skill tool (NOT Agent — Agent runs in background and causes the session to terminate). Build the args from the config file:

Skill(skill: "rfe.auto-fix", args: "--headless --announce-complete --batch-size <batch_size> <all_IDs_from_file>")

Pass --headless and --announce-complete through if set in the config. Always pass --batch-size <batch_size> using the value from tmp/speedrun-config.yaml — never omit it, never let auto-fix's own default take over. The speedrun default (5) was already pinned in Step 0; relying on it here is what makes runs reproducible.

Auto-fix handles: assessment, feasibility checks, review, auto-revision, re-assessment, splitting oversized RFEs, retry queue, and report generation. The Skill call blocks until auto-fix completes — this is correct. Do NOT stop, summarize, or skip remaining batches early — the pipeline must process every ID through all phases. Never emit a text-only response (no tool call) during pipeline execution — this terminates the CI process.

Bash discipline: Issue exactly one operation per Bash call. Never use command substitution $(...) or chain commands with ;, &&, or || — they trigger an approval prompt and are denied in headless mode. Instead, pass a value between commands by writing it to a tmp/ file with scripts/state.py and reading it back in a separate call.

After auto-fix returns, verify all RFEs were processed:

python3 scripts/check_autofix_complete.py

If incomplete (exit code 1), the output shows MISSING_IDS=RFE-006,RFE-007,.... Re-invoke auto-fix with the Skill tool using only the missing IDs:

Skill(skill: "rfe.auto-fix", args: "--headless --batch-size <batch_size> <missing_IDs>")

Repeat the verify+retry cycle until all RFEs have reviews or 3 retries have been exhausted.

Phase 3: Submit

Re-read flags (in case context was compressed):

python3 scripts/state.py read tmp/speedrun-config.yaml

Re-read ID list from disk:

python3 scripts/state.py read-ids tmp/speedrun-all-ids.txt

Collect passing IDs:

python3 scripts/collect_recommendations.py <all_IDs_from_file>

Parse the SUBMIT= line for IDs ready to submit.

If no IDs are ready to submit, skip to Phase 4.

If IDs are ready, invoke submit using the Skill tool:

Skill(skill: "rfe.submit", args: "--dry-run --headless <passing_IDs>")

Pass --dry-run and --headless through if set in the config. If not headless, /rfe.submit will show a confirmation table before writing to Jira — this is the one mandatory interaction point.

Phase 4: Summary

Re-read flags:

python3 scripts/state.py read tmp/speedrun-config.yaml

If headless, output a brief machine-readable summary. If interactive, output:

## Speedrun Complete

### Created
- RFE-NNN: <title> (Priority: Normal)

### Review Results
- Passed: N
- Failed: N
- Split: N (into M children)

### Submitted
- RHAIRFE-NNNN: <title> [created/updated/dry-run]

### Reports
- Run report: artifacts/auto-fix-runs/<timestamp>.yaml
- Review report: artifacts/auto-fix-runs/<timestamp>-report.html

### Remaining Issues
<Any RFEs that could not be auto-fixed, or "None">

$ARGUMENTS