Skip to content

spike-executor

Execute RHOAI SPIKE investigations with human-in-the-loop approval gates on OpenShift. Provides a 9-step lifecycle covering intake, plan generation, Jira ticket creation and sync, AI-powered research enrichment with hallucination detection, pytest test suite generation and execution on OpenShift clusters, rubric-based feasibility scoring with a security gate, and RFE document generation.

The skill orchestrates via a CLI (spike-executor) that handles each step as a separate command, writing artifacts to an artifacts/ directory with a consistent - naming pattern. Workflow state is persisted in .spike-state-.json for resume support. Supports both runtime evaluations (e.g., AutoGluon) and protocol library evaluations (e.g., gRPC). Integrates with Jira for ticket management (create, update, transition) and optionally requires OpenShift access for cluster test execution.

Plugin Details

Skills

Skill Description Invocable
/SPIKE-executor Execute RHOAI SPIKE investigations with human-in-the-loop approval gates

Installation

Claude Code

/plugin install spike-executor@opendatahub-skills

OpenAI Codex — add the marketplace, then enable spike-executor from the /plugins browser:

codex plugin marketplace add opendatahub-io/skills-registry

Architecture

Single-skill plugin with a 9-step CLI-driven workflow. Each step maps to a spike-executor subcommand that reads/writes from the artifacts/ directory. Breakpoints after each step pause for human review and approval before proceeding.

Key modules: plan_generator.py (Jinja2 template rendering), jira_sync.py (Jira REST API with retry, labels, comments), test_generator.py (Phase 2 plan parser + pytest suite generation), test_executor.py (pytest runner + JUnit XML parsing), scorer.py (rubric-based 0-3 scoring with domain weights and security gate), state.py (workflow persistence), cli.py (Typer CLI).

Step 5 (AI Research) is the most complex: generates a research scaffold via templates, then the LLM performs a comprehensive web research deep dive covering community health, license, architecture, performance, UBI feasibility, operator integration, security (CVE, FIPS, rootless, air-gap, supply chain), and hardware/MLOps. A validation step (validate-research) checks for hallucination via URL reachability, evidence-backed claims, TBD audit, and GitHub metrics cross-referencing.

Scoring uses a 4-domain rubric (UBI 25%, Operator 20%, Security 30%, Hardware 25%) with 14 checks, each scored 0-3. Decision bands: GO >= 80, PIVOT 55-79, NO-GO < 55. A security critical failure gate blocks GO if any security check scores 0.

Three hooks handle artifact safety: backup-artifact.sh (PreToolUse on Write), validate-artifact.sh and auto-validate-research.sh (PostToolUse on Write).