by nogunix
OpenShift/RHEL/CNV research & investigation pipeline for Claude Code — crash forensics, upgrade/compat analysis, CVE impact, component behavior. Eight composable agents — doc-search (okp-mcp + mslearn for ARO/Azure + AWS docs for ROSA/AWS), source-trace (versioned-source index, optional), github-trace (github, conditional), jira-trace (mcp-atlassian, conditional), crash-analyze (drgn/gdb), iac-author (terraform/ansible MCP — authors and statically validates the lab's IaC, provisions nothing), lab-verify (oc/terraform/linux-mcp), synthesize — connected by a universal finding format, plus periodic self-improver and upstream-adviser. Reusable investigation patterns are inlined into each agent. Working-method discipline built in: evidence-basis labels (VERIFIED/REASONED/ASSUMED), named report-acceptance gates, a causation gate for crash findings, a project-local lessons loop, a deterministic secret-safety PreToolUse hook, and a tamper-evident per-case evidence chain (append-only hash-linked chain.jsonl, auto-sealed by a PostToolUse hook, verified at synthesis and handoff), a fan-in write-lock that freezes the fact base against mid-collaboration rewrites (chain.py lock + a PreToolUse deny hook), and mechanical reference-URL liveness and verbatim-quote fidelity checks behind the report gates, a validated tool-grant boundary keeping infrastructure provisioning behind the human approval gate, a Japanese report mode whose prose is checked mechanically with textlint (ja-technical-writing) behind gate C2/prose, and click-through evidence links from every report claim to the finding behind it, verified to resolve behind gate C1/link.
Claude Code6 Skills