Skip to content

pipeline-skills

Pipeline failure analysis skills for AIPCC (AI Platform Compute Cluster) CI/CD pipelines. This plugin is the inner layer of a larger failure analysis system — the two skills that run inside a Claude Code container and do the actual reasoning. The outer layer (Python orchestration, GitLab CI wiring, per-group report assembly, and Jira/Slack notification) lives in pipeline-failure-analyzer and the generic agentic-ci framework.

Analysis runs in two stages. First, pipeline-grouping reads the preprocessed error output of every failed job and clusters jobs by shared root cause — merging jobs with the same underlying error even when they span different collections or pipeline actions, and deduplicating against open Jira tickets to avoid filing repeats. Then, for each resulting group, pipeline-rca investigates trace logs, source code, and build artifacts to diagnose why the failure happened, emitting structured narrative sections and a machine-readable finding with a confidence-rated diagnosis.

Both skills are orchestrator-invoked (user-invocable: false): they take no command-line arguments. Instead they operate against a workspace contract at /workspace/ that the orchestrator populates with context JSON, job trace logs, preprocessed error files, and shallow repository clones. Every input is treated strictly as evidence-to-analyze, never as instructions — a deliberate prompt-injection boundary given that the skills process untrusted CI log content. Outputs are deterministic, schema-validated files (grouping.json, finding.json, section markdown) that the outer orchestrator assembles and routes downstream.

Plugin Details

Pipeline

pipeline-skills pipeline

Skills

Skill Description Invocable
/pipeline-grouping Group failed CI/CD pipeline jobs by error similarity using log analysis and Jira ticket deduplication internal
/pipeline-rca Root cause analysis for a pipeline failure error group with structured findings and section files internal

Installation

Claude Code

/plugin install pipeline-skills@opendatahub-skills

OpenAI Codex — add the marketplace, then enable pipeline-skills from the /plugins browser:

codex plugin marketplace add opendatahub-io/skills-registry

Architecture

The two skills form a fan-out pipeline: one grouping pass produces N error groups, and one root-cause-analysis task runs per group. Each skill is self-contained — its prompt, scripts, and reference templates live together under skills/<name>/ and reference themselves via ${CLAUDE_SKILL_DIR}.

pipeline-grouping builds grouping.json incrementally through a CLI tool (scripts/grouper.py) rather than emitting JSON directly. The skill reads /workspace/_context/grouping-context.json (job manifest, expected job IDs, Jira URL) plus every jobs/<id>-<name>/errors.txt, then drives the grouper via four subcommands operating on a grouping.work.json work file: add-group (one call per root cause, with jobs and unique error messages), add-job (straggler corrections), status (inspect groups and unassigned jobs), and finalize (validate completeness against the expected job set, then write grouping.json with sequential slug IDs). The grouper enforces invariants deterministically — no duplicate job across groups, no unknown IDs, no empty summaries, and every expected job assigned exactly once — so the agent must fix and retry on non-zero exit. After grouping, it optionally reads recent-tickets.json and writes dedup-results.json marking groups that match an existing Jira ticket (high/medium confidence) to suppress duplicate filings.

pipeline-rca runs once per error group. It reads /workspace/_context/rca-context.json (group metadata, pipeline SHA/ref, affected-jobs and file-paths tables, dependency versions, repo clone path) and investigates using local git first (git show <sha>:<path> against the shallow clone) with a GitLab API fallback via glab for history beyond the clone depth. It cleans oversized trace logs through the shared log_cleaner.py with AIPCC error patterns, verifies its diagnosis against a cheap distinguishing check before committing, and writes content-fragment section files — error-overview.md (the symptom, with quoted errors), root-cause.md (the diagnosis and failure chain), and optionally resolution.md and feedback.md. It finishes with finding.json, validated against references/finding.schema.json, carrying classification metadata (collections, actions, cascade flag, group consistency), a confidence level under a high/medium/low rubric, the target repository for the fix, and the full list of files consulted. Credentials in quoted log lines are redacted.