CI/CD Build Diagnostics

Diagnose DotCMS GitHub Actions failures as a senior platform engineer. Use diagnostic scripts for structured evidence gathering, triage before deep-diving, and stop when confident.

Triage First (Critical — Baseline Gap)

Without this skill, agents dive straight into source code analysis — spending 20+ minutes and 100k+ tokens on what might be a known flaky test. Always triage before investigating.

Identify failure → Check known issues → THEN deep dive (only if needed)

Quick Win (30s-2min): Known issue? → Link and done.
Standard (2-10min):  Gather evidence → Hypothesize → Conclude.
Deep Dive (10+min):  Only for novel, unclear failures.

Failure type determines tools, not a fixed sequence:

Workflow Types

• cicd_1-pr.yml — PR validation with test filtering (subset may pass)
• cicd_2-merge-queue.yml — Full suite before merge (catches filtered tests)
• cicd_3-trunk.yml — Post-merge deployment (artifacts, no re-test)
• cicd_4-nightly.yml — Scheduled full run (detects flaky tests)

PR passes + merge queue fails = test filtering discrepancy.

Prerequisites

This skill must be run from within a checkout of dotCMS/core (any worktree is fine). Requires Python 3.8+ and GitHub CLI (gh) authenticated.

diagnose.py runs preflight checks automatically and will fail with actionable errors if anything is missing.

Investigation Workflow

1. Gather Evidence

All operations go through diagnose.py. It handles preflight, workspace, caching, and structured output. All cached data is reused automatically on re-runs.

# Full gather (default) — metadata, jobs, annotations, logs, error summary
python3 .claude/skills/cicd-diagnostics/diagnose.py <RUN_ID_OR_URL>

# Progressive subcommands — use when you need specific data or want to save tokens
python3 .claude/skills/cicd-diagnostics/diagnose.py <ID> --metadata      # Metadata + jobs + step detail
python3 .claude/skills/cicd-diagnostics/diagnose.py <ID> --jobs          # Jobs + step detail only
python3 .claude/skills/cicd-diagnostics/diagnose.py <ID> --annotations   # Workflow annotations only
python3 .claude/skills/cicd-diagnostics/diagnose.py <ID> --logs          # Download logs + error summary
python3 .claude/skills/cicd-diagnostics/diagnose.py <ID> --logs <JOB_ID> # Single job log + errors
python3 .claude/skills/cicd-diagnostics/diagnose.py <ID> --evidence      # Full evidence.py analysis
python3 .claude/skills/cicd-diagnostics/diagnose.py <ID> --evidence <ID> # evidence.py on single job

Recommended progression:

1. Start with full gather (no flag) for most cases
2. If the output is enough to diagnose → stop
3. If you need deeper log analysis → --evidence or --evidence <JOB_ID>
4. If you need to re-check just one thing → use the specific subcommand

Read the FULL output before proceeding to analysis. Key signals:

• Steps marked FAIL <- likely caused job failure = primary root cause
• Steps marked continue-on-error = masked secondary issues (real bugs, report separately)
• ##[error] lines that don't match any failed step = errors from continue-on-error steps

Do NOT write inline python3 -c to parse jobs JSON or extract errors. diagnose.py already provides all of this. Ad-hoc parsing causes bugs and misses continue-on-error signals.

Do NOT run ad-hoc gh run view, gh api, or gh run list commands to re-fetch data that diagnose.py already provided. The job list, step details, and error summaries are all in the output. Running separate gh commands wastes tokens, triggers permission prompts, and often misses the structured signals (like continue-on-error detection) that diagnose.py provides. Only use direct gh commands for data that diagnose.py does not cover (e.g., comparing against other runs, checking PR info, searching issues).

2. Check Known Issues (Before Deep Dive!)

This is the step the baseline skipped. Search before re-investigating from scratch:

# Search by test name or error message
gh issue list --repo dotCMS/core --search "TestClassName" --state all --limit 10 \
  --json number,title,state,labels
gh issue list --repo dotCMS/core --label "Flakey Test" --state all --limit 20

If match found → link to issue, assess if new information, and stop.

For deployment/auth errors, check external services:

import sys; from pathlib import Path
sys.path.insert(0, str(Path(".claude/skills/cicd-diagnostics/utils")))
from external_issues import extract_error_indicators, format_external_issue_report, generate_search_queries
log_content = Path("$WORKSPACE/failed-job-$JOB_ID.txt").read_text(errors='ignore')
indicators = extract_error_indicators(log_content)
print(format_external_issue_report(indicators, generate_search_queries(indicators, "DATE"), []))

3. Present Evidence for Analysis

Use evidence.py to extract structured failure data from logs:

import sys; from pathlib import Path
sys.path.insert(0, str(Path(".claude/skills/cicd-diagnostics/utils")))
from evidence import present_complete_diagnostic, get_log_stats
LOG = Path("$WORKSPACE/failed-job-$JOB_ID.txt")
print(get_log_stats(LOG))
print(present_complete_diagnostic(LOG))

This extracts: failed tests, error messages, assertion failures, stack traces, timing indicators, infrastructure events, cascade detection, known issue matches.

4. Analyze (Evidence-Based)

Form competing hypotheses and evaluate each against evidence:

• Code defect — New bug from recent PR changes?
• Flaky test — Race condition, timing, shared state?
• Infrastructure — Docker/DB/ES environment issue?
• Test filtering — PR subset passed, full suite failed?
• Cascading failure — One primary error causing secondary failures?

Before concluding root cause, verify the error actually failed the job. diagnose.py flags steps where continue-on-error is detectable from the API (step failed but later non-cleanup steps succeeded). Look at the step marked FAIL ← likely caused job failure for the actual root cause.

WARNING: continue-on-error steps may show conclusion: success in the API even when they have internal errors. GitHub masks the failure — the step reports success but error messages are still in the raw logs. For trunk deployment jobs, check WORKFLOWS.md for which steps have continue-on-error (notably CLI Deploy / deploy-jfrog). When you see ##[error] lines in logs that don't correspond to any failed step in diagnose.py output, that's a masked continue-on-error error. Report these as secondary findings — they are real bugs being hidden, not the cause of the run failure, but worth flagging.

Label each finding: FACT (logs show it), HYPOTHESIS (theory), CONFIDENCE (High/Medium/Low).

See REFERENCE.md for detailed diagnostic patterns (timing analysis, thread context, concurrency).

5. Compare Runs (If Needed)

from evidence import present_recent_runs
print(present_recent_runs("cicd_1-pr.yml", 20))  # Check if intermittent

6. Report

Write DIAGNOSIS.md to the diagnostic workspace directory (e.g., .claude/diagnostics/run-<RUN_ID>/DIAGNOSIS.md). Never write to the project root. Natural language, like a senior engineer writing to a colleague.

Always include: Executive Summary, Root Cause (with confidence), Evidence, Recommendations. Include when relevant: Known Issues, Timeline, Test Fingerprint, Impact Assessment, Competing Hypotheses.

Do not force sections that add no value. See REFERENCE.md for templates.

7. Create Issue (If Warranted)

Only when: not already tracked, new failure pattern, blocking development, actionable.

gh issue create --repo dotCMS/core --title "[CI/CD] Brief description" \
  --label "bug,ci-cd,Flakey Test" --body "$(cat $WORKSPACE/DIAGNOSIS.md)"

Key Principles

1. Triage first — Check known issues before deep investigation. Most failures are known.
2. Adaptive depth — Stop when confident. A known flaky test needs 2 min, not 20.
3. Evidence-driven — Present evidence to AI reasoning, don't hardcode classification rules.
4. Context matters — Same error means different things in different workflows.
5. Use the scripts — Workspace caching means re-runs are fast. Ad-hoc gh commands waste tokens.
6. Never skip a failed step — If diagnose.py errors, diagnose why and fix it before continuing. Do not rationalize past it ("I'll just use gh directly", "I don't really need that data"). Proceeding without evidence produces guesswork, not diagnosis.

Reference Files

• REFERENCE.md — Diagnostic patterns, report templates, collaboration examples
• WORKFLOWS.md — Workflow descriptions and CI/CD pipeline details
• LOG_ANALYSIS.md — Advanced log analysis techniques
• utils/README.md — Utility function API reference
• ISSUE_TEMPLATE.md — Issue creation template