Self-Healing CI/CD Pipelines: Why Autonomous Agents Are Coming for Your Build Failures

# What most teams do today — retries everything, including real bugs
- name: Run tests
  uses: nick-fields/retry@v2
  with:
    max_attempts: 3
    command: npm test

import anthropic
import json
 
client = anthropic.Anthropic()
 
CLASSIFY_PROMPT = """A CI job failed. Classify the failure type based on the logs below.
 
Job: {job_name}
Exit code: {exit_code}
Logs (last 100 lines):
{logs}
 
Classify as exactly one of:
- "flaky_test" — test failure with no code change relevance (timing, race condition, random data)
- "infra_transient" — network timeout, registry unavailable, runner OOM unrelated to the build itself
- "real_regression" — an actual assertion failure or compile error tied to the code change
- "dependency_issue" — lockfile conflict, missing package, version resolution failure
 
Respond with ONLY valid JSON:
{{"classification": "...", "confidence": "high"|"medium"|"low", "reasoning": "one sentence", "auto_fixable": true|false}}"""
 
 
def classify_failure(job_name: str, exit_code: int, logs: str) -> dict:
    response = client.messages.create(
        model="claude-sonnet-5",
        max_tokens=300,
        messages=[{
            "role": "user",
            "content": CLASSIFY_PROMPT.format(job_name=job_name, exit_code=exit_code, logs=logs[-4000:])
        }]
    )
    text = response.content[0].text.strip()
    if text.startswith("```"):
        text = text.split("```")[1].replace("json", "", 1).strip()
    return json.loads(text)

def handle_ci_failure(job_name: str, exit_code: int, logs: str, pr_number: int):
    verdict = classify_failure(job_name, exit_code, logs)
 
    if verdict["classification"] == "flaky_test" and verdict["confidence"] == "high":
        quarantine_test(extract_test_name(logs))
        retrigger_job(job_name)
        file_flaky_test_ticket(job_name, logs)
 
    elif verdict["classification"] == "infra_transient":
        retrigger_job(job_name, delay_seconds=60)
        # No notification — this is expected noise, not signal
 
    elif verdict["classification"] == "dependency_issue" and verdict["auto_fixable"]:
        fix_pr = attempt_dependency_fix(job_name, logs)
        if fix_pr:
            comment_on_pr(pr_number, f"Auto-fix attempted in {fix_pr} — dependency resolution issue detected")
        else:
            notify_engineer(pr_number, verdict["reasoning"])
 
    else:    # real_regression, or anything low-confidence
        block_merge(pr_number)
        notify_engineer(pr_number, f"Real failure, not auto-retrying: {verdict['reasoning']}")

# Track this metric — it tells you if you're fixing root causes or hiding them
def flaky_rate_last_30_days(job_name: str) -> float:
    failures = get_ci_failures(job_name, days=30)
    flaky = [f for f in failures if f["classification"] == "flaky_test"]
    return len(flaky) / len(failures) if failures else 0.0