ai-bug-triage

Normalize the raw CI logs to strip timestamps, ports, and process IDs, then hash stable anchors like exception type and stack frames into fingerprints for deduplication. An LLM then classifies each unique failure by category, severity, and component, and drafts a ticket for human approval.

Use deterministic fingerprinting rather than LLM comparison: normalize logs, extract stable anchors (exception type, message template, top stack frames, test name), and SHA-256 hash the canonical form. Identical failures produce identical fingerprints, and similarity scoring above 0.75 flags near-duplicates for merge review.

No. LLMs are non-deterministic, so the same pair of errors can receive different similarity scores on repeated runs. Use deterministic fingerprinting for deduplication and reserve the LLM for severity classification, root-cause hypotheses, and human-readable ticket writing.

Yes. The pipeline output is tracker-agnostic, producing title, description, labels, severity, and component fields that map to Jira, Linear, GitHub Issues, or Azure DevOps. The fingerprint is attached as a label so future runs can match against existing tickets.

Fingerprints become unstable when raw logs are hashed without normalization, because timestamps, PIDs, ports, and request IDs differ every run. Apply the normalization rules first, then hash only stable anchors; the five stability assertions in the verification step confirm the fingerprinter works.

Hosted tools like Trunk Flaky Tests, CloudBees Smart Tests, and Datadog Test Optimization already cover fingerprinting, clustering, and severity routing. Build the in-house pipeline only when you need on-prem deployment, exotic tracker integration, or an explicit AI-prompt audit trail.