A detailed walkthrough of building a four-stage pipeline that automatically converts Jira bug tickets into GitHub draft PRs using Copilot coding agent. The pipeline includes: a Lambda that mirrors Jira tickets to GitHub issues and copies attachments to S3; a triage layer using Claude Sonnet that classifies issues as auto-fixable, needs-info, or diagnosis-only; a multi-agent log analyser using GPT and Claude models with deterministic preprocessing; and finally Copilot coding agent assignment. Key lessons cover Jira webhook payload quirks, idempotency patterns, Copilot SDK auth (three separate tokens required), grounding LLMs with real repo tools, and the importance of deterministic preprocessing before model calls.
Table of contents
Stage 2: The triage layerWhat Copilot coding agent actually doesThe three-way decisionThe token mazeThe two pathsGrounding the triageGet Bristena ’s stories in your inboxSingle LLM vs orchestrated pipelineSort: