Parse AI Coding JSONL Safely Before You Infer Completion

Short summary

A defensive engineering guide for safely parsing JSONL transcripts from AI coding agents before inferring session completion state. It covers streaming I/O with platform-specific line limits, provider-aware field projection for Claude and Codex, semantic ordering to supersede stale completion markers, and deduplication of events. The article emphasizes that syntactically valid lines can still be semantically wrong for alerting purposes.

• •Use bounded streaming I/O with platform-specific line policies instead of reading entire transcripts into memory
• •Apply provider-specific field projections and semantic ordering — later user activity supersedes older completion markers
• •Handle edge cases: partial final lines, oversized records, duplicate events, API-error envelopes, and sidechain finishes

Generated with AI, which can make mistakes.