SessionLint
Built an open-source CLI that parses Claude Code transcripts and flags patterns driving agent cost, including cache invalidation on model switches, late compaction, and runaway context. It reports dollar ranges with explicitly labeled assumptions, ships a CI gate with versioned JSON and configurable severity exit codes, and includes a convergence watchdog for headless agent loops.
- TypeScript
- Bun
- Claude Code
- GitHub Actions
Overview
ccusage shows the bill. sessionlint shows patterns behind it — and helps agent loops land.
sessionlint reads your local Claude Code session history and identifies patterns
that often drive API-equivalent cost. It reports honest ranges with labeled
assumptions rather than pretending every heuristic is causal, then helps future
Audit — a read-only linter over your existing transcripts, with a
Live session — a statusline burn gauge, per-session budgets, and cheaper-model
Autonomous runs — budgets, a convergence watchdog, and morning-after reports
for Ralph/GSD-style loops, headless runs, and CI.
Written up in preyashyadav/sessionlint
- Reached 1,000+ installs with 495 passing tests.
- Private by default with no account, server, or telemetry.