AgentWorkforce/burn analyze-v0.37.0 on GitHub

[0.37.0] - 2026-04-28

Cross-harness edit-heavy session detector (#167). detectPatterns now flags sessions whose normalized edit-tool count exceeds 4× their read-tool count (with ≥ 5 edits), reaching parity across Claude / Codex / OpenCode via normalizeToolName. Reads are restricted to file-content tools (Read, NotebookRead); Grep / Glob / LS / Bash are deliberately excluded — they discover or shell out, but don't record that the model has read the file. Codex's apply_patch may bundle multiple files per call, so the same threshold flags Codex more conservatively than a file-level count would — known per-harness tunable. Emits EditHeavySession { source, sessionId, readCount, editCount, ratio, likelyRetries, cost } with ratio = +Infinity when reads is zero; likelyRetries is the sum of countRetries() over the session's turns. PatternsResult gains an editHeavySessions array; SessionPatternSummary gains editHeavyCount. The cost field is not added to totalPatternCost because the same edit-bearing turns also feed editReverts and retryLoops costs — adding them again would double-count.