slopmeter · scanned 2026-09-10
NousResearch/hermes-agent scores grade C, 66 of 100.
Commit health is what costs the grade: p90 404 lines, churn proxy 23% (−8 churn penalty).
Against established software: 59% — above 16 of the 27 hand-written public projects in slopmeter's corpus, scanned 2026-09-10. A sample, not a census.
What was parsed
| language | lines | findings / kLOC | parser |
|---|---|---|---|
| python | 518,818 | 2.6 | tree-sitter |
| ts | 343,061 | 0.7 | tree-sitter |
| js | 13,315 | 1.2 | tree-sitter |
| rust | 3,856 | 4.7 | tree-sitter |
Commits
14,619 first-parent commits on main in the window, lockfiles and build output excluded. Median 82 lines, p90 404, 1481 over 400 lines, p90 7 files per commit, churn proxy 23%.
How this was measured
Over-engineering is five patterns counted per thousand lines, parsed with go/types and tree-sitter rather than matched as text. Test coverage is logic files against whole-word mentions of their symbols in the test files beside them. Commit health reads git log --first-parent over the window, with lockfiles and build output excluded. The composite renormalises over whatever could be measured; anything that could not says so.
Run the same scan yourself — it clones into a temporary directory and removes it afterwards:
slopmeter github NousResearch/hermes-agentThe file-level findings are not published here. slopmeter prints them for the people who can act on them; this page carries the aggregates they rest on.