Harness Scouter

Harness Scouter

[English](README.md) · [한국어](README.ko.md) · [日本語](README.ja.md)

Turns the Claude Code transcripts sitting on your machine into 6 stats that measure **the quality of a coding agent harness**. It shows them as a character stat window rather than a report, and tells you how to raise the low ones.

What gets measured is the **harness**, not the model. The same model gives different results depending on how the prompts, context, hooks, and skills were put together, and this tool tries to measure that difference.

  HARNESS SCOUTER  all-time         Lv. 61  C
  2026-05-10 ~ 2026-08-14 · 23 sessions · 4 history windows · coverage 71%
  ────────────────────────────────────────────────────────────────────────────────
  Retrieval          ████████░░░░░░░░░░░░░░░░  32  D   typical   38~50  best  52
      File-finding discipline      0   n=   90
      Index-first retrieval       24   n=  372
      Evidence before edit        71   n=  328
  Verification       ███████████░░░░░░░░░░░░░  45  D   typical   43~48  best  49
      Pre-commit check freshness   3   n=   59
      No redundant checks         87   n=  149
  Delivery           █████████████████████░░░  86  A   typical   80~88  best  91
      Reached an artifact         73   n=   11  (display)
      No rework                   86   n= 1192
  Autonomy           █████████████████████░░░  86  A   typical   81~95  best  98
      No human intervention       86   n= 5473
  Discipline         █████████████████░░░░░░░  70  B   typical   68~75  best  77
      Instrumented-channel use    56   n= 2130
      No repeat gate hits         85   n=  167
  Context efficiency ██████████████████░░░░░░  74  B   typical   62~76  best  86
      Read-scope discipline       60   n=  259
      Recall of what was read     91   n= 1066
      Response brevity            76   n= 3196
      Context lightness           68   n= 5988
  ────────────────────────────────────────────────────────────────────────────────
  Overall 61.3 · C