fidelity

Use when the user runs /fidelity. Computes the plan-fidelity rate — what fraction of /mandos verdicts came back Faithful, and the Missing-vs-Scope-crept Blocker split that explains the rest — from the

SKILL.md

GitHub

Fidelity — The Plan-Fidelity Rate

Where /este asks whether the config's rules and skills were followed, /fidelity asks whether the plans they governed turned into matching code. Every /mandos verdict — Covered / Missing / Scope-crept — is recorded as it renders; this skill reduces that record into one standing rate, split by which way the drift ran.

Run

python3 ~/.claude/hooks/fidelity-scan.py

This reads ~/.skadi/mandos/history.jsonl (read-only — written by mandos-record.sh as a side effect of /mandos), writes the board channel ~/.skadi/board/fidelity.json, and renders ~/.skadi/henneth/plan-fidelity.html. It prints the headline Faithful rate.

Then ensure the Henneth window is up (boot it with /henneth if not) and surface the dashboard URL inline — the page follows the folder, so the dashboard appears on its own once rendered.

Notes

  • Read-only. The scan reads the recorded history; it never writes to a tracker or repo.
  • Sourced from /mandos, not transcripts. Unlike /este's marker scan, this rate has no signal until /mandos has actually been run a few times — an unused history reads total: 0 and the dashboard says so plainly, rather than guessing.
  • Blocker split localizes the leak. missingBlockers (execution came up short of the plan) and scopeBlockers (execution went beyond the plan) are counted separately — only Blocker-severity items gate a /mandos tier, so they are what actually explains why the rate isn't higher.
  • Standing cadence. The skill arms nothing. For a recurring read, wrap it: /loop 6h /fidelity, or a cron.
  • Tests ride beside the hook — fidelity-scan.test.sh runs offline via injected fixture history (MANDOS_HISTORY_DIR).