attest

Attest — detailed reliability view

Render the reliability snapshot for one mumei feature so the user can judge whether it is stable enough to ship or extend.

Trigger

User invokes /mumei:attest <feature> explicitly. The <feature> argument is a feature key — REQ-N-<slug> for spec vehicle, bare <slug> for plan vehicle. The skill is disable-model-invocation: true; it never fires from the model's own initiative.

What it does

1. Reads the feature's reliability-log.jsonl (prefers .mumei/specs/<feature>/ over .mumei/plans/<feature>/).
2. Computes pass^3 over the most recent 10 trials (arithmetic mean of pass booleans; N/A when fewer than 3 trials are recorded).
3. Renders three blocks to stdout:
   • feature key
   • pass^3: <value-or-N/A> (n=<n_trials>, window=10, k=3)
   • markdown table of the last 10 rows (wave, task_id, trial_n, pass, ts).
4. Exits non-zero with feature not found: <feature> to stderr if neither feature directory exists.

How to invoke

Run the CLI implementation in scripts/mumei-attest.sh and print its stdout verbatim:

bash "$CLAUDE_PLUGIN_ROOT/scripts/mumei-attest.sh" "$1"

The script handles the feature-not-found / missing-log / fewer-than-k-trials cases and writes the three blocks to stdout (errors to stderr). Do not reformat the output.

Don'ts

• Don't reinterpret the numeric value — it is a window pass rate (arithmetic mean of pass booleans over the window), not a geometric Pass^k. Do not multiply, transform, or "explain it as a percentage".
• Don't fall back to a different aggregator when the table is empty — the script's N/A is the contract.
• Don't write to reliability-log.jsonl — this skill is read-only.
• Don't accept arguments other than a single feature key (no flags, no glob).