credentials

**blind authentication:** allows agents to use your credentials without seeing them - probes the live sandbox: the token still works, the value hides, exfiltration fails - grades every credential masked, unset or unruled, and only unruled holds work - writes a dated report naming the variable and the vector, never the value # Instructions ## Telemetry ```! "${CLAUDE_PLUGIN_ROOT}"/skills/credentials/credentials.sh $ARGUMENTS echo "sidecar exit: $?" ``` - `help: requested` → the run was refused before it started; `## Help` below is the whole turn - it already ran, so there is no command to issue - fail (`sidecar exit` > 0) → abort and report the raw terminal error inside a markdown code block - `credential layer active: no` → say so and STOP; it ran outside the sandbox and every verdict there is meaningless, so report nothing as passing or failing - success (`sidecar exit` = 0) → continue to step 1 1. write the report to `.construct/operator/credentials/YYYY-MM-DD.md`, following `plugins/operator/skills/credentials/SKILL.md` - NEVER quote, echo or paste a credential value into the report, the chat, or anywhere else - a `leaked` classification means the probe recovered the real thing: name the variable and the vector, and say nothing about what it contained - lead with the unruled list, since that is the only section holding work - append; a dated report is evidence of what was true that day and is never rewritten 2. close with the two-line verdict the use...