plan-mode — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited plan-mode (Agent Skill) and scored it 92/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 2 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 2 flagged
The text {match} tells the agent to skip the normal "ask the user first" gate. Used adversarially it removes the human-in-the-loop check before destructive or sensitive actions, turning a normally-gated agent into a fire-and-forget executor.
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
Prevent premature implementation. Understand → plan → get approval → implement → validate. Do not skip the gate just because a change looks small.
The full protocol — investigation steps, issue classification, the question and plan templates, validation, and special cases — lives in PROTOCOL.md. Read it before planning or implementing.
During Plan Mode, do not mutate the project: no editing files, scaffolding, patches, package installs, lockfile updates, migrations, deploys, destructive scripts, or "trying an implementation" to learn the code.
Allowed: read files, search the repo, inspect structure, review tests / docs / configs / API contracts / schemas / migrations, run read-only or non-mutating diagnostic commands, and consult external docs when project context is insufficient.
If a tool or command might mutate state, ask first or avoid it.
Surface ambiguities, contradictions, and blockers before the plan — never bury them inside it. For each, capture: what it is, the evidence, why it matters, whether it blocks implementation, and a proposed default or resolution. If none exist, state No hard blockers found.
Ask only after investigating context, and never for what the repo, tests, docs, or provided context already answer. Group related questions, give each a recommended default, and explain briefly why each answer matters. Avoid open-ended "What should I do?" questions.
Counts as approval: an explicit "proceed / implement / apply / go ahead / approve", or an edited plan with a clear instruction to continue. Does not count: answering one question, commenting on the plan, asking for more detail or alternatives, or "is this enough?". When approval is ambiguous, ask for a direct confirmation.
If new evidence contradicts the approved plan mid-implementation, pause and return to planning with what changed, the evidence, the impact, and a revised recommendation.
Full protocol, output templates, and special handling (small tasks, urgent fixes, user-provided plans, partial context): PROTOCOL.md.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.