interview-framework — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited interview-framework (Agent Skill) and scored it 83/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 4 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 4 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.
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.
Treat every normal-mode interview governed by this framework as a grill. Run the approval-gated interview for start, triage, research, requirements, design, and tasks. Treat this skill and its references as the single source of truth for interview behavior. Phase commands supply exploration territory and artifact context; they do not redefine the algorithm.
Quick mode bypasses interview questions only. It still requires current discovery, contract loading, bypass receipts, delegation checks, and artifact-agent load parity.
Before each new or resumed interview:
${CLAUDE_PLUGIN_ROOT}/references/normal-mode-gates.md.SKILL.md, references/algorithm.md, references/domain-modeling.md, every selected skill body, and every selected skill resource required for the current work. Load references/examples.md only when an example is needed.phase_gate.py record-skill-load.Block when this skill or the core algorithm reference cannot be loaded. Warn and continue when a domain skill fails to load. Put unresolved material conflicts in the first critical frontier.
Grill only a decision that meets both conditions:
Inspect facts with read-only tools or an Explore agent. Exclude setup choices, administrative preferences, status questions, facts the repository can answer, and low-impact polish. Treat a prescribed task action in a loaded domain skill as reference material during preload; do not execute it until the phase has approval and delegation begins.
Before building the tree, read the goal, state, .progress.md, prior phase artifacts, the configured .index/index.md, and the applicable CONTEXT.md reached through CONTEXT-MAP.md when present. Open only relevant indexed entries. Inspect code, configuration, tests, and existing specs for every discoverable fact. Run independent read-only lookups in parallel; a pending fact blocks only the nodes that depend on it.
Build a design tree from the phase territory. Each node contains a stable decision ID, dependencies, known evidence, viable options, recommendation, tradeoffs, and material consequences. Track nodes as open, investigating, resolved, or explicitly out of scope. The frontier contains every open critical decision whose prerequisites are resolved.
Ask the whole currently unblocked critical frontier. Use as many AskUserQuestion calls as needed, with at most four questions per call. Batch independent decisions together.
Before every AskUserQuestion call, call open-frontier for every decision ID in that batch.
After each response:
classify-reply on the whole reply before applying any part of it.record-answer.Ask the whole current frontier in one round. Number each question (Q1, Q2, and so on). Use AskUserQuestion for the round when the tool is available. If AskUserQuestion is unavailable, render the same numbered round in the response and wait for the answers.
Turn an Other response into a specific dependent question in the next frontier. Never use a generic follow-up. Add branches exposed by concrete answers or contradictions, and remove branches that evidence resolves.
Each question must:
(Recommended) unless the options are symmetric.Give a recommended answer with a short rationale. Provide 2-4 meaningful options. Require that the design-tree frontier is empty before final confirmation. Continue only when the user confirms the resulting shared understanding through the explicit approval choice.
See references/algorithm.md for the complete state machine.
Apply references/domain-modeling.md during every grill. Challenge terms that conflict with the applicable CONTEXT.md, replace fuzzy or overloaded words with a proposed canonical term, and use boundary or edge-case scenarios to test the model. Record resolved domain terms promptly. Keep implementation details out of CONTEXT.md. This interview framework does not create ADRs; design.md remains the specification's technical-decision record.
Classify the entire reply before applying it.
Apply text that answers one or more active decisions. Persist answered decisions and keep the rest open. A substantive answer can include control words without losing its decision content.
These replies do not answer any active decision by themselves:
apply the changescontinueproceedgo aheadKeep the active frontier open and ask it again. Do not infer defaults or approval from a control-only reply.
Treat bare skip, after an active question, as authorization to default the remaining phase interview. Call skip with explicit defaults and assumptions; this moves to awaiting_confirmation, not a delegable terminal state. Continue to final approval and confirm decision ID skip-confirmation. A sentence that contains skip plus substantive decision text is a substantive reply, not bare skip.
When the critical frontier is exhausted or skipped:
await-confirmation with a stable confirmation decision ID and the proposed approach.AskUserQuestion:Approve and delegate (Recommended)Revise decisionsCancelconfirm --source approve-and-delegate, run check-delegation, and delegate immediately in the same response. Do not ask another question or stop between approval and delegation.When the user requests revisions, call one revise transition with every affected --decision-id before updating answers. Recompute any dependent frontier, return to final approval using the same confirmation ID, and keep the same interview record until the brief is approved again.
Artifact review is a separate approval gate after delegation. apply the changes during artifact review means revise the artifact using the supplied feedback, redisplay the walkthrough, and remain in artifact approval. It never approves the artifact or advances the phase.
Use phase_gate.py transitions after each state change. Append every completed frontier round to .progress.md without treating that Markdown as enforcement state:
### <Phase> Grill - Round <N>
- Facts resolved: <fact and evidence>
- Decisions: <decision-id> -> <answer>
- Out of scope: <explicitly excluded branch or none>
- Domain language: <canonical term and definition or none>
- Frontier after round: <remaining unblocked decisions or empty>For triage, store enforcement state in the epic .epic-state.json. For spec phases, use .ralph-state.json.
references/algorithm.md - Critical-frontier state machine and reply handling.references/domain-modeling.md - Required context discovery, language challenges, scenarios, and glossary updates.references/examples.md - Optional examples for frontier, partial-answer, skip, approval, and artifact revision cases.~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.