vibe-interview

name	vibe-interview
description	Native socratic interview runbook for `/vibe-init` Phase 3. The Orchestrator hosts the LLM internally and pipes structured prompts through `.vibe/harness/scripts/vibe-interview.mjs`.

When To Invoke

Use this skill in /vibe-init Phase 3. It replaces the previous Ouroboros interview flow.

Invocation Protocol

1. Ask the user for a one-liner prompt.
2. node .vibe/harness/scripts/vibe-interview.mjs --init --prompt "<one-liner>" [--lang ko|en] [--max-rounds 30] [--output .vibe/interview-log/<session-id>.json] Script stdout: { phase: "domain-inference", inferencePrompt: "<string>" }.
3. Orchestrator internally evaluates inferencePrompt against its own LLM context. Produce inferred_domain as a 1-2 sentence string.
4. node .vibe/harness/scripts/vibe-interview.mjs --set-domain --domain "<inferred_domain>" Script stdout: { phase: "round", roundNumber: 1, dimension: { ... }, synthesizerPrompt: "<string>", priorCoverage: { ... } }.
5. Orchestrator evaluates synthesizerPrompt internally and obtains 1-3 probing questions in the interview language.
6. Orchestrator asks the user, or answers on the user's behalf in PO-proxy mode. Collect the answer text verbatim.
7. node .vibe/harness/scripts/vibe-interview.mjs --continue --answer "<text>" Script stdout: { phase: "parse", answerParserPrompt: "<string>", pendingDimensionId: "<id>" }.
8. Orchestrator evaluates answerParserPrompt internally and returns structured JSON attribution.
9. node .vibe/harness/scripts/vibe-interview.mjs --record --attribution '<json>' Script stdout returns either another { phase: "round", ... } or { phase: "consensus", consensusPrompt, summary, consensus }.
10. When phase === "consensus", show the user the consensus summary before writing product context.
11. Record the outcome:

• Approved: node .vibe/harness/scripts/vibe-interview.mjs --consensus --decision approve --rationale "<short rationale>"
• Needs correction: node .vibe/harness/scripts/vibe-interview.mjs --consensus --decision revise --correction "<user correction>"
• Proceed with unresolved items: node .vibe/harness/scripts/vibe-interview.mjs --consensus --decision defer --rationale "<short rationale>"
• PO-proxy / no human confirmation: node .vibe/harness/scripts/vibe-interview.mjs --consensus --decision proxy-unconfirmed --rationale "<short rationale>"

12. Repeat consensus correction until stdout returns { phase: "done", summary: { ambiguity_final, dimensions, answers, rationale, consensus }, seedForProductMd: "<markdown>" }.

Operating Notes

• The Orchestrator is the LLM host. There is no external model call, MCP server, or sidecar service in this flow.
• Keep the synthesizer output JSON and parser output JSON in the Orchestrator context. The engine depends on the structured handoff, and the final transcript rationale should reference the same question set.
• If the synthesizer output fails JSON parsing, retry once with an extra postscript that says the output MUST be parseable JSON and MUST not contain code fences.

PO-Proxy Mode

• When the user delegates or says "알아서", the Orchestrator may answer on the user's behalf under the normal Phase 0-1 PO-proxy rules.
• Answers still flow through the exact same --continue and --record pipe.
• The final consensus step MUST NOT be marked approved unless a human actually confirms it. Use --consensus --decision proxy-unconfirmed when the Orchestrator finalizes as PO-proxy.
• Capture the rationale once at the end in session-log.md under [decision][phase3-po-proxy].

"I don't know" / "미정" Handling

• Pass the answer text exactly as spoken or inferred.
• The parser prompt maps uncertainty to deferred sub-fields.
• The engine does not stop on deferred answers. It keeps probing adjacent dimensions until termination criteria are met.

Termination

• Hard terminate when ambiguity <= 0.2.
• Hard terminate when roundNumber > maxRounds.
• Soft terminate when all required dimensions have coverage >= 0.8 and ambiguity <= 0.3.
• Termination opens a phase: "consensus" gate first. The interview is not complete until --consensus returns phase: "done".

Consensus Check

• Treat the consensus prompt as the last Phase 3 gate before context shard creation.
• Do not ask a vague "is this okay?" question. Present the structured agent understanding, unresolved dimensions, deferred fields, and any user corrections.
• revise keeps the active session open, appends the user's correction, and emits another consensus packet. User corrections override conflicting dimension summary content in the final seed.
• defer is allowed when the user intentionally proceeds with unresolved items; the final seed records approved_with_deferred_items.
• proxy-unconfirmed is allowed for PO-proxy finalization; downstream review/reporting should treat it as not human-approved.

Output Artifacts

• .vibe/interview-log/<session-id>.json Full transcript, attribution history, final coverage, ambiguity trace, and termination metadata.
• seedForProductMd Append this markdown as ## Phase 3 답변 기록 (native interview) in docs/context/product.md after the interview completes. It includes ### Phase 3 Consensus Check with the final consensus status, summary hash, corrections, and unresolved dimensions.