plan-interviewer（计划面试官）

Plan Interviewer — Orchestrator

*Productivity — Multi-agent orchestration. Transform shallow feature requests into well-specified plans through multi-round interviews.*

Core Question: "What haven't we thought of yet?"

Inputs Required

• A feature request, problem description, or existing plan/spec file to refine
• Access to AskUserQuestion tool for multi-round interviews

Output

• Spec file at user-designated path (default: .agents/spec.md)

Chain Position

Previous: none | Next: task-breakdown or system-architecture

Skill Deference

• Have a FEATURE or PRODUCT to specify? → Use this skill.
• Have a declining METRIC to diagnose (X is at Y, should be Z)? → Use problem-analysis (from strategy-skills) instead.

Re-run triggers: When requirements change significantly, when new technical constraints emerge, or when a previous spec leads to implementation blockers.

Multi-Agent Architecture

Agent Roster

Execution Layers

Layer 1 (parallel):
  codebase-scanner-agent ──┐
  artifact-reader-agent ───┘── run simultaneously

Layer 2 (sequential):
  challenger-agent ─────────── uses Layer 1 context to validate premise
    → interviewer-agent ────── conducts interview (INTERACTIVE — uses AskUserQuestion)
      → synthesis-agent ────── writes the specification
        → critic-agent ─────── final quality review

Dispatch Protocol

1. Layer 1 dispatch — send brief to codebase-scanner-agent and artifact-reader-agent in parallel. They extract what the codebase and existing artifacts already answer.
2. Challenger dispatch — send brief + Layer 1 outputs to challenger-agent. It produces exactly 3 challenge questions and a PROCEED/REFRAME/DEFER recommendation.
3. User confirmation — present challenger's questions to the user. If REFRAME, present the alternative framing. If user confirms proceed, continue. If DEFER, stop and explain.
4. Interviewer dispatch — send all context to interviewer-agent. It conducts 3-7 rounds of interview via AskUserQuestion, covering all 5 dimensions, skipping questions already answered by Layer 1.
5. Synthesis dispatch — send all outputs to synthesis-agent to write the spec.
6. Critic review — send spec to critic-agent.
7. Revision loop — if critic returns FAIL, re-dispatch affected agents. Maximum 2 rounds.
8. Save — write spec to user-designated path (default: .agents/spec.md).

Routing Rules

Critical Gates

Before delivering, the critic-agent verifies ALL of these pass:

• All 5 interview dimensions explored (Technical, UX, Business Logic, Architecture, Security)
• Spec contains concrete decisions, not open-ended options
• Edge cases documented with explicit handling strategies
• Open questions list is empty or explicitly deferred with rationale

If any gate fails: the critic identifies which agent must fix it and the orchestrator re-dispatches.

Single-Agent Fallback

When context window is constrained or the feature is simple:

1. Skip multi-agent dispatch
2. Check for existing codebase context and artifacts
3. Challenge the premise (3 questions: right problem? what if nothing? what exists?)
4. Conduct interview using AskUserQuestion (2-4 questions per round, all 5 dimensions)
5. Write specification using the Spec File Format below
6. Run Critical Gates as self-review
7. Save to .agents/spec.md

AskUserQuestion Tool Usage

This skill is interactive. The interviewer-agent uses AskUserQuestion for every question round.

Structure questions for maximum clarity:

• 2-4 questions per invocation
• Each question has 2-4 concrete options
• Options represent real tradeoffs, not obvious choices
• Include brief description explaining implications of each option
• State which option you recommend and why

Example:

Question: "When a background sync fails, how should we handle it?"
Options:
1. Silent retry (3x with backoff) — User unaware, but may see stale data
2. Toast notification — User informed but may be annoyed
3. Badge indicator — Subtle, user can investigate when ready
4. Block UI until resolved — Safest but worst UX
Recommended: Option 1 — most sync failures are transient; fall back to option 2 if retries exhaust.

Interview Dimensions (all 5 required)

Technical Implementation

• Data flow edge cases (What happens when X fails mid-operation?)
• State management boundaries (Where does this state live?)
• Concurrency considerations (What if two users do this simultaneously?)
• Migration paths, performance cliffs

UX & Interaction Design

• Error recovery flows, loading/empty/error states
• Undo/redo expectations, accessibility edge cases
• Mobile-first considerations

Business Logic & Domain

• Edge case ownership, audit/compliance needs
• Multi-tenancy implications, internationalization

Architecture & Tradeoffs

• Consistency vs availability, coupling concerns
• Testability, observability
• Implementation foresight (decisions to resolve NOW, not at 2am during build)

Security & Privacy

• Trust boundaries, data exposure, rate limiting

Spec File Format

# [Feature Name] Specification

## Problem Statement
[1-2 paragraphs on what we're solving and why]

## Decided Approach
[High-level approach with key architectural decisions]

## Key Decisions
| Decision | Choice | Rationale |
|----------|--------|-----------|
| [Topic]  | [What] | [Why]     |

## Edge Cases
- **[Scenario]**: [How we handle it]

## Out of Scope
- [What we're explicitly NOT doing]

## Open Questions
- [ ] [Anything still unresolved]

## Implementation Notes
[Technical details, gotchas, dependencies]

Anti-Patterns

Worked Example

User: "I want to build a habit tracker app."

Layer 1 (parallel):

• codebase-scanner-agent → NO CODEBASE (greenfield)
• artifact-reader-agent → NO ARTIFACTS

Challenger (3 questions):

1. Right problem? "Outcome: users build consistent habits. A habit tracker is the most direct path."
2. What if nothing? "Users currently use spreadsheets — friction causes abandonment."
3. What exists? "Greenfield — no existing code." → Recommendation: PROCEED

Interviewer (5 rounds, AskUserQuestion):

• Round 1 (Scope): Personal only, mobile-first PWA, 4-week MVP
• Round 2 (Users): Health-conscious 25-35, switching from spreadsheets
• Round 3 (Features): Add habit → check off → see streak. Configurable reminders.
• Round 4 (Edge cases): Streak resets on miss (history preserved), user's local timezone
• Round 5 (Technical): Email magic link auth, offline sync on reconnect

Synthesis: Spec with 14 concrete decisions and 3 deferred items.

Critic: PASS — all 4 gates pass, all 5 dimensions covered.

Artifact saved to .agents/spec.md.

Before Starting

Step 0: Product Context

Check for .agents/product-context.md. If available, read for product and audience context before interviewing. If missing, plan-interviewer works standalone.

Optional Artifacts

Artifact Template

On re-run: rename existing artifact to spec.v[N].md and create new with incremented version.

Save to user-designated path (default: .agents/spec.md) using the Spec File Format above.

References

• references/question-bank.md — Extended list of probing questions by domain