| name | multi-agent-discussion-advisor |
| description | Guide discussion-phase multi-agent decision making for vibe coding projects by providing role design and a launch specification for the host coding agent to start multi-agent review. Use when users need cross-role synthesis to evaluate options and converge on one recommendation across product, requirement, architecture, or hard-problem discussions, including preflight plan review before execution starts. Do not use for execution planning, business runtime orchestration, or direct implementation. |
Multi-Agent Discussion Advisor
Purpose
Use this skill to design discussion roles and emit a launch signal package before execution starts.
The output must include:
- a role design for current scenario (who should participate and why)
- role-level goals (what each agent must prove or challenge)
- explicit sub-agent creation briefs (ready for host-agent tool-triggered delegation)
- synthesis contract for consolidating actual sub-agent findings
Activation Cues
Activate when any condition is true:
- High-uncertainty discussion needs cross-role synthesis to converge on one方案, even if user does not specify exact roles.
- User asks whether multiple roles/agents are needed for requirement or architecture discussion.
- User asks to assess current project product capabilities and technical architecture design from multiple roles.
- User explicitly asks: "评估下当前项目的产品功能,技术架构设计方案" / "需要多角色综合分析确定方案" / "多角色讨论后定方案".
- User explicitly asks to run real multi-agent review (for example: "开多个agent评审" / "用多agent讨论后给结论").
- A draft plan or sprint structure exists and the user wants a preflight review before execution starts.
- The same design point is repeatedly disputed and needs cross-role alignment.
- A proposed solution path is blocked and the team needs reframing suggestions.
- User explicitly asks for collaboration strategy guidance in discussion phase.
Do not activate when any condition is true:
- The request is already execution planning, task breakdown, or milestone tracking.
- The request is direct coding implementation or runtime scheduling.
- The issue can be closed by a local test/log check without cross-role discussion.
Scope Limit
This skill does:
- discussion-phase collaboration pattern guidance
- role design and launch-signal design for discussion analysis
- preflight plan review guidance before execution starts
- minimal-sufficient role selection guidance
- risk-checkpoint and handoff-note guidance
This skill does not:
- replace runtime orchestrator logic in the target business system
- generate or execute business runtime dispatch commands
- own sub-agent scheduling; host coding agent owns actual sub-agent launch
- replace SDD execution planning or plan lifecycle governance
- persist project memory or reusable experience cards by itself
Ownership Boundary
sdd-plan-maintainer owns execution planning and lifecycle governance.
layered-project-memory owns project continuity memory.
experience-capture owns reusable cross-task experience cards.
multi-agent-discussion-advisor owns discussion-phase collaboration guidance.
Boundary rule:
- This skill defines how sub-agents should be launched for analysis-only discussion work.
- The host coding agent executes the actual multi-agent launch using its tool surface.
- If user request shifts to execution planning, route to
sdd-plan-maintainer.
Script Decision
Use the skill-bundled scripts/smoke.sh for structure and contract checks only.
No business runtime script is required in v1.
Workflow Contract
- Frame the discussion problem: goal, constraints, uncertainties.
- If the mode is
preflight, review plan assumptions, owner boundaries, and evaluator readiness before execution starts.
- Design a scenario-fit role set and role-level goals.
- Produce explicit sub-agent creation briefs and launch signal package from the role design.
- Hand off launch signal package to host coding agent for actual sub-agent launch (use
references/subagent-protocol.md).
- Collect sub-agent findings (if launched), identify consensus and disagreements.
- Produce advisory card with evidence, risk checkpoints, and handoff note.
- If user asks to execute implementation, stop this skill flow and hand off to execution planning skill.
Hard Rules
- Advisory only: no business runtime orchestration commands.
- Examples are references, not exhaustive case limits.
- Keep trigger logic cue-based and explainable; do not add in-skill scoring engines.
- Prefer minimal sufficient roles over role inflation.
- Do not write memory/experience records unless user explicitly asks and target skill is invoked.
- Triggered flow must include role design + role goals + explicit sub-agent creation briefs.
- User is not required to provide a full role list; the skill must infer minimal sufficient roles from uncertainty signals.
- If user explicitly asks for real multi-agent discussion, do not answer with role-play-only output; provide launch signal package for host-agent to start at least 2 sub-agents unless tooling is unavailable.
- Each delegated sub-agent brief must have a concrete scope and required output format.
- In
preflight mode, the skill may emit approve, revise, or block review guidance, but it must not mutate plan lifecycle state.
Resource Map
- Read
references/principles.md for abstract design principles.
- Read
references/trigger-signals.md for trigger and non-trigger signals.
- Read
references/process-protocol.md for discussion workflow protocol.
- Read
references/collaboration-patterns.md for reusable collaboration patterns.
- Read
references/subagent-protocol.md for delegation protocol and synthesis rules.
- Read
references/role-selection.md for minimal sufficient role selection.
- Read
references/output-contract.md for advisory-card output schema.
- Read
references/platform-adapters.md for Codex/Claude discussion-phase adapter hints.
- Read
references/regression-cases.md for trigger-boundary regression cases.
- Read
references/examples/README.md for example index and usage notes.
- Use the skill-bundled
scripts/smoke.sh for deterministic smoke checks.