Countdown - Validation pré-implémentation

Workflow de validation avant implémentation. Exécute Oracle (Opus) pour la planification, puis Go Poll avec agents de révision (Haiku) pour valider avant le codage.

Spar Skills Guide Bot
DevOpsAvancé1 vues0 installations08/03/2026
Claude CodeCursor
workflow-automationplanning-validationquality-gatesmulti-agent-reviewimplementation-strategy

name: countdown description: "Countdown - pre-implementation validation. Runs oracle (Opus) for planning, then Go Poll with review agents (Haiku) to validate before coding starts."

Countdown

The pre-launch sequence for validating a plan before implementation begins.

Usage

/countdown [issue-number]

What This Does

Countdown is the pre-implementation workflow:

  1. Plan - Oracle creates/refines implementation plan
  2. Go Poll - Review agents validate the plan
  3. Iterate - Fix issues until all agents report GO
  4. Launch - Begin implementation only when validated

Sequence

Phase 1: Oracle Planning

Oracle creates the implementation plan at:

~/.claude/plans/reovim/{issue_number}-{subject}.md

Phase 2: Go Poll

Review agents validate the plan:

| Agent | Model | Focus | |-------|-------|-------| | mission-control | Haiku | Plan completeness, phases, acceptance criteria | | telemetry | Haiku | Test strategy, coverage targets | | flight-director | Sonnet | Architecture, Unix philosophy, layer boundaries |

A+ Skip Rule

If an agent gives ALL A+ in round N, skip that agent in round N+1.

Output Structure

tmp/{ISSUE}/
└── countdown/
    └── round-{N}/
        ├── mission-control.md
        ├── telemetry.md
        └── flight-director.md

Grading Scale

| Grade | Meaning | |-------|---------| | A+ | Exemplary plan (skip next round) | | A | Ready for implementation | | B | Minor gaps - refine plan | | C | Significant gaps - more planning needed | | F | Major issues - rethink approach |

Success Criteria

Must achieve A+ from ALL THREE review agents before implementation.

If any agent gives below A+:

  1. Refine the plan based on feedback
  2. Re-run /countdown
  3. Repeat until all agents report "go"

Instructions

When invoked, you MUST:

  1. Determine the issue number from context (git branch, user input, or ask)
  2. Check if a plan exists at ~/.claude/plans/reovim/{issue}-*.md
  3. If no plan exists or plan needs refinement:
    • Launch oracle agent (Opus) to create/refine the plan
  4. Determine the round number (start at 1, increment for re-reviews)
  5. Check previous round grades - skip agents with A+ from prior rounds
  6. Create the output directory: tmp/{ISSUE}/countdown/round-{N}/
  7. Launch review agents IN PARALLEL using Task tool with model: "haiku"
  8. After all agents complete, summarize the grades in a table
  9. If any grade is below A, list what needs to be fixed in the plan

Phase 1: Oracle (if needed)

Create/refine implementation plan for issue #{ISSUE}.

You are Oracle - the far-seeing architect.

1. Understand the requirements from the issue
2. Explore the codebase to understand current architecture
3. Design a phased implementation plan

Write the plan to: ~/.claude/plans/reovim/{ISSUE}-{subject}.md

Include:
- Summary and approach
- Phases with specific files and changes
- Risks and mitigations
- Test strategy
- Acceptance criteria

Phase 2: Review Agents

Agent 1 (mission-control):

Countdown review for issue #{ISSUE} (Round {N}).

You are Mission Control validating the plan before launch.

Focus: Plan completeness and clarity.

1. Read the plan at ~/.claude/plans/reovim/{ISSUE}-*.md
2. Verify all phases have clear acceptance criteria
3. Check dependencies between phases
4. Ensure scope is well-bounded

Write your report to: tmp/{ISSUE}/countdown/round-{N}/mission-control.md

End with: "Mission Control: GO / NO-GO" and grade (A+/A/B/C/F).

Agent 2 (telemetry):

Countdown review for issue #{ISSUE} (Round {N}).

You are Telemetry validating the test strategy.

Focus: Test coverage planning.

1. Read the plan at ~/.claude/plans/reovim/{ISSUE}-*.md
2. Verify test strategy covers happy paths, errors, edge cases
3. Check that critical functionality has test targets
4. Identify any testing gaps

Write your report to: tmp/{ISSUE}/countdown/round-{N}/telemetry.md

End with: "Telemetry: GO / NO-GO" and grade (A+/A/B/C/F).

Agent 3 (flight-director):

Countdown review for issue #{ISSUE} (Round {N}).

You are Flight Director validating the architecture.

Focus: Unix philosophy and layer boundaries.

1. Read the plan at ~/.claude/plans/reovim/{ISSUE}-*.md
2. Verify mechanism vs policy separation
3. Check layer boundaries are respected
4. Assess complexity and simplicity

Write your report to: tmp/{ISSUE}/countdown/round-{N}/flight-director.md

End with: "Flight Director: GO / NO-GO" and grade (A+/A/B/C/F).

CRITICAL:

  • Launch oracle with model: "opus" for planning
  • Launch review agents with model: "{model}" for validation
    • model: "haiku" for fast, efficient reviews
    • model: "sonnet" for comprehensive, precise reviews
  • Skip agents that received ALL A+ in previous rounds
  • Do NOT start implementation until all agents report GO

Skills similaires