deep-dive-task

name	deep-dive-task
description	Multi-AI consultation for structured TASK document creation + Team-based implementation. Phase 1-2: Parallel AI consultation + self-review → Implementation-ready document. Phase 3: Agent team implementation with Critic + Explorer teammates → Validated code.
allowed-tools	Read, Write, Bash, TaskOutput, Grep, Glob, AskUserQuestion

Deep Dive TASK - Multi-AI Consultation Workflow

Transform user problems into implementation-ready TASK documents through parallel AI consultation and rigorous self-review.

Output: 1 task = 1 file (docs/tasks/active/TASK-{slug}.md)

---

Prerequisites

1. Read context (first found wins):
   • cat nova-context.md (project root, next to CLAUDE.md)
   • If not found: read CLAUDE.md instead (project instructions as context)
2. Check API keys: python3 ~/.claude/skills/llm-review/llm_client.py --check-keys

---

AI Models (HARDCODED — do NOT change)

Phase	Gemini	OpenAI
Consultation	gemini-3-flash-preview	gpt-5.4
Review	gemini-3.1-pro-preview	gpt-5.4 (Responses API, reasoning=high)

---

LLM Client

python3 ~/.claude/skills/llm-review/llm_client.py \
  --provider [gemini|openai] --phase [consultation|review] \
  --prompt "$PROMPT" --output /tmp/result.json

# Parse: jq -r '.success' / jq -r '.response' / jq '.tokens'

---

Phase 1: Consultation

1.1 Problem Structuring

Ask user: issue, goal, related files, constraints. Create initial file: docs/tasks/active/TASK-{slug}.md with Problem + Consultation sections.

1.2 Parallel AI Consultation (180s timeout)

Run gemini & openai in parallel (& + wait), --phase consultation.

1.3 Claude Self-Analysis (parallel)

Explore codebase → root cause → 2-3 approaches → save /tmp/claude_analysis_*.md

1.4 User Discussion (parallel)

AskUserQuestion for unclear requirements. Timeout 90s → auto-proceed.

1.5 Integration & TASK Definition

1. Parse API responses
2. Critical review (do NOT blindly accept):
   • Each suggestion: Accept/Reject/Defer with reasoning
   • Compare against codebase reality, flag contradictions
3. Update the TASK file: add Implementation Plan section

Document rules: DO NOT dump full AI responses. Summarize with decisions. Focus on actionable insights.

1.6 User Approval (90s timeout → auto-proceed)

---

Phase 2: Deep Review

2.1 High-Effort Feedback (600s timeout)

Send TASK to both providers (--phase review, --prompt-file).

2.2 Self-Review (parallel)

Line-by-line: technical accuracy, codebase consistency, scalability, maintainability.

2.3 Finalize Document

1. Parse feedback, Accept/Reject/Defer each suggestion
2. Update the same TASK file: set Status to ACTIVE, add Key Decisions + Result (expected)

Final document = Agent-to-Agent communication:

• Execution-focused (WHAT + HOW, not discussion)
• Readable by another agent in < 5 minutes
• Structured for immediate execution

2.4 Final Confirmation (90s timeout → auto-approve)

---

Error Recovery

Scenario	Recovery
Phase 1 rejection	Revise (max 2 iterations)
Partial API fail	Use available responses
All APIs fail	Claude-only or retry
Phase 2 fail	Self-review only

---

Deliverable

Single file: docs/tasks/active/TASK-{slug}.md

Lifecycle:

• Created in docs/tasks/active/ during Phase 1
• Stays in active/ during implementation
• Moved to docs/tasks/done/YYYY-MM/ when completed

Note: For simple tasks, direct implementation may be faster.

Details: templates.md, reference.md