Claude-Code-Workflow/.codex/skills/issue-discover/phases/03-discover-by-prompt.md

Phase 3: Discover by Prompt

来源: commands/issue/discover-by-prompt.md

Overview

Prompt-driven issue discovery with intelligent planning. Instead of fixed perspectives, this command analyzes user intent via Gemini, plans exploration strategy dynamically, and executes iterative multi-agent exploration with ACE semantic search.

Core workflow: Prompt Analysis → ACE Context → Gemini Planning → Iterative Exploration → Cross-Analysis → Issue Generation

Core Difference from Phase 2 (Discover):

• Phase 2: Pre-defined perspectives (bug, security, etc.), parallel execution
• Phase 3: User-driven prompt, Gemini-planned strategy, iterative exploration

Prerequisites

• User prompt describing what to discover
• ccw cli available (for Gemini planning)
• ccw issue CLI available

Auto Mode

When --yes or -y: Auto-continue all iterations, skip confirmations.

Arguments

Argument	Required	Type	Default	Description
prompt	Yes	String	-	Natural language description of what to find
--scope	No	String	**/*	File pattern to explore
--depth	No	String	standard	standard (3 iterations) or deep (5+ iterations)
--max-iterations	No	Integer	5	Maximum exploration iterations
--plan-only	No	Flag	false	Stop after Gemini planning, show plan
-y, --yes	No	Flag	false	Skip all confirmations

Use Cases

Scenario	Example Prompt
API Contract	"Check if frontend calls match backend endpoints"
Error Handling	"Find inconsistent error handling patterns"
Migration Gap	"Compare old auth with new auth implementation"
Feature Parity	"Verify mobile has all web features"
Schema Drift	"Check if TypeScript types match API responses"
Integration	"Find mismatches between service A and service B"

Execution Steps

Step 3.1: Prompt Analysis & Initialization

// Parse arguments
const { prompt, scope, depth, maxIterations } = parseArgs(args);

// Generate discovery ID
const discoveryId = `DBP-${formatDate(new Date(), 'YYYYMMDD-HHmmss')}`;

// Create output directory
const outputDir = `.workflow/issues/discoveries/${discoveryId}`;
await mkdir(outputDir, { recursive: true });
await mkdir(`${outputDir}/iterations`, { recursive: true });

// Detect intent type from prompt
const intentType = detectIntent(prompt);
// Returns: 'comparison' | 'search' | 'verification' | 'audit'

// Initialize discovery state
await writeJson(`${outputDir}/discovery-state.json`, {
  discovery_id: discoveryId,
  type: 'prompt-driven',
  prompt: prompt,
  intent_type: intentType,
  scope: scope || '**/*',
  depth: depth || 'standard',
  max_iterations: maxIterations || 5,
  phase: 'initialization',
  created_at: new Date().toISOString(),
  iterations: [],
  cumulative_findings: [],
  comparison_matrix: null
});

Step 3.2: ACE Context Gathering

// Extract keywords from prompt for semantic search
const keywords = extractKeywords(prompt);

// Use ACE to understand codebase structure
const aceQueries = [
  `Project architecture and module structure for ${keywords.join(', ')}`,
  `Where are ${keywords[0]} implementations located?`,
  `How does ${keywords.slice(0, 2).join(' ')} work in this codebase?`
];

const aceResults = [];
for (const query of aceQueries) {
  const result = await mcp__ace-tool__search_context({
    project_root_path: process.cwd(),
    query: query
  });
  aceResults.push({ query, result });
}

// Build context package for Gemini (kept in memory)
const aceContext = {
  prompt_keywords: keywords,
  codebase_structure: aceResults[0].result,
  relevant_modules: aceResults.slice(1).map(r => r.result),
  detected_patterns: extractPatterns(aceResults)
};

ACE Query Strategy by Intent Type:

Intent	ACE Queries
comparison	"frontend API calls", "backend API handlers", "API contract definitions"
search	"{keyword} implementations", "{keyword} usage patterns"
verification	"expected behavior for {feature}", "test coverage for {feature}"
audit	"all {category} patterns", "{category} security concerns"

Step 3.3: Gemini Strategy Planning

// Build Gemini planning prompt with ACE context
const planningPrompt = `
PURPOSE: Analyze discovery prompt and create exploration strategy based on codebase context
TASK:
• Parse user intent from prompt: "${prompt}"
• Use codebase context to identify specific modules and files to explore
• Create exploration dimensions with precise search targets
• Define comparison matrix structure (if comparison intent)
• Set success criteria and iteration strategy
MODE: analysis
CONTEXT: @${scope || '**/*'} | Discovery type: ${intentType}

## Codebase Context (from ACE semantic search)
${JSON.stringify(aceContext, null, 2)}

EXPECTED: JSON exploration plan:
{
  "intent_analysis": { "type": "${intentType}", "primary_question": "...", "sub_questions": [...] },
  "dimensions": [{ "name": "...", "description": "...", "search_targets": [...], "focus_areas": [...], "agent_prompt": "..." }],
  "comparison_matrix": { "dimension_a": "...", "dimension_b": "...", "comparison_points": [...] },
  "success_criteria": [...],
  "estimated_iterations": N,
  "termination_conditions": [...]
}
CONSTRAINTS: Use ACE context to inform targets | Focus on actionable plan
`;

// Execute Gemini planning
Bash({
  command: `ccw cli -p "${planningPrompt}" --tool gemini --mode analysis`,
  run_in_background: true,
  timeout: 300000
});

// Parse and validate
const explorationPlan = await parseGeminiPlanOutput(geminiResult);

Gemini Planning Output Schema:

{
  "intent_analysis": {
    "type": "comparison|search|verification|audit",
    "primary_question": "string",
    "sub_questions": ["string"]
  },
  "dimensions": [
    {
      "name": "frontend",
      "description": "Client-side API calls and error handling",
      "search_targets": ["src/api/**", "src/hooks/**"],
      "focus_areas": ["fetch calls", "error boundaries", "response parsing"],
      "agent_prompt": "Explore frontend API consumption patterns..."
    }
  ],
  "comparison_matrix": {
    "dimension_a": "frontend",
    "dimension_b": "backend",
    "comparison_points": [
      {"aspect": "endpoints", "frontend_check": "fetch URLs", "backend_check": "route paths"},
      {"aspect": "methods", "frontend_check": "HTTP methods used", "backend_check": "methods accepted"},
      {"aspect": "payloads", "frontend_check": "request body structure", "backend_check": "expected schema"},
      {"aspect": "responses", "frontend_check": "response parsing", "backend_check": "response format"},
      {"aspect": "errors", "frontend_check": "error handling", "backend_check": "error responses"}
    ]
  },
  "success_criteria": ["All API endpoints mapped", "Discrepancies identified with file:line"],
  "estimated_iterations": 3,
  "termination_conditions": ["All comparison points verified", "Confidence > 0.8"]
}

Step 3.4: Iterative Agent Exploration (with ACE)

let iteration = 0;
let cumulativeFindings = [];
let sharedContext = { aceDiscoveries: [], crossReferences: [] };
let shouldContinue = true;

while (shouldContinue && iteration < maxIterations) {
  iteration++;
  const iterationDir = `${outputDir}/iterations/${iteration}`;
  await mkdir(iterationDir, { recursive: true });

  // ACE-assisted iteration planning
  const iterationAceQueries = iteration === 1
    ? explorationPlan.dimensions.map(d => d.focus_areas[0])
    : deriveQueriesFromFindings(cumulativeFindings);

  const iterationAceResults = [];
  for (const query of iterationAceQueries) {
    const result = await mcp__ace-tool__search_context({
      project_root_path: process.cwd(),
      query: `${query} in ${explorationPlan.scope}`
    });
    iterationAceResults.push({ query, result });
  }

  sharedContext.aceDiscoveries.push(...iterationAceResults);

  // Plan this iteration
  const iterationPlan = planIteration(iteration, explorationPlan, cumulativeFindings, iterationAceResults);

  // Step 1: Spawn dimension agents (parallel creation)
  const dimensionAgents = [];

  iterationPlan.dimensions.forEach(dimension => {
    const agentId = spawn_agent({
      message: buildDimensionPromptWithACE(dimension, iteration, cumulativeFindings, iterationAceResults, iterationDir)
    });
    dimensionAgents.push({ agentId, dimension });
  });

  // Step 2: Batch wait for all dimension agents
  const dimensionAgentIds = dimensionAgents.map(a => a.agentId);
  const iterationResults = wait({
    ids: dimensionAgentIds,
    timeout_ms: 600000  // 10 minutes
  });

  // Step 3: Check for timeouts
  if (iterationResults.timed_out) {
    console.log(`Iteration ${iteration}: some agents timed out, using completed results`);
  }

  // Step 4: Close all dimension agents
  dimensionAgentIds.forEach(id => close_agent({ id }));

  // Collect and analyze iteration findings
  const iterationFindings = await collectIterationFindings(iterationDir, iterationPlan.dimensions);

  // Cross-reference findings between dimensions
  if (iterationPlan.dimensions.length > 1) {
    const crossRefs = findCrossReferences(iterationFindings, iterationPlan.dimensions);
    sharedContext.crossReferences.push(...crossRefs);
  }

  cumulativeFindings.push(...iterationFindings);

  // Decide whether to continue
  const convergenceCheck = checkConvergence(iterationFindings, cumulativeFindings, explorationPlan);
  shouldContinue = !convergenceCheck.converged;

  // Update state
  await updateDiscoveryState(outputDir, {
    iterations: [...state.iterations, {
      number: iteration,
      findings_count: iterationFindings.length,
      ace_queries: iterationAceQueries.length,
      cross_references: sharedContext.crossReferences.length,
      new_discoveries: convergenceCheck.newDiscoveries,
      confidence: convergenceCheck.confidence,
      continued: shouldContinue
    }],
    cumulative_findings: cumulativeFindings
  });
}

Iteration Loop:

┌─────────────────────────────────────────────────────────────┐
│                    Iteration Loop                           │
├─────────────────────────────────────────────────────────────┤
│  1. Plan: What to explore this iteration                    │
│     └─ Based on: previous findings + unexplored areas       │
│                                                             │
│  2. Execute: Spawn agents for this iteration                │
│     └─ Each agent: explore → collect → return summary       │
│     └─ Lifecycle: spawn_agent → batch wait → close_agent    │
│                                                             │
│  3. Analyze: Process iteration results                      │
│     └─ New findings? Gaps? Contradictions?                  │
│                                                             │
│  4. Decide: Continue or terminate                           │
│     └─ Terminate if: max iterations OR convergence OR       │
│                      high confidence on all questions       │
└─────────────────────────────────────────────────────────────┘

Step 3.5: Cross-Analysis & Synthesis

// For comparison intent, perform cross-analysis
if (intentType === 'comparison' && explorationPlan.comparison_matrix) {
  const comparisonResults = [];

  for (const point of explorationPlan.comparison_matrix.comparison_points) {
    const dimensionAFindings = cumulativeFindings.filter(f =>
      f.related_dimension === explorationPlan.comparison_matrix.dimension_a &&
      f.category.includes(point.aspect)
    );

    const dimensionBFindings = cumulativeFindings.filter(f =>
      f.related_dimension === explorationPlan.comparison_matrix.dimension_b &&
      f.category.includes(point.aspect)
    );

    const discrepancies = findDiscrepancies(dimensionAFindings, dimensionBFindings, point);

    comparisonResults.push({
      aspect: point.aspect,
      dimension_a_count: dimensionAFindings.length,
      dimension_b_count: dimensionBFindings.length,
      discrepancies: discrepancies,
      match_rate: calculateMatchRate(dimensionAFindings, dimensionBFindings)
    });
  }

  await writeJson(`${outputDir}/comparison-analysis.json`, {
    matrix: explorationPlan.comparison_matrix,
    results: comparisonResults,
    summary: {
      total_discrepancies: comparisonResults.reduce((sum, r) => sum + r.discrepancies.length, 0),
      overall_match_rate: average(comparisonResults.map(r => r.match_rate)),
      critical_mismatches: comparisonResults.filter(r => r.match_rate < 0.5)
    }
  });
}

const prioritizedFindings = prioritizeFindings(cumulativeFindings, explorationPlan);

Step 3.6: Issue Generation & Summary

// Convert high-confidence findings to issues
const issueWorthy = prioritizedFindings.filter(f =>
  f.confidence >= 0.7 || f.priority === 'critical' || f.priority === 'high'
);

const issues = issueWorthy.map(finding => ({
  id: `ISS-${discoveryId}-${finding.id}`,
  title: finding.title,
  description: finding.description,
  source: { discovery_id: discoveryId, finding_id: finding.id, dimension: finding.related_dimension },
  file: finding.file,
  line: finding.line,
  priority: finding.priority,
  category: finding.category,
  confidence: finding.confidence,
  status: 'discovered',
  created_at: new Date().toISOString()
}));

await writeJsonl(`${outputDir}/discovery-issues.jsonl`, issues);

// Update final state
await updateDiscoveryState(outputDir, {
  phase: 'complete',
  updated_at: new Date().toISOString(),
  results: {
    total_iterations: iteration,
    total_findings: cumulativeFindings.length,
    issues_generated: issues.length,
    comparison_match_rate: comparisonResults
      ? average(comparisonResults.map(r => r.match_rate))
      : null
  }
});

// Prompt user for next action
await AskUserQuestion({
  questions: [{
    question: `Discovery complete: ${issues.length} issues from ${cumulativeFindings.length} findings across ${iteration} iterations. What next?`,
    header: "Next Step",
    multiSelect: false,
    options: [
      { label: "Export to Issues (Recommended)", description: `Export ${issues.length} issues for planning` },
      { label: "Review Details", description: "View comparison analysis and iteration details" },
      { label: "Run Deeper", description: "Continue with more iterations" },
      { label: "Skip", description: "Complete without exporting" }
    ]
  }]
});

Dimension Agent Prompt Template

function buildDimensionPromptWithACE(dimension, iteration, previousFindings, aceResults, outputDir) {
  const relevantAceResults = aceResults.filter(r =>
    r.query.includes(dimension.name) || dimension.focus_areas.some(fa => r.query.includes(fa))
  );

  return `
## TASK ASSIGNMENT

### MANDATORY FIRST STEPS (Agent Execute)
1. **Read role definition**: ~/.codex/agents/cli-explore-agent.md (MUST read first)
2. Read: .workflow/project-tech.json
3. Read: .workflow/project-guidelines.json

---

## Task Objective
Explore ${dimension.name} dimension for issue discovery (Iteration ${iteration})

## Context
- Dimension: ${dimension.name}
- Description: ${dimension.description}
- Search Targets: ${dimension.search_targets.join(', ')}
- Focus Areas: ${dimension.focus_areas.join(', ')}

## ACE Semantic Search Results (Pre-gathered)
${JSON.stringify(relevantAceResults.map(r => ({ query: r.query, files: r.result.slice(0, 5) })), null, 2)}

**Use ACE for deeper exploration**: mcp__ace-tool__search_context available.

${iteration > 1 ? `
## Previous Findings to Build Upon
${summarizePreviousFindings(previousFindings, dimension.name)}

## This Iteration Focus
- Explore areas not yet covered
- Verify/deepen previous findings
- Follow leads from previous discoveries
` : ''}

## MANDATORY FIRST STEPS
1. Read schema: ~/.codex/workflows/cli-templates/schemas/discovery-finding-schema.json
2. Review ACE results above for starting points
3. Explore files identified by ACE

## Exploration Instructions
${dimension.agent_prompt}

## Output Requirements
**1. Write JSON file**: ${outputDir}/${dimension.name}.json
- findings: [{id, title, category, description, file, line, snippet, confidence, related_dimension}]
- coverage: {files_explored, areas_covered, areas_remaining}
- leads: [{description, suggested_search}]
- ace_queries_used: [{query, result_count}]

**2. Return summary**: Total findings, key discoveries, recommended next areas
`;
}

Output File Structure

.workflow/issues/discoveries/
└── {DBP-YYYYMMDD-HHmmss}/
    ├── discovery-state.json          # Session state with iteration tracking
    ├── iterations/
    │   ├── 1/
    │   │   └── {dimension}.json      # Dimension findings
    │   ├── 2/
    │   │   └── {dimension}.json
    │   └── ...
    ├── comparison-analysis.json      # Cross-dimension comparison (if applicable)
    └── discovery-issues.jsonl        # Generated issue candidates

Configuration Options

Flag	Default	Description
--scope	**/*	File pattern to explore
--depth	standard	standard (3 iterations) or deep (5+ iterations)
--max-iterations	5	Maximum exploration iterations
--tool	gemini	Planning tool (gemini/qwen)
--plan-only	false	Stop after Gemini planning, show plan

Schema References

Schema	Path	Used By
Discovery State	discovery-state-schema.json	Orchestrator (state tracking)
Discovery Finding	discovery-finding-schema.json	Dimension agents (output)
Exploration Plan	exploration-plan-schema.json	Gemini output validation (memory only)

Error Handling

Error	Message	Resolution
Gemini planning failed	CLI error	Retry with qwen fallback
ACE search failed	No results	Fall back to file glob patterns
No findings after iterations	Convergence at 0	Report clean status
Agent timeout	Exploration too large	Narrow scope, reduce iterations
Agent lifecycle error	Resource leak	Ensure close_agent in error paths

Examples

# Single module deep dive
issue-discover --action discover-by-prompt "Find all potential issues in auth" --scope=src/auth/**

# API contract comparison
issue-discover --action discover-by-prompt "Check if API calls match implementations" --scope=src/**

# Plan only mode
issue-discover --action discover-by-prompt "Find inconsistent patterns" --plan-only

Post-Phase Update

After prompt-driven discovery:

• Findings aggregated across iterations with confidence scores
• Comparison analysis generated (if comparison intent)
• Issue candidates written to discovery-issues.jsonl
• Report: total iterations, findings, issues, match rate
• Recommend next step: Export → issue-resolve (plan solutions)