CSV Wave Pipeline

Auto Mode

When --yes or -y: Auto-confirm task decomposition, skip interactive validation, use defaults.

CSV Wave Pipeline

Usage

$csv-wave-pipeline "Implement user authentication with OAuth, JWT, and 2FA"
$csv-wave-pipeline -c 4 "Refactor payment module with Stripe and PayPal"
$csv-wave-pipeline -y "Build notification system with email and SMS"
$csv-wave-pipeline --continue "auth-20260228"

Flags:

• -y, --yes: Skip all confirmations (auto mode)
• -c, --concurrency N: Max concurrent agents within each wave (default: 4)
• --continue: Resume existing session

Output Directory: .workflow/.csv-wave/{session-id}/ Core Output: tasks.csv (master state) + results.csv (final) + discoveries.ndjson (shared exploration) + context.md (human-readable report)

Overview

Wave-based batch execution using spawn_agents_on_csv with cross-wave context propagation. Tasks are grouped into dependency waves; each wave executes concurrently, and its results feed into the next wave.

Core workflow: Decompose → Compute Waves → Execute Wave-by-Wave → Aggregate

┌─────────────────────────────────────────────────────────────────────────┐
│                    CSV BATCH EXECUTION WORKFLOW                          │
├─────────────────────────────────────────────────────────────────────────┤
│                                                                          │
│  Phase 1: Requirement → CSV                                              │
│     ├─ Parse requirement into subtasks (3-10 tasks)                      │
│     ├─ Identify dependencies (deps column)                               │
│     ├─ Compute dependency waves (topological sort → depth grouping)      │
│     ├─ Generate tasks.csv with wave column                               │
│     └─ User validates task breakdown (skip if -y)                        │
│                                                                          │
│  Phase 2: Wave Execution Engine                                          │
│     ├─ For each wave (1..N):                                             │
│     │   ├─ Build wave CSV (filter rows for this wave)                    │
│     │   ├─ Inject previous wave findings into prev_context column        │
│     │   ├─ spawn_agents_on_csv(wave CSV)                                 │
│     │   ├─ Collect results, merge into master tasks.csv                  │
│     │   └─ Check: any failed? → skip dependents or retry                 │
│     └─ discoveries.ndjson shared across all waves (append-only)          │
│                                                                          │
│  Phase 3: Results Aggregation                                            │
│     ├─ Export final results.csv                                          │
│     ├─ Generate context.md with all findings                             │
│     ├─ Display summary: completed/failed/skipped per wave                │
│     └─ Offer: view results | retry failed | done                         │
│                                                                          │
└─────────────────────────────────────────────────────────────────────────┘

CSV Schema

tasks.csv (Master State)

id,title,description,test,acceptance_criteria,scope,hints,execution_directives,deps,context_from,wave,status,findings,files_modified,tests_passed,acceptance_met,error
"1","Setup auth module","Create auth directory structure and base files","Verify directory exists and base files export expected interfaces","auth/ dir created; index.ts and types.ts export AuthProvider interface","src/auth/**","Follow monorepo module pattern || package.json;src/shared/types.ts","","","","1","","","","","",""
"2","Implement OAuth","Add OAuth provider integration with Google and GitHub","Unit test: mock OAuth callback returns valid token; Integration test: verify redirect URL generation","OAuth login redirects to provider; callback returns JWT; supports Google and GitHub","src/auth/oauth/**","Use passport.js strategy pattern || src/auth/index.ts;docs/oauth-flow.md","Run npm test -- --grep oauth before completion","1","1","2","","","","","",""
"3","Add JWT tokens","Implement JWT generation and validation","Unit test: sign/verify round-trip; Edge test: expired token returns 401","generateToken() returns valid JWT; verifyToken() rejects expired/tampered tokens","src/auth/jwt/**","Use jsonwebtoken library; Set default expiry 1h || src/config/auth.ts","Ensure tsc --noEmit passes","1","1","2","","","","","",""
"4","Setup 2FA","Add TOTP-based 2FA with QR code generation","Unit test: TOTP verify with correct code; Test: QR data URL is valid","QR code generates scannable image; TOTP verification succeeds within time window","src/auth/2fa/**","Use speakeasy + qrcode libraries || src/auth/oauth/strategy.ts;src/auth/jwt/token.ts","Run full test suite: npm test","2;3","1;2;3","3","","","","","",""

Columns:

Per-Wave CSV (Temporary)

Each wave generates a temporary wave-{N}.csv with an extra prev_context column:

id,title,description,test,acceptance_criteria,scope,hints,execution_directives,deps,context_from,wave,prev_context
"2","Implement OAuth","Add OAuth integration","Unit test: mock OAuth callback returns valid token","OAuth login redirects to provider; callback returns JWT","src/auth/oauth/**","Use passport.js strategy pattern || src/auth/index.ts;docs/oauth-flow.md","Run npm test -- --grep oauth","1","1","2","[Task 1] Created auth/ with index.ts and types.ts"
"3","Add JWT tokens","Implement JWT","Unit test: sign/verify round-trip; Edge test: expired token returns 401","generateToken() returns valid JWT; verifyToken() rejects expired/tampered tokens","src/auth/jwt/**","Use jsonwebtoken library; Set default expiry 1h || src/config/auth.ts","Ensure tsc --noEmit passes","1","1","2","[Task 1] Created auth/ with index.ts and types.ts"

The prev_context column is built from context_from by looking up completed tasks' findings in the master CSV.

Output Artifacts

Session Structure

.workflow/.csv-wave/{session-id}/
├── tasks.csv                  # Master state (updated per wave)
├── results.csv                # Final results export
├── discoveries.ndjson         # Shared discovery board (all agents)
├── context.md                 # Human-readable report
└── wave-{N}.csv               # Temporary per-wave input (cleaned up)

Implementation

Session Initialization

const getUtc8ISOString = () => new Date(Date.now() + 8 * 60 * 60 * 1000).toISOString()

// Parse flags
const AUTO_YES = $ARGUMENTS.includes('--yes') || $ARGUMENTS.includes('-y')
const continueMode = $ARGUMENTS.includes('--continue')
const concurrencyMatch = $ARGUMENTS.match(/(?:--concurrency|-c)\s+(\d+)/)
const maxConcurrency = concurrencyMatch ? parseInt(concurrencyMatch[1]) : 4

// Clean requirement text (remove flags)
const requirement = $ARGUMENTS
  .replace(/--yes|-y|--continue|--concurrency\s+\d+|-c\s+\d+/g, '')
  .trim()

const slug = requirement.toLowerCase()
  .replace(/[^a-z0-9\u4e00-\u9fa5]+/g, '-')
  .substring(0, 40)
const dateStr = getUtc8ISOString().substring(0, 10).replace(/-/g, '')
const sessionId = `cwp-${slug}-${dateStr}`
const sessionFolder = `.workflow/.csv-wave/${sessionId}`

// Continue mode: find existing session
if (continueMode) {
  const existing = Bash(`ls -t .workflow/.csv-wave/ 2>/dev/null | head -1`).trim()
  if (existing) {
    sessionId = existing
    sessionFolder = `.workflow/.csv-wave/${sessionId}`
    // Read existing tasks.csv, find incomplete waves, resume from there
    const existingCsv = Read(`${sessionFolder}/tasks.csv`)
    // → jump to Phase 2 with remaining waves
  }
}

Bash(`mkdir -p ${sessionFolder}`)

Phase 1: Requirement → CSV

Objective: Decompose requirement into tasks, compute dependency waves, generate tasks.csv.

Steps:

1. Decompose Requirement

   // Use ccw cli to decompose requirement into subtasks
   Bash({
     command: `ccw cli -p "PURPOSE: Decompose requirement into 3-10 atomic tasks for batch agent execution. Each task must include implementation description, test cases, and acceptance criteria.
   

TASK: • Parse requirement into independent subtasks • Identify dependencies between tasks (which must complete before others) • Identify context flow (which tasks need previous tasks' findings) • For each task, define concrete test cases (unit/integration/edge) • For each task, define measurable acceptance criteria (what defines 'done') • Each task must be executable by a single agent with file read/write access MODE: analysis CONTEXT: @**/* EXPECTED: JSON object with tasks array. Each task: {id: string, title: string, description: string, test: string, acceptance_criteria: string, scope: string, hints: string, execution_directives: string, deps: string[], context_from: string[]}.

• description: what to implement (specific enough for an agent to execute independently)
• test: what tests to write and how to verify (e.g. 'Unit test: X returns Y; Edge test: handles Z')
• acceptance_criteria: measurable conditions that define done (e.g. 'API returns 200; token expires after 1h')
• scope: target file/directory glob (e.g. 'src/auth/**') — tasks in same wave MUST have non-overlapping scopes
• hints: implementation tips + reference files, format ' || ;' (e.g. 'Use strategy pattern || src/base/Strategy.ts;docs/design.md')
• execution_directives: commands to run for verification or tool constraints (e.g. 'Run npm test --bail; Ensure tsc passes')
• deps: task IDs that must complete first
• context_from: task IDs whose findings are needed CONSTRAINTS: 3-10 tasks | Each task is atomic | No circular deps | test and acceptance_criteria must be concrete and verifiable | Same-wave tasks must have non-overlapping scopes

REQUIREMENT: ${requirement}" --tool gemini --mode analysis --rule planning-breakdown-task-steps`, run_in_background: true }) // Wait for CLI completion via hook callback // Parse JSON from CLI output → decomposedTasks[]

2. **Compute Waves** (Topological Sort → Depth Grouping)

```javascript
function computeWaves(tasks) {
  // Build adjacency: task.deps → predecessors
  const taskMap = new Map(tasks.map(t => [t.id, t]))
  const inDegree = new Map(tasks.map(t => [t.id, 0]))
  const adjList = new Map(tasks.map(t => [t.id, []]))

  for (const task of tasks) {
    for (const dep of task.deps) {
      if (taskMap.has(dep)) {
        adjList.get(dep).push(task.id)
        inDegree.set(task.id, inDegree.get(task.id) + 1)
      }
    }
  }

  // BFS-based topological sort with depth tracking
  const queue = []  // [taskId, depth]
  const waveAssignment = new Map()

  for (const [id, deg] of inDegree) {
    if (deg === 0) {
      queue.push([id, 1])
      waveAssignment.set(id, 1)
    }
  }

  let maxWave = 1
  let idx = 0
  while (idx < queue.length) {
    const [current, depth] = queue[idx++]
    for (const next of adjList.get(current)) {
      const newDeg = inDegree.get(next) - 1
      inDegree.set(next, newDeg)
      const nextDepth = Math.max(waveAssignment.get(next) || 0, depth + 1)
      waveAssignment.set(next, nextDepth)
      if (newDeg === 0) {
        queue.push([next, nextDepth])
        maxWave = Math.max(maxWave, nextDepth)
      }
    }
  }

  // Detect cycles: any task without wave assignment
  for (const task of tasks) {
    if (!waveAssignment.has(task.id)) {
      throw new Error(`Circular dependency detected involving task ${task.id}`)
    }
  }

  return { waveAssignment, maxWave }
}

const { waveAssignment, maxWave } = computeWaves(decomposedTasks)

3. Generate tasks.csv

   const header = 'id,title,description,test,acceptance_criteria,scope,hints,execution_directives,deps,context_from,wave,status,findings,files_modified,tests_passed,acceptance_met,error'
   const rows = decomposedTasks.map(task => {
     const wave = waveAssignment.get(task.id)
     return [
       task.id,
       csvEscape(task.title),
       csvEscape(task.description),
       csvEscape(task.test),
       csvEscape(task.acceptance_criteria),
       csvEscape(task.scope),
       csvEscape(task.hints),
       csvEscape(task.execution_directives),
       task.deps.join(';'),
       task.context_from.join(';'),
       wave,
       'pending',  // status
       '',         // findings
       '',         // files_modified
       '',         // tests_passed
       '',         // acceptance_met
       ''          // error
     ].map(cell => `"${String(cell).replace(/"/g, '""')}"`).join(',')
   })
   
   Write(`${sessionFolder}/tasks.csv`, [header, ...rows].join('\n'))
   

4. User Validation (skip if AUTO_YES)

   if (!AUTO_YES) {
     // Display task breakdown with wave assignment
     console.log(`\n## Task Breakdown (${decomposedTasks.length} tasks, ${maxWave} waves)\n`)
     for (let w = 1; w <= maxWave; w++) {
       const waveTasks = decomposedTasks.filter(t => waveAssignment.get(t.id) === w)
       console.log(`### Wave ${w} (${waveTasks.length} tasks, concurrent)`)
       waveTasks.forEach(t => console.log(`  - [${t.id}] ${t.title}`))
     }
   
     const answer = AskUserQuestion({
       questions: [{
         question: "Approve task breakdown?",
         header: "Validation",
         multiSelect: false,
         options: [
           { label: "Approve", description: "Proceed with wave execution" },
           { label: "Modify", description: `Edit ${sessionFolder}/tasks.csv manually, then --continue` },
           { label: "Cancel", description: "Abort" }
         ]
       }]
     })  // BLOCKS
   
     if (answer.Validation === "Modify") {
       console.log(`Edit: ${sessionFolder}/tasks.csv\nResume: $csv-wave-pipeline --continue`)
       return
     } else if (answer.Validation === "Cancel") {
       return
     }
   }
   

Success Criteria:

• tasks.csv created with valid schema and wave assignments
• No circular dependencies
• User approved (or AUTO_YES)

Phase 2: Wave Execution Engine

Objective: Execute tasks wave-by-wave via spawn_agents_on_csv. Each wave sees previous waves' results.

Steps:

1. Wave Loop

   const failedIds = new Set()
   const skippedIds = new Set()
   
   for (let wave = 1; wave <= maxWave; wave++) {
     console.log(`\n## Wave ${wave}/${maxWave}\n`)
   
     // 1. Read current master CSV
     const masterCsv = parseCsv(Read(`${sessionFolder}/tasks.csv`))
   
     // 2. Filter tasks for this wave
     const waveTasks = masterCsv.filter(row => parseInt(row.wave) === wave)
   
     // 3. Skip tasks whose deps failed
     const executableTasks = []
     for (const task of waveTasks) {
       const deps = task.deps.split(';').filter(Boolean)
       if (deps.some(d => failedIds.has(d) || skippedIds.has(d))) {
         skippedIds.add(task.id)
         // Update master CSV: mark as skipped
         updateMasterCsvRow(sessionFolder, task.id, {
           status: 'skipped',
           error: 'Dependency failed or skipped'
         })
         console.log(`  [${task.id}] ${task.title} → SKIPPED (dependency failed)`)
         continue
       }
       executableTasks.push(task)
     }
   
     if (executableTasks.length === 0) {
       console.log(`  No executable tasks in wave ${wave}`)
       continue
     }
   
     // 4. Build prev_context for each task
     for (const task of executableTasks) {
       const contextIds = task.context_from.split(';').filter(Boolean)
       const prevFindings = contextIds
         .map(id => {
           const prevRow = masterCsv.find(r => r.id === id)
           if (prevRow && prevRow.status === 'completed' && prevRow.findings) {
             return `[Task ${id}: ${prevRow.title}] ${prevRow.findings}`
           }
           return null
         })
         .filter(Boolean)
         .join('\n')
       task.prev_context = prevFindings || 'No previous context available'
     }
   
     // 5. Write wave CSV
     const waveHeader = 'id,title,description,test,acceptance_criteria,scope,hints,execution_directives,deps,context_from,wave,prev_context'
     const waveRows = executableTasks.map(t =>
       [t.id, t.title, t.description, t.test, t.acceptance_criteria, t.scope, t.hints, t.execution_directives, t.deps, t.context_from, t.wave, t.prev_context]
         .map(cell => `"${String(cell).replace(/"/g, '""')}"`)
         .join(',')
     )
     Write(`${sessionFolder}/wave-${wave}.csv`, [waveHeader, ...waveRows].join('\n'))
   
     // 6. Execute wave
     console.log(`  Executing ${executableTasks.length} tasks (concurrency: ${maxConcurrency})...`)
   
     const waveResult = spawn_agents_on_csv({
       csv_path: `${sessionFolder}/wave-${wave}.csv`,
       id_column: "id",
       instruction: buildInstructionTemplate(sessionFolder, wave),
       max_concurrency: maxConcurrency,
       max_runtime_seconds: 600,
       output_csv_path: `${sessionFolder}/wave-${wave}-results.csv`,
       output_schema: {
         type: "object",
         properties: {
           id: { type: "string" },
           status: { type: "string", enum: ["completed", "failed"] },
           findings: { type: "string" },
           files_modified: { type: "array", items: { type: "string" } },
           tests_passed: { type: "boolean" },
           acceptance_met: { type: "string" },
           error: { type: "string" }
         },
         required: ["id", "status", "findings", "tests_passed"]
       }
     })
     // ↑ Blocks until all agents in this wave complete
   
     // 7. Merge results into master CSV
     const waveResults = parseCsv(Read(`${sessionFolder}/wave-${wave}-results.csv`))
     for (const result of waveResults) {
       updateMasterCsvRow(sessionFolder, result.id, {
         status: result.status,
         findings: result.findings || '',
         files_modified: (result.files_modified || []).join(';'),
         tests_passed: String(result.tests_passed ?? ''),
         acceptance_met: result.acceptance_met || '',
         error: result.error || ''
       })
   
       if (result.status === 'failed') {
         failedIds.add(result.id)
         console.log(`  [${result.id}] ${result.title} → FAILED: ${result.error}`)
       } else {
         console.log(`  [${result.id}] ${result.title} → COMPLETED`)
       }
     }
   
     // 8. Cleanup temporary wave CSV
     Bash(`rm -f "${sessionFolder}/wave-${wave}.csv"`)
   
     console.log(`  Wave ${wave} done: ${waveResults.filter(r => r.status === 'completed').length} completed, ${waveResults.filter(r => r.status === 'failed').length} failed`)
   }
   

2. Instruction Template Builder

   function buildInstructionTemplate(sessionFolder, wave) {
     return `
   

TASK ASSIGNMENT

MANDATORY FIRST STEPS

1. Read shared discoveries: ${sessionFolder}/discoveries.ndjson (if exists, skip if not)
2. Read project context: .workflow/project-tech.json (if exists)

Your Task

Task ID: {id} Title: {title} Description: {description} Scope: {scope}

Implementation Hints & Reference Files

{hints}

Format: ` || ;`. Read ALL reference files (after ||) before starting implementation. Apply tips (before ||) as implementation guidance.

Execution Directives

{execution_directives}

Commands to run for verification, tool restrictions, or environment requirements. Follow these constraints during and after implementation.

Test Cases

{test}

Acceptance Criteria

{acceptance_criteria}

Previous Tasks' Findings (Context)

{prev_context}

Execution Protocol

1. Read references: Parse {hints} — read all files listed after `||` to understand existing patterns
2. Read discoveries: Load ${sessionFolder}/discoveries.ndjson for shared exploration findings
3. Use context: Apply previous tasks' findings from prev_context above
4. Stay in scope: ONLY create/modify files within {scope} — do NOT touch files outside this boundary
5. Apply hints: Follow implementation tips from {hints} (before `||`)
6. Execute: Implement the task as described
7. Write tests: Implement the test cases defined above
8. Run directives: Execute commands from {execution_directives} to verify your work
9. Verify acceptance: Ensure all acceptance criteria are met before reporting completion
10. Share discoveries: Append exploration findings to shared board: ```bash echo '{"ts":"","worker":"{id}","type":"","data":{...}}' >> ${sessionFolder}/discoveries.ndjson ```
11. Report result: Return JSON via report_agent_job_result

Discovery Types to Share

• `code_pattern`: {name, file, description} — reusable patterns found
• `integration_point`: {file, description, exports[]} — module connection points
• `convention`: {naming, imports, formatting} — code style conventions
• `blocker`: {issue, severity, impact} — blocking issues encountered

Output (report_agent_job_result)

Return JSON: { "id": "{id}", "status": "completed" | "failed", "findings": "Key discoveries and implementation notes (max 500 chars)", "files_modified": ["path1", "path2"], "tests_passed": true | false, "acceptance_met": "Summary of which acceptance criteria were met/unmet", "error": "" }

IMPORTANT: Set status to "completed" ONLY if:

• All test cases pass
• All acceptance criteria are met Otherwise set status to "failed" with details in error field. ` }

3. Master CSV Update Helper

   function updateMasterCsvRow(sessionFolder, taskId, updates) {
     const csvPath = `${sessionFolder}/tasks.csv`
     const content = Read(csvPath)
     const lines = content.split('\n')
     const header = lines[0].split(',')
   
     for (let i = 1; i < lines.length; i++) {
       const cells = parseCsvLine(lines[i])
       if (cells[0] === taskId || cells[0] === `"${taskId}"`) {
         // Update specified columns
         for (const [col, val] of Object.entries(updates)) {
           const colIdx = header.indexOf(col)
           if (colIdx >= 0) {
             cells[colIdx] = `"${String(val).replace(/"/g, '""')}"`
           }
         }
         lines[i] = cells.join(',')
         break
       }
     }
   
     Write(csvPath, lines.join('\n'))
   }
   

Success Criteria:

• All waves executed in order
• Each wave's results merged into master CSV before next wave starts
• Dependent tasks skipped when predecessor failed
• discoveries.ndjson accumulated across all waves

Phase 3: Results Aggregation

Objective: Generate final results and human-readable report.

Steps:

1. Export results.csv

   const masterCsv = Read(`${sessionFolder}/tasks.csv`)
   // results.csv = master CSV (already has all results populated)
   Write(`${sessionFolder}/results.csv`, masterCsv)
   

2. Generate context.md

   const tasks = parseCsv(masterCsv)
   const completed = tasks.filter(t => t.status === 'completed')
   const failed = tasks.filter(t => t.status === 'failed')
   const skipped = tasks.filter(t => t.status === 'skipped')
   
   const contextContent = `# CSV Batch Execution Report
   

Session: ${sessionId} Requirement: ${requirement} Completed: ${getUtc8ISOString()} Waves: ${maxWave} | Concurrency: ${maxConcurrency}

Summary

Wave Execution

${Array.from({ length: maxWave }, (_, i) => i + 1).map(w => { const waveTasks = tasks.filter(t => parseInt(t.wave) === w) return ### Wave ${w} ${waveTasks.map(t => - [${t.id}] ${t.title}: ${t.status}${t.tests_passed ? ' ✓tests' : ''}${t.error ? ' — ' + t.error : ''} ${t.findings ? 'Findings: ' + t.findings : ''}).join('\n')} }).join('\n\n')}

Task Details

${tasks.map(t => `### ${t.id}: ${t.title}

Description: ${t.description}

Test Cases: ${t.test || 'N/A'}

Acceptance Criteria: ${t.acceptance_criteria || 'N/A'}

Hints: ${t.hints || 'N/A'}

Execution Directives: ${t.execution_directives || 'N/A'}

Findings: ${t.findings || 'N/A'}

Files Modified: ${t.files_modified || 'none'} `).join('\n---\n')}

All Modified Files

${[...new Set(tasks.flatMap(t => (t.files_modified || '').split(';')).filter(Boolean))].map(f => '- ' + f).join('\n') || 'None'} `

Write(${sessionFolder}/context.md, contextContent)

3. **Display Summary**

```javascript
console.log(`
## Execution Complete

- **Session**: ${sessionId}
- **Waves**: ${maxWave}
- **Completed**: ${completed.length}/${tasks.length}
- **Failed**: ${failed.length}
- **Skipped**: ${skipped.length}

**Results**: ${sessionFolder}/results.csv
**Report**: ${sessionFolder}/context.md
**Discoveries**: ${sessionFolder}/discoveries.ndjson
`)

4. Offer Next Steps (skip if AUTO_YES)

   if (!AUTO_YES && failed.length > 0) {
     const answer = AskUserQuestion({
       questions: [{
         question: `${failed.length} tasks failed. Next action?`,
         header: "Next Step",
         multiSelect: false,
         options: [
           { label: "Retry Failed", description: `Re-execute ${failed.length} failed tasks with updated context` },
           { label: "View Report", description: "Display context.md" },
           { label: "Done", description: "Complete session" }
         ]
       }]
     })  // BLOCKS
   
     if (answer['Next Step'] === "Retry Failed") {
       // Reset failed tasks to pending, re-run Phase 2 for their waves
       for (const task of failed) {
         updateMasterCsvRow(sessionFolder, task.id, { status: 'pending', error: '' })
       }
       // Also reset skipped tasks whose deps are now retrying
       for (const task of skipped) {
         updateMasterCsvRow(sessionFolder, task.id, { status: 'pending', error: '' })
       }
       // Re-execute Phase 2 (loop will skip already-completed tasks)
       // → goto Phase 2
     } else if (answer['Next Step'] === "View Report") {
       console.log(Read(`${sessionFolder}/context.md`))
     }
   }
   

Success Criteria:

• results.csv exported
• context.md generated
• Summary displayed to user

Shared Discovery Board Protocol

All agents across all waves share discoveries.ndjson. This eliminates redundant codebase exploration.

Lifecycle:

• Created by the first agent to write a discovery
• Carries over across waves — never cleared
• Agents append via echo '...' >> discoveries.ndjson

Format: NDJSON, each line is a self-contained JSON:

{"ts":"2026-02-28T10:00:00+08:00","worker":"1","type":"code_pattern","data":{"name":"repository-pattern","file":"src/repos/Base.ts","description":"Abstract CRUD repository"}}
{"ts":"2026-02-28T10:01:00+08:00","worker":"2","type":"integration_point","data":{"file":"src/auth/index.ts","description":"Auth module entry","exports":["authenticate","authorize"]}}

Discovery Types:

Protocol Rules:

1. Read board before own exploration → skip covered areas
2. Write discoveries immediately via echo >> → don't batch
3. Deduplicate — check existing entries; skip if same type + dedup key exists
4. Append-only — never modify or delete existing lines

Wave Computation Details

Algorithm

Kahn's BFS topological sort with depth tracking:

Input:  tasks[] with deps[]
Output: waveAssignment (taskId → wave number)

1. Build in-degree map and adjacency list from deps
2. Enqueue all tasks with in-degree 0 at wave 1
3. BFS: for each dequeued task at wave W:
   - For each dependent task D:
     - Decrement D's in-degree
     - D.wave = max(D.wave, W + 1)
     - If D's in-degree reaches 0, enqueue D
4. Any task without wave assignment → circular dependency error

Wave Properties

• Wave 1: No dependencies — all tasks in wave 1 are fully independent
• Wave N: All dependencies are in waves 1..(N-1) — guaranteed completed before wave N starts
• Within a wave: Tasks are independent of each other → safe for concurrent execution

Example

Task A (no deps)      → Wave 1
Task B (no deps)      → Wave 1
Task C (deps: A)      → Wave 2
Task D (deps: A, B)   → Wave 2
Task E (deps: C, D)   → Wave 3

Execution:
  Wave 1: [A, B]  ← concurrent
  Wave 2: [C, D]  ← concurrent, sees A+B findings
  Wave 3: [E]     ← sees A+B+C+D findings

Context Propagation Flow

Wave 1 agents:
  ├─ Execute tasks (no prev_context)
  ├─ Write findings to report_agent_job_result
  └─ Append discoveries to discoveries.ndjson

        ↓ merge results into master CSV

Wave 2 agents:
  ├─ Read discoveries.ndjson (exploration sharing)
  ├─ Read prev_context column (wave 1 findings from context_from)
  ├─ Execute tasks with full upstream context
  ├─ Write findings to report_agent_job_result
  └─ Append new discoveries to discoveries.ndjson

        ↓ merge results into master CSV

Wave 3 agents:
  ├─ Read discoveries.ndjson (accumulated from waves 1+2)
  ├─ Read prev_context column (wave 1+2 findings from context_from)
  ├─ Execute tasks
  └─ ...

Two context channels:

1. CSV findings (structured): context_from column → prev_context injection — task-specific directed context
2. NDJSON discoveries (broadcast): discoveries.ndjson — general exploration findings available to all

Error Handling

Core Rules

1. Start Immediately: First action is session initialization, then Phase 1
2. Wave Order is Sacred: Never execute wave N before wave N-1 completes and results are merged
3. CSV is Source of Truth: Master tasks.csv holds all state — always read before wave, always write after
4. Context Propagation: prev_context built from master CSV, not from memory
5. Discovery Board is Append-Only: Never clear, modify, or recreate discoveries.ndjson
6. Skip on Failure: If a dependency failed, skip the dependent task (don't attempt)
7. Cleanup Temp Files: Remove wave-{N}.csv after results are merged
8. DO NOT STOP: Continuous execution until all waves complete or all remaining tasks are skipped

Best Practices

1. Task Granularity: 3-10 tasks optimal; too many = overhead, too few = no parallelism benefit
2. Minimize Cross-Wave Deps: More tasks in wave 1 = more parallelism
3. Specific Descriptions: Agent sees only its CSV row + prev_context — make description self-contained
4. Context From ≠ Deps: deps = execution order constraint; context_from = information flow. A task can have context_from without deps (it just reads previous findings but doesn't require them to be done first in its wave)
5. Concurrency Tuning: -c 1 for serial execution (maximum context sharing); -c 8 for I/O-bound tasks

Usage Recommendations