2504 Design D170c10c

Technical Design: Bite-Sized Task Planning

Metadata

• Feature: bite-sized-planning
• Status: DRAFT
• Created: 2026-02-04
• GitHub Issues: #65, #119

1. Overview

1.1 Summary

Add step-level granularity to ZERG task planning. Each task can contain ordered steps following TDD methodology. Workers execute steps strictly in sequence with exit-code verification. Includes auto-detection of project formatters to ensure clean commits.

1.2 Goals

• Enable bite-sized, mechanically executable task steps
• Enforce TDD workflow (test-first, verify failure, implement, verify pass)
• Auto-detect and run formatters before commits
• Track step-level progress in status UI
• Adapt detail level based on worker familiarity

1.3 Non-Goals

• LLM-generated code snippets (AST patterns only)
• Regex matching on command output
• Worker deviation from step sequence

2. Architecture

2.1 High-Level Design

┌─────────────────┐     ┌──────────────────┐     ┌─────────────────┐
│  /zerg:design   │────▶│  StepGenerator   │────▶│ task-graph.json │
│  --detail high  │     │  + ASTAnalyzer   │     │ (with steps[])  │
└─────────────────┘     └──────────────────┘     └─────────────────┘
                                │
                                ▼
                        ┌──────────────────┐
                        │ FormatterDetector│
                        │ (pyproject, etc) │
                        └──────────────────┘

┌─────────────────┐     ┌──────────────────┐     ┌─────────────────┐
│     Worker      │────▶│  StepExecutor    │────▶│   Heartbeat     │
│   (strict)      │     │  (verify exit)   │     │ (step progress) │
└─────────────────┘     └──────────────────┘     └─────────────────┘

┌─────────────────┐     ┌──────────────────┐
│  /zerg:status   │────▶│ StepProgressUI   │
│                 │     │ [✅✅🔄⏳⏳]     │
└─────────────────┘     └──────────────────┘

2.2 Component Breakdown

2.3 Data Flow

1. User runs /zerg:design --detail high
2. StepGenerator loads task-graph template
3. For each task:
   • ASTAnalyzer extracts patterns from related files
   • StepGenerator creates TDD step sequence
   • FormatterDetector adds format step with detected tool
4. Output: task-graph.json with steps[] per task
5. Worker loads task, StepExecutor runs each step
6. Heartbeat updates with current step number
7. /zerg:status renders step progress visually

3. Detailed Design

3.1 Step Schema (JSON Schema Extension)

{
  "step": {
    "type": "object",
    "required": ["step", "action", "run"],
    "properties": {
      "step": {
        "type": "integer",
        "minimum": 1,
        "description": "Step number in sequence"
      },
      "action": {
        "type": "string",
        "enum": ["write_test", "verify_fail", "implement", "verify_pass", "format", "commit"],
        "description": "Step action type"
      },
      "file": {
        "type": "string",
        "description": "Primary file for this step"
      },
      "code_snippet": {
        "type": "string",
        "description": "AST-generated code hint (high detail only)"
      },
      "run": {
        "type": "string",
        "description": "Command to execute"
      },
      "verify": {
        "type": "string",
        "enum": ["exit_code", "exit_code_nonzero", "none"],
        "default": "exit_code",
        "description": "Verification method"
      }
    }
  }
}

3.2 TDD Step Template

For each task, generate 6 steps:

3.3 FormatterDetector API

@dataclass
class FormatterConfig:
    name: str              # e.g., "ruff"
    format_cmd: str        # e.g., "ruff format {files}"
    fix_cmd: str | None    # e.g., "ruff check --fix {files}"
    file_patterns: list[str]  # e.g., ["*.py"]

class FormatterDetector:
    def detect(self, repo_path: Path) -> list[FormatterConfig]:
        """Detect formatters from project config files."""

    def detect_for_files(self, files: list[str]) -> FormatterConfig | None:
        """Get formatter for specific file extensions."""

Detection sources:

• pyproject.toml: [tool.ruff], [tool.black], [tool.isort]
• package.json: prettier, eslint
• .prettierrc, .prettierrc.json, .prettierrc.yaml
• rustfmt.toml, .rustfmt.toml
• .clang-format
• Presence of go.mod → gofmt

3.4 ASTAnalyzer API

class ASTAnalyzer:
    def __init__(self, cache: ASTCache):
        self.cache = cache

    def extract_patterns(self, file_path: Path) -> CodePatterns:
        """Extract import patterns, base classes, utilities from a file."""

    def generate_test_snippet(self,
                              target_file: Path,
                              function_name: str) -> str:
        """Generate test snippet matching project conventions."""

    def generate_impl_snippet(self,
                              target_file: Path,
                              based_on: list[Path]) -> str:
        """Generate implementation snippet using patterns from similar files."""

3.5 AdaptiveDetail Logic

@dataclass
class DetailMetrics:
    file_modifications: dict[str, int]  # file -> modification count
    task_results: dict[str, bool]       # task_id -> success

class AdaptiveDetail:
    def should_reduce_detail(self,
                             task: Task,
                             metrics: DetailMetrics,
                             config: PlanningConfig) -> bool:
        # Check file familiarity
        for file in task.files.modify + task.files.create:
            if metrics.file_modifications.get(file, 0) >= config.adaptive_familiarity_threshold:
                return True

        # Check success rate in same directory
        related_tasks = [t for t in metrics.task_results
                        if self._same_area(t, task)]
        if related_tasks:
            success_rate = sum(related_tasks.values()) / len(related_tasks)
            if success_rate >= config.adaptive_success_threshold:
                return True

        return False

3.6 StepExecutor Protocol

class StepExecutor:
    def execute_task(self, task: Task, worker_id: int) -> TaskResult:
        if not task.steps:
            return self._execute_classic(task)

        for step in task.steps:
            self._update_heartbeat(worker_id, task.id, step.step, len(task.steps))

            if step.action == "write_test":
                self._write_file(step.file, step.code_snippet)
            elif step.action == "implement":
                self._write_file(step.file, step.code_snippet)

            if step.run:
                result = subprocess.run(step.run, shell=True)

                if step.verify == "exit_code" and result.returncode != 0:
                    return TaskResult(success=False, failed_step=step.step)
                elif step.verify == "exit_code_nonzero" and result.returncode == 0:
                    return TaskResult(success=False, failed_step=step.step,
                                     error="Expected failure but got success")

        return TaskResult(success=True)

4. Key Decisions

4.1 Decision: Exit Code Only for Verification

Context: Original proposal used regex matching on command output.

Options Considered:

1. Regex on stdout — flexible but brittle across tools
2. Exit code only — universal and reliable
3. Hybrid — exit code default, optional regex

Decision: Exit code only (Option 2)

Rationale: Test frameworks, linters, and build tools universally use exit codes. Regex patterns would need maintenance per tool version.

4.2 Decision: Strict Step Protocol

Context: Workers could follow steps strictly or use them as guidance.

Options Considered:

1. Strict — follow exactly, no deviation
2. Guidance — suggestions only, worker autonomy
3. Verified checkpoints — can deviate but must pass each step

Decision: Strict protocol (Option 1)

Rationale: Bite-sized steps are designed to be mechanical. Deviation introduces unpredictability. If steps are wrong, fix the generator.

4.3 Decision: AST-Based Snippet Generation

Context: Code snippets could be templates, AST-generated, or LLM-generated.

Options Considered:

1. Template — placeholders like {class_name}
2. AST-aware — extract real patterns from codebase
3. LLM — Claude generates snippets during design

Decision: AST-aware (Option 2)

Rationale: Templates are generic. LLM adds latency and cost to design phase. AST extraction produces realistic snippets without external calls.

5. Implementation Plan

5.1 Phase Summary

5.2 File Ownership

5.3 Dependency Graph

BITE-L1-001 (FormatterDetector) ─┬─▶ BITE-L2-001 (StepGenerator) ─┬─▶ BITE-L3-001 (design.py)
BITE-L1-002 (PlanningConfig)    ─┤                                │
BITE-L1-003 (Schema)            ─┘                                ├─▶ BITE-L3-002 (status.py)
                                                                  │
BITE-L2-002 (ASTAnalyzer)      ──▶ BITE-L2-001                   ├─▶ BITE-L3-003 (worker.md)
                                                                  │
BITE-L2-003 (AdaptiveDetail)   ──▶ BITE-L2-001                   │
                                                                  │
BITE-L2-004 (StepExecutor)     ─────────────────────────────────▶┘
                                                                  │
                                                                  ▼
                                                            BITE-L4-001 (Unit Tests)
                                                                  │
                                                                  ▼
                                                            BITE-L4-002 (Integration)
                                                                  │
                                                                  ▼
                                                            BITE-L5-001 (Quality)

6. Risk Assessment

7. Testing Strategy

7.1 Unit Tests

• StepGenerator: TDD sequence generation, step ordering
• FormatterDetector: Detection from each config file type
• ASTAnalyzer: Pattern extraction, snippet generation
• AdaptiveDetail: Threshold logic, metric tracking

7.2 Integration Tests

• End-to-end: design → rush → worker step execution
• Formatter integration: pre-commit hooks pass
• Status UI: step progress rendering

8. Parallel Execution Notes

8.1 Recommended Workers

• Minimum: 2 workers
• Optimal: 3 workers
• Maximum: 4 workers (L1 has 3 parallel tasks)

8.2 Estimated Duration

• Single worker: ~170m
• With 3 workers: ~85m
• Speedup: 2x

9. Approval