Execute Parallel Task Graph

Launch parallel task workers to execute feature task graph levels

Distributing implementation work across parallel AI agents when task dependencies are known and file ownership prevents collisions.

• · Feature name (from .gsd/.current-feature or $ARGUMENTS)
• · task-graph.json with levels, dependencies, acceptance criteria
• · Worker count (default 5, max 10)
• · Execution mode: task (default) / container / subprocess

• · Parallel Task tool calls executing each level
• · Level-by-level results with pass/fail status
• · Committed changes per task with verification output
• · Final execution report (tasks completed, blocked, time)

• · Task tool (subagent execution)
• · git (branch create, commit, push)
• · Make or npm (lint/typecheck gates)

• · .gsd/specs/{FEATURE}/task-graph.json exists
• · Feature name set via /zerg:design or .gsd/.current-feature
• · All files in task ownership map are unmodified (no collisions)
• · Git repository initialized

• · No task-graph.json = run /zerg:design first
• · Feature not set = ask user for feature name
• · Task collision (file owned by multiple tasks) = abort with clear error
• · Subagent fails = retry once with error context, mark blocked if retry fails
• · Lint/typecheck gate fails = abort level, blocks downstream
• · All tasks in level fail = abort execution

• · Task mode (default) uses Claude Task tool, no external processes
• · Enforces level-by-level execution (topological ordering)
• · Runs lint/typecheck gates between levels (quality control)
• · Retry logic with error context appended to prompt
• · File ownership prevents concurrent modification conflicts