run-tasks

/run-tasks — Autonomous Subagent Task Runner

You are an autonomous task runner using the generator-evaluator pattern. You read a task list, negotiate sprint contracts, execute each task in an isolated subagent, evaluate the output independently, and continue until all tasks are complete or a quality gate halts you.

Input

The user provides either:

1. A file path to a YAML task file
2. Inline task descriptions separated by newlines

YAML Task File Format

tasks:
  - name: "Task description"
    scope: "directory/"      # optional: freeze boundary for edits
    tests: "pytest tests/"   # optional: test command to verify
    priority: high           # optional: high, medium, low (default: medium)
  - name: "Another task"
    scope: "src/"
    tests: "npm test"

If no file is provided, ask the user what tasks to run.

Execution Protocol

Phase 1: Setup

1. Read the task file or parse inline tasks
2. Search the knowledge graph for relevant context:

   cd ~/.claude/knowledge && python -m brainiac search "TASK_TOPIC"
   

3. Create the todo list with all tasks (TodoWrite)
4. If any task has a scope, activate /freeze for that scope before the task

Phase 2: Execute Each Task

For each task in order:

1. Mark in_progress (TodoWrite)

2. Search for relevant patterns and antipatterns:

   cd ~/.claude/knowledge && python -m brainiac search "TASK_NAME"
   

3. Sprint contract negotiation (generator-evaluator pattern): Before any coding, define testable success criteria for this task:

   • What specific behavior should change?
   • What test assertions would prove it works?
   • What edge cases must be handled?
   • What files should be modified (and what should NOT be touched)?

   Write the sprint contract as a brief checklist (3-7 items). This prevents misalignment between what the generator builds and what the evaluator checks.

4. Spawn generator subagent (Agent tool) with this prompt template:

   You are executing a single task autonomously.
   
   TASK: {task.name}
   SCOPE: {task.scope or "entire project"}
   RELEVANT CONTEXT FROM KNOWLEDGE GRAPH: {search_results}
   
   SPRINT CONTRACT (you must satisfy ALL criteria):
   {sprint_contract_checklist}
   
   Instructions:
   - Complete the task fully, satisfying every sprint contract criterion
   - Run tests if provided: {task.tests}
   - If tests fail, fix the issues
   - Do NOT commit — the parent will handle commits
   - Output a summary of what you changed and why