task-runner

Task Runner

Overview

Load the plan, scrutinize it, execute tasks in batches, and pause for architect review between batches.

Core principle: Batch execution with checkpoints for architect oversight.

Announce at start: "I'm applying the task-runner skill to execute this implementation plan."

The Prime Directive

NO PLAN EXECUTION WITHOUT CRITICAL REVIEW OF EACH TASK FIRST

No exceptions. No workarounds. No shortcuts.

Plans contain errors, ambiguities, and outdated assumptions. Every task must be critically examined before execution begins. Blindly following a plan is not "efficiency" -- it is investing time in incorrect work.

When to Use

digraph applicability {
    "Written plan exists?" [shape=diamond];
    "Tasks independent?" [shape=diamond];
    "Human reviews between batches?" [shape=diamond];
    "task-runner" [shape=box];
    "delegated-execution" [shape=box];
    "Manual execution" [shape=box];

    "Written plan exists?" -> "Tasks independent?" [label="yes"];
    "Written plan exists?" -> "Manual execution" [label="no - write plan first"];
    "Tasks independent?" -> "Human reviews between batches?" [label="yes"];
    "Tasks independent?" -> "Manual execution" [label="no - tightly coupled"];
    "Human reviews between batches?" -> "task-runner" [label="yes - architect reviews"];
    "Human reviews between batches?" -> "delegated-execution" [label="no - stay in session"];
}

Use task-runner when:

• A written implementation plan exists (from task-planning or equivalent)
• Tasks should execute in batches with human review between them
• The architect wants to inspect progress and make corrections between batches
• Working in a parallel session (separate from the planning session)

Use delegated-execution instead when:

• Staying in the same session (no context switch)
• Prefer automated two-stage review (spec + quality) over human review
• Tasks are independent and can be dispatched to fresh subagents

Use manual execution when:

• No written plan exists yet
• Tasks are tightly coupled and need continuous human guidance
• Exploratory work where a plan would be premature

The Entry Protocol

Before beginning plan execution, verify ALL of these:

1. Plan exists and is accessible - You have a path to a written plan file
2. Plan has been scrutinized - You have critically read every task and flagged concerns
3. Concerns are addressed - Questions and ambiguities have been raised and answered
4. Tests pass at baseline - Existing tests pass before any changes are made
5. Branch is clean - Working in a feature branch (not main/master), no uncommitted changes
6. Worktree is configured - Using git worktree for isolation (godmode:workspace-isolation)
7. Task checklist is populated - All tasks loaded into a tracking checklist

If any gate fails, STOP. Do not proceed until every gate passes.

Cognitive Traps

Rationalization	What Is Actually True
"The plan looks fine, skip review"	Plans contain errors. Review catches them before time is wasted on incorrect work.
"I'll review while executing"	Reviewing during execution means you are already committed. Review before starting.
"This task is self-explanatory"	Self-explanatory tasks conceal assumptions. Thirty seconds of review prevents thirty minutes of rework.
"Batching is slow, I'll do everything at once"	Batches exist for human review. Skipping review lets errors compound across all tasks.
"Tests are slow, skip verification this round"	Unverified batches hide failures. Debugging later costs more.
"The plan says X but Y seems better"	Raise it with your human partner. Silently deviating from the plan creates confusion.

The Process

Step 1: Load and Scrutinize Plan

1. Read the plan file
2. Review critically -- flag any questions or concerns about each task
3. If concerns: Raise with your human partner before starting
4. If no concerns: Build a task checklist and proceed

Step 2: Execute Batch

Default batch size: 3 tasks

For each task:

1. Mark as in_progress
2. Follow each step precisely (plan provides granular steps)
3. Run verifications as specified
4. Mark as completed

Step 3: Report

When batch completes:

• Show what was implemented
• Show verification output
• Say: "Ready for feedback."

Step 4: Continue

Based on feedback:

• Apply changes if needed
• Execute next batch
• Repeat until all tasks are done

Step 5: Finalize

After all tasks are complete and verified:

• Announce: "I'm applying the merge-protocol skill to finalize this work."
• REQUIRED SUB-SKILL: Use godmode:merge-protocol
• Follow that skill to verify tests, present options, execute choice

When to Stop and Ask

STOP executing immediately when:

• You hit a blocker mid-batch (missing dependency, test failure, unclear instruction)
• The plan has critical gaps that prevent starting
• You do not understand an instruction
• Verification fails repeatedly

Ask for clarification rather than guessing.

When to Revisit Earlier Steps

Return to Scrutiny (Step 1) when:

• Partner updates the plan based on your feedback
• The fundamental approach needs rethinking

Do not force through blockers -- stop and ask.

Guardrails

Never:

• Execute a task you have not critically reviewed first
• Skip verification steps to "save time"
• Silently deviate from the plan without raising it with your human partner
• Start implementation on main/master branch without explicit user consent

Always:

• Review each task before marking it in_progress
• Run verifications as specified in the plan after each task
• Stop and ask when blocked (do not guess or improvise)
• Report the completed batch and wait for feedback before continuing

Batch Execution Diagram

digraph batch_flow {
    rankdir=TB;

    gate [label="Entry Protocol\n(all checks pass?)", shape=diamond];
    gate_fail [label="STOP\nResolve before proceeding", shape=box, style=filled, fillcolor="#ffcccc"];
    load [label="Load plan\nBuild task checklist", shape=box];
    review_task [label="Scrutinize next task", shape=box];
    concerns [label="Concerns?", shape=diamond];
    raise [label="Raise with\nhuman partner", shape=box, style=filled, fillcolor="#ffffcc"];
    execute [label="Mark in_progress\nExecute task steps\nRun verifications", shape=box];
    mark_done [label="Mark completed", shape=box];
    batch_done [label="Batch complete?\n(default: 3 tasks)", shape=diamond];
    report [label="Report results\nSay: Ready for feedback", shape=box, style=filled, fillcolor="#ccffcc"];
    feedback [label="Wait for\nhuman feedback", shape=box];
    apply [label="Apply adjustments\nif needed", shape=box];
    more [label="More tasks?", shape=diamond];
    finish [label="Invoke merge-protocol\nskill", shape=box, style=filled, fillcolor="#ccccff"];

    gate -> gate_fail [label="no"];
    gate -> load [label="yes"];
    load -> review_task;
    review_task -> concerns;
    concerns -> raise [label="yes"];
    raise -> review_task [label="resolved"];
    concerns -> execute [label="no"];
    execute -> mark_done;
    mark_done -> batch_done;
    batch_done -> review_task [label="no"];
    batch_done -> report [label="yes"];
    report -> feedback;
    feedback -> apply;
    apply -> more;
    more -> review_task [label="yes - next batch"];
    more -> finish [label="no - all done"];
}

Reminders

• Scrutinize the plan critically before starting
• Follow plan steps precisely
• Do not skip verifications
• Reference skills when the plan calls for them
• Between batches: report and wait
• Stop when blocked, do not guess
• Never start implementation on main/master without explicit user consent

Connections

Required workflow skills:

• godmode:workspace-isolation - REQUIRED: Set up isolated workspace before starting
• godmode:task-planning - Creates the plan this skill executes
• godmode:merge-protocol - Finalize development after all tasks

During execution, tasks may invoke:

• godmode:test-first - When plan tasks involve writing new code
• godmode:fault-diagnosis - When a task encounters unexpected failures
• godmode:completion-gate - Verify each batch before reporting

Alternative workflow:

• godmode:delegated-execution - Use for same-session execution with automated two-stage review instead of human-in-loop batches