Atomic Execution Engine

Prevents context degradation during long execution phases by breaking work down into atomic, verification-gated tasks using structured XML. Integrated via /steal from the get-shit-done methodology.

Triggers

"execute roadmap", "start building phase", "implement this plan", "atomic execution"

Core Mechanics

When entering Execution phase, do not build monolithic features linearly in one context window. Instead:

1. XML Plan Extraction

Convert the current phase of the ROADMAP.md (or the design.md spec) into 1 or more atomic XML execution plans. If the user hasn't provided one, draft it and execute it.

Format each atomic task precisely as:

<task type="auto">
  <name>Short descriptive name</name>
  <files>path/to/affected/file.ext</files>
  <action>
    Explicit, step-by-step instructions.
    Dependencies and logic.
  </action>
  <verify>Command to run to verify correctness (e.g. tests or curl)</verify>
  <done>Definition of done</done>
</task>

2. Isolated Execution

Execute ONE <task> block at a time. Maintain peak reasoning by not polluting the context window with previous tasks. If a task requires subagents or isolated worktrees, trigger /git-worktree-swarm.

3. Verification Gate

After code is written, explicitly run the command specified in <verify>. If it passes the <done> criteria, commit immediately using micro-commit. If it fails, fix the code within the same context before proceeding to the next <task>.

4. State Management

Update STATE.md after every completed XML task block to persist progress across sessions without bloating the context window. Always log:

• Decisions made
• Blockers encountered
• Current position in ROADMAP.md

Reference Paths

• spec-driven-dev (Pre-requisite for Execution via P107)
• micro-commit (Post-Execution step)