SuperAPI AI Directory

delegate

Agents & models

View repository

Updated 10-16 03:04

About this skill

Delegate

The parent owns the complete requested outcome, decomposition, decisions, review, validation, and final response. GPT-5.6 Luna workers perform substantive implementation in narrow slices. The parent may directly make a clearly correct, localized, low-risk edit when delegation and review would cost more than making and validating it.

Define the next slice

Inspect the relevant repository state and instructions, and resolve consequential product or authority decisions that the worker is not authorized to make. Decompose broad features, refactors, migrations, and multi-surface requirements; never transfer the end-to-end requirement to one worker.

Each slice needs one concrete outcome, bounded behavior and file ownership, constraints, dependencies, acceptance criteria, and independently reviewable validation. Size by behavior, not line count. A source change and directly coupled regression test may share a slice when separating them would make validation meaningless; this does not justify a broad assignment.

Dispatch with complete context

Use one primary implementer per slice with these exact settings:

  • model: "gpt-5.6-luna"
  • reasoning_effort: "max"
  • fork_turns: "none"

The self-contained prompt must reproduce the original user request verbatim and include, without paraphrasing, compressing, or omitting, every known requirement, constraint, acceptance criterion, decision, failure or error output, repository fact, and integration dependency that can affect this slice. Include exact files to read in full when needed. Omit unrelated background, but do not replace result-affecting context with a vague summary.

Include the working directory, slice contract, validation, and how this narrow assignment contributes to the larger outcome without owning it. Tell the worker to inspect before editing, preserve unrelated work, avoid commits or publication unless authorized, and return changed files, validation, and unresolved issues.

Use one writer for overlapping files or behavior. Parallelize only independent slices whose ownership cannot overlap; never edit worker-owned files concurrently. If the required model or reasoning setting is unavailable, report the limitation and preserve the current state instead of silently substituting.

Verify and finish

Inspect each worker's diff and validation against its contract and repository instructions. Independently review and validate a slice before dispatching anything dependent on it; a success report alone is insufficient.

Send material defects or missing coverage back to the same worker as a precise, bounded correction with the evidence and required validation. The parent may apply an obvious localized correction directly and validate it; return broader or judgment-heavy rework to the worker. Continue through the authorized slices until the complete requested behavior and required checks pass, or a real blocker prevents further progress.

Report the verified outcome, important decisions, changed files, validation, and remaining limitations across all slices.