prompt-to-loop-engineering

Prompt to Loop Engineering

Skill version: 3.1.0 Normative contract: Loop Engineering KB v4.0.2 Self-design graph: loop_spec.json

Mandatory execution protocol

For every invocation, the agent:

1. MUST read loop_spec.json before designing the result.
2. MUST preserve the raw request and run scripts/normalize_design_request.py to materialize a separate effective Loop_design_request. Missing capability booleans are false; missing tools are unavailable.
3. MUST use the effective request—not the raw request—to generate exactly one loop_design_result and save it as JSON.
4. MUST run scripts/validate_design_result.py with the result, raw request, effective request, and normalization report before returning the result.
5. MUST NOT emit spec_ready when validation fails. Correct and revalidate the design, or return a non-executable disposition with the validation errors preserved.
6. MUST NOT execute the user task, invoke a generated node or tool, advance a generated edge, or report runtime success during design validation.
7. MUST NOT create or require an independent Runtime Engine. Codex is the host executor when the user explicitly asks to use or continue a generated scaffold.

Example validation command:

python scripts/validate_design_result.py path/to/loop_design_result.json \
  --request path/to/effective_request.json \
  --raw-request path/to/raw_request.json \
  --normalization-report path/to/request_normalization_report.json

Example normalization command:

python scripts/normalize_design_request.py path/to/raw_request.json \
  --output path/to/effective_request.json \
  --report path/to/request_normalization_report.json

This Skill contains no Runtime Engine and must never scaffold one. Runtime capabilities are constraints supplied by the caller so generated LoopSpec and Agent Config Scaffold files match the real Codex session and project permissions.

v2.0.0 — Release-Hardened Contract Alignment