ai-goal · diff
git:20260901.dca22e5 to git:20260901.deb68d3
34 added, 15 removed. Audit A to A.
---
name: ai-goal
- description: Úsalo cuando el usuario quiere construir en bucle autónomo — "loop this feature", "run these overnight", "write the goal for this" — fijando el contrato del loop: qué consume, qué gates cierra, cuándo para y qué reporta, sin reimplementar nunca el bucle en sí.
+ description: >-
+ Use when the user wants to build in an autonomous loop — "loop this feature", "run these
+ overnight", "write the goal for this" — pinning the loop contract: what it consumes, which
+ gates it closes, when it stops, and what it reports, without ever reimplementing the loop
+ itself (the surface's native goal mode runs it). Not for one-shot verification — use
+ /ai-verify. Not for writing the plan the loop executes — use /ai-plan.
license: MIT
---
- # ai-goal — el contrato del bucle (Loop Engineering)
+ # ai-goal — the loop contract (Loop Engineering)
- El bucle NO se reimplementa: lo corre el goal nativo de la superficie. Esta skill fija el contrato del bucle a partir de tres skills fuente que cubren el método completo: setup de feature con dos paradas humanas (mock y aprobación), condición de goal para un feature, y batch en cola nocturna.
+ The loop is NOT reimplemented: the surface's native goal mode runs it. This skill pins the
+ loop contract from three source skills that cover the full method: feature setup with two
+ human stops (mock and approval), the goal condition for one feature, and an overnight
+ multi-feature queue.
- - Setup del feature: carpeta, mock aprobado, spec re-entrante → [new-feature/SKILL.md](new-feature/SKILL.md)
- - La condición `/goal` para un feature: pointer, reporting clause, met condition → [goal-writer/SKILL.md](goal-writer/SKILL.md)
- - Cola de varios features con subagent builder + adversary verificador → [feature-batch/SKILL.md](feature-batch/SKILL.md)
- - El batch usa el agente verificador adversary (frío, sin herramientas de edición) → [agents/adversary.md](agents/adversary.md)
- - Fuente: Loop-Engineering, repo de demo Loop Salon; método del bucle en sus skills `.agents/skills/` (attributed — sin autor ni URL declarados en la fuente; candidato issue H4). Otra copia en Downloads, `loop-engineering-2/`, contiene solo ai-slop-detector (ruta por ai-design, no método de bucle).
+ - Feature setup: folder, approved mock, re-entrant spec → [new-feature/SKILL.md](new-feature/SKILL.md)
+ - The goal condition for one feature: pointer, reporting clause, met condition → [goal-writer/SKILL.md](goal-writer/SKILL.md)
+ - Multi-feature queue with a builder subagent + adversary verifier → [feature-batch/SKILL.md](feature-batch/SKILL.md)
- Lo que añade ai-engineering (la costura):
+ Source: Loop-Engineering, Loop Salon demo repo (attributed — no author or URL declared
+ upstream; open license issue H4).
- 1. El bucle NO se reimplementa: lo corre el goal nativo de la superficie. ai-goal fija el contrato.
- 2. Consume plan.html, cierra gates de spec.html con receipt.
- 3. Las 3 paradas del §5.1 como condiciones de salida: contrato aprobado / loop guard / destructivo+presupuesto.
- 4. Presupuesto de config.toml.
- 5. Reporte = ai-visual-recap.
- 6. Criterio «suficiente» = ponytail/KISS.
+ ## The ai-engineering seam
+
+ 1. The loop is NOT reimplemented: the surface's native goal mode runs it. ai-goal fixes the
+ contract.
+ 2. It consumes `.ai-engineering/plan.html` and closes `.ai-engineering/spec.html` gates with
+ receipts (`ai-eng spec run`).
+ 3. The three stops of blueprint §5.1 become exit conditions: approved contract / loop guard
+ deny / destructive action above budget.
+ 4. The token/turn budget comes from `.ai-engineering/config.toml` (`[budget]`), not from the
+ conversation.
+ 5. The closing report is produced by ai-visual-recap (`.ai-engineering/recap.html`).
+ 6. The "enough" criterion is KISS/YAGNI: the smallest thing that closes the gate.
+
+ ## Routing
+
+ In scope: autonomous construction loops, overnight batches, goal contracts with gates and
+ stop conditions. Not for: diagnosing a failure (/ai-debug), verifying a finished diff
+ (/ai-verify), deciding what to build (/ai-brainstorm then /ai-plan).