feat(ml): cierre del bucle de mejora comfyui-skill (genera→juzga→bump)
Tres funciones nuevas que cierran el lazo skill→generación→juicio→promoción del grupo comfyui-skill (issue 0087): - comfyui_bump_skill_version (impura): promueve una versión nueva SOLO si el score del panel-juez sube (gate objetivo). Snapshot versions/vN.json pre-mutación, deep-merge de recipe_patch, semver↑, línea en growth_log.jsonl. force=True salta el gate. No usa datetime.now(). - comfyui_update_skill_score (impura): media incremental de score_mean/score_n reescribiendo recipe.json in-place (sin snapshot ni growth_log). - comfyui_generate_with_skill_oneshot (pipeline): one-shot load→build→submit→ wait→fetch→judge→score_mean. recipe_patch prueba variantes sin guardar score. Compone 7 funciones del registry. Tests offline: 11 passed (gate, semver, deep-merge, media incremental, errores). Página madre docs/capabilities/comfyui-skill.md: +3 funciones, sección "Bucle de mejora" con diagrama, fronteras de scoring actualizadas. Demo real verificada: skill seed portrait_cinematic_sd15 (SD1.5) generó imagen SFW real, el panel la juzgó, una variante puntuó más alto (4.787 > 4.7276) y el gate promovió v1.0.0→v1.1.0 con el judge_run_id como evidencia. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
@@ -0,0 +1,90 @@
|
||||
---
|
||||
name: comfyui_bump_skill_version
|
||||
kind: function
|
||||
lang: py
|
||||
domain: ml
|
||||
version: "1.0.0"
|
||||
purity: impure
|
||||
signature: "def comfyui_bump_skill_version(slug: str, change: str, *, score_before: float, score_after: float, judge_run_id: str = None, recipe_patch: dict = None, force: bool = False, bump: str = \"minor\", library_dir: str = None, timestamp: float = None) -> dict"
|
||||
description: "Promueve una version nueva de una skill ComfyUI SOLO si el score sube (gate objetivo del bucle de mejora, grupo comfyui-skill). Si score_after <= score_before y no force, rechaza con ok=False sin tocar nada. Si pasa: snapshot pre-mutacion versions/vN.json, aplica recipe_patch (deep-merge) a recipe.json, sube el semver (minor default), y appende a growth_log.jsonl una linea {version,date,change,score_before,score_after,judge_run_id,diff}. library_dir default ~/ComfyUI/skills_library. Slug inexistente -> ok=False. No usa datetime.now (deriva la fecha del timestamp recibido o time.time). Impura: disco. Nunca lanza."
|
||||
error_type: error_py_core
|
||||
tags: [comfyui, comfyui-skill, ml, skill, versioning, growth]
|
||||
uses_functions: []
|
||||
uses_types: []
|
||||
returns: []
|
||||
returns_optional: false
|
||||
imports: []
|
||||
params:
|
||||
- name: slug
|
||||
desc: "Slug de la skill (su carpeta en la libreria)."
|
||||
- name: change
|
||||
desc: "Descripcion de una linea del cambio que motiva el bump (va al growth_log)."
|
||||
- name: score_before
|
||||
desc: "Score 0-10 de la version actual (del panel comfyui-judge). keyword-only."
|
||||
- name: score_after
|
||||
desc: "Score 0-10 de la variante candidata. Debe ser > score_before salvo force=True. keyword-only."
|
||||
- name: judge_run_id
|
||||
desc: "Identificador de la corrida del juez que justifica el bump (evidencia trazable). keyword-only."
|
||||
- name: recipe_patch
|
||||
desc: "Dict con los cambios a aplicar sobre la receta (deep-merge). Ej. {'params': {'steps': 32}}. keyword-only."
|
||||
- name: force
|
||||
desc: "Si True, salta el gate y promueve aunque el score no mejore. keyword-only."
|
||||
- name: bump
|
||||
desc: "Parte del semver a subir: 'minor' (default), 'major' o 'patch'. keyword-only."
|
||||
- name: library_dir
|
||||
desc: "Raiz de la libreria. Default ~/ComfyUI/skills_library. keyword-only."
|
||||
- name: timestamp
|
||||
desc: "Epoch en segundos para la fecha del growth_log; None = time.time(). keyword-only."
|
||||
output: "dict {ok, slug, old_version, new_version, snapshot_file, growth_entry, recipe_path, error}. ok=False con error si el gate rechaza el bump, si la skill no existe, o si falla la escritura; nunca lanza."
|
||||
tested: true
|
||||
tests: [test_semver_helper, test_deep_merge_no_pisa_otras_claves, test_golden_promueve_cuando_score_sube, test_edge_major_y_patch, test_error_gate_bloquea_si_no_mejora, test_error_force_salta_gate, test_error_skill_inexistente]
|
||||
test_file_path: "python/functions/ml/tests/test_comfyui_bump_skill_version.py"
|
||||
file_path: "python/functions/ml/comfyui_bump_skill_version.py"
|
||||
---
|
||||
|
||||
# comfyui_bump_skill_version
|
||||
|
||||
Pieza de cierre del bucle de mejora del grupo [`comfyui-skill`](../../../docs/capabilities/comfyui-skill.md):
|
||||
una skill genera, el panel [`comfyui-judge`](../../../docs/capabilities/comfyui-judge.md) la
|
||||
puntúa, y esta función promueve una versión nueva **solo si el score sube**. El juez decide, no
|
||||
el humano. Es el "crecimiento por composición" del issue 0087 aplicado a la generación de imágenes.
|
||||
|
||||
## Ejemplo
|
||||
|
||||
```python
|
||||
import sys, os
|
||||
sys.path.insert(0, os.path.join(os.environ["HOME"], "fn_registry", "python", "functions"))
|
||||
from ml.comfyui_bump_skill_version import comfyui_bump_skill_version
|
||||
|
||||
# La variante con steps=32 puntuó 7.4 vs 6.5 de la versión vigente → se promueve.
|
||||
res = comfyui_bump_skill_version(
|
||||
"portrait_cinematic_sd15", "subir steps 28→32 (mejor detalle facial)",
|
||||
score_before=6.5, score_after=7.4,
|
||||
judge_run_id="judge_abc123", recipe_patch={"params": {"steps": 32}},
|
||||
)
|
||||
print(res["old_version"], "→", res["new_version"]) # 1.0.0 → 1.1.0
|
||||
# recipe.json ahora en 1.1.0 con steps=32; versions/vN.json conserva la 1.0.0;
|
||||
# growth_log.jsonl tiene una línea con score_before/after + judge_run_id.
|
||||
```
|
||||
|
||||
## Cuando usarla
|
||||
|
||||
Tras juzgar una variante de una skill: si el `score` del panel supera al de la versión vigente,
|
||||
llama a esta función para **promover** la mejora (snapshot + semver + growth_log). Si el score no
|
||||
sube, NO la llames (o se rechaza por el gate) — esa es justo la garantía del bucle. Pásale el
|
||||
`judge_run_id` de `comfyui_generate_with_skill_oneshot` como evidencia trazable.
|
||||
|
||||
## Gotchas
|
||||
|
||||
- **Gate duro**: `score_after <= score_before` sin `force=True` devuelve `{ok:False}` y NO toca
|
||||
nada (ni snapshot, ni growth_log, ni versión). Es el comportamiento deseado, no un error.
|
||||
- **`force=True`** salta el gate (promueve aunque empeore) y marca `growth_entry.forced=True`.
|
||||
Úsalo solo para correcciones manuales, no en el bucle automático.
|
||||
- **`recipe_patch` es deep-merge**: dicts anidados (`params`) se fusionan; listas (`loras`,
|
||||
`blocks`) se reemplazan enteras (no se concatenan).
|
||||
- **Snapshot pre-mutación**: `versions/vN.json` guarda la receta ANTES del patch; el `recipe.json`
|
||||
queda con la versión nueva. Recupera la vieja con `comfyui_load_skill(slug, version=N)`.
|
||||
- **Fecha sin `datetime.now()`**: se deriva de `timestamp` o `time.time()` vía `time.strftime`
|
||||
(compatible con entornos que prohíben `datetime.now()`).
|
||||
- **No genera ni juzga**: solo promueve la receta. Generar + puntuar es trabajo de
|
||||
`comfyui_generate_with_skill_oneshot` + `comfyui_judge_image`.
|
||||
@@ -0,0 +1,224 @@
|
||||
"""comfyui_bump_skill_version — promueve una nueva versión de una *skill* ComfyUI.
|
||||
|
||||
Cierra el bucle de mejora del grupo `comfyui-skill`: una skill genera, el panel
|
||||
`comfyui-judge` la puntúa y, **solo si el score sube**, esta función promociona una
|
||||
versión nueva de la receta. Es la pieza de "crecimiento por composición" del issue 0087
|
||||
aplicada a la generación de imágenes — el catálogo de recetas crece registrando mejoras
|
||||
medibles, no inflando recetas a ciegas.
|
||||
|
||||
Qué hace, en orden:
|
||||
|
||||
1. **Gate objetivo**: si ``score_after <= score_before`` y no se pasa ``force=True``,
|
||||
rechaza con ``{ok: False}`` sin tocar nada. El juez (no el humano) decide.
|
||||
2. **Snapshot pre-mutación**: escribe ``versions/vN.json`` con la receta ACTUAL antes de
|
||||
cambiarla (backup recuperable con ``comfyui_load_skill(slug, version=N)``).
|
||||
3. **Aplica ``recipe_patch``** (deep-merge) sobre ``recipe.json``.
|
||||
4. **Sube el semver** de la receta (``minor`` por defecto: 1.0.0 → 1.1.0).
|
||||
5. **Append a ``growth_log.jsonl``**: una línea
|
||||
``{version, date, change, score_before, score_after, judge_run_id, diff}``.
|
||||
|
||||
`library_dir` por defecto ``~/ComfyUI/skills_library``. Slug inexistente → ``{ok: False}``.
|
||||
|
||||
Impura: lee y escribe archivos en disco. Sin red. No usa ``datetime.now()`` (prohibido en
|
||||
algunos entornos) — la fecha se deriva del ``timestamp`` recibido o de ``time.time()``.
|
||||
"""
|
||||
|
||||
import json
|
||||
import os
|
||||
import time
|
||||
|
||||
DEFAULT_LIBRARY = "~/ComfyUI/skills_library"
|
||||
|
||||
|
||||
def _lib_dir(library_dir):
|
||||
return os.path.expanduser(library_dir or DEFAULT_LIBRARY)
|
||||
|
||||
|
||||
def _bump_semver(version: str, part: str) -> str:
|
||||
"""Sube un semver ``X.Y.Z`` por ``major``/``minor``/``patch``.
|
||||
|
||||
Tolerante a versiones mal formadas: si ``version`` no parsea como tres enteros,
|
||||
parte de ``0.0.0`` antes de aplicar el incremento.
|
||||
"""
|
||||
try:
|
||||
nums = (list(map(int, str(version).split("."))) + [0, 0, 0])[:3]
|
||||
major, minor, patch = nums
|
||||
except (ValueError, TypeError):
|
||||
major, minor, patch = 0, 0, 0
|
||||
if part == "major":
|
||||
return f"{major + 1}.0.0"
|
||||
if part == "patch":
|
||||
return f"{major}.{minor}.{patch + 1}"
|
||||
# minor (default)
|
||||
return f"{major}.{minor + 1}.0"
|
||||
|
||||
|
||||
def _deep_merge(base: dict, patch: dict) -> dict:
|
||||
"""Merge recursivo de ``patch`` sobre ``base`` sin mutar los originales.
|
||||
|
||||
Las claves cuyo valor es dict en ambos se fusionan en profundidad; cualquier otro
|
||||
tipo (incluida una lista, p.ej. ``loras``/``blocks``) se reemplaza entero.
|
||||
"""
|
||||
out = dict(base)
|
||||
for key, val in (patch or {}).items():
|
||||
if isinstance(val, dict) and isinstance(out.get(key), dict):
|
||||
out[key] = _deep_merge(out[key], val)
|
||||
else:
|
||||
out[key] = val
|
||||
return out
|
||||
|
||||
|
||||
def _next_version_index(versions_dir: str) -> int:
|
||||
"""Siguiente N para ``versions/vN.json`` (1 + cuántos snapshots ya hay)."""
|
||||
try:
|
||||
existing = [f for f in os.listdir(versions_dir)
|
||||
if f.startswith("v") and f.endswith(".json")]
|
||||
except OSError:
|
||||
existing = []
|
||||
return len(existing) + 1
|
||||
|
||||
|
||||
def comfyui_bump_skill_version(
|
||||
slug: str,
|
||||
change: str,
|
||||
*,
|
||||
score_before: float,
|
||||
score_after: float,
|
||||
judge_run_id: str = None,
|
||||
recipe_patch: dict = None,
|
||||
force: bool = False,
|
||||
bump: str = "minor",
|
||||
library_dir: str = None,
|
||||
timestamp: float = None,
|
||||
) -> dict:
|
||||
"""Promueve una versión nueva de una skill si el score mejora (gate objetivo).
|
||||
|
||||
Args:
|
||||
slug: slug de la skill (su carpeta en la librería).
|
||||
change: descripción de una línea del cambio que motiva el bump (va al growth_log).
|
||||
score_before: score de la versión actual (del panel-juez). keyword-only.
|
||||
score_after: score de la variante candidata. Debe ser ``> score_before`` salvo
|
||||
``force=True``. keyword-only.
|
||||
judge_run_id: identificador de la corrida del juez que justifica el bump
|
||||
(evidencia trazable). keyword-only.
|
||||
recipe_patch: dict con los cambios a aplicar sobre la receta (deep-merge). Por
|
||||
ejemplo ``{"params": {"steps": 32}}``. keyword-only.
|
||||
force: si True, salta el gate y promueve aunque el score no mejore. keyword-only.
|
||||
bump: parte del semver a subir: ``minor`` (default), ``major`` o ``patch``.
|
||||
keyword-only.
|
||||
library_dir: raíz de la librería. Default ``~/ComfyUI/skills_library``. keyword-only.
|
||||
timestamp: epoch en segundos para la fecha del growth_log; None = ``time.time()``.
|
||||
keyword-only.
|
||||
|
||||
Returns:
|
||||
dict ``{ok, slug, old_version, new_version, snapshot_file, growth_entry,
|
||||
recipe_path, error}``. ``ok=False`` con ``error`` si el gate rechaza el bump, si
|
||||
la skill no existe, o si falla la escritura; nunca lanza.
|
||||
"""
|
||||
if not slug or not isinstance(slug, str):
|
||||
return {"ok": False, "slug": slug, "old_version": "", "new_version": "",
|
||||
"snapshot_file": "", "growth_entry": None, "recipe_path": "",
|
||||
"error": "slug requerido (string no vacío)"}
|
||||
|
||||
# 1. Gate objetivo: el juez decide, no el humano.
|
||||
try:
|
||||
sb = float(score_before)
|
||||
sa = float(score_after)
|
||||
except (TypeError, ValueError):
|
||||
return {"ok": False, "slug": slug, "old_version": "", "new_version": "",
|
||||
"snapshot_file": "", "growth_entry": None, "recipe_path": "",
|
||||
"error": f"score_before/score_after deben ser numéricos "
|
||||
f"(recibido {score_before!r}, {score_after!r})"}
|
||||
if sa <= sb and not force:
|
||||
return {"ok": False, "slug": slug, "old_version": "", "new_version": "",
|
||||
"snapshot_file": "", "growth_entry": None, "recipe_path": "",
|
||||
"error": f"gate: score_after ({sa}) no supera score_before ({sb}); "
|
||||
f"no se promueve (usa force=True para forzar)"}
|
||||
|
||||
lib = _lib_dir(library_dir)
|
||||
skill_dir = os.path.join(lib, slug)
|
||||
recipe_path = os.path.join(skill_dir, "recipe.json")
|
||||
versions_dir = os.path.join(skill_dir, "versions")
|
||||
|
||||
if not os.path.isfile(recipe_path):
|
||||
return {"ok": False, "slug": slug, "old_version": "", "new_version": "",
|
||||
"snapshot_file": "", "growth_entry": None, "recipe_path": recipe_path,
|
||||
"error": f"skill no encontrada: {slug!r} (sin recipe.json en {skill_dir})"}
|
||||
|
||||
try:
|
||||
with open(recipe_path, encoding="utf-8") as fh:
|
||||
recipe = json.load(fh)
|
||||
except (OSError, json.JSONDecodeError) as exc:
|
||||
return {"ok": False, "slug": slug, "old_version": "", "new_version": "",
|
||||
"snapshot_file": "", "growth_entry": None, "recipe_path": recipe_path,
|
||||
"error": f"no se pudo leer la receta: {exc}"}
|
||||
|
||||
old_version = recipe.get("version", "0.0.0")
|
||||
new_version = _bump_semver(old_version, bump)
|
||||
ts = float(timestamp) if timestamp is not None else time.time()
|
||||
date = time.strftime("%Y-%m-%d", time.gmtime(ts))
|
||||
|
||||
try:
|
||||
os.makedirs(versions_dir, exist_ok=True)
|
||||
# 2. Snapshot pre-mutación: preserva la receta actual antes de cambiarla.
|
||||
n = _next_version_index(versions_dir)
|
||||
snapshot_file = os.path.join(versions_dir, f"v{n}.json")
|
||||
with open(snapshot_file, "w", encoding="utf-8") as fh:
|
||||
json.dump(recipe, fh, indent=2, ensure_ascii=False)
|
||||
|
||||
# 3. Aplicar el patch (deep-merge) + 4. subir el semver.
|
||||
new_recipe = _deep_merge(recipe, recipe_patch or {})
|
||||
new_recipe["version"] = new_version
|
||||
with open(recipe_path, "w", encoding="utf-8") as fh:
|
||||
json.dump(new_recipe, fh, indent=2, ensure_ascii=False)
|
||||
|
||||
# 5. Bitácora append-only del crecimiento.
|
||||
growth_entry = {
|
||||
"version": new_version,
|
||||
"date": date,
|
||||
"ts": int(ts),
|
||||
"change": change,
|
||||
"score_before": sb,
|
||||
"score_after": sa,
|
||||
"judge_run_id": judge_run_id or "",
|
||||
"diff": recipe_patch or {},
|
||||
"forced": bool(force and sa <= sb),
|
||||
}
|
||||
with open(os.path.join(skill_dir, "growth_log.jsonl"), "a", encoding="utf-8") as fh:
|
||||
fh.write(json.dumps(growth_entry, ensure_ascii=False) + "\n")
|
||||
except OSError as exc:
|
||||
return {"ok": False, "slug": slug, "old_version": old_version, "new_version": "",
|
||||
"snapshot_file": "", "growth_entry": None, "recipe_path": recipe_path,
|
||||
"error": f"fallo de escritura: {exc}"}
|
||||
|
||||
return {"ok": True, "slug": slug, "old_version": old_version,
|
||||
"new_version": new_version, "snapshot_file": snapshot_file,
|
||||
"growth_entry": growth_entry, "recipe_path": recipe_path, "error": ""}
|
||||
|
||||
|
||||
# Alias con el nombre completo del ID para descubrimiento por convención.
|
||||
bump_skill_version = comfyui_bump_skill_version
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
import sys
|
||||
|
||||
# Demo offline contra una librería temporal.
|
||||
demo_lib = "/tmp/skills_bump_demo"
|
||||
sdir = os.path.join(demo_lib, "demo_skill")
|
||||
os.makedirs(os.path.join(sdir, "versions"), exist_ok=True)
|
||||
with open(os.path.join(sdir, "recipe.json"), "w", encoding="utf-8") as f:
|
||||
json.dump({"schema_version": 1, "slug": "demo_skill", "version": "1.0.0",
|
||||
"base_workflow": "txt2img", "params": {"steps": 28}}, f, indent=2)
|
||||
|
||||
# Gate bloquea cuando no mejora.
|
||||
blocked = comfyui_bump_skill_version("demo_skill", "subir steps", score_before=7.0,
|
||||
score_after=6.5, library_dir=demo_lib)
|
||||
print("gate bloquea:", blocked["ok"], "->", blocked["error"], file=sys.stderr)
|
||||
|
||||
# Promueve cuando mejora.
|
||||
ok = comfyui_bump_skill_version("demo_skill", "subir steps a 32", score_before=6.5,
|
||||
score_after=7.4, judge_run_id="judge_abc",
|
||||
recipe_patch={"params": {"steps": 32}},
|
||||
library_dir=demo_lib)
|
||||
print(json.dumps(ok, indent=2, ensure_ascii=False))
|
||||
@@ -0,0 +1,64 @@
|
||||
---
|
||||
name: comfyui_update_skill_score
|
||||
kind: function
|
||||
lang: py
|
||||
domain: ml
|
||||
version: "1.0.0"
|
||||
purity: impure
|
||||
signature: "def comfyui_update_skill_score(slug: str, new_score: float, *, library_dir: str = None) -> dict"
|
||||
description: "Acumula el score de un juicio en la media de una skill ComfyUI (score_mean/score_n) por media incremental, reescribiendo recipe.json en sitio. A diferencia de comfyui_save_skill NO crea snapshot versions/vN.json ni entrada en growth_log: el score es telemetria de la version vigente, no una version nueva (los snapshots los reserva comfyui_bump_skill_version). library_dir default ~/ComfyUI/skills_library. Slug inexistente o score no numerico -> ok=False. Impura: disco. Nunca lanza."
|
||||
error_type: error_py_core
|
||||
tags: [comfyui, comfyui-skill, ml, skill, scoring]
|
||||
uses_functions: []
|
||||
uses_types: []
|
||||
returns: []
|
||||
returns_optional: false
|
||||
imports: []
|
||||
params:
|
||||
- name: slug
|
||||
desc: "Slug de la skill (su carpeta en la libreria)."
|
||||
- name: new_score
|
||||
desc: "Score 0-10 del ultimo juicio (p.ej. comfyui_judge_image(...)['score'])."
|
||||
- name: library_dir
|
||||
desc: "Raiz de la libreria. Default ~/ComfyUI/skills_library. keyword-only."
|
||||
output: "dict {ok, slug, score_mean, score_n, prev_score_mean, prev_score_n, error}. En exito score_mean/score_n son los valores actualizados. Slug inexistente / score no numerico / fallo de escritura -> ok=False; nunca lanza."
|
||||
tested: true
|
||||
tests: [test_golden_media_incremental, test_edge_sin_campos_previos, test_error_skill_inexistente, test_error_score_no_numerico]
|
||||
test_file_path: "python/functions/ml/tests/test_comfyui_update_skill_score.py"
|
||||
file_path: "python/functions/ml/comfyui_update_skill_score.py"
|
||||
---
|
||||
|
||||
# comfyui_update_skill_score
|
||||
|
||||
Acumulador de telemetría del grupo [`comfyui-skill`](../../../docs/capabilities/comfyui-skill.md).
|
||||
Cada juicio del panel [`comfyui-judge`](../../../docs/capabilities/comfyui-judge.md) sobre una
|
||||
imagen de la skill se incorpora a la media `score_mean`/`score_n` de la receta, reescribiendo
|
||||
`recipe.json` en sitio. Es lo que el pipeline `comfyui_generate_with_skill_oneshot` llama por dentro.
|
||||
|
||||
## Ejemplo
|
||||
|
||||
```python
|
||||
import sys, os
|
||||
sys.path.insert(0, os.path.join(os.environ["HOME"], "fn_registry", "python", "functions"))
|
||||
from ml.comfyui_update_skill_score import comfyui_update_skill_score
|
||||
|
||||
# Tras juzgar una imagen de la skill con score 7.4:
|
||||
res = comfyui_update_skill_score("portrait_cinematic_sd15", 7.4)
|
||||
print(res["score_mean"], res["score_n"]) # media acumulada y nº de juicios
|
||||
```
|
||||
|
||||
## Cuando usarla
|
||||
|
||||
Después de puntuar una imagen generada por una skill (su receta canónica, sin variantes ad-hoc),
|
||||
para que `score_mean` refleje la calidad media observada. Normalmente NO se llama directo: lo hace
|
||||
`comfyui_generate_with_skill_oneshot` cuando `judge=True` y no hay `recipe_patch`. Llámala a mano
|
||||
solo si juzgas por separado.
|
||||
|
||||
## Gotchas
|
||||
|
||||
- **In-place, sin snapshot**: reescribe `recipe.json` y nada más. No toca `versions/` ni
|
||||
`growth_log` — eso lo reserva `comfyui_bump_skill_version` para versiones reales.
|
||||
- **Media incremental** `mean_k = mean_{k-1} + (x - mean_{k-1})/k`: estable numéricamente, no
|
||||
necesita re-leer el histórico. `score_mean` se redondea a 4 decimales.
|
||||
- **No acumules variantes**: si juzgas una variante de prueba (un `recipe_patch` no guardado), NO
|
||||
la metas en la media de la skill canónica — contaminaría `score_mean`. El pipeline ya lo evita.
|
||||
@@ -0,0 +1,115 @@
|
||||
"""comfyui_update_skill_score — acumula el score de un juicio en una *skill* ComfyUI.
|
||||
|
||||
Cada vez que el panel `comfyui-judge` puntúa una imagen generada por una skill, esta
|
||||
función incorpora ese score a la media acumulada de la receta (``score_mean`` / ``score_n``)
|
||||
mediante **media incremental**, reescribiendo ``recipe.json`` en sitio.
|
||||
|
||||
A diferencia de `comfyui_save_skill`, NO crea un snapshot ``versions/vN.json`` ni añade
|
||||
entrada al ``growth_log``: el score es telemetría acumulada de la versión vigente, no una
|
||||
versión nueva. Los snapshots y el growth_log los reserva `comfyui_bump_skill_version` para
|
||||
promociones reales (cuando el score mejora y se sube el semver). Así ``versions/`` refleja
|
||||
versiones de receta y no se ensucia con una entrada por cada generación.
|
||||
|
||||
`library_dir` por defecto ``~/ComfyUI/skills_library``. Slug inexistente → ``{ok: False}``.
|
||||
|
||||
Impura: lee y reescribe ``recipe.json`` en disco. Sin red.
|
||||
"""
|
||||
|
||||
import json
|
||||
import os
|
||||
|
||||
DEFAULT_LIBRARY = "~/ComfyUI/skills_library"
|
||||
|
||||
|
||||
def _lib_dir(library_dir):
|
||||
return os.path.expanduser(library_dir or DEFAULT_LIBRARY)
|
||||
|
||||
|
||||
def comfyui_update_skill_score(
|
||||
slug: str,
|
||||
new_score: float,
|
||||
*,
|
||||
library_dir: str = None,
|
||||
) -> dict:
|
||||
"""Acumula ``new_score`` en la media de la skill (media incremental, in-place).
|
||||
|
||||
Args:
|
||||
slug: slug de la skill (su carpeta en la librería).
|
||||
new_score: score 0-10 del último juicio (p.ej. ``comfyui_judge_image(...)["score"]``).
|
||||
library_dir: raíz de la librería. Default ``~/ComfyUI/skills_library``. keyword-only.
|
||||
|
||||
Returns:
|
||||
dict ``{ok, slug, score_mean, score_n, prev_score_mean, prev_score_n, error}``.
|
||||
En éxito ``ok=True`` y ``score_mean``/``score_n`` son los valores actualizados. Si
|
||||
la skill no existe, el score no es numérico, o falla la escritura, ``ok=False`` con
|
||||
``error``; nunca lanza.
|
||||
"""
|
||||
if not slug or not isinstance(slug, str):
|
||||
return {"ok": False, "slug": slug, "score_mean": 0.0, "score_n": 0,
|
||||
"prev_score_mean": 0.0, "prev_score_n": 0,
|
||||
"error": "slug requerido (string no vacío)"}
|
||||
try:
|
||||
score = float(new_score)
|
||||
except (TypeError, ValueError):
|
||||
return {"ok": False, "slug": slug, "score_mean": 0.0, "score_n": 0,
|
||||
"prev_score_mean": 0.0, "prev_score_n": 0,
|
||||
"error": f"new_score debe ser numérico (recibido {new_score!r})"}
|
||||
|
||||
lib = _lib_dir(library_dir)
|
||||
recipe_path = os.path.join(lib, slug, "recipe.json")
|
||||
if not os.path.isfile(recipe_path):
|
||||
return {"ok": False, "slug": slug, "score_mean": 0.0, "score_n": 0,
|
||||
"prev_score_mean": 0.0, "prev_score_n": 0,
|
||||
"error": f"skill no encontrada: {slug!r} (sin recipe.json)"}
|
||||
|
||||
try:
|
||||
with open(recipe_path, encoding="utf-8") as fh:
|
||||
recipe = json.load(fh)
|
||||
except (OSError, json.JSONDecodeError) as exc:
|
||||
return {"ok": False, "slug": slug, "score_mean": 0.0, "score_n": 0,
|
||||
"prev_score_mean": 0.0, "prev_score_n": 0,
|
||||
"error": f"no se pudo leer la receta: {exc}"}
|
||||
|
||||
try:
|
||||
prev_mean = float(recipe.get("score_mean", 0.0) or 0.0)
|
||||
prev_n = int(recipe.get("score_n", 0) or 0)
|
||||
except (TypeError, ValueError):
|
||||
prev_mean, prev_n = 0.0, 0
|
||||
|
||||
new_n = prev_n + 1
|
||||
# Media incremental: mean_k = mean_{k-1} + (x_k - mean_{k-1}) / k.
|
||||
new_mean = prev_mean + (score - prev_mean) / new_n
|
||||
recipe["score_mean"] = round(new_mean, 4)
|
||||
recipe["score_n"] = new_n
|
||||
|
||||
try:
|
||||
with open(recipe_path, "w", encoding="utf-8") as fh:
|
||||
json.dump(recipe, fh, indent=2, ensure_ascii=False)
|
||||
except OSError as exc:
|
||||
return {"ok": False, "slug": slug, "score_mean": prev_mean, "score_n": prev_n,
|
||||
"prev_score_mean": prev_mean, "prev_score_n": prev_n,
|
||||
"error": f"fallo de escritura: {exc}"}
|
||||
|
||||
return {"ok": True, "slug": slug, "score_mean": recipe["score_mean"],
|
||||
"score_n": new_n, "prev_score_mean": prev_mean, "prev_score_n": prev_n,
|
||||
"error": ""}
|
||||
|
||||
|
||||
# Alias con el nombre completo del ID para descubrimiento por convención.
|
||||
update_skill_score = comfyui_update_skill_score
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
import sys
|
||||
|
||||
demo_lib = "/tmp/skills_score_demo"
|
||||
sdir = os.path.join(demo_lib, "demo_skill")
|
||||
os.makedirs(sdir, exist_ok=True)
|
||||
with open(os.path.join(sdir, "recipe.json"), "w", encoding="utf-8") as f:
|
||||
json.dump({"schema_version": 1, "slug": "demo_skill", "version": "1.0.0",
|
||||
"base_workflow": "txt2img", "score_mean": 0.0, "score_n": 0}, f, indent=2)
|
||||
|
||||
for s in (7.0, 8.0, 6.0):
|
||||
res = comfyui_update_skill_score("demo_skill", s, library_dir=demo_lib)
|
||||
print(f"score={s} -> mean={res['score_mean']} n={res['score_n']}", file=sys.stderr)
|
||||
print(json.dumps(res, indent=2, ensure_ascii=False))
|
||||
@@ -0,0 +1,127 @@
|
||||
"""Tests offline para comfyui_bump_skill_version (impura, sin red/GPU).
|
||||
|
||||
Verifican el contrato del bucle de mejora contra una librería temporal:
|
||||
- golden: score sube → promueve, snapshot pre-mutación, semver subido, recipe_patch aplicado,
|
||||
growth_log con una línea bien formada,
|
||||
- edge: bump major/patch + deep-merge anidado sin pisar otras claves,
|
||||
- error: gate (score no mejora sin force) → ok=False; force salta el gate; skill inexistente.
|
||||
"""
|
||||
|
||||
import json
|
||||
import os
|
||||
import sys
|
||||
|
||||
sys.path.insert(0, os.path.dirname(__file__))
|
||||
sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", ".."))
|
||||
|
||||
import pytest # noqa: E402
|
||||
|
||||
from ml.comfyui_bump_skill_version import comfyui_bump_skill_version, _bump_semver, _deep_merge # noqa: E402
|
||||
|
||||
|
||||
def _seed_skill(lib, slug="demo_skill", version="1.0.0", **extra):
|
||||
sdir = os.path.join(lib, slug)
|
||||
os.makedirs(os.path.join(sdir, "versions"), exist_ok=True)
|
||||
recipe = {"schema_version": 1, "slug": slug, "version": version,
|
||||
"base_workflow": "txt2img", "checkpoint": "dreamshaper_8.safetensors",
|
||||
"params": {"steps": 28, "cfg": 6.0}, **extra}
|
||||
with open(os.path.join(sdir, "recipe.json"), "w", encoding="utf-8") as f:
|
||||
json.dump(recipe, f)
|
||||
return sdir
|
||||
|
||||
|
||||
def test_semver_helper():
|
||||
assert _bump_semver("1.0.0", "minor") == "1.1.0"
|
||||
assert _bump_semver("1.2.3", "major") == "2.0.0"
|
||||
assert _bump_semver("1.2.3", "patch") == "1.2.4"
|
||||
assert _bump_semver("nan", "minor") == "0.1.0" # versión mal formada → 0.0.0 base
|
||||
|
||||
|
||||
def test_deep_merge_no_pisa_otras_claves():
|
||||
base = {"params": {"steps": 28, "cfg": 6.0}, "checkpoint": "a"}
|
||||
out = _deep_merge(base, {"params": {"steps": 32}})
|
||||
assert out == {"params": {"steps": 32, "cfg": 6.0}, "checkpoint": "a"}
|
||||
assert base["params"]["steps"] == 28 # no muta el original
|
||||
|
||||
|
||||
def test_golden_promueve_cuando_score_sube(tmp_path):
|
||||
lib = str(tmp_path)
|
||||
_seed_skill(lib)
|
||||
res = comfyui_bump_skill_version(
|
||||
"demo_skill", "subir steps a 32", score_before=6.5, score_after=7.4,
|
||||
judge_run_id="judge_xyz", recipe_patch={"params": {"steps": 32}}, library_dir=lib)
|
||||
|
||||
assert res["ok"] is True
|
||||
assert res["old_version"] == "1.0.0"
|
||||
assert res["new_version"] == "1.1.0"
|
||||
|
||||
# snapshot pre-mutación conserva la receta vieja (steps 28).
|
||||
with open(res["snapshot_file"], encoding="utf-8") as f:
|
||||
snap = json.load(f)
|
||||
assert snap["params"]["steps"] == 28
|
||||
assert snap["version"] == "1.0.0"
|
||||
|
||||
# recipe.json mutado: patch aplicado + semver subido, cfg intacto.
|
||||
with open(os.path.join(lib, "demo_skill", "recipe.json"), encoding="utf-8") as f:
|
||||
cur = json.load(f)
|
||||
assert cur["params"]["steps"] == 32
|
||||
assert cur["params"]["cfg"] == 6.0
|
||||
assert cur["version"] == "1.1.0"
|
||||
|
||||
# growth_log: una línea con la evidencia.
|
||||
with open(os.path.join(lib, "demo_skill", "growth_log.jsonl"), encoding="utf-8") as f:
|
||||
lines = [json.loads(x) for x in f if x.strip()]
|
||||
assert len(lines) == 1
|
||||
entry = lines[0]
|
||||
assert entry["version"] == "1.1.0"
|
||||
assert entry["score_before"] == 6.5
|
||||
assert entry["score_after"] == 7.4
|
||||
assert entry["judge_run_id"] == "judge_xyz"
|
||||
assert entry["diff"] == {"params": {"steps": 32}}
|
||||
|
||||
|
||||
def test_edge_major_y_patch(tmp_path):
|
||||
lib = str(tmp_path)
|
||||
_seed_skill(lib, slug="s_major", version="1.4.2")
|
||||
r1 = comfyui_bump_skill_version("s_major", "rework", score_before=5.0, score_after=8.0,
|
||||
bump="major", library_dir=lib)
|
||||
assert r1["new_version"] == "2.0.0"
|
||||
|
||||
_seed_skill(lib, slug="s_patch", version="1.4.2")
|
||||
r2 = comfyui_bump_skill_version("s_patch", "tweak", score_before=5.0, score_after=5.1,
|
||||
bump="patch", library_dir=lib)
|
||||
assert r2["new_version"] == "1.4.3"
|
||||
|
||||
|
||||
def test_error_gate_bloquea_si_no_mejora(tmp_path):
|
||||
lib = str(tmp_path)
|
||||
sdir = _seed_skill(lib)
|
||||
res = comfyui_bump_skill_version("demo_skill", "no mejora", score_before=7.0,
|
||||
score_after=6.5, library_dir=lib)
|
||||
assert res["ok"] is False
|
||||
assert "gate" in res["error"]
|
||||
# No se tocó nada: ni snapshot ni growth_log ni cambio de versión.
|
||||
assert not os.path.exists(os.path.join(sdir, "growth_log.jsonl"))
|
||||
with open(os.path.join(sdir, "recipe.json"), encoding="utf-8") as f:
|
||||
assert json.load(f)["version"] == "1.0.0"
|
||||
|
||||
|
||||
def test_error_force_salta_gate(tmp_path):
|
||||
lib = str(tmp_path)
|
||||
_seed_skill(lib)
|
||||
res = comfyui_bump_skill_version("demo_skill", "forzado", score_before=7.0,
|
||||
score_after=6.5, force=True, library_dir=lib)
|
||||
assert res["ok"] is True
|
||||
assert res["new_version"] == "1.1.0"
|
||||
assert res["growth_entry"]["forced"] is True
|
||||
|
||||
|
||||
def test_error_skill_inexistente(tmp_path):
|
||||
res = comfyui_bump_skill_version("no_existe", "x", score_before=1.0, score_after=2.0,
|
||||
library_dir=str(tmp_path))
|
||||
assert res["ok"] is False
|
||||
assert "no encontrada" in res["error"]
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
sys.exit(pytest.main([__file__, "-q"]))
|
||||
@@ -0,0 +1,75 @@
|
||||
"""Tests offline para comfyui_update_skill_score (impura, sin red).
|
||||
|
||||
- golden: media incremental correcta tras varios juicios; reescribe in-place sin crear
|
||||
snapshots ni growth_log,
|
||||
- edge: arranca desde score_mean/score_n ausentes (skill recién creada),
|
||||
- error: skill inexistente / score no numérico → ok=False.
|
||||
"""
|
||||
|
||||
import json
|
||||
import os
|
||||
import sys
|
||||
|
||||
sys.path.insert(0, os.path.dirname(__file__))
|
||||
sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", ".."))
|
||||
|
||||
import pytest # noqa: E402
|
||||
|
||||
from ml.comfyui_update_skill_score import comfyui_update_skill_score # noqa: E402
|
||||
|
||||
|
||||
def _seed(lib, slug="demo", **extra):
|
||||
sdir = os.path.join(lib, slug)
|
||||
os.makedirs(sdir, exist_ok=True)
|
||||
recipe = {"schema_version": 1, "slug": slug, "version": "1.0.0",
|
||||
"base_workflow": "txt2img", **extra}
|
||||
with open(os.path.join(sdir, "recipe.json"), "w", encoding="utf-8") as f:
|
||||
json.dump(recipe, f)
|
||||
return sdir
|
||||
|
||||
|
||||
def test_golden_media_incremental(tmp_path):
|
||||
lib = str(tmp_path)
|
||||
sdir = _seed(lib, score_mean=0.0, score_n=0)
|
||||
|
||||
r1 = comfyui_update_skill_score("demo", 7.0, library_dir=lib)
|
||||
assert r1["ok"] and r1["score_n"] == 1 and r1["score_mean"] == 7.0
|
||||
|
||||
r2 = comfyui_update_skill_score("demo", 8.0, library_dir=lib)
|
||||
assert r2["score_n"] == 2 and r2["score_mean"] == 7.5
|
||||
|
||||
r3 = comfyui_update_skill_score("demo", 6.0, library_dir=lib)
|
||||
assert r3["score_n"] == 3
|
||||
assert abs(r3["score_mean"] - 7.0) < 1e-9 # (7+8+6)/3
|
||||
|
||||
# in-place: ni versions/ ni growth_log creados por este helper.
|
||||
assert not os.path.exists(os.path.join(sdir, "versions"))
|
||||
assert not os.path.exists(os.path.join(sdir, "growth_log.jsonl"))
|
||||
|
||||
with open(os.path.join(sdir, "recipe.json"), encoding="utf-8") as f:
|
||||
cur = json.load(f)
|
||||
assert cur["score_n"] == 3
|
||||
|
||||
|
||||
def test_edge_sin_campos_previos(tmp_path):
|
||||
lib = str(tmp_path)
|
||||
_seed(lib, slug="fresh") # sin score_mean/score_n
|
||||
res = comfyui_update_skill_score("fresh", 9.0, library_dir=lib)
|
||||
assert res["ok"] and res["score_mean"] == 9.0 and res["score_n"] == 1
|
||||
assert res["prev_score_n"] == 0
|
||||
|
||||
|
||||
def test_error_skill_inexistente(tmp_path):
|
||||
res = comfyui_update_skill_score("nope", 5.0, library_dir=str(tmp_path))
|
||||
assert res["ok"] is False and "no encontrada" in res["error"]
|
||||
|
||||
|
||||
def test_error_score_no_numerico(tmp_path):
|
||||
lib = str(tmp_path)
|
||||
_seed(lib, slug="x", score_mean=0.0, score_n=0)
|
||||
res = comfyui_update_skill_score("x", "siete", library_dir=lib)
|
||||
assert res["ok"] is False and "numérico" in res["error"]
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
sys.exit(pytest.main([__file__, "-q"]))
|
||||
Reference in New Issue
Block a user