feat(qw2): hook LoopDetector dans api_stream + extension replay_state
replay_state enrichi de _screenshot_history (5 dernières images PIL) et _action_history (5 dernières signatures action). report_action_result : - met à jour les deux anneaux après chaque action - évalue le LoopDetector (singleton lazy avec _clip_embedder serveur) - si detected → bascule paused_need_help avec pause_reason="loop_detected" et bus event lea:loop_detected (signal + evidence) Tous les chemins d'erreur (embedder absent, OOM, exception) loggent et laissent le replay continuer — aucun blocage par la couche détection. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
@@ -3942,6 +3942,82 @@ async def report_action_result(report: ReplayResultReport):
|
||||
f"— worker VLM autorisé à reprendre"
|
||||
)
|
||||
|
||||
# ===================================================================
|
||||
# QW2 — LoopDetector : alimentation des anneaux + évaluation
|
||||
# ===================================================================
|
||||
# On n'évalue que si le replay est encore "running" — inutile de
|
||||
# pauser quelque chose de déjà completed/error/paused.
|
||||
if replay_state["status"] == "running":
|
||||
# Snapshot image (PIL) dans l'anneau
|
||||
try:
|
||||
from PIL import Image
|
||||
ss_raw = screenshot_after or replay_state.get("last_screenshot")
|
||||
img = None
|
||||
if isinstance(ss_raw, str) and ss_raw:
|
||||
if os.path.isfile(ss_raw):
|
||||
img = Image.open(ss_raw).copy() # détache du file handle
|
||||
else:
|
||||
# Possible base64 — décoder
|
||||
try:
|
||||
import base64
|
||||
import io as _io
|
||||
img_bytes = base64.b64decode(ss_raw, validate=False)
|
||||
img = Image.open(_io.BytesIO(img_bytes)).copy()
|
||||
except Exception:
|
||||
img = None
|
||||
if img is not None:
|
||||
replay_state.setdefault("_screenshot_history", []).append(img)
|
||||
replay_state["_screenshot_history"] = replay_state["_screenshot_history"][-5:]
|
||||
except Exception as e:
|
||||
logger.debug("LoopDetector: snapshot historique échoué: %s", e)
|
||||
|
||||
# Snapshot signature de l'action courante
|
||||
try:
|
||||
_act_pos = report.actual_position or {}
|
||||
action_sig = {
|
||||
"type": (original_action or {}).get("type")
|
||||
or replay_state.get("_last_action_type", ""),
|
||||
"x_pct": _act_pos.get("x_pct") if isinstance(_act_pos, dict)
|
||||
else (original_action or {}).get("x_pct"),
|
||||
"y_pct": _act_pos.get("y_pct") if isinstance(_act_pos, dict)
|
||||
else (original_action or {}).get("y_pct"),
|
||||
}
|
||||
replay_state.setdefault("_action_history", []).append(action_sig)
|
||||
replay_state["_action_history"] = replay_state["_action_history"][-5:]
|
||||
except Exception as e:
|
||||
logger.debug("LoopDetector: snapshot action_sig échoué: %s", e)
|
||||
|
||||
# Évaluation (silencieux si rien)
|
||||
try:
|
||||
verdict = _get_loop_detector().evaluate(
|
||||
replay_state,
|
||||
screenshots=replay_state.get("_screenshot_history", []),
|
||||
actions=replay_state.get("_action_history", []),
|
||||
)
|
||||
if verdict.detected:
|
||||
replay_state["status"] = "paused_need_help"
|
||||
replay_state["pause_reason"] = "loop_detected"
|
||||
replay_state["pause_message"] = (
|
||||
f"Léa semble bloquée — {verdict.signal} "
|
||||
f"(détail: {verdict.evidence})"
|
||||
)
|
||||
logger.warning(
|
||||
"LoopDetector: replay %s mis en pause — signal=%s evidence=%s",
|
||||
replay_state["replay_id"], verdict.signal, verdict.evidence,
|
||||
)
|
||||
# Bus event d'observabilité (logger pattern QW1)
|
||||
try:
|
||||
logger.info(
|
||||
"[BUS] lea:loop_detected replay=%s signal=%s evidence=%s",
|
||||
replay_state["replay_id"],
|
||||
verdict.signal,
|
||||
verdict.evidence,
|
||||
)
|
||||
except Exception as _e_bus:
|
||||
logger.debug("emit lea:loop_detected échec: %s", _e_bus)
|
||||
except Exception as e:
|
||||
logger.warning("LoopDetector: évaluation échouée (non bloquant): %s", e)
|
||||
|
||||
return {
|
||||
"status": "recorded",
|
||||
"action_id": action_id,
|
||||
|
||||
Reference in New Issue
Block a user