Fix tool-call loop leaks, raise step cap, add message queueing

core/config.py

@@ -14,8 +14,17 @@ class AppConfig:
     providers: Dict[str, Any]
     default_provider: str
 
+# The installed location of the harness (parent of core/), not the caller's
+# CWD. `motion` can be pointed at any workspace directory, so config lookup
+# must not depend on where it happens to be invoked from.
+_REPO_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+
+
 class ConfigManager:
-    CONFIG_PATHS = ["config.yml", "config.example.yml"]
+    CONFIG_PATHS = [
+        os.path.join(_REPO_DIR, "config.yml"),
+        os.path.join(_REPO_DIR, "config.example.yml"),
+    ]
 
     def __init__(self, config_path: str = ""):
         if config_path:

core/learning.py

@@ -1,10 +1,22 @@
+import hashlib
 import os
 import asyncio
 from typing import List, Dict, Any, Optional
 from dataclasses import dataclass
 from datetime import datetime
 from core.providers import ModelConfig, ProviderFactory
-from memory.db import MemoryDB, MemoryChunk
+from memory.db import MemoryDB, MemoryChunk, EMBEDDING_DIM
+
+
+def _fallback_embedding(text: str, dim: int = EMBEDDING_DIM) -> List[float]:
+    """Deterministic, always-non-zero embedding for when no embedding
+    provider is available. Mirrors MotionAgent.get_embedding's fallback so
+    behavior is consistent across the codebase."""
+    h = hashlib.sha256(text.encode()).digest()
+    raw = [float(b) / 255.0 for b in h]
+    vec = (raw * ((dim // len(raw)) + 1))[:dim]
+    norm = sum(v * v for v in vec) ** 0.5 or 1.0
+    return [v / norm for v in vec]
 
 @dataclass
 class Trajectory:
@@ -14,15 +26,26 @@ class Trajectory:
     final_result: str
     success: bool
 
+# The installed location of the harness (parent of core/), not the caller's
+# CWD - auto-synthesized skills accumulate here regardless of which project
+# directory `motion` is currently pointed at.
+_REPO_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
+
+
 class SkillSynthesizer:
     """
     The 'Crystallization' engine. 
     Turns successful tool-call trajectories into reusable .md skills.
     """
-    def __init__(self, model_config: ModelConfig, db: MemoryDB, skills_dir: str = "skills"):
+    def __init__(self, model_config: ModelConfig, db: MemoryDB, skills_dir: Optional[str] = None, embedding_provider=None):
         self.provider = ProviderFactory.get_provider(model_config)
         self.db = db
-        self.skills_dir = skills_dir
+        self.skills_dir = skills_dir or os.path.join(_REPO_DIR, "skills")
+        # Used to compute a real embedding for synthesized skills so they can
+        # actually participate in semantic recall. Expected to expose an
+        # async get_embedding(text) -> list[float] (e.g. a MotionAgent
+        # instance, which already has a safe hash-based fallback built in).
+        self.embedding_provider = embedding_provider
 
         if not os.path.exists(self.skills_dir):
             os.makedirs(self.skills_dir)
@@ -60,10 +83,21 @@ async def synthesize(self, trajectory: Trajectory) -> Optional[str]:
             with open(file_path, "w", encoding="utf-8") as f:
                 f.write(skill_content)
 
-            # Also index the skill in the MemoryDB for semantic recall
+            # Also index the skill in the MemoryDB for semantic recall. A
+            # zero-vector embedding is never valid here: cosine
+            # similarity/distance is undefined for a zero-norm vector, which
+            # crashes semantic search rather than just being unhelpful.
+            embedding = None
+            if self.embedding_provider is not None:
+                try:
+                    embedding = await self.embedding_provider.get_embedding(skill_content)
+                except Exception:
+                    embedding = None
+            if not embedding:
+                embedding = _fallback_embedding(skill_content)
             self.db.add_memory(MemoryChunk(
                 content=skill_content,
-                embedding=[0.0] * 128, # Embedding would be generated by a real provider
+                embedding=embedding,
                 metadata={"file": file_path, "type": "SKILL"},
                 mem_type="DOC"
             ))

core/orchestrator.py

@@ -10,7 +10,7 @@
 import logging
 
 from core.providers import ModelConfig, ProviderFactory
-from main import MotionAgent
+from main import MotionAgent, REPO_DIR
 
 logger = logging.getLogger(__name__)
 
@@ -189,7 +189,7 @@ async def _execute_task(self, request: TaskRequest, model_override: Optional[Mod
 
             try:
                 config = model_override or self.default_config
-                agent = MotionAgent(config, memory_path=f"memory_{request.task_id}.db")
+                agent = MotionAgent(config, memory_path=os.path.join(REPO_DIR, f"memory_{request.task_id}.db"))
 
                 # Record user turn
                 task.conversation.append({"role": "user", "content": request.prompt})
@@ -208,6 +208,8 @@ async def on_stream_chunk(chunk: str) -> None:
                     request.prompt,
                     target="user",
                     on_stream_chunk=on_stream_chunk,
+                    workspace=self.workspace_path,
+                    agent_mode="build",
                 )
 
                 # Fallback for non-streaming providers
@@ -233,7 +235,7 @@ async def on_stream_chunk(chunk: str) -> None:
                 except Exception:
                     pass
                 try:
-                    db_path = f"memory_{request.task_id}.db"
+                    db_path = os.path.join(REPO_DIR, f"memory_{request.task_id}.db")
                     if os.path.exists(db_path):
                         os.remove(db_path)
                 except Exception: