Implement config-driven models (Phase 1): graph MODELS dict, instantiate applies, per-request overrides

- Graph definitions (v3, v4) now declare MODELS mapping role → model string
- engine.py extracts MODELS and applies to nodes during instantiation
- frame_engine.process_message() accepts model_overrides for per-request swaps
  (restored via try/finally after processing)
- 11/11 engine tests green

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Nico 2026-04-03 18:05:21 +02:00
parent ecfbc86676
commit cf42951b77
4 changed files with 87 additions and 43 deletions

View File

@ -63,6 +63,7 @@ def _graph_from_module(mod) -> dict:
"conditions": getattr(mod, "CONDITIONS", {}),
"audit": getattr(mod, "AUDIT", {}),
"engine": getattr(mod, "ENGINE", "imperative"),
"models": getattr(mod, "MODELS", {}),
}
@ -79,7 +80,13 @@ def instantiate_nodes(graph: dict, send_hud, process_manager: ProcessManager = N
nodes[role] = cls(send_hud=send_hud, process_manager=process_manager)
else:
nodes[role] = cls(send_hud=send_hud)
log.info(f"[engine] {role} = {impl_name} ({cls.__name__})")
# Apply model from graph config (overrides class default)
model = graph.get("models", {}).get(role)
if model and hasattr(nodes[role], "model"):
nodes[role].model = model
log.info(f"[engine] {role} = {impl_name} ({cls.__name__}) model={model}")
else:
log.info(f"[engine] {role} = {impl_name} ({cls.__name__})")
return nodes

View File

@ -173,57 +173,76 @@ class FrameEngine:
# --- Main entry point ---
async def process_message(self, text: str, dashboard: list = None) -> dict:
async def process_message(self, text: str, dashboard: list = None,
model_overrides: dict = None) -> dict:
"""Process a message through the frame pipeline.
Returns {response, controls, memorizer, frames, trace}."""
Returns {response, controls, memorizer, frames, trace}.
self._begin_trace(text)
model_overrides: optional {role: model} to override node models for this request only.
"""
# Apply per-request model overrides (restored after processing)
saved_models = {}
if model_overrides:
for role, model in model_overrides.items():
node = self.nodes.get(role)
if node and hasattr(node, "model"):
saved_models[role] = node.model
node.model = model
# Handle ACTION: prefix
if text.startswith("ACTION:"):
return await self._handle_action(text, dashboard)
try:
self._begin_trace(text)
# Setup
envelope = Envelope(
text=text, user_id=self.identity,
session_id="test", timestamp=time.strftime("%Y-%m-%d %H:%M:%S"),
)
self.sensor.note_user_activity()
if dashboard is not None:
self.sensor.update_browser_dashboard(dashboard)
self.history.append({"role": "user", "content": text})
# Handle ACTION: prefix
if text.startswith("ACTION:"):
return await self._handle_action(text, dashboard)
# --- Frame 1: Input ---
mem_ctx = self._build_context(dashboard)
rec = self._begin_frame(1, "input", input_summary=text[:100])
# Setup
envelope = Envelope(
text=text, user_id=self.identity,
session_id="test", timestamp=time.strftime("%Y-%m-%d %H:%M:%S"),
)
self.sensor.note_user_activity()
if dashboard is not None:
self.sensor.update_browser_dashboard(dashboard)
self.history.append({"role": "user", "content": text})
command = await self.nodes["input"].process(
envelope, self.history, memory_context=mem_ctx,
identity=self.identity, channel=self.channel)
# --- Frame 1: Input ---
mem_ctx = self._build_context(dashboard)
rec = self._begin_frame(1, "input", input_summary=text[:100])
a = command.analysis
cmd_summary = f"intent={a.intent} language={a.language} tone={a.tone} complexity={a.complexity}"
command = await self.nodes["input"].process(
envelope, self.history, memory_context=mem_ctx,
identity=self.identity, channel=self.channel)
# Check reflex condition
is_reflex = self._check_condition("reflex", command=command)
if is_reflex:
self._end_frame(rec, output_summary=cmd_summary,
route="output (reflex)", condition="reflex=True")
await self._send_hud({"node": "runtime", "event": "reflex_path",
"detail": f"{a.intent}/{a.complexity}"})
return await self._run_reflex(command, mem_ctx)
else:
next_node = "pa" if self.has_pa else ("director" if self.has_director else "thinker")
self._end_frame(rec, output_summary=cmd_summary,
route=next_node, condition=f"reflex=False")
a = command.analysis
cmd_summary = f"intent={a.intent} language={a.language} tone={a.tone} complexity={a.complexity}"
# --- Frame 2+: Pipeline ---
if self.has_pa:
return await self._run_expert_pipeline(command, mem_ctx, dashboard)
elif self.has_director:
return await self._run_director_pipeline(command, mem_ctx, dashboard)
else:
return await self._run_thinker_pipeline(command, mem_ctx, dashboard)
# Check reflex condition
is_reflex = self._check_condition("reflex", command=command)
if is_reflex:
self._end_frame(rec, output_summary=cmd_summary,
route="output (reflex)", condition="reflex=True")
await self._send_hud({"node": "runtime", "event": "reflex_path",
"detail": f"{a.intent}/{a.complexity}"})
return await self._run_reflex(command, mem_ctx)
else:
next_node = "pa" if self.has_pa else ("director" if self.has_director else "thinker")
self._end_frame(rec, output_summary=cmd_summary,
route=next_node, condition=f"reflex=False")
# --- Frame 2+: Pipeline ---
if self.has_pa:
return await self._run_expert_pipeline(command, mem_ctx, dashboard)
elif self.has_director:
return await self._run_director_pipeline(command, mem_ctx, dashboard)
else:
return await self._run_thinker_pipeline(command, mem_ctx, dashboard)
finally:
# Restore original models after per-request overrides
for role, original_model in saved_models.items():
node = self.nodes.get(role)
if node:
node.model = original_model
# --- Pipeline variants ---

View File

@ -62,4 +62,13 @@ CONDITIONS = {
"has_tool_output": "thinker.tool_used is not empty",
}
MODELS = {
"input": "google/gemini-2.0-flash-001",
"director": "anthropic/claude-haiku-4.5",
"thinker": "google/gemini-2.0-flash-001",
"interpreter": "google/gemini-2.0-flash-001",
"output": "google/gemini-2.0-flash-001",
"memorizer": "google/gemini-2.0-flash-001",
}
AUDIT = {}

View File

@ -68,4 +68,13 @@ CONDITIONS = {
"has_tool_output": "expert.tool_used is not empty",
}
MODELS = {
"input": "google/gemini-2.0-flash-001",
"pa": "anthropic/claude-haiku-4.5",
"expert_eras": "google/gemini-2.0-flash-001",
"interpreter": "google/gemini-2.0-flash-001",
"output": "google/gemini-2.0-flash-001",
"memorizer": "google/gemini-2.0-flash-001",
}
AUDIT = {}