agent-runtime/agent/runtime.py
Nico 4c412d3c4b v0.14.4: Interpreter wired in v2, tool_call convention, Haiku models, UI fix
- Wire Interpreter into v2 pipeline (after Thinker tool_output, before Output)
- Rename tool_exec -> tool_call everywhere (consistent convention across v1/v2)
- Switch Director v1+v2 to anthropic/claude-haiku-4.5 (was opus, reserved)
- Fix UI apply_machine_ops crash when states are strings instead of dicts
- Fix runtime_test.py async poll to match on message ID (prevent stale results)
- Add traceback to pipeline error logging

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-03-29 06:06:13 +02:00

387 lines
18 KiB
Python

"""Runtime: wires all nodes together into a processing pipeline."""
import asyncio
import json
import logging
import time
from pathlib import Path
from typing import Callable
from .types import Envelope, Command, InputAnalysis, ThoughtResult, DirectorPlan
from .process import ProcessManager
from .engine import load_graph, instantiate_nodes, list_graphs, get_graph_for_cytoscape
log = logging.getLogger("runtime")
TRACE_FILE = Path(__file__).parent.parent / "trace.jsonl"
# Default graph — can be switched at runtime
_active_graph_name = "v1-current"
class OutputSink:
"""Collects output. Optionally streams to attached WebSocket."""
def __init__(self):
self.ws = None
self.response: str = ""
self.controls: list = []
self.done: bool = False
def attach(self, ws):
self.ws = ws
def detach(self):
self.ws = None
def reset(self):
self.response = ""
self.controls = []
self.done = False
async def send_delta(self, text: str):
self.response += text
if self.ws:
try:
await self.ws.send_text(json.dumps({"type": "delta", "content": text}))
except Exception:
pass
async def send_controls(self, controls: list):
self.controls = controls
if self.ws:
try:
await self.ws.send_text(json.dumps({"type": "controls", "controls": controls}))
except Exception:
pass
async def send_hud(self, data: dict):
if self.ws:
try:
await self.ws.send_text(json.dumps({"type": "hud", **data}))
except Exception:
pass
async def send_done(self):
self.done = True
if self.ws:
try:
await self.ws.send_text(json.dumps({"type": "done"}))
except Exception:
pass
class Runtime:
def __init__(self, user_claims: dict = None, origin: str = "",
broadcast: Callable = None, graph_name: str = None):
self.sink = OutputSink()
self.history: list[dict] = []
self.MAX_HISTORY = 40
self._broadcast = broadcast or (lambda e: None)
# Load graph and instantiate nodes
gname = graph_name or _active_graph_name
self.graph = load_graph(gname)
self.process_manager = ProcessManager(send_hud=self._send_hud)
nodes = instantiate_nodes(self.graph, send_hud=self._send_hud,
process_manager=self.process_manager)
# Bind nodes by role (pipeline code references these)
self.input_node = nodes["input"]
self.thinker = nodes["thinker"]
self.output_node = nodes["output"]
self.ui_node = nodes["ui"]
self.memorizer = nodes["memorizer"]
self.director = nodes["director"]
self.sensor = nodes["sensor"]
self.interpreter = nodes.get("interpreter") # v2 only
# Detect v2 graph: director has decide(), thinker takes DirectorPlan
self.is_v2 = hasattr(self.director, "decide")
self.sensor.start(
get_memo_state=lambda: self.memorizer.state,
get_server_controls=lambda: self.ui_node.current_controls,
)
claims = user_claims or {}
log.info(f"[runtime] user_claims: {claims}")
self.identity = claims.get("name") or claims.get("preferred_username") or claims.get("username") or "unknown"
log.info(f"[runtime] resolved identity: {self.identity}")
self.channel = origin or "unknown"
self.memorizer.state["user_name"] = self.identity
self.memorizer.state["situation"] = f"authenticated on {self.channel}" if origin else "local session"
def attach_ws(self, ws):
"""Attach a WebSocket for real-time streaming."""
self.sink.attach(ws)
log.info("[runtime] WS attached")
def detach_ws(self):
"""Detach WebSocket. Runtime keeps running."""
self.sink.detach()
log.info("[runtime] WS detached")
def update_identity(self, user_claims: dict, origin: str = ""):
"""Update identity from WS auth claims."""
claims = user_claims or {}
self.identity = claims.get("name") or claims.get("preferred_username") or claims.get("username") or self.identity
if origin:
self.channel = origin
self.memorizer.state["user_name"] = self.identity
self.memorizer.state["situation"] = f"authenticated on {self.channel}"
log.info(f"[runtime] identity updated: {self.identity}")
async def _send_hud(self, data: dict):
await self.sink.send_hud(data)
trace_entry = {"ts": time.strftime("%Y-%m-%d %H:%M:%S.") + f"{time.time() % 1:.3f}"[2:], **data}
try:
with open(TRACE_FILE, "a", encoding="utf-8") as f:
f.write(json.dumps(trace_entry, ensure_ascii=False) + "\n")
if TRACE_FILE.exists() and TRACE_FILE.stat().st_size > 500_000:
lines = TRACE_FILE.read_text(encoding="utf-8").strip().split("\n")
TRACE_FILE.write_text("\n".join(lines[-500:]) + "\n", encoding="utf-8")
except Exception as e:
log.error(f"trace write error: {e}")
self._broadcast(trace_entry)
async def _stream_text(self, text: str):
"""Stream pre-formed text to the client as deltas."""
self.sink.reset()
chunk_size = 12
for i in range(0, len(text), chunk_size):
await self.sink.send_delta(text[i:i + chunk_size])
await self.sink.send_done()
async def _run_output_and_ui(self, thought, mem_ctx):
"""Run Output and UI nodes in parallel. Returns (response_text, controls)."""
self.sink.reset()
output_task = asyncio.create_task(
self.output_node.process(thought, self.history, self.sink, memory_context=mem_ctx))
ui_task = asyncio.create_task(
self.ui_node.process(thought, self.history, memory_context=mem_ctx))
response, controls = await asyncio.gather(output_task, ui_task)
if controls:
await self.sink.send_controls(controls)
return response
async def handle_action(self, action: str, data: dict = None):
"""Handle a structured UI action (button click etc.)."""
self.sensor.note_user_activity()
# Try machine transition first (go: target — no LLM needed)
handled, transition_result = self.ui_node.try_machine_transition(action)
if handled:
await self._send_hud({"node": "ui", "event": "machine_transition",
"action": action, "detail": transition_result})
# Re-render all controls (machines + state + buttons)
controls = self.ui_node.get_machine_controls()
# Include non-machine buttons and labels
for ctrl in self.ui_node.current_controls:
if not ctrl.get("machine_id"):
controls.append(ctrl)
self.ui_node.current_controls = controls
await self.sink.send_controls(controls)
await self._send_hud({"node": "ui", "event": "controls", "controls": controls})
await self._stream_text(transition_result)
self.history.append({"role": "user", "content": f"[clicked {action}]"})
self.history.append({"role": "assistant", "content": transition_result})
return
# Try local UI action next (inc, dec, toggle — no LLM needed)
result, controls = await self.ui_node.process_local_action(action, data)
if result is not None:
# Local action handled — send controls update + short response
if controls:
await self.sink.send_controls(controls)
await self._stream_text(result)
self.history.append({"role": "user", "content": f"[clicked {action}]"})
self.history.append({"role": "assistant", "content": result})
return
# Complex action — needs Thinker reasoning
action_desc = f"ACTION: {action}"
if data:
action_desc += f" | data: {json.dumps(data)}"
self.history.append({"role": "user", "content": action_desc})
sensor_lines = self.sensor.get_context_lines()
director_line = self.director.get_context_line()
mem_ctx = self.memorizer.get_context_block(sensor_lines=sensor_lines, ui_state=self.ui_node.state)
mem_ctx += f"\n\n{director_line}"
command = Command(
analysis=InputAnalysis(intent="action", topic=action, complexity="simple"),
source_text=action_desc)
if self.is_v2:
plan = await self.director.decide(command, self.history, memory_context=mem_ctx)
thought = await self.thinker.process(command, plan, self.history, memory_context=mem_ctx)
if self.interpreter and thought.tool_used and thought.tool_output:
interpreted = await self.interpreter.interpret(
thought.tool_used, thought.tool_output, action_desc)
thought.response = interpreted.summary
else:
thought = await self.thinker.process(command, self.history, memory_context=mem_ctx)
response = await self._run_output_and_ui(thought, mem_ctx)
self.history.append({"role": "assistant", "content": response})
await self.memorizer.update(self.history)
if not self.is_v2:
await self.director.update(self.history, self.memorizer.state)
if len(self.history) > self.MAX_HISTORY:
self.history = self.history[-self.MAX_HISTORY:]
def _format_dashboard(self, dashboard: list) -> str:
"""Format dashboard controls into a context string for Thinker.
Compares browser-reported state against server-side controls to detect mismatches."""
server_controls = self.ui_node.current_controls
server_buttons = [c.get("label", "") for c in server_controls if c.get("type") == "button"]
browser_buttons = [c.get("label", "") for c in dashboard if c.get("type") == "button"] if dashboard else []
lines = []
# Mismatch detection (S3* audit)
if server_buttons and not browser_buttons:
lines.append(f"WARNING: Server sent {len(server_buttons)} controls but dashboard shows NONE.")
lines.append(f" Expected buttons: {', '.join(server_buttons)}")
lines.append(" Controls failed to render or were lost. You MUST re-emit them in ACTIONS.")
elif server_buttons and set(server_buttons) != set(browser_buttons):
lines.append(f"WARNING: Dashboard mismatch.")
lines.append(f" Server sent: {', '.join(server_buttons)}")
lines.append(f" Browser shows: {', '.join(browser_buttons) or 'nothing'}")
lines.append(" Re-emit correct controls in ACTIONS if needed.")
if not dashboard:
lines.append("Dashboard: empty (user sees nothing)")
else:
lines.append("Dashboard (what user currently sees):")
for ctrl in dashboard:
ctype = ctrl.get("type", "unknown")
if ctype == "button":
lines.append(f" - Button: {ctrl.get('label', '?')}")
elif ctype == "label":
lines.append(f" - Label: {ctrl.get('text', '?')} = {ctrl.get('value', '?')}")
elif ctype == "table":
cols = ctrl.get("columns", [])
rows = len(ctrl.get("data", []))
lines.append(f" - Table: {', '.join(cols)} ({rows} rows)")
else:
lines.append(f" - {ctype}: {ctrl.get('label', ctrl.get('text', '?'))}")
return "\n".join(lines)
async def handle_message(self, text: str, dashboard: list = None):
# Detect ACTION: prefix from API/test runner
if text.startswith("ACTION:"):
parts = text.split("|", 1)
action = parts[0].replace("ACTION:", "").strip()
data = None
if len(parts) > 1:
try:
data = json.loads(parts[1].replace("data:", "").strip())
except (json.JSONDecodeError, Exception):
pass
return await self.handle_action(action, data)
envelope = Envelope(
text=text,
user_id="bob",
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})
# Check Sensor flags (idle return, workspace mismatch)
sensor_flags = self.sensor.consume_flags()
sensor_lines = self.sensor.get_context_lines()
director_line = self.director.get_context_line()
mem_ctx = self.memorizer.get_context_block(sensor_lines=sensor_lines, ui_state=self.ui_node.state)
mem_ctx += f"\n\n{director_line}"
machine_summary = self.ui_node.get_machine_summary()
if machine_summary:
mem_ctx += f"\n\n{machine_summary}"
if dashboard is not None:
mem_ctx += f"\n\n{self._format_dashboard(dashboard)}"
# Inject sensor flags into context
if sensor_flags:
flag_lines = ["Sensor flags:"]
for f in sensor_flags:
if f["type"] == "idle_return":
flag_lines.append(f" - User returned after {f['away_duration']} away. Welcome them back briefly, mention what's on their dashboard.")
elif f["type"] == "workspace_mismatch":
flag_lines.append(f" - Workspace mismatch detected: {f['detail']}. Check if controls need re-emitting.")
mem_ctx += "\n\n" + "\n".join(flag_lines)
command = await self.input_node.process(
envelope, self.history, memory_context=mem_ctx,
identity=self.identity, channel=self.channel)
# Reflex path: trivial social messages skip Thinker entirely
if command.analysis.intent == "social" and command.analysis.complexity == "trivial":
await self._send_hud({"node": "runtime", "event": "reflex_path",
"detail": f"{command.analysis.intent}/{command.analysis.complexity}"})
thought = ThoughtResult(response=command.source_text, actions=[])
response = await self._run_output_and_ui(thought, mem_ctx)
self.history.append({"role": "assistant", "content": response})
await self.memorizer.update(self.history)
if not self.is_v2:
await self.director.update(self.history, self.memorizer.state)
if len(self.history) > self.MAX_HISTORY:
self.history = self.history[-self.MAX_HISTORY:]
return
if self.is_v2:
# v2 flow: Director decides, Thinker executes, Interpreter reads results
plan = await self.director.decide(command, self.history, memory_context=mem_ctx)
thought = await self.thinker.process(command, plan, self.history, memory_context=mem_ctx)
# Interpreter: factual summary of tool results (no hallucination)
if self.interpreter and thought.tool_used and thought.tool_output:
interpreted = await self.interpreter.interpret(
thought.tool_used, thought.tool_output, text)
# Replace thinker's response with interpreter's factual summary
thought.response = interpreted.summary
else:
# v1 flow: optional Director pre-planning for complex requests
is_complex = command.analysis.complexity == "complex"
is_data_request = (command.analysis.intent in ("request", "action")
and any(k in text.lower()
for k in ["daten", "data", "database", "db", "tabelle", "table",
"query", "abfrage", "untersuche", "investigate", "explore",
"analyse", "analyze", "umsatz", "revenue", "billing",
"abrechnung", "customer", "kunde", "geraete", "device",
"objekt", "object", "how many", "wieviele", "welche"]))
needs_planning = is_complex or (is_data_request and len(text.split()) > 8)
if needs_planning:
plan = await self.director.plan(self.history, self.memorizer.state, text)
if plan:
director_line = self.director.get_context_line()
mem_ctx = self.memorizer.get_context_block(sensor_lines=sensor_lines, ui_state=self.ui_node.state)
mem_ctx += f"\n\n{director_line}"
if machine_summary:
mem_ctx += f"\n\n{machine_summary}"
if dashboard is not None:
mem_ctx += f"\n\n{self._format_dashboard(dashboard)}"
thought = await self.thinker.process(command, self.history, memory_context=mem_ctx)
self.director.current_plan = ""
# Output (voice) and UI (screen) run in parallel
response = await self._run_output_and_ui(thought, mem_ctx)
self.history.append({"role": "assistant", "content": response})
await self.memorizer.update(self.history)
if not self.is_v2:
await self.director.update(self.history, self.memorizer.state)
if len(self.history) > self.MAX_HISTORY:
self.history = self.history[-self.MAX_HISTORY:]