a2bc6347fc
Architecture: - Graph engine (engine.py) loads graph definitions, instantiates nodes - Versioned nodes: input_v1, thinker_v1, output_v1, memorizer_v1, director_v1 - NODE_REGISTRY for dynamic node lookup by name - Graph API: /api/graph/active, /api/graph/list, /api/graph/switch - Graph definition: graphs/v1_current.py (7 nodes, 13 edges, 3 edge types) S3* Audit system: - Workspace mismatch detection (server vs browser controls) - Code-without-tools retry (Thinker wrote code but no tool calls) - Intent-without-action retry (request intent but Thinker only produced text) - Dashboard feedback: browser sends workspace state on every message - Sensor continuous comparison on 5s tick State machines: - create_machine / add_state / reset_machine / destroy_machine via function calling - Local transitions (go:) resolve without LLM round-trip - Button persistence across turns Database tools: - query_db tool via pymysql to MariaDB K3s pod (eras2_production) - Table rendering in workspace (tab-separated parsing) - Director pre-planning with Opus for complex data requests - Error retry with corrected SQL Frontend: - Cytoscape.js pipeline graph with real-time node animations - Overlay scrollbars (CSS-only, no reflow) - Tool call/result trace events - S3* audit events in trace Testing: - 167 integration tests (11 test suites) - 22 node-level unit tests (test_nodes/) - Three test levels: node unit, graph integration, scenario Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
89 lines
2.9 KiB
Python
89 lines
2.9 KiB
Python
"""LLM helper: OpenRouter calls, token estimation, context fitting."""
|
|
|
|
import json
|
|
import logging
|
|
import os
|
|
from typing import Any
|
|
|
|
import httpx
|
|
|
|
log = logging.getLogger("runtime")
|
|
|
|
API_KEY = os.environ.get("OPENROUTER_API_KEY", "")
|
|
OPENROUTER_URL = "https://openrouter.ai/api/v1/chat/completions"
|
|
|
|
|
|
async def llm_call(model: str, messages: list[dict], stream: bool = False,
|
|
tools: list[dict] = None) -> Any:
|
|
"""Single LLM call via OpenRouter.
|
|
Returns full text, (client, response) for streaming, or (text, tool_calls) when tools are used."""
|
|
headers = {"Authorization": f"Bearer {API_KEY}", "Content-Type": "application/json"}
|
|
body = {"model": model, "messages": messages, "stream": stream}
|
|
if tools:
|
|
body["tools"] = tools
|
|
|
|
client = httpx.AsyncClient(timeout=60)
|
|
if stream:
|
|
resp = await client.send(client.build_request("POST", OPENROUTER_URL, headers=headers, json=body), stream=True)
|
|
return client, resp
|
|
|
|
resp = await client.post(OPENROUTER_URL, headers=headers, json=body)
|
|
await client.aclose()
|
|
data = resp.json()
|
|
if "choices" not in data:
|
|
log.error(f"LLM error: {data}")
|
|
error_msg = f"[LLM error: {data.get('error', {}).get('message', 'unknown')}]"
|
|
return (error_msg, []) if tools else error_msg
|
|
|
|
msg = data["choices"][0]["message"]
|
|
content = msg.get("content", "") or ""
|
|
tool_calls = msg.get("tool_calls", [])
|
|
|
|
if tools:
|
|
return content, tool_calls
|
|
return content
|
|
|
|
|
|
def estimate_tokens(text: str) -> int:
|
|
"""Rough token estimate: 1 token ~ 4 chars."""
|
|
return len(text) // 4
|
|
|
|
|
|
def fit_context(messages: list[dict], max_tokens: int, protect_last: int = 4) -> list[dict]:
|
|
"""Trim oldest messages (after system prompt) to fit token budget.
|
|
Always keeps: system prompt(s) at start + last `protect_last` messages."""
|
|
if not messages:
|
|
return messages
|
|
|
|
system_msgs = []
|
|
rest = []
|
|
for m in messages:
|
|
if not rest and m["role"] == "system":
|
|
system_msgs.append(m)
|
|
else:
|
|
rest.append(m)
|
|
|
|
protected = rest[-protect_last:] if len(rest) > protect_last else rest
|
|
middle = rest[:-protect_last] if len(rest) > protect_last else []
|
|
|
|
fixed_tokens = sum(estimate_tokens(m["content"]) for m in system_msgs + protected)
|
|
|
|
if fixed_tokens >= max_tokens:
|
|
result = system_msgs + protected
|
|
total = sum(estimate_tokens(m["content"]) for m in result)
|
|
while total > max_tokens and len(result) > 2:
|
|
removed = result.pop(1)
|
|
total -= estimate_tokens(removed["content"])
|
|
return result
|
|
|
|
remaining = max_tokens - fixed_tokens
|
|
kept_middle = []
|
|
for m in reversed(middle):
|
|
t = estimate_tokens(m["content"])
|
|
if remaining - t < 0:
|
|
break
|
|
kept_middle.insert(0, m)
|
|
remaining -= t
|
|
|
|
return system_msgs + kept_middle + protected
|