agent-runtime/agent/nodes/director_v2.py

"""Director Node v2: always-on brain — decides what Thinker should execute."""

import json
import logging

from .base import Node
from ..llm import llm_call
from ..types import Command, DirectorPlan

log = logging.getLogger("runtime")


class DirectorV2Node(Node):
    name = "director_v2"
    model = "anthropic/claude-haiku-4.5"
    max_context_tokens = 4000

    SYSTEM = """You are the Director — the brain of this cognitive agent runtime.
You receive the user's message (already classified by Input) and conversation history.
Your job: decide WHAT to do and HOW, then produce an action plan for the Thinker (a fast executor).

The Thinker has these tools:
- query_db(query, database) — SQL SELECT/DESCRIBE/SHOW on MariaDB
  Databases: eras2_production (heating/energy, 693 customers), plankiste_test (Kita planning)
  Tables are lowercase: kunden, objekte, geraete, nutzeinheit, geraeteverbraeuche, etc.
- emit_actions(actions) — show buttons [{label, action, payload?}]
- set_state(key, value) — persistent key-value store
- emit_display(items) — per-response formatted data [{type, label, value?, style?}]
- create_machine(id, initial, states) — persistent interactive UI with navigation
  states is a dict: {{"state_name": {{"actions": [{{"label":"Go","action":"go","payload":"target","go":"target"}}], "display": [{{"type":"text","label":"Title","value":"Content"}}]}}}}
- add_state(id, state, buttons, content) — add a state to existing machine. id=machine name, state=new state name, buttons=[{{label,action,go}}], content=["text"]
- reset_machine(id) — reset machine to initial state. id=machine name
- destroy_machine(id) — remove machine. id=machine name

Your output is a JSON plan:
{{
  "goal": "what we're trying to achieve",
  "steps": ["Step 1: ...", "Step 2: ..."],
  "present_as": "table | summary | machine",
  "tool_sequence": [
    {{"tool": "query_db", "args": {{"query": "SELECT ...", "database": "eras2_production"}}}},
    {{"tool": "emit_display", "args": {{"items": [...]}}}}
  ],
  "reasoning": "why this approach",
  "response_hint": "how Thinker should phrase the response (if no tools needed)",
  "mode": "casual | building | debugging | exploring",
  "style": "brief directive for response style"
}}

Rules:
- NEVER guess column or table names. If you don't know the schema, your FIRST step MUST be DESCRIBE or SHOW TABLES. Only write SELECT queries using columns you have seen in a prior DESCRIBE result or in conversation history.
- For simple social/greeting: empty tool_sequence, set response_hint instead.
- For data questions: plan the SQL queries. Be specific — the Thinker is not smart.
- For UI requests: plan the exact tool calls with full args.
- Max 5 tools in sequence. Keep it focused.
- mode/style guide the Output node's voice.

Output ONLY valid JSON. No markdown fences, no explanation."""

    def __init__(self, send_hud):
        super().__init__(send_hud)
        self.directive: dict = {
            "mode": "casual",
            "style": "be helpful and concise",
        }

    def get_context_line(self) -> str:
        d = self.directive
        return f"Director: {d['mode']} mode. {d['style']}."

    async def decide(self, command: Command, history: list[dict],
                     memory_context: str = "") -> DirectorPlan:
        """Analyze input and produce an action plan for Thinker v2."""
        await self.hud("thinking", detail="deciding action plan")

        a = command.analysis
        messages = [
            {"role": "system", "content": self.SYSTEM},
        ]
        if memory_context:
            messages.append({"role": "system", "content": memory_context})

        for msg in history[-12:]:
            messages.append(msg)

        input_ctx = (
            f"User message analysis:\n"
            f"- Who: {a.who} | Intent: {a.intent} | Complexity: {a.complexity}\n"
            f"- Topic: {a.topic} | Tone: {a.tone} | Language: {a.language}\n"
            f"- Context: {a.context}\n"
            f"- Original: {command.source_text}"
        )
        messages.append({"role": "user", "content": input_ctx})
        messages = self.trim_context(messages)

        await self.hud("context", messages=messages, tokens=self.last_context_tokens,
                       max_tokens=self.max_context_tokens, fill_pct=self.context_fill_pct)

        raw = await llm_call(self.model, messages)
        log.info(f"[director_v2] raw: {raw[:300]}")

        plan = self._parse_plan(raw, command)

        # Update style directive (for Output node)
        if hasattr(plan, '_raw') and plan._raw:
            raw_data = plan._raw
            if raw_data.get("mode"):
                self.directive["mode"] = raw_data["mode"]
            if raw_data.get("style"):
                self.directive["style"] = raw_data["style"]

        await self.hud("decided", goal=plan.goal, tools=len(plan.tool_sequence),
                       direct=plan.is_direct_response)
        return plan

    def _parse_plan(self, raw: str, command: Command) -> DirectorPlan:
        """Parse LLM output into DirectorPlan, with fallback."""
        text = raw.strip()
        if text.startswith("```"):
            text = text.split("\n", 1)[1] if "\n" in text else text[3:]
            if text.endswith("```"):
                text = text[:-3]
            text = text.strip()

        try:
            data = json.loads(text)
            plan = DirectorPlan(
                goal=data.get("goal", ""),
                steps=data.get("steps", []),
                present_as=data.get("present_as", "summary"),
                tool_sequence=data.get("tool_sequence", []),
                reasoning=data.get("reasoning", ""),
                response_hint=data.get("response_hint", ""),
            )
            # Stash raw for directive extraction
            plan._raw = data
            return plan
        except (json.JSONDecodeError, Exception) as e:
            log.error(f"[director_v2] parse failed: {e}, raw: {text[:200]}")
            # Fallback: direct response
            plan = DirectorPlan(
                goal=f"respond to: {command.source_text[:50]}",
                steps=[],
                present_as="summary",
                tool_sequence=[],
                reasoning=f"parse failed: {e}",
                response_hint=f"Respond naturally to: {command.source_text}",
            )
            plan._raw = {}
            return plan