trading-agent/agents/graph.py

"""The agent runner: one persona researches with tools under a hard step + wall-clock
budget, then finalizes structured picks. Hand-rolled (not create_react_agent) for precise
budget control and guaranteed partial-output salvage on timeouts — the two things that
matter for unattended overnight runs.

Flow: research loop (LLM <-> tools, compaction as needed) -> finalize (structured JSON
picks). Finalize is ALWAYS attempted, even after a timeout, so a stuck agent still yields
any picks it formed rather than nothing.
"""

from __future__ import annotations

import json
import time
from datetime import datetime, timezone

from langchain_core.messages import HumanMessage, SystemMessage, ToolMessage
from loguru import logger

from agents import compaction
from agents.schemas import AgentOutput, AgentStatus, Pick
from config.personas import Persona
from config.settings import get_settings
from llm.client import build_llm
from tools import build_toolset

RESEARCH_KICKOFF = (
    "Today is {date}. This is DEEP overnight research — be exhaustive, not a quick scan. "
    "Choose {n} strong stocks that fit your style.\n\n"
    "1. Call get_portfolio first (your current holdings).\n"
    "2. Build a candidate list of 8-12 tickers in your style, then investigate EACH one: "
    "get_quote, get_bars (trend/momentum), get_news (catalysts), get_fundamentals "
    "(quality/valuation), and web_search / web_fetch for context and second opinions.\n"
    "3. Use run_python for ALL math (position sizing, %s, ratios) — never do arithmetic yourself.\n"
    "4. Compare candidates, then narrow to your best {n} and pressure-test each thesis "
    "(what would make it wrong?) before finalizing.\n"
    "Work one tool call at a time. Do NOT rush to finalize — keep researching until you have "
    "genuinely vetted a broad set of names.\n"
    "You may go LONG (bet a name rises) or SHORT (bet it falls). You may trade the STOCK or a "
    "single-leg OPTION: buy a CALL for a bullish view or a PUT for a bearish view. For options, "
    "call get_option_chain(underlying, 'call' or 'put') and copy an exact contract_symbol."
)

DEEPEN_NUDGE = (
    "You're not finished. Keep researching before finalizing: evaluate additional candidate "
    "tickers you haven't examined yet, pull bars/news/fundamentals for them, verify with "
    "web_search, and stress-test your leading ideas against the alternatives. Continue using tools."
)

FINALIZE_PROMPT = (
    "Research complete. Return a JSON object of the form {{\"picks\": [ ... ]}} containing about "
    "{n} picks. Each pick object has these fields:\n"
    '{{"ticker": "SYM", "direction": "long", "conviction": 0.72, "thesis": "...", '
    '"catalysts": ["..."], "time_horizon": "swing", "requested_weight_pct": 0.08, '
    '"entry_type": "market", "stop_loss_pct": 0.07, "take_profit_pct": 0.15, '
    '"risk_notes": "...", "data_refs": ["..."]}}\n'
    "For an OPTION pick (long call/put) also add: \"instrument\": \"option\", "
    "\"option_right\": \"call\" or \"put\", \"contract_symbol\": \"<exact OCC symbol from "
    "get_option_chain>\", \"expiry\": \"YYYY-MM-DD\", \"strike\": <number>. Omit these for stocks.\n"
    "Rules: go long OR short as your thesis dictates (direction 'long' = bet it rises, "
    "'short' = bet it falls). requested_weight_pct is a fraction of the portfolio (0.08 = 8%). "
    "Base every number on your research. Output ONLY the JSON object."
)

# Deliberately NOT the persona prompt: the persona says "always use tools / never estimate",
# which makes a reasoning model defer to tools and emit an empty final channel at finalize
# time. A clean "research is done" system makes gpt-oss commit to the JSON answer.
FINALIZE_SYSTEM = (
    "You are finalizing stock picks from COMPLETED research. Do not ask for tools or more "
    "data — research is done. Output ONLY the requested JSON object, nothing else."
)


def _short(args: dict) -> str:
    try:
        return ", ".join(f"{k}={str(v)[:30]}" for k, v in args.items())
    except Exception:
        return str(args)[:60]


def _prompt_tokens(ai, messages: list) -> int:
    """Best available count of tokens the last prompt consumed (for compaction timing)."""
    um = getattr(ai, "usage_metadata", None) or {}
    if um.get("input_tokens"):
        return um["input_tokens"]
    tu = (getattr(ai, "response_metadata", {}) or {}).get("token_usage") or {}
    if tu.get("prompt_tokens"):
        return tu["prompt_tokens"]
    return compaction.estimate_tokens(messages)


def _extract_between(text: str, open_ch: str, close_ch: str) -> str | None:
    start, end = text.find(open_ch), text.rfind(close_ch)
    if start != -1 and end != -1 and end > start:
        return text[start : end + 1]
    return None


def parse_picks(text: str, default_stop: float, default_take: float) -> tuple[list[Pick], list[str]]:
    """Parse the finalize output into validated Picks, collecting per-item errors.

    Accepts an object {"picks": [...]}, a bare array, or those embedded in surrounding text
    / code fences. Robust to reasoning-model chatter around the JSON.
    """
    t = text.strip()
    if t.startswith("```"):
        t = t.strip("`")
        if t.lower().startswith("json"):
            t = t[4:]
        t = t.strip()

    data = None
    for candidate in (t, _extract_between(t, "{", "}"), _extract_between(t, "[", "]")):
        if not candidate:
            continue
        try:
            parsed = json.loads(candidate)
        except Exception:
            continue
        if isinstance(parsed, dict) and isinstance(parsed.get("picks"), list):
            data = parsed["picks"]
            break
        if isinstance(parsed, list):
            data = parsed
            break
    if data is None:
        return [], ["finalize: no JSON picks found in model output"]

    picks, errors = [], []
    for item in data:
        if not isinstance(item, dict):
            errors.append(f"finalize: skipped non-object item {item!r}")
            continue
        item.setdefault("stop_loss_pct", default_stop)
        item.setdefault("take_profit_pct", default_take)
        try:
            picks.append(Pick(**item))
        except Exception as e:
            errors.append(f"finalize: invalid pick {item.get('ticker', '?')}: {e}")
    return picks, errors


def _finalize(persona: Persona, running_summary: str, reasoning_trace: list[str]):
    """Ask the model for structured picks using a clean finalize context (not the persona's
    tool-pushing prompt), fed the research findings we gathered. Retries once."""
    findings = "\n".join(reasoning_trace[-50:])
    if running_summary:
        findings = f"Summary:\n{running_summary}\n\nDetails:\n{findings}"
    human = (
        f"You are {persona.display_name}. Choose picks that fit your style "
        f"(typical stop ~{persona.default_stop_pct * 100:.0f}%, "
        f"take ~{persona.default_take_pct * 100:.0f}%, horizon {persona.preferred_horizon}).\n\n"
        f"RESEARCH FINDINGS:\n{findings or '(sparse — use your best judgment)'}\n\n"
        f"{FINALIZE_PROMPT.format(n=persona.target_pick_count)}"
    )
    llm = build_llm(max_tokens=4096)  # no tools, no json_mode (json_mode returns empty on oMLX)
    errors: list[str] = []
    for _ in range(2):
        try:
            resp = llm.invoke(
                [SystemMessage(content=FINALIZE_SYSTEM), HumanMessage(content=human)]
            )
            content = resp.content if isinstance(resp.content, str) else str(resp.content)
            picks, errs = parse_picks(content, persona.default_stop_pct, persona.default_take_pct)
            if picks:
                return picks, errs
            errors = errs
        except Exception as e:
            errors = [f"finalize llm error: {e}"]
    return [], errors


def run_agent(
    persona: Persona,
    *,
    deadline_ts: float,
    step_budget: int,
    min_steps: int = 0,
    checkpointer=None,
    model_name: str = "",
) -> AgentOutput:
    start = time.time()
    started_at = datetime.now(timezone.utc)
    tools = build_toolset(persona)
    tools_by_name = {t.name: t for t in tools}
    tool_llm = build_llm(tools=tools)
    plain_llm = build_llm()  # no tools — used for summarization
    ctx_len = get_settings().llm_context_length

    date_str = datetime.now().astimezone().strftime("%A %Y-%m-%d")
    messages: list = [
        SystemMessage(content=persona.system_prompt),
        HumanMessage(content=RESEARCH_KICKOFF.format(date=date_str, n=persona.target_pick_count)),
    ]
    running_summary = ""
    reasoning_trace: list[str] = []
    steps_used = 0
    ended_reason = "natural"

    def snapshot(status: AgentStatus, picks: list[Pick] | None = None, errors=None) -> AgentOutput:
        return AgentOutput(
            agent_name=persona.display_name,
            persona_id=persona.id,
            started_at=started_at,
            ended_at=datetime.now(timezone.utc),
            status=status,
            steps_used=steps_used,
            wall_seconds=round(time.time() - start, 1),
            market_summary=running_summary,
            picks=picks or [],
            reasoning_trace=list(reasoning_trace),
            errors=errors or [],
        )

    # --- research loop ---
    while steps_used < step_budget:
        if time.time() >= deadline_ts:
            ended_reason = "timeout"
            break
        steps_used += 1
        try:
            ai = tool_llm.invoke(messages)
        except Exception as e:
            reasoning_trace.append(f"[error] llm.invoke: {e}")
            ended_reason = "error"
            logger.warning("{}: llm.invoke failed: {}", persona.id, e)
            break
        messages.append(ai)

        tool_calls = getattr(ai, "tool_calls", None)
        if not tool_calls:
            # Force depth: if the model wants to stop early, nudge it to keep researching
            # until it has done at least min_steps (still bounded by the wall-clock deadline).
            if steps_used < min_steps and time.time() < deadline_ts:
                messages.append(HumanMessage(content=DEEPEN_NUDGE))
                reasoning_trace.append(f"[nudge] deepen research ({steps_used}/{min_steps} steps)")
                continue
            if ai.content:
                reasoning_trace.append(f"[note] {str(ai.content)[:200]}")
            ended_reason = "natural"
            break

        for tc in tool_calls:
            name, args = tc["name"], tc.get("args", {}) or {}
            tool = tools_by_name.get(name)
            try:
                result = tool.invoke(args) if tool else f"ERROR: unknown tool {name}"
            except Exception as e:
                result = f"ERROR: {type(e).__name__}: {e}"
            messages.append(ToolMessage(content=str(result), tool_call_id=tc["id"]))
            reasoning_trace.append(f"{name}({_short(args)}) -> {str(result)[:160]}")

        if compaction.should_compact(_prompt_tokens(ai, messages), ctx_len):
            messages, running_summary = compaction.compact(plain_llm, messages, running_summary)

        if checkpointer:  # incremental save after each round
            checkpointer.write(snapshot(AgentStatus.PARTIAL))

    if steps_used >= step_budget and ended_reason == "natural":
        ended_reason = "budget"

    # --- finalize (always attempted, even on timeout, to salvage picks) ---
    picks: list[Pick] = []
    errors: list[str] = []
    if ended_reason != "error":
        picks, errors = _finalize(persona, running_summary, reasoning_trace)

    if ended_reason == "error":
        status = AgentStatus.ERROR
    elif ended_reason == "timeout":
        status = AgentStatus.TIMEOUT
    elif picks:
        status = AgentStatus.COMPLETE
    else:
        status = AgentStatus.PARTIAL

    out = snapshot(status, picks=picks, errors=errors)
    if checkpointer:
        checkpointer.write(out)
    logger.info(
        "{}: status={} steps={} picks={} wall={}s",
        persona.id, status.value, steps_used, len(picks), out.wall_seconds,
    )
    return out