api/config.py

"""Settings, read from environment variables (see .env.example)."""

from __future__ import annotations

import os
from dataclasses import dataclass
from functools import lru_cache
from pathlib import Path

REPO_ROOT = Path(__file__).resolve().parent.parent


def _env(name: str, default: str) -> str:
    value = os.environ.get(name)
    return value if value not in (None, "") else default


@dataclass(frozen=True)
class Settings:
    # Model: one model for every tester in Stage 1, via any OpenAI-compatible endpoint (OpenRouter by default).
    llm_api_key: str
    llm_base_url: str
    llm_model: str  # scene generation (/simulate/generate)
    code_llm_model: str  # ROS 2 code generation: OpenCode's model, or the direct backend's
    llm_timeout: float

    # Storage
    data_dir: Path
    testers_file: Path
    log_db: Path

    # MuJoCo generation
    menagerie_dir: Path
    mjcf_max_attempts: int
    mjcf_sim_seconds: float
    validate_timeout: float

    # ROS 2 generation
    codegen_backend: str  # "opencode" or "direct"
    opencode_bin: str
    opencode_timeout: float
    code_max_attempts: int

    # Simulation runs (Modal or RunPod serverless, results in Cloudflare R2)
    sim_backend: str  # "modal" or "runpod"
    modal_app: str
    runpod_api_key: str
    runpod_endpoint_id: str
    run_execution_timeout_s: int
    run_max_duration_s: float
    run_burst_warn: int  # log a warning when a tester starts more runs than this within 10 minutes
    report_secret: str  # signs report links
    public_base_url: str  # how testers reach this API (for report links); empty = use the request's host
    # Spend: estimated from each job's run time; new jobs are refused once the month's estimate reaches the limit.
    monthly_credit_usd: float  # 0 = no limit
    # ...and each tester gets an equal share of it, enforced before every job against the job's worst case.
    compute_testers: int  # how many testers share the compute credit
    compute_per_tester_usd: float  # 0 = monthly_credit_usd / compute_testers
    storage_limit_gb: float  # new jobs are refused while the R2 bucket holds more than this (0 = no limit)
    job_prices: dict  # USD per second by hardware: t4, l4, cpu, runpod
    job_overhead_s: float  # billed seconds per job beyond the worker's own timing (start-up, idle shutdown)
    isaac_daily_limit: int  # Isaac Sim renders per day, across all testers
    auto_review: bool  # review every finished simulation video with Cosmos Reason
    # AI budget in rupees for the test group: one shared pool, an equal allocation per tester. Only billable
    # AI calls count (OpenRouter-exempt ones are recorded but not deducted); GPU/CPU jobs have their own
    # dollar limit above (Modal's free credit).
    usd_inr: float
    budget_total_inr: float
    budget_per_user_inr: float
    confirm_above_inr: float  # ask before running a task estimated to cost more than this
    admin_testers: tuple  # testers who see the team budget view
    hf_namespace: str  # Hugging Face user/org that dataset exports are pushed to (with the worker's HF_TOKEN)
    foxglove_org: str  # optional Foxglove organization slug for the embedded viewer
    # Prompt enhancement through OpenCode with a free OpenCode Zen model (never OpenRouter, never billed).
    prompt_enhance: bool
    enhance_model: str  # "<opencode provider>/<model>"
    enhance_timeout: float
    zen_api_key: str  # OpenCode Zen key, if the free model asks for one
    # Debugging: the first repair goes to a free OpenCode Zen model first (paid tier only if that call fails).
    free_debug: bool
    debug_model: str  # "opencode/<model>"
    owner_api_key: str  # persistent key for the owner (bypasses budgets, doesn't count as tester)

    @property
    def compute_share_usd(self) -> float:
        """Each tester's monthly GPU/CPU allowance in dollars (0 = no per-tester limit)."""
        if self.compute_per_tester_usd > 0:
            return self.compute_per_tester_usd
        if self.monthly_credit_usd > 0 and self.compute_testers > 0:
            return round(self.monthly_credit_usd / self.compute_testers, 6)
        return 0.0

    @property
    def interfaces_file(self) -> Path:
        return self.data_dir / "ros2_interfaces.json"

    @property
    def mjcf_schema_file(self) -> Path:
        return self.data_dir / "mjcf_schema.json"


# Estimated USD per second of container time, GPU plus the CPU and memory that come with it
# (Modal list prices, Sept 2026; override with JOB_PRICES="t4=...,l4=...,cpu=...,runpod=...").
DEFAULT_PRICES = {"t4": 0.00019, "l4": 0.00025, "cpu": 0.00004, "runpod": 0.00031}


def _prices(spec: str) -> dict:
    prices = dict(DEFAULT_PRICES)
    for part in filter(None, (p.strip() for p in spec.split(","))):
        key, _, value = part.partition("=")
        prices[key.strip()] = float(value)
    return prices


@lru_cache
def get_settings() -> Settings:
    data_dir = Path(_env("ROSDIFF_DATA_DIR", str(REPO_ROOT / "data"))).resolve()
    return Settings(
        llm_api_key=_env("LLM_API_KEY", ""),
        llm_base_url=_env("LLM_BASE_URL", "https://openrouter.ai/api/v1").rstrip("/"),
        llm_model=_env("LLM_MODEL", "z-ai/glm-5.2"),
        # DeepSeek V4 Flash: 79.0% on SWE-bench Verified (V4 Pro: 80.6%) at about $0.10/$0.20 per M tokens on
        # OpenRouter (Sept 2026) - enough for a test phase. For more quality: deepseek/deepseek-v4-pro.
        code_llm_model=_env("CODE_LLM_MODEL", "deepseek/deepseek-v4-flash"),
        llm_timeout=float(_env("LLM_TIMEOUT", "180")),
        data_dir=data_dir,
        testers_file=Path(_env("ROSDIFF_TESTERS_FILE", str(REPO_ROOT / "testers.yaml"))),
        log_db=Path(_env("ROSDIFF_LOG_DB", str(data_dir / "generations.sqlite3"))),
        menagerie_dir=Path(_env("MENAGERIE_DIR", str(data_dir / "sources" / "mujoco_menagerie"))),
        mjcf_max_attempts=int(_env("MJCF_MAX_ATTEMPTS", "3")),
        mjcf_sim_seconds=float(_env("MJCF_SIM_SECONDS", "2.0")),
        validate_timeout=float(_env("VALIDATE_TIMEOUT", "60")),
        codegen_backend=_env("CODEGEN_BACKEND", "opencode"),
        opencode_bin=_env("OPENCODE_BIN", "opencode"),
        opencode_timeout=float(_env("OPENCODE_TIMEOUT", "600")),
        code_max_attempts=int(_env("CODE_MAX_ATTEMPTS", "3")),
        sim_backend=_env("SIM_BACKEND", "modal").lower(),
        modal_app=_env("MODAL_APP", "rosdiff-sim"),
        runpod_api_key=_env("RUNPOD_API_KEY", ""),
        runpod_endpoint_id=_env("RUNPOD_ENDPOINT_ID", ""),
        run_execution_timeout_s=int(_env("RUN_EXECUTION_TIMEOUT_S", "600")),
        run_max_duration_s=float(_env("RUN_MAX_DURATION_S", "60")),
        run_burst_warn=int(_env("RUN_BURST_WARN", "10")),
        report_secret=_env("REPORT_SECRET", ""),
        public_base_url=_env("PUBLIC_BASE_URL", "").rstrip("/"),
        monthly_credit_usd=float(_env("MONTHLY_CREDIT_LIMIT_USD", "29.98")),
        compute_testers=int(_env("COMPUTE_TESTERS", "4")),
        compute_per_tester_usd=float(_env("COMPUTE_PER_TESTER_USD", "0")),
        storage_limit_gb=float(_env("STORAGE_LIMIT_GB", "9")),
        job_prices=_prices(_env("JOB_PRICES", "")),
        job_overhead_s=float(_env("JOB_OVERHEAD_S", "45")),
        isaac_daily_limit=int(_env("ISAAC_DAILY_LIMIT", "10")),
        auto_review=_env("AUTO_REVIEW", "false").lower() in ("1", "true", "yes"),
        hf_namespace=_env("HF_NAMESPACE", "rosdiff"),
        usd_inr=float(_env("USD_INR", "88")),
        budget_total_inr=float(_env("BUDGET_TOTAL_INR", "500")),
        budget_per_user_inr=float(_env("BUDGET_PER_USER_INR", "125")),
        confirm_above_inr=float(_env("CONFIRM_ABOVE_INR", "10")),
        admin_testers=tuple(n.strip() for n in _env("ADMIN_TESTERS", "").split(",") if n.strip()),
        foxglove_org=_env("FOXGLOVE_ORG", ""),
        prompt_enhance=_env("PROMPT_ENHANCE", "true").lower() in ("1", "true", "yes"),
        enhance_model=_env("ENHANCE_MODEL", "opencode/deepseek-v4-flash-free"),
        enhance_timeout=float(_env("ENHANCE_TIMEOUT", "90")),
        zen_api_key=_env("ZEN_API_KEY", ""),
        free_debug=_env("FREE_DEBUG", "true").lower() in ("1", "true", "yes"),
        debug_model=_env("DEBUG_MODEL", "opencode/deepseek-v4-flash-free"),
        owner_api_key=_env("OWNER_API_KEY", ""),
    )