agnes-the-ai-analyst/services/scheduler/__main__.py

"""Scheduler service — replaces systemd timers.

Lightweight sidecar that fires scheduled jobs over HTTP against the main
app. Authenticates with ``SCHEDULER_API_TOKEN`` (shared-secret synthetic
admin — see ``app.auth.scheduler_token``); falls back to no-auth in
LOCAL_DEV_MODE.

Schedules are strings parsed by ``src.scheduler.is_table_due`` — accepts
"every 15m", "every 1h", "daily 03:00", "daily 07:00,13:00".

Why every job is HTTP and nothing runs in-process: the scheduler container
shares ``/data/state/system.duckdb`` with the app container, but DuckDB
permits only one writer per file across processes. An in-process call
from the scheduler raced the app's long-lived handle and 500-ed on
``Could not set lock on file``. Going through HTTP makes the app the sole
writer; the scheduler is reduced to a pure cron clock.

Usage: python -m services.scheduler
"""

import logging
import os
import signal
import time
from datetime import datetime, timezone

import httpx

from app.logging_config import setup_logging
from src.scheduler import is_table_due

setup_logging(__name__)
logger = logging.getLogger(__name__)

API_URL = os.environ.get("API_URL", "http://localhost:8000")
SCHEDULER_API_TOKEN = os.environ.get("SCHEDULER_API_TOKEN", "")

_token_warning_emitted = False


def _get_auth_token() -> str:
    """Return the bearer token for API calls.

    Production: ``SCHEDULER_API_TOKEN`` is a shared secret generated by the
    Terraform startup script and written to ``/opt/agnes/.env``. Both the
    ``app`` and ``scheduler`` containers source the same .env via Docker
    Compose ``env_file:``, so the secret is symmetric. The app validates
    incoming Bearer tokens against this env var (constant-time compare in
    ``app.auth.scheduler_token``) and resolves matches to a synthetic
    ``scheduler@system.local`` user that is a member of the Admin group.

    Dev / LOCAL_DEV_MODE: leave it unset. The scheduler returns the empty
    string and calls the API without an ``Authorization`` header — the
    API's dev-bypass auto-authenticates the request as the dev user.
    """
    global _token_warning_emitted
    if SCHEDULER_API_TOKEN:
        return SCHEDULER_API_TOKEN
    if not _token_warning_emitted:
        logger.warning(
            "SCHEDULER_API_TOKEN is not set — calling the API without "
            "Authorization. Required in production; in LOCAL_DEV_MODE "
            "the dev-bypass auto-authenticates and this is fine."
        )
        _token_warning_emitted = True
    return ""


# --- Env parsing ------------------------------------------------------------

_DEFAULTS = {
    "SCHEDULER_DATA_REFRESH_INTERVAL": 15 * 60,   # seconds
    "SCHEDULER_HEALTH_CHECK_INTERVAL": 5 * 60,
    "SCHEDULER_SCRIPT_RUN_INTERVAL":   1 * 60,
    "SCHEDULER_TICK_SECONDS":          30,
    # LLM pipeline cadences (#176, #179 review). Defaults preserve the
    # 10m / 15m / 17m coprime offset so the three jobs don't fire on the
    # same tick and stack their API + DB load. The verification-detector
    # default (900s) is also the source of truth for the health-check
    # staleness grace window in app/api/health.py — single env var drives
    # both, so an operator changing the cadence moves both.
    "SCHEDULER_SESSION_COLLECTOR_INTERVAL":     10 * 60,
    "SCHEDULER_VERIFICATION_DETECTOR_INTERVAL": 15 * 60,
    "SCHEDULER_CORPORATE_MEMORY_INTERVAL":      17 * 60,
}


def _read_positive_int(name: str) -> int:
    """Read an env var as a positive integer or fall back to the default.

    Treats unset env (``None``) as "use default". Treats explicitly empty
    string (``""``) as an operator typo and raises — silently defaulting
    on a literal ``FOO=`` in the env_file would mask configuration bugs.
    """
    raw = os.environ.get(name)
    if raw is None:
        if name not in _DEFAULTS:
            raise ValueError(f"Unknown scheduler env var: {name}")
        return _DEFAULTS[name]
    if raw == "":
        raise ValueError(f"{name}='' must be a positive integer (seconds)")
    try:
        value = int(raw)
    except (TypeError, ValueError):
        raise ValueError(f"{name}={raw!r} must be a positive integer (seconds)")
    if value <= 0:
        raise ValueError(f"{name}={value} must be > 0 (seconds)")
    return value


def _seconds_to_schedule(seconds: int) -> str:
    """Convert a seconds value to the closest 'every Nm' / 'every Nh' string.

    Uses ceiling division so a non-multiple-of-60 input never produces a
    schedule that fires MORE often than the operator configured (90s →
    'every 2m', not 'every 1m'). Sub-minute inputs clamp to 'every 1m'
    because the schedule grammar has minute-level resolution.
    """
    if seconds % 3600 == 0 and seconds >= 3600:
        return f"every {seconds // 3600}h"
    # Ceiling division: -(-x // y) is the standard trick.
    minutes = max(1, -(-seconds // 60))
    return f"every {minutes}m"


def resolved_tick_seconds() -> int:
    """Read + validate SCHEDULER_TICK_SECONDS in isolation (test helper)."""
    return _read_positive_int("SCHEDULER_TICK_SECONDS")


def build_jobs() -> list[tuple[str, str, str, str, int]]:
    """Build the JOBS list from env, applying defaults and validation.

    Tuple shape: (name, schedule_string, endpoint, method, http_timeout_sec).
    Marketplaces stays hardcoded — promoting it to env is out of #77 scope.
    """
    refresh = _read_positive_int("SCHEDULER_DATA_REFRESH_INTERVAL")
    health  = _read_positive_int("SCHEDULER_HEALTH_CHECK_INTERVAL")
    scripts = _read_positive_int("SCHEDULER_SCRIPT_RUN_INTERVAL")
    sess    = _read_positive_int("SCHEDULER_SESSION_COLLECTOR_INTERVAL")
    verify  = _read_positive_int("SCHEDULER_VERIFICATION_DETECTOR_INTERVAL")
    corpmem = _read_positive_int("SCHEDULER_CORPORATE_MEMORY_INTERVAL")
    tick    = _read_positive_int("SCHEDULER_TICK_SECONDS")
    smallest = min(refresh, health, scripts, sess, verify, corpmem)
    if tick > smallest:
        raise ValueError(
            f"SCHEDULER_TICK_SECONDS={tick} must be <= the smallest job "
            f"interval ({smallest}s) so jobs don't consistently miss their "
            f"cadence by up to one tick"
        )
    return [
        ("data-refresh",          _seconds_to_schedule(refresh), "/api/sync/trigger",                    "POST", 120),
        ("health-check",          _seconds_to_schedule(health),  "/api/health",                          "GET",   30),
        ("script-runner",         _seconds_to_schedule(scripts), "/api/scripts/run-due",                 "POST", 600),
        ("marketplaces",          "daily 03:00",                 "/api/marketplaces/sync-all",           "POST", 900),
        # LLM pipeline (#176, #179 review). Cadences are deliberately offset
        # (10m / 15m / 17m by default — all coprime modulo the 30s tick) so
        # the three LLM-driven jobs don't fire on the same tick and stack
        # their API + DB load. Driven by env so an operator can throttle
        # without a code change; the verification-detector cadence is the
        # single source of truth for the health-check staleness grace
        # window in app/api/health.py (which uses 2x the cadence).
        ("session-collector",     _seconds_to_schedule(sess),    "/api/admin/run-session-collector",     "POST", 300),
        ("verification-detector", _seconds_to_schedule(verify),  "/api/admin/run-verification-detector", "POST", 900),
        ("corporate-memory",      _seconds_to_schedule(corpmem), "/api/admin/run-corporate-memory",      "POST", 900),
    ]

_running = True


def _signal_handler(sig, frame):
    global _running
    logger.info(f"Received signal {sig}, shutting down...")
    _running = False


def _call_api(endpoint: str, method: str, timeout_sec: int) -> bool:
    """Call the main app API. Returns True on success."""
    url = f"{API_URL}{endpoint}"
    headers = {}
    token = _get_auth_token()
    if token:
        headers["Authorization"] = f"Bearer {token}"
    try:
        if method == "POST":
            resp = httpx.post(url, headers=headers, timeout=timeout_sec)
        else:
            resp = httpx.get(url, headers=headers, timeout=timeout_sec)
        if resp.status_code < 400:
            logger.info(f"Job {endpoint}: {resp.status_code}")
            return True
        else:
            logger.warning(f"Job {endpoint}: HTTP {resp.status_code} - {resp.text[:200]}")
            return False
    except Exception as e:
        logger.error(f"Job {endpoint} failed: {e}")
        try:
            from src.observability import get_posthog
            get_posthog().capture_exception(
                e,
                distinct_id="system",
                properties={"job": endpoint, "method": method, "component": "scheduler"},
            )
        except Exception:
            logger.exception("PostHog capture_exception failed in scheduler")
        return False


def run():
    signal.signal(signal.SIGTERM, _signal_handler)
    signal.signal(signal.SIGINT, _signal_handler)

    jobs = build_jobs()
    tick = resolved_tick_seconds()
    logger.info(
        "Scheduler started. API_URL=%s, %d jobs, tick=%ds. Schedules: %s",
        API_URL, len(jobs), tick,
        {name: schedule for name, schedule, *_ in jobs},
    )

    last_run: dict[str, str | None] = {name: None for name, *_ in jobs}

    while _running:
        now_iso = datetime.now(timezone.utc).isoformat()
        for name, schedule, endpoint, method, timeout_sec in jobs:
            if not is_table_due(schedule, last_run[name]):
                continue
            logger.info("Running job: %s (%s)", name, schedule)
            ok = _call_api(endpoint, method, timeout_sec)
            if ok:
                last_run[name] = now_iso
        time.sleep(tick)

    logger.info("Scheduler stopped.")


if __name__ == "__main__":
    run()