commit 05bb4193adc27174c6b7de79c1f1101d4e0b1b5b
Author: nick2day <nick2day@doubleday.haus>
Date:   Sun Mar 22 01:36:19 2026 -0400

    Initial commit: FGS metal discovery standalone agent
    
    FastAPI service replacing the 77-node n8n pipeline. Implements full
    discovery pipeline: 29 rotating SearXNG queries, nomic-embed-text
    scoring against Last.fm taste centroid, Mistral-nemo prefilter and
    curator with provenance validation, SQLite dedup, writes to
    metal-picks.json for the existing FGS dashboard.
    
    Runs as systemd service on port 8766 (fgs-agent.home via Caddy).
    n8n reduced to a 2-node schedule trigger → HTTP POST.
    
    Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

diff --git a/.gitignore b/.gitignore
new file mode 100644
index 0000000..3c56fbd
--- /dev/null
+++ b/.gitignore
@@ -0,0 +1,5 @@
+venv/
+__pycache__/
+*.pyc
+.env
+dedup.db
diff --git a/agent.py b/agent.py
new file mode 100644
index 0000000..51a15d1
--- /dev/null
+++ b/agent.py
@@ -0,0 +1,721 @@
+"""
+fgs-agent — First Gain Stage metal discovery pipeline
+Standalone FastAPI service; n8n just POSTs to /api/fgs/run on a schedule.
+
+Endpoints:
+  POST /api/fgs/run      — trigger a discovery run (async, returns run_id)
+  GET  /api/fgs/status   — current run status
+  GET  /api/fgs/picks    — read metal-picks.json
+  GET  /api/fgs/health   — liveness check
+"""
+
+import asyncio
+import json
+import logging
+import math
+import re
+import time
+import uuid
+from datetime import datetime, timezone
+from typing import Any, Optional
+
+import aiosqlite
+import httpx
+import uvicorn
+from fastapi import BackgroundTasks, FastAPI
+from fastapi.responses import JSONResponse
+from pydantic import BaseModel
+from pydantic_settings import BaseSettings, SettingsConfigDict
+
+logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
+log = logging.getLogger("fgs-agent")
+
+
+# ── Config ──────────────────────────────────────────────────────────────────
+
+class Settings(BaseSettings):
+    model_config = SettingsConfigDict(env_file=".env", env_file_encoding="utf-8")
+
+    lastfm_username: str = "nick2day"
+    lastfm_api_key: str = ""
+    searxng_url: str = "http://localhost:8080"
+    ollama_url: str = "http://localhost:11434"
+    picks_path: str = "/home/nick2day/.n8n/metal-picks.json"
+    dedup_db: str = "/home/nick2day/fgs-agent/dedup.db"
+    port: int = 8766
+
+
+settings = Settings()
+
+
+# ── SEARXNG QUERY BANK ───────────────────────────────────────────────────────
+# All 29 queries — dynamic year, rotating subgenre index
+
+SUBGENRES = [
+    "death metal", "black metal", "doom metal", "thrash metal", "sludge metal",
+    "progressive metal", "blackened death metal", "post-metal", "funeral doom", "grindcore",
+    "melodic death metal", "technical death metal", "atmospheric black metal", "war metal",
+    "drone metal", "stoner metal", "gothic metal", "power metal", "viking metal", "brutal death metal",
+]
+
+def build_queries() -> list[str]:
+    year = datetime.now().year
+    # rotate subgenre every 6 hours so successive runs hit different niches
+    idx = int(time.time() // 21600) % len(SUBGENRES)
+    sg = SUBGENRES[idx]
+    sg2 = SUBGENRES[(idx + 1) % len(SUBGENRES)]
+    sg3 = SUBGENRES[(idx + 2) % len(SUBGENRES)]
+    return [
+        # Subgenre new releases
+        f'{sg} new album release {year}',
+        f'{sg2} new album {year}',
+        f'{sg3} full album stream {year}',
+        # Bandcamp label searches
+        f'site:bandcamp.com "sentient ruin" new {year}',
+        f'site:bandcamp.com "20 buck spin" new {year}',
+        f'site:bandcamp.com "prosthetic records" {year}',
+        f'site:bandcamp.com "unique leader records" {year}',
+        f'site:bandcamp.com "redefining darkness" {year}',
+        f'site:bandcamp.com "profound lore" {year}',
+        f'site:bandcamp.com "season of mist" {year}',
+        f'site:bandcamp.com "century media" {year}',
+        f'site:bandcamp.com "nuclear blast" {year}',
+        # Bandcamp genre pages
+        f'site:bandcamp.com {sg} album {year}',
+        f'site:bandcamp.com {sg2} album {year}',
+        # Metal reviews / coverage
+        f'site:metal-archives.com new album review {year}',
+        f'site:metalinjection.net new album {year}',
+        f'site:cvltnation.com new album {year}',
+        f'site:decibelmag.com new album stream {year}',
+        f'site:invisibleoranges.com album review {year}',
+        f'site:brooklynvegan.com {sg} new album {year}',
+        # Obscure/underground
+        f'underground {sg} demo release {year}',
+        f'independent {sg} album bandcamp {year}',
+        f'new {sg} EP release {year}',
+        # Michigan / regional bias
+        f'Michigan metal band new album {year}',
+        f'Detroit metal band new release {year}',
+        # Best-of / roundup (useful for discovery)
+        f'best {sg} albums {year}',
+        f'new {sg} releases {year} bandcamp',
+        # Last.fm tag pages
+        f'site:last.fm {sg} new {year}',
+        # Full-album YouTube
+        f'"{sg}" "full album" {year} new band',
+    ]
+
+
+# ── NOISE FILTERS ────────────────────────────────────────────────────────────
+
+NOISE_PATTERNS = [
+    re.compile(r'wikipedia\.org', re.I),
+    re.compile(r'\bin heavy metal music\b', re.I),
+    re.compile(r'album release calendar', re.I),
+    re.compile(r'most anticipated', re.I),
+    re.compile(r'\byear in review\b', re.I),
+    re.compile(r'^top \d+', re.I),
+    re.compile(r'albums you need to hear', re.I),
+    re.compile(r'upcoming.*releases', re.I),
+    re.compile(r'\bto release\b.*album', re.I),
+    re.compile(r'\bannounce[sd]?\b.*album', re.I),
+    re.compile(r'\bdelayed\b', re.I),
+    re.compile(r'best.*releases', re.I),
+    re.compile(r'\d+ metal albums.*20\d\d', re.I),
+    re.compile(r'shatner', re.I),
+    re.compile(r'mark your calendars', re.I),
+    re.compile(r'\bvideo game\b', re.I),
+    re.compile(r'\bsoundtrack\b', re.I),
+    re.compile(r'\btribute\b', re.I),
+    re.compile(r'\bcovers album\b', re.I),
+    re.compile(r'heavy metal music - wikipedia', re.I),
+    re.compile(r'in \d{4} - wikipedia', re.I),
+]
+
+
+def is_noise(title: str, url: str) -> bool:
+    text = f"{title} {url}"
+    return any(p.search(text) for p in NOISE_PATTERNS)
+
+
+def parse_title(title: str) -> tuple[str, str]:
+    """Extract (artist, album) from common title formats."""
+    # "Album Title, by Artist Name" — Bandcamp format
+    m = re.match(r'^(.+?),\s+by\s+(.+)$', title, re.I)
+    if m:
+        return m.group(2).strip(), m.group(1).strip()
+    # "Artist - Album" or "Artist – Album"
+    m = re.match(r'^([^-–]+?)\s*[-–]\s*(.+)$', title)
+    if m and len(m.group(1)) < 60:
+        return m.group(1).strip(), m.group(2).strip()
+    return '', title.strip()
+
+
+def artist_from_url(url: str) -> str:
+    """Extract artist slug from Bandcamp URL."""
+    m = re.match(r'https?://([^.]+)\.bandcamp\.com', url, re.I)
+    if m:
+        return m.group(1).replace('-', ' ')
+    return ''
+
+
+# ── DEDUP ────────────────────────────────────────────────────────────────────
+
+async def init_dedup_db():
+    async with aiosqlite.connect(settings.dedup_db) as db:
+        await db.execute(
+            "CREATE TABLE IF NOT EXISTS seen (key TEXT PRIMARY KEY, ts INTEGER)"
+        )
+        await db.commit()
+
+
+async def filter_unseen(candidates: list[dict]) -> list[dict]:
+    """Remove already-seen albums. Returns only fresh candidates."""
+    async with aiosqlite.connect(settings.dedup_db) as db:
+        fresh = []
+        for c in candidates:
+            key = f"{c.get('artist','').lower()}::{c.get('album','').lower()}"
+            if not key or key == '::':
+                key = c.get('url', c.get('album', ''))
+            row = await db.execute("SELECT 1 FROM seen WHERE key=?", (key,))
+            if not await row.fetchone():
+                fresh.append(c)
+        return fresh
+
+
+async def mark_seen(picks: list[dict]):
+    async with aiosqlite.connect(settings.dedup_db) as db:
+        ts = int(time.time())
+        for p in picks:
+            key = f"{p.get('artist','').lower()}::{p.get('album','').lower()}"
+            if key and key != '::':
+                await db.execute(
+                    "INSERT OR IGNORE INTO seen (key, ts) VALUES (?, ?)", (key, ts)
+                )
+        await db.commit()
+
+
+# ── PIPELINE STAGES ──────────────────────────────────────────────────────────
+
+async def fetch_lastfm(client: httpx.AsyncClient) -> dict:
+    """Fetch top artists + recent tracks from Last.fm."""
+    base = "https://ws.audioscrobbler.com/2.0/"
+    params_top = {
+        "method": "user.gettopartists",
+        "user": settings.lastfm_username,
+        "api_key": settings.lastfm_api_key,
+        "period": "12month",
+        "limit": "100",
+        "format": "json",
+    }
+    params_recent = {
+        "method": "user.getrecenttracks",
+        "user": settings.lastfm_username,
+        "api_key": settings.lastfm_api_key,
+        "limit": "1000",
+        "format": "json",
+    }
+    top_r, recent_r = await asyncio.gather(
+        client.get(base, params=params_top, timeout=15),
+        client.get(base, params=params_recent, timeout=15),
+    )
+    return {
+        "top": top_r.json() if top_r.status_code == 200 else {},
+        "recent": recent_r.json() if recent_r.status_code == 200 else {},
+    }
+
+
+def build_taste_profile(lastfm: dict) -> dict:
+    """Build taste centroid from Last.fm data."""
+    top_artists_raw = lastfm.get("top", {}).get("topartists", {}).get("artist", [])
+    recent_tracks_raw = lastfm.get("recent", {}).get("recenttracks", {}).get("track", [])
+
+    top_artists = [
+        {"name": a["name"], "playcount": int(a.get("playcount", 0)), "rank": int(a.get("@attr", {}).get("rank", 99))}
+        for a in top_artists_raw[:100]
+    ]
+
+    recent_map: dict[str, int] = {}
+    for t in recent_tracks_raw:
+        name = (t.get("artist") or {}).get("#text", "")
+        if name:
+            recent_map[name] = recent_map.get(name, 0) + 1
+    recent_artists = sorted(recent_map.items(), key=lambda x: -x[1])[:20]
+
+    return {
+        "topArtists": top_artists,
+        "recentArtists": [{"name": n, "plays": p} for n, p in recent_artists],
+    }
+
+
+async def search_candidates(client: httpx.AsyncClient) -> list[dict]:
+    """Run all SearXNG queries concurrently, collect and deduplicate results."""
+    queries = build_queries()
+    log.info(f"Running {len(queries)} SearXNG queries")
+
+    async def one_query(q: str) -> list[dict]:
+        try:
+            r = await client.get(
+                f"{settings.searxng_url}/search",
+                params={"q": q, "format": "json", "time_range": "month"},
+                timeout=20,
+            )
+            if r.status_code != 200:
+                return []
+            data = r.json()
+            results = data.get("results", [])[:15]  # top 15 per query
+            out = []
+            for res in results:
+                title = res.get("title", "")
+                url = res.get("url", "")
+                if is_noise(title, url):
+                    continue
+                artist, album = parse_title(title)
+                if not artist:
+                    artist = artist_from_url(url)
+                out.append({
+                    "title": title,
+                    "url": url,
+                    "artist": artist,
+                    "album": album,
+                    "source": q[:60],
+                    "content": (res.get("content") or "")[:200],
+                })
+            return out
+        except Exception as e:
+            log.warning(f"SearXNG query failed: {q[:40]} — {e}")
+            return []
+
+    results = await asyncio.gather(*[one_query(q) for q in queries])
+    all_results = [r for batch in results for r in batch]
+
+    # Deduplicate by URL
+    seen_urls: set[str] = set()
+    unique = []
+    for r in all_results:
+        u = r["url"]
+        if u and u not in seen_urls:
+            seen_urls.add(u)
+            unique.append(r)
+
+    log.info(f"Collected {len(unique)} unique candidates from search")
+    return unique
+
+
+async def embed_text(client: httpx.AsyncClient, text: str) -> list[float]:
+    try:
+        r = await client.post(
+            f"{settings.ollama_url}/api/embeddings",
+            json={"model": "nomic-embed-text", "prompt": text},
+            timeout=30,
+        )
+        if r.status_code == 200:
+            return r.json().get("embedding", [])
+    except Exception as e:
+        log.warning(f"Embed failed: {e}")
+    return []
+
+
+def cosine_similarity(a: list[float], b: list[float]) -> float:
+    if not a or not b or len(a) != len(b):
+        return 0.0
+    dot = sum(x * y for x, y in zip(a, b))
+    mag_a = math.sqrt(sum(x * x for x in a))
+    mag_b = math.sqrt(sum(x * x for x in b))
+    if mag_a == 0 or mag_b == 0:
+        return 0.0
+    return dot / (mag_a * mag_b)
+
+
+async def score_candidates(
+    client: httpx.AsyncClient, candidates: list[dict], taste: dict
+) -> list[dict]:
+    """Embed candidates and score against taste centroid."""
+    if not candidates:
+        return []
+
+    # Build taste centroid from top artists
+    top_names = [a["name"] for a in taste.get("topArtists", [])[:30]]
+    centroid_text = "Heavy metal artist: " + ", ".join(top_names)
+
+    log.info("Embedding taste centroid")
+    centroid_vec = await embed_text(client, centroid_text)
+
+    if not centroid_vec:
+        log.warning("Centroid embedding failed — scoring by 0")
+        for c in candidates:
+            c["embedScore"] = 0.0
+        return candidates
+
+    log.info(f"Scoring {len(candidates)} candidates against taste centroid")
+    scored = []
+    # Batch embed in groups to avoid hammering Ollama
+    for i, c in enumerate(candidates):
+        text = f"Metal album: {c['artist']} - {c['album']}" if c.get("artist") else f"Metal album: {c['album']}"
+        vec = await embed_text(client, text)
+        c["embedScore"] = round(cosine_similarity(centroid_vec, vec), 4) if vec else 0.0
+        if i % 20 == 0:
+            log.info(f"  scored {i}/{len(candidates)}")
+        scored.append(c)
+
+    scored.sort(key=lambda x: -x.get("embedScore", 0))
+    return scored
+
+
+async def prefilter_candidates(
+    client: httpx.AsyncClient, candidates: list[dict]
+) -> list[dict]:
+    """Use mistral-nemo to prefilter — keep actual album releases, drop noise."""
+    if not candidates:
+        return []
+
+    # Send in batches of 40
+    batch = candidates[:80]
+    items_for_prompt = [
+        {"index": i, "artist": c.get("artist", ""), "album": c.get("album", ""), "url": c.get("url", "")}
+        for i, c in enumerate(batch)
+    ]
+
+    prompt = f"""You are a metal music expert. Review these {len(items_for_prompt)} items from web searches.
+Return ONLY a JSON array of items that are actual metal album or EP releases (not news articles, not Wikipedia, not lists).
+For each item kept, include: index (integer), artist (string), album (string), subgenre (string), confidence (0.0-1.0).
+
+Items:
+{json.dumps(items_for_prompt, indent=2)}
+
+Return ONLY valid JSON array starting with ["""
+
+    body = {
+        "model": "mistral-nemo:latest",
+        "messages": [{"role": "user", "content": prompt}],
+        "stream": False,
+        "options": {"temperature": 0.1, "num_predict": 4000},
+    }
+
+    try:
+        r = await client.post(
+            f"{settings.ollama_url}/api/chat",
+            json=body,
+            timeout=120,
+        )
+        raw = r.json().get("message", {}).get("content", "") if r.status_code == 200 else ""
+    except Exception as e:
+        log.error(f"Prefilter failed: {e}")
+        return candidates[:40]  # fall through with top candidates
+
+    # Extract JSON from response
+    parsed = extract_json_array(raw)
+    if not parsed:
+        log.warning("Prefilter returned no parseable JSON — using all candidates")
+        return candidates[:40]
+
+    log.info(f"Prefilter kept {len(parsed)}/{len(batch)} candidates")
+
+    # Map back to full candidate objects with prefilter data
+    enriched = []
+    for p in parsed:
+        idx = p.get("index")
+        if idx is None or idx >= len(batch):
+            continue
+        orig = batch[idx].copy()
+        orig["artist"] = p.get("artist") or orig.get("artist", "")
+        orig["album"] = p.get("album") or orig.get("album", "")
+        orig["prefilter"] = {
+            "subgenre": p.get("subgenre", "unknown"),
+            "confidence": p.get("confidence", 0.5),
+        }
+        enriched.append(orig)
+
+    return enriched
+
+
+async def curate_picks(
+    client: httpx.AsyncClient, candidates: list[dict], taste: dict
+) -> list[dict]:
+    """Use mistral-nemo to produce final scored picks with full metadata."""
+    if not candidates:
+        return []
+
+    top = sorted(candidates, key=lambda x: -x.get("embedScore", 0))[:20]
+    top_artists = [a["name"] for a in taste.get("topArtists", [])[:20]]
+    recent_artists = [a["name"] for a in taste.get("recentArtists", [])[:10]]
+    year = datetime.now().year
+
+    items_for_prompt = [
+        {
+            "index": i,
+            "artist": c.get("artist", ""),
+            "album": c.get("album", ""),
+            "subgenre": (c.get("prefilter") or {}).get("subgenre", "unknown"),
+            "url": c.get("url", ""),
+            "source": c.get("source", ""),
+        }
+        for i, c in enumerate(top)
+    ]
+
+    prompt = f"""You are a heavy metal curator reviewing NEW {year} releases found via web search.
+
+IMPORTANT: Only score albums from the list below. Do NOT recommend albums not in this list.
+Do NOT recommend classic albums. These must be actual {year} releases found at the provided URLs.
+If you are not sure an item is a real new release, skip it.
+
+nick2day's taste profile:
+- Top artists (12-month): {', '.join(top_artists)}
+- Recent listens: {', '.join(recent_artists)}
+
+New {year} releases to evaluate (use ONLY these):
+{json.dumps(items_for_prompt, indent=2)}
+
+For each item worth recommending, output a JSON object with:
+  index (integer, from above), artist, album, score (0-100, based on fit to taste profile),
+  subgenre, obscurity (underground/indie/major), similar_to (2-3 artists from taste profile),
+  why (1 sentence), review_angle (unique angle), tags (string array), url (copy from above)
+
+Skip items without a clear artist + album. Skip classics or non-{year} releases.
+Return ONLY a JSON array starting with ["""
+
+    body = {
+        "model": "mistral-nemo:latest",
+        "messages": [{"role": "user", "content": prompt}],
+        "stream": False,
+        "options": {"temperature": 0.3, "num_predict": 8000},
+    }
+
+    # Build index → candidate map for provenance validation
+    top_by_index = {i: c for i, c in enumerate(top)}
+    known_urls = {c.get("url", "") for c in top if c.get("url")}
+
+    raw = ""
+    for attempt in range(3):
+        try:
+            r = await client.post(
+                f"{settings.ollama_url}/api/chat",
+                json=body,
+                timeout=180,
+            )
+            raw = r.json().get("message", {}).get("content", "") if r.status_code == 200 else ""
+            picks = extract_json_array(raw)
+            if picks:
+                # Validate provenance: url must match a search result OR index must be valid
+                validated = []
+                for p in picks:
+                    if p.get("score", 0) < 60:
+                        continue
+                    idx = p.get("index")
+                    url = p.get("url", "")
+                    if idx is not None and idx in top_by_index:
+                        # Merge URL from original candidate if LLM dropped it
+                        if not url:
+                            p["url"] = top_by_index[idx].get("url", "")
+                        validated.append(p)
+                    elif url and url in known_urls:
+                        validated.append(p)
+                    else:
+                        log.debug(f"Dropped hallucinated pick: {p.get('artist')} — {p.get('album')}")
+                log.info(f"Curator returned {len(validated)} validated picks (attempt {attempt+1})")
+                if validated:
+                    return validated
+        except Exception as e:
+            log.error(f"Curator attempt {attempt+1} failed: {e}")
+        await asyncio.sleep(2)
+
+    log.warning("Curator failed all attempts")
+    return []
+
+
+def extract_json_array(text: str) -> list:
+    """Robustly extract a JSON array from LLM output."""
+    if not text:
+        return []
+    # Strip markdown fences
+    cleaned = re.sub(r'```(?:json)?\s*', '', text).strip()
+    # Try direct parse
+    try:
+        data = json.loads(cleaned)
+        if isinstance(data, list):
+            return data
+        if isinstance(data, dict):
+            for v in data.values():
+                if isinstance(v, list):
+                    return v
+    except Exception:
+        pass
+    # Find first [ to last ]
+    start = cleaned.find('[')
+    end = cleaned.rfind(']')
+    if start != -1 and end > start:
+        try:
+            data = json.loads(cleaned[start:end+1])
+            if isinstance(data, list):
+                return data
+        except Exception:
+            pass
+    return []
+
+
+async def write_picks(new_picks: list[dict], run_stats: dict):
+    """Append new picks to metal-picks.json."""
+    try:
+        with open(settings.picks_path, "r") as f:
+            store = json.load(f)
+    except Exception:
+        store = {"allPicks": [], "runs": [], "lastUpdated": ""}
+
+    # Add run record
+    run = {
+        "timestamp": datetime.now(timezone.utc).isoformat(),
+        "source": "fgs-agent",
+        "picksAdded": len(new_picks),
+        "candidatesSearched": run_stats.get("candidates", 0),
+        "candidatesFiltered": run_stats.get("filtered", 0),
+    }
+    store["runs"] = ([run] + store.get("runs", []))[:10]
+    store["allPicks"] = (new_picks + store.get("allPicks", []))[:200]
+    store["lastUpdated"] = run["timestamp"]
+
+    with open(settings.picks_path, "w") as f:
+        json.dump(store, f, indent=2)
+
+    log.info(f"Wrote {len(new_picks)} new picks to {settings.picks_path}")
+
+
+# ── RUN STATE ────────────────────────────────────────────────────────────────
+
+run_state: dict[str, Any] = {
+    "status": "idle",
+    "run_id": None,
+    "started": None,
+    "finished": None,
+    "picks_added": 0,
+    "candidates_found": 0,
+    "error": None,
+    "log": [],
+}
+_run_lock = asyncio.Lock()
+
+
+async def pipeline_run(run_id: str):
+    global run_state
+
+    def log_step(msg: str):
+        log.info(msg)
+        run_state["log"].append({"ts": datetime.now(timezone.utc).isoformat(), "msg": msg})
+
+    async with _run_lock:
+        run_state.update({
+            "status": "running",
+            "run_id": run_id,
+            "started": datetime.now(timezone.utc).isoformat(),
+            "finished": None,
+            "picks_added": 0,
+            "candidates_found": 0,
+            "error": None,
+            "log": [],
+        })
+
+    try:
+        async with httpx.AsyncClient() as client:
+            log_step("Fetching Last.fm data")
+            lastfm = await fetch_lastfm(client)
+            taste = build_taste_profile(lastfm)
+            log_step(f"Taste profile: {len(taste['topArtists'])} top artists, {len(taste['recentArtists'])} recent")
+
+            log_step("Searching candidates")
+            candidates = await search_candidates(client)
+            run_state["candidates_found"] = len(candidates)
+
+            if not candidates:
+                log_step("No candidates found — ending run")
+                run_state["status"] = "done"
+                run_state["finished"] = datetime.now(timezone.utc).isoformat()
+                return
+
+            log_step("Filtering seen albums")
+            fresh = await filter_unseen(candidates)
+            log_step(f"{len(fresh)} fresh candidates (of {len(candidates)} total)")
+
+            if not fresh:
+                log_step("All candidates already seen — ending run")
+                run_state["status"] = "done"
+                run_state["finished"] = datetime.now(timezone.utc).isoformat()
+                return
+
+            log_step("Scoring by embedding similarity")
+            scored = await score_candidates(client, fresh, taste)
+
+            log_step("Prefiltering with Mistral")
+            prefiltered = await prefilter_candidates(client, scored)
+            log_step(f"{len(prefiltered)} candidates passed prefilter")
+
+            log_step("Curating picks with Mistral")
+            picks = await curate_picks(client, prefiltered, taste)
+            log_step(f"{len(picks)} picks scored ≥60")
+
+            if picks:
+                # Mark all prefiltered candidates as seen (not just picks) to avoid re-processing
+                await mark_seen(prefiltered)
+                await write_picks(picks, {
+                    "candidates": len(candidates),
+                    "filtered": len(prefiltered),
+                })
+
+            run_state.update({
+                "status": "done",
+                "finished": datetime.now(timezone.utc).isoformat(),
+                "picks_added": len(picks),
+            })
+            log_step(f"Run complete — {len(picks)} new picks added")
+
+    except Exception as e:
+        log.exception(f"Pipeline run failed: {e}")
+        run_state.update({
+            "status": "error",
+            "finished": datetime.now(timezone.utc).isoformat(),
+            "error": str(e),
+        })
+
+
+# ── API ───────────────────────────────────────────────────────────────────────
+
+app = FastAPI(title="fgs-agent", description="First Gain Stage metal discovery")
+
+
+@app.on_event("startup")
+async def startup():
+    await init_dedup_db()
+    log.info(f"fgs-agent started on port {settings.port}")
+
+
+@app.post("/api/fgs/run")
+async def trigger_run(background_tasks: BackgroundTasks):
+    if run_state["status"] == "running":
+        return JSONResponse({"status": "already_running", "run_id": run_state["run_id"]}, status_code=409)
+    run_id = str(uuid.uuid4())[:8]
+    background_tasks.add_task(pipeline_run, run_id)
+    return {"status": "started", "run_id": run_id}
+
+
+@app.get("/api/fgs/status")
+async def get_status():
+    return run_state
+
+
+@app.get("/api/fgs/picks")
+async def get_picks():
+    try:
+        with open(settings.picks_path, "r") as f:
+            return json.load(f)
+    except Exception as e:
+        return JSONResponse({"error": str(e)}, status_code=500)
+
+
+@app.get("/api/fgs/health")
+async def health():
+    return {"status": "ok", "agent": "fgs-agent"}
+
+
+# ── Entry point ───────────────────────────────────────────────────────────────
+
+if __name__ == "__main__":
+    uvicorn.run("agent:app", host="0.0.0.0", port=settings.port, reload=False)
diff --git a/fgs-agent.service b/fgs-agent.service
new file mode 100644
index 0000000..bcb950f
--- /dev/null
+++ b/fgs-agent.service
@@ -0,0 +1,15 @@
+[Unit]
+Description=FGS Metal Discovery Agent
+After=network.target
+
+[Service]
+Type=simple
+User=nick2day
+WorkingDirectory=/home/nick2day/fgs-agent
+EnvironmentFile=/home/nick2day/fgs-agent/.env
+ExecStart=/home/nick2day/fgs-agent/venv/bin/python agent.py
+Restart=on-failure
+RestartSec=10
+
+[Install]
+WantedBy=multi-user.target
diff --git a/requirements.txt b/requirements.txt
new file mode 100644
index 0000000..176d4f9
--- /dev/null
+++ b/requirements.txt
@@ -0,0 +1,7 @@
+fastapi==0.115.6
+uvicorn[standard]==0.32.1
+httpx==0.28.1
+pydantic-settings==2.7.0
+python-dotenv==1.0.1
+aiosqlite==0.20.0
+numpy==1.26.4