notebook

Local replacement for NotebookLM. Given a folder of documents and an objective, agy reads every document and produces one objective-driven summary per document, plus a relevance INDEX.md and a cited RESUMEN_MAESTRO.md. The point is to keep Claude's context cheap: agy does all the document reading; you only read the two small final files.

Raw user request: $ARGUMENTS

Phase 0 — Parse + list + classify + cache (ONE Bash call)

Parse $ARGUMENTS: split on the first |. Left side = folder, right side = objective. If there is no |, the longest leading token that resolves to an existing directory is the folder and the rest is the objective. If the folder is missing, ask once: "¿Qué carpeta querés analizar?" and stop.

Run ONE Bash call (a Python helper). It lists supported files, classifies each as text (PDF with a real text layer → pre-extract) or vision (scanned/image → agy OCR), and applies an incremental cache: a document is marked cached (skipped) when its summary already exists AND its size+mtime AND the objective are unchanged since the last run. The cache key includes a hash of the objective, so changing the objective re-summarizes everything.

python - "$FOLDER_ABS" "$OUTDIR" "$OBJETIVO" <<'PYEOF'
import sys, os, re, glob, hashlib
import fitz  # PyMuPDF
folder, outdir, objetivo = sys.argv[1], sys.argv[2], (sys.argv[3] if len(sys.argv) > 3 else "")
os.makedirs(os.path.join(outdir, "_text"), exist_ok=True)
objhash = hashlib.sha1(objetivo.strip().encode("utf-8")).hexdigest()[:8]
MAXV, CHUNK, GROUP_MAX, CHAR_BUDGET = 20, 15, 4, 24000   # scans >MAXV pages -> CHUNK-page subs; uncached text docs -> groups of <=GROUP_MAX docs and <=CHAR_BUDGET chars (1 agy call -> many summaries)
cache_path = os.path.join(outdir, "_cache.tsv"); prev = {}
if os.path.exists(cache_path):
    for ln in open(cache_path, encoding="utf-8"):
        pp = ln.rstrip("\n").split("\t")
        if len(pp) == 2: prev[pp[0]] = pp[1]
exts = (".pdf",".docx",".doc",".png",".jpg",".jpeg",".webp",".gif")
files = sorted(f for f in glob.glob(os.path.join(folder,"*")) if f.lower().endswith(exts))
def slug(s):
    s = re.sub(r"[^a-z0-9]+","-", os.path.splitext(os.path.basename(s))[0].lower()).strip("-")
    return s[:60] or "doc"
def mkrow(nn, mode, src, tpath, summ, key):     # incremental cache per output file
    if os.path.exists(os.path.join(outdir, summ)) and prev.get(summ) == key:
        return (nn, "cached", src, "-", summ, key)
    return (nn, mode, src, tpath, summ, key)
rows=[]; small=[]   # small = UNCACHED text docs to pack into groups: (nn, sl, tpath, srcabs, key, nchars)
for i,f in enumerate(files,1):
    nn=f"{i:03d}"; sl=slug(f); st=os.stat(f); key=f"{st.st_size}:{int(st.st_mtime)}:{objhash}"
    is_pdf=f.lower().endswith(".pdf"); mode="vision"; tpath="-"; pages=0; d=None; nchars=0
    if is_pdf:
        try:
            d=fitz.open(f); pages=d.page_count; txt="\n".join(p.get_text() for p in d)
            if pages and len(txt.strip())/pages >= 200:
                mode="text"; nchars=len(txt.strip()); tpath=os.path.join(outdir,"_text",f"{nn}-{sl}.txt")
                open(tpath,"w",encoding="utf-8").write(txt)
        except Exception:
            mode,pages,d="vision",0,None
    if mode=="vision" and is_pdf and pages>MAXV and d is not None:
        os.makedirs(os.path.join(outdir,"_chunks"),exist_ok=True)   # oversized scan -> page-range chunks
        for ci,startp in enumerate(range(0,pages,CHUNK),1):
            endp=min(startp+CHUNK,pages)
            cpath=os.path.join(outdir,"_chunks",f"{nn}-{sl}-p{startp+1:03d}-{endp:03d}.pdf")
            sub=fitz.open(); sub.insert_pdf(d,from_page=startp,to_page=endp-1); sub.save(cpath); sub.close()
            summ=f"{nn}-{sl}-p{startp+1:03d}-{endp:03d}.resumen.md"
            rows.append(mkrow(nn,"vision",cpath,"-",summ,f"{key}:c{ci}"))
    elif mode=="text":
        summ=f"{nn}-{sl}.resumen.md"
        if os.path.exists(os.path.join(outdir,summ)) and prev.get(summ)==key:
            rows.append((nn,"cached",os.path.abspath(f),tpath,summ,key))     # already summarized -> skip
        else:
            small.append((nn, sl, tpath, os.path.abspath(f), key, nchars))   # pack into a group below
    else:
        rows.append(mkrow(nn,mode,os.path.abspath(f),tpath,f"{nn}-{sl}.resumen.md",key))
    if d is not None: d.close()
# greedy-pack UNCACHED text docs into groups (<=GROUP_MAX docs, <=CHAR_BUDGET chars) to cut agy calls.
# Each group = ONE agy call that writes one summary file PER member (see Mode: notebook-group).
batches=[]; cur=[]; cc=0
for it in small:                                    # it = (nn, sl, tpath, srcabs, key, nchars)
    if cur and (len(cur)>=GROUP_MAX or cc+it[5]>CHAR_BUDGET):
        batches.append(cur); cur=[]; cc=0
    cur.append(it); cc+=it[5]
if cur: batches.append(cur)
for gi,b in enumerate(batches,1):
    if len(b)==1:                                   # lone text doc -> 1-per-call (no group overhead)
        nn,sl,tpath,srcabs,key,_=b[0]
        rows.append((nn,"text",srcabs,tpath,f"{nn}-{sl}.resumen.md",key)); continue
    g=f"G{gi:02d}"
    texts="|".join(x[2] for x in b)                 # member text paths
    names="|".join(f"{x[0]}-{x[1]}" for x in b)     # member display names
    summs="|".join(f"{x[0]}-{x[1]}.resumen.md" for x in b)  # one output file PER member
    gkey="|".join(x[4] for x in b)                  # per-member cache keys (pipe-joined)
    rows.append((g,"group",texts,names,summs,gkey))
rows.sort(key=lambda r: r[0])
with open(os.path.join(outdir,"_manifest.tsv"),"w",encoding="utf-8") as m:
    for r in rows: m.write("\t".join(r)+"\n")
nc=sum(1 for r in rows if r[1]=="cached"); nt=sum(1 for r in rows if r[1]=="text")
nv=sum(1 for r in rows if r[1]=="vision"); ng=sum(1 for r in rows if r[1]=="group")
print(f"DOCS={len(files)} ROWS={len(rows)} CACHED={nc} TEXT={nt} VISION={nv} GROUPS={ng} OUTDIR={outdir}")
for r in rows: print("\t".join(r))
PYEOF