718 lines
34 KiB
Python
718 lines
34 KiB
Python
"""Guide-Generierung als Konsens-Pipeline.
|
|
|
|
Gliederung: Auswahl der Bausteine (deterministisch je Format) → 3 Vorschläge
|
|
(Grace), die Bausteine NUMMERN-basiert in Kapitel ordnen → ein Judge merged die
|
|
Vorschläge zu einer kohärenten Reihenfolge.
|
|
Schreiben: Writer je Baustein. Lese-Prüfung: Check→Fix (eine Runde),
|
|
Folgerunden prüfen nur ersetzte Sections; danach bleiben Beanstandungen stehen.
|
|
Schritt-Dateien bleiben liegen → Abbruch erhält Fortschritt, ▶ setzt am offenen Schritt fort.
|
|
"""
|
|
|
|
import asyncio
|
|
import json
|
|
import logging
|
|
import math
|
|
from datetime import datetime, timezone
|
|
from pathlib import Path
|
|
|
|
from agents import run_agent
|
|
from bausteine import _pdfs_konvertieren, quelle_ordner
|
|
from config import (
|
|
DEFAULT_PROVIDER, FORMAT_ZWECK, KONSENS_GRACE,
|
|
LESBARKEIT_AKTIV, TEMPLATES_DIR,
|
|
)
|
|
import lesbarkeit
|
|
from database import list_guides, update_guide, list_bausteine, list_subbausteine, set_guide_content
|
|
from fsutil import atomic_write_json, atomic_write_text
|
|
from jsonio import read_json_file as _json_datei
|
|
from paths import bausteine_path, guide_content_path, project_dir, subbausteine_path
|
|
from pipeline import (
|
|
CANCELLED, FAILED, GenContext, _claude_error, _extra,
|
|
_fail, _gather_error, _gather_fortschritt, _log, _prompt, _race,
|
|
_semaphore, _set_progress, _set_step, _timeout, clear_guide_cancelled,
|
|
is_guide_cancelled, run_single_slot,
|
|
)
|
|
from textkit import (
|
|
_eindeutige_titel, _lade_bausteine, _parse_fragment, _split_chunks,
|
|
_titel, _titel_aufloesen, _titel_index,
|
|
)
|
|
|
|
log = logging.getLogger("creator.guide")
|
|
|
|
GUIDE_STEPS = ("Gliederung", "Inhalte", "Inhalts-Check", "Schreiben", "Lese-Prüfung")
|
|
|
|
# Writer skalieren mit der Section-Zahl: 1 Writer je ~30 Sections (gedeckelt).
|
|
# Kleine Pakete vermeiden Lazy-Output bei langen Listen und begrenzen den Schaden
|
|
# eines fehlgeschlagenen Writers.
|
|
# Inhalte/Inhalts-Check/Lese-Prüfung laufen in Paketen von ~GUIDE_CHUNK Bausteinen je Agent.
|
|
# Nur der Writer (Schreiben) bleibt 1 Agent je Baustein (variable Längen, kein Kürzen).
|
|
GUIDE_CHUNK = 10
|
|
|
|
# Lese-Prüfung: nur EINE Runde (Check + Fix). Folgerunden brachten kaum Mehrwert
|
|
# (1 Agent je Baustein prüft ohnehin fein), kosten aber extra Agenten.
|
|
LESE_RUNDEN = 1
|
|
|
|
|
|
async def _load_subbausteine(topic: str) -> dict[str, list[dict]]:
|
|
"""Subbausteine je Baustein — DB-first ({titel, stufe, relevanz}), Fallback Sidecar-Datei.
|
|
Fehlt beides → {} (Guide nimmt alles)."""
|
|
out: dict[str, list[dict]] = {}
|
|
for r in await list_subbausteine(topic):
|
|
if r["status"] == "konsens" and r["sub_titel"] and r["stufe"] in ("einfach", "mittel", "schwer"):
|
|
out.setdefault(r["baustein"], []).append(
|
|
{"titel": r["sub_titel"], "stufe": r["stufe"], "relevanz": r["relevanz"]})
|
|
if out:
|
|
return out
|
|
data = _json_datei(subbausteine_path(topic))
|
|
if not isinstance(data, dict):
|
|
return {}
|
|
for titel, subs in data.items():
|
|
if not isinstance(subs, list):
|
|
continue
|
|
gut = [s for s in subs if isinstance(s, dict) and str(s.get("titel", "")).strip()
|
|
and s.get("stufe") in ("einfach", "mittel", "schwer")]
|
|
if gut:
|
|
out[titel] = gut
|
|
return out
|
|
|
|
|
|
def _zuteilung_subs(chunk: list[dict], entries: dict[int, str], subs_by_titel: dict[str, list[dict]]) -> str:
|
|
"""Wie _zuteilung_text, aber listet pro Baustein seine Subbausteine mit Stufe."""
|
|
lines: list[str] = []
|
|
for ch in chunk:
|
|
lines.append(f"KAPITEL: {ch['title']}")
|
|
for num in ch["nums"]:
|
|
lines.append(f"- {entries[num]}")
|
|
for s in subs_by_titel.get(_titel(entries[num]), []):
|
|
lines.append(f" [{s['stufe']}] {s['titel']}")
|
|
return "\n".join(lines)
|
|
|
|
|
|
def _guide_files(content_path: Path) -> dict:
|
|
d, stem = content_path.parent, content_path.stem
|
|
return {
|
|
"gliederung_slots": [d / f"{stem}.gliederung-{i}.json" for i in (1, 2, 3)],
|
|
"gliederung": d / f"{stem}.gliederung.json", # Judge-Ausgabe
|
|
# chunk-/lese-check-/fix-Dateien sind dynamisch:
|
|
# {stem}.chunk-i.md, {stem}.lese-check-r{n}-{i}.json, {stem}.fix-r{n}-{i}.md
|
|
}
|
|
|
|
|
|
def guide_slot_dateien(content_path: Path) -> list[Path]:
|
|
"""Alle Schritt-Dateien eines Guides (für den Frischstart)."""
|
|
return [p for p in content_path.parent.glob(f"{content_path.stem}.*") if p != content_path]
|
|
|
|
|
|
def _fertig_path(content_path: Path) -> Path:
|
|
return content_path.parent / f"{content_path.stem}.fertig"
|
|
|
|
|
|
def guide_fertig_step(content_path: Path) -> int:
|
|
"""Höchster VOLL abgeschlossener Schritt-Index (Marker je Thema+Format). -1 = keiner.
|
|
Existiert die Content-Datei, sind alle Schritte fertig."""
|
|
if content_path.exists():
|
|
return len(GUIDE_STEPS) - 1
|
|
try:
|
|
return int(_fertig_path(content_path).read_text(encoding="utf-8").strip())
|
|
except (OSError, ValueError):
|
|
return -1
|
|
|
|
|
|
def _set_fertig(content_path: Path, step: int) -> None:
|
|
"""Marker auf `step` setzen — monoton (nur erhöhen), außer beim Re-Run-Reset (force)."""
|
|
if step > guide_fertig_step(content_path):
|
|
atomic_write_text(_fertig_path(content_path), str(step))
|
|
|
|
|
|
def _reset_fertig(content_path: Path, step: int) -> None:
|
|
"""Marker hart auf `step` setzen (für Re-Run ab Schritt; step kann sinken)."""
|
|
if step < 0:
|
|
_fertig_path(content_path).unlink(missing_ok=True)
|
|
else:
|
|
atomic_write_text(_fertig_path(content_path), str(step))
|
|
|
|
|
|
# Slot-Datei-Globs je Schritt (Index = GUIDE_STEPS). Stem-verankert, kollisionsfrei.
|
|
_STEP_GLOBS = (
|
|
("gliederung*",), # 0 Gliederung (inkl. Auswahl-Filter)
|
|
("inhalt-chunk-*",), # 1 Inhalte
|
|
("inhalt-check-*", "inhalt-fix-*"), # 2 Inhalts-Check
|
|
("chunk-*",), # 3 Schreiben
|
|
("lese-check-*", "fix-r*"), # 4 Lese-Prüfung
|
|
)
|
|
|
|
|
|
def _reset_guide_ab_step(content_path: Path, step: int) -> None:
|
|
"""Re-Run ab Schritt: Content + alle Slot-Dateien der Schritte ≥ step löschen.
|
|
Frühere Schritte bleiben → der Resume baut ab `step` neu (alles darunter wiederverwendet)."""
|
|
content_path.unlink(missing_ok=True) # nicht mehr „done" → kein Frischstart-Wipe
|
|
d, stem = content_path.parent, content_path.stem
|
|
for globs in _STEP_GLOBS[step:]:
|
|
for pat in globs:
|
|
for p in d.glob(f"{stem}.{pat}"):
|
|
p.unlink(missing_ok=True)
|
|
_reset_fertig(content_path, step - 1) # Schritte < step gelten als fertig
|
|
|
|
|
|
def _lese_probleme_schema(data):
|
|
"""{"ok": true} → [] · {"probleme": [{"section", "problem"}]} → Liste · sonst None."""
|
|
if not isinstance(data, dict):
|
|
return None
|
|
if data.get("ok") is True:
|
|
return []
|
|
p = data.get("probleme")
|
|
if not isinstance(p, list) or not p:
|
|
return None
|
|
out = []
|
|
for x in p:
|
|
if not isinstance(x, dict) or not isinstance(x.get("section"), str) or not isinstance(x.get("problem"), str):
|
|
return None
|
|
out.append({"section": x["section"].strip(), "problem": x["problem"].strip()})
|
|
return out or None
|
|
|
|
|
|
def _resolve_gliederung(data, entries: dict[int, str], soll_min: int, soll_max: int) -> list[dict] | None:
|
|
"""{"kapitel": [{"titel", "nummern": [1, 3, 7]}]} → [{"title", "nums"}].
|
|
|
|
Nummern sind die IDs aus `entries` (1-basiert, wie dem Agenten präsentiert).
|
|
`soll_min`/`soll_max` = erlaubte Spanne gewählter Bausteine (mit kleiner Toleranz).
|
|
"""
|
|
if not isinstance(data, dict) or not isinstance(data.get("kapitel"), list):
|
|
return None
|
|
gueltig = set(entries)
|
|
chapters: list[dict] = []
|
|
seen: set[int] = set()
|
|
total = unknown = 0
|
|
for ch in data["kapitel"]:
|
|
if not isinstance(ch, dict) or not isinstance(ch.get("nummern"), list):
|
|
return None
|
|
nums = []
|
|
for t in ch["nummern"]:
|
|
total += 1
|
|
num = t if isinstance(t, int) and not isinstance(t, bool) else None
|
|
if num is None or num not in gueltig:
|
|
unknown += 1
|
|
elif num not in seen:
|
|
nums.append(num)
|
|
seen.add(num)
|
|
if nums:
|
|
chapters.append({"title": str(ch.get("titel", "")).strip() or "Kapitel", "nums": nums})
|
|
if not chapters or total == 0:
|
|
return None
|
|
if (total - unknown) / total < 0.85:
|
|
return None
|
|
if len(seen) < 0.9 * soll_min or len(seen) > 1.1 * soll_max:
|
|
return None
|
|
return chapters
|
|
|
|
|
|
def _fallback_gliederung(entries: dict[int, str]) -> list[dict]:
|
|
"""Deterministische Gliederung, wenn die Agenten keine liefern: ein Kapitel mit
|
|
allen gewählten Bausteinen in Reihenfolge. Garantiert vollständige Abdeckung."""
|
|
return [{"title": "Inhalte", "nums": list(entries)}]
|
|
|
|
|
|
def _mit_resten(plan: list[dict], entries: dict[int, str]) -> list[dict]:
|
|
"""Stellt sicher, dass JEDER gewählte Baustein im Plan steht — fehlende landen in
|
|
einem Kapitel „Weitere" (gegen Agenten/Judge, die Bausteine weglassen)."""
|
|
drin = {num for ch in plan for num in ch.get("nums", [])}
|
|
fehlen = [num for num in entries if num not in drin]
|
|
return [*plan, {"title": "Weitere", "nums": fehlen}] if fehlen else plan
|
|
|
|
|
|
async def _generate_sections(
|
|
guide_id: str, topic: str, format_name: str, entries: dict[int, str],
|
|
facts: str, instructions: str, provider: str,
|
|
content_path: Path,
|
|
) -> list[dict] | None:
|
|
def is_cancelled() -> bool:
|
|
return is_guide_cancelled(guide_id)
|
|
|
|
ctx = GenContext(topic=topic, provider=provider, is_cancelled=is_cancelled, guide_id=guide_id)
|
|
spec = (TEMPLATES_DIR / "Format" / "Section.md").read_text(encoding="utf-8")
|
|
files = _guide_files(content_path)
|
|
zweck = FORMAT_ZWECK[format_name]
|
|
|
|
# Subbausteine je Baustein (DB-first) — früh geladen: steuert Auswahl + Sub-Filter je Format.
|
|
# Fehlt sie → {} (Fallback: Guide nimmt alles).
|
|
subs_raw = await _load_subbausteine(topic)
|
|
|
|
def _hat_relevanz(num, art):
|
|
return any(isinstance(s, dict) and s.get("relevanz") == art for s in subs_raw.get(_titel(entries[num]), []))
|
|
|
|
# Auswahl (Teil von Gliederung). Deterministisch je Format:
|
|
# Guide=relevante Bausteine · FullGuide=alle · Rest=reine Rand-Bausteine (keine im Guide).
|
|
if format_name == "FullGuide":
|
|
auswahl = list(entries)
|
|
elif format_name == "Rest":
|
|
auswahl = [num for num in entries if not _hat_relevanz(num, "relevant")]
|
|
else: # Guide
|
|
auswahl = [num for num in entries if _hat_relevanz(num, "relevant")]
|
|
if not auswahl:
|
|
_log(topic, "Guide: keine Relevanz-Daten — alle Bausteine genommen")
|
|
auswahl = list(entries)
|
|
if not auswahl:
|
|
await _fail(guide_id, "Keine passenden Bausteine für dieses Format")
|
|
return None
|
|
|
|
sel_entries = {num: entries[num] for num in auswahl}
|
|
soll = len(sel_entries)
|
|
# Nummerierte Liste (ID = Baustein-Nummer aus entries) — Agenten/Judge ordnen per Nummer.
|
|
sel_liste = "\n".join(f"{num}. {t}" for num, t in sel_entries.items())
|
|
|
|
# Schritt 0: Gliederung — bis zu 3 Vorschläge (Grace), ein Judge merged sie. Bricht NIE ab:
|
|
# 0 gültige → Code-Fallback, 1 → direkt, ≥2 → Judge (mit Vorschlag als Rückfall).
|
|
# Gültiges gliederung.json überspringt den Schritt.
|
|
plan = _resolve_gliederung(_json_datei(files["gliederung"]), sel_entries, soll, soll)
|
|
if plan is None:
|
|
await _set_step(guide_id, 0, "Gliederungs-Vorschläge (3 Agenten)…")
|
|
files["gliederung"].unlink(missing_ok=True)
|
|
vorschlaege: list[list[dict]] = []
|
|
offen = []
|
|
for i, path in enumerate(files["gliederung_slots"], 1):
|
|
res = _resolve_gliederung(_json_datei(path), sel_entries, soll, soll)
|
|
if res is not None:
|
|
vorschlaege.append(res)
|
|
else:
|
|
offen.append((i, path))
|
|
if len(vorschlaege) < 3 and offen:
|
|
slots = [
|
|
{
|
|
"key": f"{guide_id}-gliederung-{i}",
|
|
"prompt": _prompt(
|
|
"Guide-Gliederung",
|
|
topic=topic, format_name=format_name, bausteine=sel_liste,
|
|
out_path=path, extra=_extra(instructions),
|
|
),
|
|
"role": "guide", "capabilities": "files",
|
|
"payload": (lambda result, p=path: _resolve_gliederung(_json_datei(p), sel_entries, soll, soll)),
|
|
}
|
|
for i, path in offen
|
|
]
|
|
# Quorum 1: nimm, was kommt — kein Mindestzwang, kein Abbruch.
|
|
neue = await _race(
|
|
topic, "Gliederung", slots, 1, _timeout("plan", soll),
|
|
provider, cancelled=is_cancelled, grace=KONSENS_GRACE,
|
|
)
|
|
if is_cancelled():
|
|
return None
|
|
vorschlaege += neue or []
|
|
|
|
if not vorschlaege:
|
|
_log(topic, "Gliederung: kein gültiger Vorschlag — deterministischer Fallback")
|
|
plan = _fallback_gliederung(sel_entries)
|
|
elif len(vorschlaege) == 1:
|
|
plan = vorschlaege[0] # ein Vorschlag → kein Judge nötig
|
|
else:
|
|
await _set_step(guide_id, 0, "Gliederungen zusammenführen…")
|
|
bloecke = "\n\n".join(
|
|
f"### Vorschlag {i}\n"
|
|
+ "\n".join(f"KAPITEL: {ch['title']}\n Nummern: {', '.join(str(num) for num in ch['nums'])}" for ch in v)
|
|
for i, v in enumerate(vorschlaege, 1)
|
|
)
|
|
status, plan = await run_single_slot(
|
|
ctx, "Gliederungs-Judge",
|
|
key=f"{guide_id}-gliederung-judge",
|
|
prompt=_prompt(
|
|
"Guide-Gliederung-Judge",
|
|
topic=topic, format_name=format_name, zweck=zweck, n=len(vorschlaege),
|
|
bausteine=sel_liste, gliederungen=bloecke,
|
|
out_path=files["gliederung"], extra=_extra(instructions),
|
|
),
|
|
role="judge", capabilities="files",
|
|
payload=lambda result: _resolve_gliederung(_json_datei(files["gliederung"]), sel_entries, soll, soll),
|
|
timeout=_timeout("plan_judge", soll),
|
|
)
|
|
if status == CANCELLED:
|
|
return None
|
|
if status == FAILED or plan is None:
|
|
_log(topic, "Gliederung-Judge ohne Ergebnis — bester Vorschlag bleibt")
|
|
plan = vorschlaege[0]
|
|
|
|
# Garantie: jeder gewählte Baustein steht im Plan (gegen weglassende Agenten/Judges).
|
|
plan = _mit_resten(plan, sel_entries)
|
|
_set_fertig(content_path, 0) # Gliederung steht
|
|
|
|
# Grobe Chunks (~GUIDE_CHUNK Bausteine je Agent) für Inhalte, Inhalts-Check und Lese-Prüfung.
|
|
# Der Writer baut darunter pro Baustein (eigene feine Chunks, s.u.) → variable Längen.
|
|
total_sections = sum(len(c["nums"]) for c in plan)
|
|
chunks = _split_chunks(plan, max(1, math.ceil(total_sections / GUIDE_CHUNK)))
|
|
# Subbausteine je Baustein, je Format gefiltert: Guide=relevant · FullGuide=alle · Rest=Rand.
|
|
# Subs ohne Relevanz-Feld (Alt-Themen) zählen als „nicht Rand" (→ Guide/FullGuide).
|
|
if format_name == "FullGuide":
|
|
subs_by_titel = {t: list(subs) for t, subs in subs_raw.items()}
|
|
elif format_name == "Rest":
|
|
subs_by_titel = {t: [s for s in subs if s.get("relevanz") == "rand"] for t, subs in subs_raw.items()}
|
|
else: # Guide
|
|
subs_by_titel = {t: [s for s in subs if s.get("relevanz") != "rand"] for t, subs in subs_raw.items()}
|
|
subs_by_titel = {t: subs for t, subs in subs_by_titel.items() if subs}
|
|
zuteilungen = [_zuteilung_subs(chunk, entries, subs_by_titel) for chunk in chunks]
|
|
chunk_sizes = [sum(len(c["nums"]) for c in chunk) for chunk in chunks]
|
|
writer_count = len(chunks)
|
|
idx = _titel_index(entries)
|
|
|
|
# Schritt 2: Inhalte je Baustein identifizieren — pro Chunk ein Agent (Marker-Output, Resume).
|
|
inhalt_paths = [content_path.parent / f"{content_path.stem}.inhalt-chunk-{i}.md" for i in range(1, writer_count + 1)]
|
|
offen = [i for i, p in enumerate(inhalt_paths) if not p.exists()]
|
|
if offen:
|
|
async def melde(d, t): await _set_step(guide_id, 1, f"Sammle Inhalte {d}/{t}…")
|
|
results = await _gather_fortschritt([
|
|
run_agent(
|
|
f"{guide_id}-inhalt-{i + 1}",
|
|
_prompt(
|
|
"Guide-Inhalt",
|
|
topic=topic, zuteilung=zuteilungen[i], facts=facts,
|
|
out_path=inhalt_paths[i], extra=_extra(instructions),
|
|
),
|
|
_timeout("inhalt", chunk_sizes[i]), provider=provider, role="guide", capabilities="full",
|
|
)
|
|
for i in offen
|
|
], writer_count, melde, start=writer_count - len(offen))
|
|
if is_cancelled():
|
|
return None
|
|
if not any(p.exists() for p in inhalt_paths):
|
|
await _fail(guide_id, _gather_error("Inhalts-Fehler", list(results)))
|
|
return None
|
|
|
|
inhalt_by_num: dict[int, str] = {}
|
|
for p in inhalt_paths:
|
|
if not p.exists():
|
|
continue
|
|
for sec in _parse_fragment(p.read_text(encoding="utf-8")):
|
|
num = _titel_aufloesen(idx, sec["titel"])
|
|
if num is not None and num not in inhalt_by_num and sec["md"].strip():
|
|
inhalt_by_num[num] = sec["md"]
|
|
if not inhalt_by_num:
|
|
await _fail(guide_id, "Keine Inhalte identifiziert")
|
|
return None
|
|
if all(p.exists() for p in inhalt_paths):
|
|
_set_fertig(content_path, 1) # Inhalte vollständig
|
|
|
|
inhalt_chunk_nums = [[num for ch in chunk for num in ch["nums"] if num in inhalt_by_num] for chunk in chunks]
|
|
|
|
# Schritt 3: Inhalte prüfen (pro Chunk ein Check) + beanstandete einmal überarbeiten.
|
|
check_paths = [content_path.parent / f"{content_path.stem}.inhalt-check-{i}.json" for i in range(1, writer_count + 1)]
|
|
offen_checks = [i for i, p in enumerate(check_paths) if inhalt_chunk_nums[i] and _lese_probleme_schema(_json_datei(p)) is None]
|
|
if offen_checks:
|
|
await _set_step(guide_id, 2, "Prüfe Inhalte…")
|
|
slots = [{
|
|
"key": f"{guide_id}-inhalt-check-{i + 1}",
|
|
"prompt": _prompt(
|
|
"Guide-Inhalt-Check",
|
|
topic=topic, format_name=format_name,
|
|
sections="\n\n".join(f"SECTION: {_titel(entries[num])}\n{inhalt_by_num[num]}" for num in inhalt_chunk_nums[i]),
|
|
out_path=check_paths[i], extra=_extra(instructions),
|
|
),
|
|
"role": "judge", "capabilities": "files",
|
|
"payload": (lambda result, p=check_paths[i]: _lese_probleme_schema(_json_datei(p))),
|
|
} for i in offen_checks]
|
|
n_checks = len(slots)
|
|
upd = lambda n: asyncio.create_task(_set_step(guide_id, 2, f"Prüfe Inhalte {n}/{n_checks}…"))
|
|
await _race(topic, "Inhalts-Prüfung", slots, len(slots), _timeout("inhalt_check", max(chunk_sizes)), provider, on_update=upd, cancelled=is_cancelled)
|
|
if is_cancelled():
|
|
return None
|
|
|
|
probleme_by_num: dict[int, str] = {}
|
|
for i, p in enumerate(check_paths):
|
|
geltung = set(inhalt_chunk_nums[i])
|
|
for item in (_lese_probleme_schema(_json_datei(p)) or []):
|
|
num = _titel_aufloesen(idx, item["section"])
|
|
if num in geltung and num in inhalt_by_num and num not in probleme_by_num:
|
|
probleme_by_num[num] = item["problem"]
|
|
|
|
if probleme_by_num:
|
|
_log(topic, f"Inhalts-Prüfung: {len(probleme_by_num)} Baustein(e) beanstandet")
|
|
await _set_step(guide_id, 2, f"Überarbeite {len(probleme_by_num)} Inhalt(e)…")
|
|
fix_chunks = [[num for num in nums if num in probleme_by_num] for nums in inhalt_chunk_nums]
|
|
fix_paths = [content_path.parent / f"{content_path.stem}.inhalt-fix-{i + 1}.md" for i in range(writer_count)]
|
|
fix_offen = [i for i, nums in enumerate(fix_chunks) if nums and not fix_paths[i].exists()]
|
|
results = await asyncio.gather(*[
|
|
run_agent(
|
|
f"{guide_id}-inhalt-fix-{i + 1}",
|
|
_prompt(
|
|
"Guide-Inhalt-Fix",
|
|
topic=topic, facts=facts,
|
|
auftraege="\n\n".join(
|
|
f"SECTION: {_titel(entries[num])}\nPROBLEM: {probleme_by_num[num]}\nAKTUELL:\n{inhalt_by_num[num]}"
|
|
for num in fix_chunks[i]
|
|
),
|
|
out_path=fix_paths[i], extra=_extra(instructions),
|
|
),
|
|
_timeout("inhalt", len(fix_chunks[i])), provider=provider, role="guide", capabilities="full",
|
|
)
|
|
for i in fix_offen
|
|
], return_exceptions=True)
|
|
if is_cancelled():
|
|
return None
|
|
for p in fix_paths:
|
|
if not p.exists():
|
|
continue
|
|
for sec in _parse_fragment(p.read_text(encoding="utf-8")):
|
|
num = _titel_aufloesen(idx, sec["titel"])
|
|
if num in probleme_by_num and sec["md"].strip():
|
|
inhalt_by_num[num] = sec["md"]
|
|
|
|
_set_fertig(content_path, 2) # Inhalts-Check durch
|
|
|
|
# Schritt 4: Schreiben — Writer formuliert die geprüften Inhalte aus (Resume).
|
|
# FEINE Chunks: genau 1 Baustein je Writer → variable Längen, kein Budget-Rationieren.
|
|
def inhalte_text(chunk) -> str:
|
|
nums = [num for ch in chunk for num in ch["nums"] if num in inhalt_by_num]
|
|
return "\n\n".join(f"<!-- section: {_titel(entries[num])} -->\n{inhalt_by_num[num]}" for num in nums)
|
|
|
|
w_chunks = [[{"title": ch["title"], "nums": [num]}] for ch in plan for num in ch["nums"]]
|
|
w_zuteil = [_zuteilung_subs(c, entries, subs_by_titel) for c in w_chunks]
|
|
paths = [content_path.parent / f"{content_path.stem}.chunk-{i}.md" for i in range(1, len(w_chunks) + 1)]
|
|
offen = [i for i, p in enumerate(paths) if not p.exists()]
|
|
if offen:
|
|
async def melde(d, t): await _set_step(guide_id, 3, f"Schreibe Sections {d}/{t}…")
|
|
results = await _gather_fortschritt([
|
|
run_agent(
|
|
f"{guide_id}-w{i + 1}",
|
|
_prompt(
|
|
"Guide-Writer",
|
|
topic=topic, format_name=format_name, zuteilung=w_zuteil[i],
|
|
inhalte=inhalte_text(w_chunks[i]),
|
|
facts=facts, spec=spec, out_path=paths[i], extra=_extra(instructions),
|
|
),
|
|
_timeout("writer", 1), provider=provider, role="guide", capabilities="full",
|
|
)
|
|
for i in offen
|
|
], len(w_chunks), melde, start=len(w_chunks) - len(offen))
|
|
if is_cancelled():
|
|
return None
|
|
for i, r in zip(offen, results):
|
|
if isinstance(r, BaseException):
|
|
_log(topic, f"Writer {i + 1}: {type(r).__name__}: {r}")
|
|
elif r[0] != 0:
|
|
_log(topic, f"Writer {i + 1}: {_claude_error('Fehler', *r)}")
|
|
elif not paths[i].exists():
|
|
_log(topic, f"Writer {i + 1}: keine Ausgabedatei erstellt")
|
|
if not any(p.exists() for p in paths):
|
|
await _fail(guide_id, _gather_error("Writer-Fehler", list(results)))
|
|
return None
|
|
|
|
by_num: dict[int, dict] = {}
|
|
for p in paths:
|
|
if not p.exists():
|
|
continue
|
|
for sec in _parse_fragment(p.read_text(encoding="utf-8")):
|
|
num = _titel_aufloesen(idx, sec["titel"])
|
|
if num is None:
|
|
_log(topic, f"Writer lieferte unbekannte Section '{sec['titel'][:40]}' (ignoriert)")
|
|
elif num not in by_num:
|
|
by_num[num] = sec
|
|
if not by_num:
|
|
await _fail(guide_id, "Keine Sections in der Writer-Ausgabe gefunden")
|
|
return None
|
|
if all(p.exists() for p in paths):
|
|
_set_fertig(content_path, 3) # Schreiben vollständig
|
|
|
|
# Schritt 3: Lese-Prüfungs-Loop — Check pro Writer-Paket, Fix nur für
|
|
# beanstandete Sections; Folgerunden prüfen NUR die ersetzten Sections.
|
|
# Nach dem Runden-Cap bleiben offene Beanstandungen stehen.
|
|
chunk_nums = [[num for ch in chunk for num in ch["nums"] if num in by_num] for chunk in chunks]
|
|
|
|
def sections_text(nums: list[int]) -> str:
|
|
return "\n\n".join(f"SECTION: {_titel(entries[num])}\n{by_num[num]['md']}" for num in nums)
|
|
|
|
def auftraege_text(nums: list[int], probleme: dict[int, str]) -> str:
|
|
return "\n\n".join(
|
|
f"SECTION: {_titel(entries[num])}\nPROBLEM: {probleme[num]}\nAKTUELLER INHALT:\n{by_num[num]['md']}"
|
|
for num in nums
|
|
)
|
|
|
|
scope = chunk_nums
|
|
for runde in range(1, LESE_RUNDEN + 1):
|
|
check_paths = [content_path.parent / f"{content_path.stem}.lese-check-r{runde}-{i}.json" for i in range(1, writer_count + 1)]
|
|
offen_checks = [i for i, p in enumerate(check_paths) if scope[i] and _lese_probleme_schema(_json_datei(p)) is None]
|
|
if offen_checks:
|
|
await _set_step(guide_id, 4, "Prüfe Lesbarkeit…")
|
|
slots = [{
|
|
"key": f"{guide_id}-lese-check-r{runde}-{i + 1}",
|
|
"prompt": _prompt(
|
|
"Guide-Lese-Check",
|
|
topic=topic, format_name=format_name, spec=spec,
|
|
sections=sections_text(scope[i]),
|
|
out_path=check_paths[i], extra=_extra(instructions),
|
|
),
|
|
"role": "judge", "capabilities": "files",
|
|
"payload": (lambda result, p=check_paths[i]: _lese_probleme_schema(_json_datei(p))),
|
|
} for i in offen_checks]
|
|
n_checks = len(slots)
|
|
upd = lambda n: asyncio.create_task(_set_step(guide_id, 4, f"Prüfe Lesbarkeit {n}/{n_checks}…"))
|
|
res = await _race(topic, f"Lese-Prüfung r{runde}", slots, len(slots), _timeout("lese_check", max(chunk_sizes)), provider, on_update=upd, cancelled=is_cancelled)
|
|
if is_cancelled():
|
|
return None
|
|
if res is None:
|
|
# Nicht hart failen: bei 1 Agent je Baustein darf ein einzelner Check-Ausfall
|
|
# nicht den ganzen Guide kippen. Sections sind geschrieben; Lese-Prüfung ist Politur.
|
|
_log(topic, f"Lese-Prüfung Runde {runde} ohne vollständiges Ergebnis — Stand bleibt")
|
|
break
|
|
|
|
probleme_by_num: dict[int, str] = {}
|
|
for i, p in enumerate(check_paths):
|
|
geltung = set(scope[i])
|
|
for item in (_lese_probleme_schema(_json_datei(p)) or []):
|
|
num = _titel_aufloesen(idx, item["section"])
|
|
if num in geltung and num in by_num and num not in probleme_by_num:
|
|
probleme_by_num[num] = item["problem"]
|
|
|
|
# Deterministisches Lesbarkeits-Gate: zu schwere Sections in dieselbe
|
|
# Überarbeitung einreihen (LLM-Beanstandung hat Vorrang). Gate aus → no-op.
|
|
if LESBARKEIT_AKTIV:
|
|
md_by_num = {num: by_num[num]["md"] for nums in scope for num in nums if num in by_num}
|
|
hinweise = await asyncio.to_thread(lesbarkeit.bewerte_sections, md_by_num)
|
|
if hinweise:
|
|
_log(topic, f"Lesbarkeit: {len(hinweise)} Section(s) zu schwer")
|
|
for num, hinweis in hinweise.items():
|
|
probleme_by_num.setdefault(num, hinweis)
|
|
|
|
if not probleme_by_num:
|
|
break
|
|
|
|
_log(topic, f"Lese-Prüfung Runde {runde}: {len(probleme_by_num)} Section(s) beanstandet")
|
|
await _set_step(guide_id, 4, f"Überarbeite {len(probleme_by_num)} Section(s) (Runde {runde})…")
|
|
fix_chunks = [[num for num in nums if num in probleme_by_num] for nums in chunk_nums]
|
|
fix_paths = [content_path.parent / f"{content_path.stem}.fix-r{runde}-{i + 1}.md" for i in range(writer_count)]
|
|
fix_offen = [i for i, nums in enumerate(fix_chunks) if nums and not fix_paths[i].exists()]
|
|
results = await asyncio.gather(*[
|
|
run_agent(
|
|
f"{guide_id}-fix-r{runde}-w{i + 1}",
|
|
_prompt(
|
|
"Guide-Sections-Fix",
|
|
topic=topic, format_name=format_name, facts=facts, spec=spec,
|
|
auftraege=auftraege_text(fix_chunks[i], probleme_by_num),
|
|
out_path=fix_paths[i], extra=_extra(instructions),
|
|
),
|
|
_timeout("writer", len(fix_chunks[i])), provider=provider, role="guide", capabilities="full",
|
|
)
|
|
for i in fix_offen
|
|
], return_exceptions=True)
|
|
if is_cancelled():
|
|
return None
|
|
for i, r in zip(fix_offen, results):
|
|
if isinstance(r, BaseException) or (not isinstance(r, BaseException) and r[0] != 0):
|
|
_log(topic, f"Sections-Fix {i + 1} (Runde {runde}) fehlgeschlagen — Original bleibt")
|
|
ersetzt: set[int] = set()
|
|
for p in fix_paths:
|
|
if not p.exists():
|
|
continue
|
|
for sec in _parse_fragment(p.read_text(encoding="utf-8")):
|
|
num = _titel_aufloesen(idx, sec["titel"])
|
|
if num in probleme_by_num and sec["md"].strip():
|
|
by_num[num] = sec
|
|
ersetzt.add(num)
|
|
_log(topic, f"Lese-Prüfung Runde {runde}: {len(ersetzt)} Section(s) überarbeitet")
|
|
if not ersetzt:
|
|
break
|
|
if runde == LESE_RUNDEN:
|
|
_log(topic, f"Lese-Prüfung: 1 Runde — Überarbeitung bleibt ungeprüft")
|
|
break
|
|
scope = [[num for num in nums if num in ersetzt] for nums in chunk_nums]
|
|
_set_fertig(content_path, 4) # Lese-Prüfung durch
|
|
|
|
# Prüfbar = Format hat Prüfung UND Baustein hat ≥1 relevanten Subbaustein.
|
|
# Guide ist immer prüfbar (auch ohne Relevanz-Daten, Fallback = alles).
|
|
def _pruefbar(num):
|
|
if format_name == "Guide":
|
|
return True
|
|
if format_name == "FullGuide":
|
|
return any(isinstance(s, dict) and s.get("relevanz") == "relevant"
|
|
for s in subs_raw.get(_titel(entries[num]), []))
|
|
return False # Rest u.a. → reine Lese-Sections
|
|
|
|
await _set_progress(guide_id, "Setze zusammen…")
|
|
chapters: list[dict] = []
|
|
for ch in plan:
|
|
sections = [
|
|
{"num": num, "title": _titel(entries[num]), "md": by_num[num]["md"],
|
|
"kompakt": by_num[num].get("kompakt", ""), "pruefbar": _pruefbar(num)}
|
|
for num in ch["nums"] if num in by_num
|
|
]
|
|
if sections:
|
|
chapters.append({"title": ch["title"], "sections": sections})
|
|
geplant = {num for ch in plan for num in ch["nums"]}
|
|
missing = sorted(geplant - set(by_num))
|
|
if missing:
|
|
_log(topic, f"Sections fehlen in der Writer-Ausgabe: {[_titel(entries[n]) for n in missing]}")
|
|
if not chapters:
|
|
await _fail(guide_id, "Keine Sections in der Writer-Ausgabe gefunden")
|
|
return None
|
|
return chapters
|
|
|
|
|
|
async def reconcile_guides() -> None:
|
|
"""DB↔Dateisystem abgleichen: status=done ohne Content-Datei → error.
|
|
|
|
Läuft beim Server-Start (nach init_db) — fängt Crashes zwischen
|
|
Datei-Write und Status-Update ab.
|
|
"""
|
|
for g in await list_guides():
|
|
if g["status"] == "done" and not guide_content_path(g["topic"], g["format"]).exists():
|
|
log.warning("[%s] Guide %s: done ohne Content-Datei — auf error gesetzt", g["topic"], g["id"])
|
|
now = datetime.now(timezone.utc).isoformat()
|
|
await update_guide(g["id"], status="error", error_msg="Inhalt fehlt — neu generieren", updated_at=now)
|
|
|
|
|
|
async def generate_guide(guide_id: str, topic: str, format_name: str, instructions: str = "", provider: str = DEFAULT_PROVIDER, ab_step: int | None = None) -> None:
|
|
async with _semaphore:
|
|
now = datetime.now(timezone.utc).isoformat()
|
|
await update_guide(guide_id, status="generating", progress="Starte…", updated_at=now)
|
|
|
|
content_path = guide_content_path(topic, format_name)
|
|
content_path.parent.mkdir(parents=True, exist_ok=True)
|
|
project = quelle_ordner(topic) # Ordner-Quelle (projekt/uni/link) → Pfad, sonst None
|
|
|
|
try:
|
|
if is_guide_cancelled(guide_id):
|
|
return
|
|
|
|
if project:
|
|
await asyncio.to_thread(_pdfs_konvertieren, project)
|
|
|
|
# Re-Run ab Schritt: Content + Slots ab `ab_step` löschen, Rest bleibt → Resume baut ab dort.
|
|
# Sonst „Neu erstellen": fertiger Guide → kompletter Frischstart.
|
|
# Sonst sind Schritt-Dateien Reste eines Abbruchs/Fehlers → Resume.
|
|
if ab_step is not None:
|
|
_reset_guide_ab_step(content_path, ab_step)
|
|
elif content_path.exists():
|
|
for p_alt in guide_slot_dateien(content_path):
|
|
p_alt.unlink(missing_ok=True)
|
|
|
|
bs = await list_bausteine(topic, status="konsens")
|
|
if bs:
|
|
alle = {i: (f"{b['titel']} — {b['beschreibung']}" if b["beschreibung"] else b["titel"])
|
|
for i, b in enumerate(bs, 1)}
|
|
else: # Fallback: bausteine.md (Alt-Themen)
|
|
bp = bausteine_path(topic)
|
|
alle = _lade_bausteine(bp.read_text(encoding="utf-8")) if bp.exists() else {}
|
|
if not alle:
|
|
await _fail(guide_id, "Keine Bausteine gefunden")
|
|
return
|
|
entries = _eindeutige_titel(alle)
|
|
facts = _prompt("Guide-Fakten-Projekt", project=project) if project else _prompt("Guide-Fakten-Thema")
|
|
chapters = await _generate_sections(
|
|
guide_id, topic, format_name, entries,
|
|
facts, instructions, provider, content_path,
|
|
)
|
|
if chapters is None or is_guide_cancelled(guide_id):
|
|
return
|
|
content = {"topic": topic, "format": format_name, "chapters": chapters}
|
|
|
|
atomic_write_json(content_path, content, indent=1) # Brücke (Resume/Fallback)
|
|
await set_guide_content(topic, format_name, json.dumps(content, ensure_ascii=False))
|
|
|
|
now = datetime.now(timezone.utc).isoformat()
|
|
await update_guide(guide_id, status="done", progress=None, step=None, updated_at=now)
|
|
|
|
except asyncio.TimeoutError:
|
|
await _fail(guide_id, "Timeout bei der Generierung")
|
|
except FileNotFoundError:
|
|
await _fail(guide_id, "Bausteine fehlen")
|
|
except Exception as e:
|
|
log.exception("[%s] Guide-Generierung fehlgeschlagen (%s)", topic, guide_id)
|
|
await _fail(guide_id, str(e)[:2000])
|
|
finally:
|
|
clear_guide_cancelled(guide_id)
|