This commit is contained in:
Team3
2026-07-04 16:50:50 +02:00
parent 8488737303
commit 92c69c1561
7 changed files with 298 additions and 37 deletions

View File

@@ -597,6 +597,32 @@ def _sink_json(result, path: Path, schema):
return val
def _sink_subs(result, path: Path):
"""Finder reply as TEXT (marker format), persisted to `path` for audit/diagnosis.
File fallback: a tool-capable agent (thema web mode) that wrote the file despite the
text instruction still counts — same tolerance as _sink_or_file."""
text = _reply_text(result).strip()
d = _parse_subblocks(text)
if d:
atomic_write_text(path, text)
return d
return _parse_subblocks(_read(path)) or None
def _finder_material(folder, sources: list[str] | None, queries: list[str]) -> tuple[str, str, str]:
"""→ (material, backing, caps) for the Subblock-Research prompt. uni/projekt: corpus
excerpts INLINE — the finder had neither path nor excerpts and hunted the material per
call via glob/grep/bash (measured: 313 tool rounds, 52141 s vs 1535 s single-shot).
thema (or excerpt miss): web research as before, fail-open."""
ev = _evidence_pack(folder, sources, queries) if folder else ""
if ev:
material = "\n" + _prompt("Blocks-Source-Inline", excerpts=ev) + "\n"
backing = ("Backed by the SOURCE EXCERPTS above, not invented — leave out any "
"sub-point the excerpts do not support.")
return material, backing, "none"
return "", "Backed, not invented. Verify uncertain points via web search.", ("files" if folder else "full")
def _build_research_prompt(topic: str, out_path: Path, instructions: str, type: str, folder: Path | None, fokus: str = "", section: str = "") -> str:
if section:
# Section mode (uni/projekt): text directly in the prompt → small context, no file reading.
@@ -760,15 +786,15 @@ async def _subblocks_block(ctx: GenContext, set_p, files: dict, entries: dict, i
return ("\n\nBEREITS ERFASST — liste diese NICHT erneut. Finde nur, was FEHLT:\n" + "\n".join(known))
# ONE finder round (3 slots, quorum 2) → count of NEW sub norms; None = no result/cancel.
async def _one_round(label, subset, assignment, paths, keys, known, extra_instr):
async def _one_round(label, subset, assignment, paths, keys, known, extra_instr, material, backing, round_caps):
chunk_idx = _title_index({num: title_by_num[num] for num in subset})
for p in paths:
p.unlink(missing_ok=True)
slots = [{
"key": k,
"prompt": _prompt("Subblock-Research", topic=topic, assignment=assignment, known=known, out_path=p, extra=_extra(extra_instr)),
"role": "quick", "capabilities": caps,
"payload": (lambda result, p=p: _parse_subblocks(_read(p)) or None),
"prompt": _prompt("Subblock-Research", topic=topic, assignment=assignment, known=known, material=material, backing=backing, extra=_extra(extra_instr)),
"role": "quick", "capabilities": round_caps,
"payload": (lambda result, p=p: _sink_subs(result, p)),
} for k, p in zip(keys, paths)]
agent_texts = await _race(topic, label, slots, 2, _timeout("subblock", len(subset)), provider, cancelled=is_cancelled, grace=CONSENSUS_GRACE)
if is_cancelled() or not agent_texts:
@@ -818,6 +844,10 @@ async def _subblocks_block(ctx: GenContext, set_p, files: dict, entries: dict, i
# Phase "Subblocks find": per package loop until 0 new subs / time cap.
async def _find(c, chunk):
assignment = "\n".join(f"- {entries[num]}" for num in chunk)
# Material once per package (excerpts are round-invariant; one query per block
# keeps the coverage guarantee of _evidence_pack).
material, backing, round_caps = await asyncio.to_thread(
_finder_material, folder, sources, [str(entries[num]) for num in chunk])
start = time.monotonic()
round_n = 0
while not is_cancelled():
@@ -825,7 +855,7 @@ async def _subblocks_block(ctx: GenContext, set_p, files: dict, entries: dict, i
bekannt = await _known_block(chunk) if round_n > 1 else ""
paths = [work_dir / f"subblock-c{c}-r{round_n}-{i}.md" for i in (1, 2, 3)]
keys = [f"blocks-{topic}-{ns}subblock-c{c}-r{round_n}-{i}" for i in (1, 2, 3)]
new = await _one_round(f"{lbl}Subblocks package {c} R{round_n}", chunk, assignment, paths, keys, bekannt, instructions)
new = await _one_round(f"{lbl}Subblocks package {c} R{round_n}", chunk, assignment, paths, keys, bekannt, instructions, material, backing, round_caps)
if new is None:
if is_cancelled():
return False
@@ -1067,9 +1097,11 @@ async def _subblocks_block(ctx: GenContext, set_p, files: dict, entries: dict, i
focus = (instructions + "\n\nDieser Block hat bisher nur sehr wenige belegte "
"Subbausteine. Suche gezielt nach WEITEREN belegbaren Kernaspekten, die "
"oben fehlen. Nimm NUR auf, was die Quellen wirklich hergeben — nicht aufblähen.")
material, backing, round_caps = await asyncio.to_thread(
_finder_material, folder, sources, [str(entries[num]) for num in lacking])
paths = [work_dir / f"subblock-x{k}-c{c}-{i}.md" for i in (1, 2, 3)]
keys = [f"blocks-{topic}-{ns}subblock-x{k}-c{c}-{i}" for i in (1, 2, 3)]
new = await _one_round(f"{lbl}Subblocks catch-up {c} X{k}", lacking, assignment, paths, keys, known, focus)
new = await _one_round(f"{lbl}Subblocks catch-up {c} X{k}", lacking, assignment, paths, keys, known, focus, material, backing, round_caps)
if not new:
return
await _select(lacking, keep_consensus=True)
@@ -1412,12 +1444,14 @@ async def _luecken_runde(ctx: GenContext, files: dict, title: str, luecken: list
"Aspekten des Blocks — nichts anderes:\n" + "\n".join(f"- {l}" for l in luecken))
known = ("\n\nBEREITS ERFASST — liste diese NICHT erneut:\n"
+ "\n".join(f"- {s}" for s in have)) if have else ""
material, backing, caps = await asyncio.to_thread(
_finder_material, folder, sources, [title] + list(luecken))
paths = [work_dir / f"luecken-{ns}r1-{i}.md" for i in (1, 2, 3)]
slots = [{
"key": f"blocks-{topic}-{ns}luecken-r1-{i}",
"prompt": _prompt("Subblock-Research", topic=topic, assignment=f"- {title}", known=known, out_path=p, extra=_extra(focus)),
"role": "quick", "capabilities": "files" if folder else "full",
"payload": (lambda result, p=p: _parse_subblocks(_read(p)) or None),
"prompt": _prompt("Subblock-Research", topic=topic, assignment=f"- {title}", known=known, material=material, backing=backing, extra=_extra(focus)),
"role": "quick", "capabilities": caps,
"payload": (lambda result, p=p: _sink_subs(result, p)),
} for i, p in zip((1, 2, 3), paths)]
agent_texts = await _race(topic, f"{lbl}Lücken-Nachfass", slots, 2,
_timeout("subblock", 1), provider, cancelled=is_cancelled, grace=CONSENSUS_GRACE)

View File

@@ -104,6 +104,20 @@ def _naming_schema(data, count: int) -> tuple[int | None, str | None, bool] | No
return n, name, False
def _sanierung_schema(data) -> tuple[str, str] | None:
"""{"title": …, "description": …} → (title, description), beides getrimmt; Titel über
80 Zeichen verfällt (leerer String = kein Vorschlag, Feld bleibt wie es ist)."""
if not isinstance(data, dict):
return None
t = str(data.get("title") or "").strip()
d = str(data.get("description") or "").strip()
if not t and not d:
return None
if len(t) > 80:
t = ""
return t, d
def _name_verankert(name: str, rows: list[dict], ctoks: set[str] | None) -> bool:
"""Abstraction guard: a free-formed title must be anchored — in the corpus (uni/projekt)
or in the members' own words (thema). Unanchored names drift to textbook canon
@@ -470,6 +484,16 @@ def _hat_anker(title: str, ctoks: set[str]) -> bool:
return False
def _sanierung_noetig(p: dict, ctoks: set[str] | None) -> bool:
"""QA-messbare Befund-Formen am Entstehungsort: Titel ohne Korpus-Anker (QA: fremd —
misst Token-Anker, nicht Semantik) oder leere Beschreibung (QA: hygiene). Gilt auch für
Singleton-Cluster, die das Naming sonst überspringen — Reader-Rohtitel gingen wörtlich
bis done_block durch (gemessen: 'k-Coloring' statt Korpus-Form 'k-Color')."""
if not (p.get("description") or "").strip():
return True
return bool(ctoks) and not _hat_anker(p.get("title", ""), ctoks)
async def _anker_beleg(ctx: GenContext, flow: Flow, kandidaten: list[tuple[str, dict]]) -> set[str]:
"""Evidence judge for quorum titles WITHOUT any corpus anchor — two readers naming the
same famous canon independently beat the quorum although the material never mentions it
@@ -703,9 +727,59 @@ async def _name_one(ctx: GenContext, flow: Flow, c):
p["title"] = custom or w["title"]
p["description"] = w.get("description") or p.get("description", "")
await db.kanban_set_payload(topic, BOARD, cid, p)
await _saniere_one(ctx, flow, cid, p)
await db.kanban_advance(topic, BOARD, cid, "naming_check")
async def _saniere_one(ctx: GenContext, flow: Flow, cid: str, p: dict) -> None:
"""Fremd-/Hygiene-Sanierung am Entstehungsort (nur Korpus-Quellen): Titel ohne
Korpus-Anker auf die Oberflächenform des Materials umschreiben, leere Beschreibung
belegt nachfassen. Läuft NACH der Member-Wahl auf dem finalen Payload — greift damit
auch für Singleton-Cluster und unverankerte Naming-Gewinner. Vorschläge werden
deterministisch validiert (_hat_anker, dieselbe Messlatte wie QA-fremd); fail-open:
unverankerter Vorschlag oder Judge-Ausfall lässt die Karte unverändert weiter."""
from qa import _distinctive
topic = flow.topic
folder = source_folder(topic)
if folder is None:
return # thema: kein Korpus, keine Oberflächenform zum Verankern
ctoks = flow.state.get("korpus_tokens")
if ctoks is None: # Resume: das Konsens-Gate dieses Flows lief nie
ctoks = flow.state["korpus_tokens"] = await asyncio.to_thread(_korpus_tokens, folder)
if not _sanierung_noetig(p, ctoks):
return
titel, beschr = p.get("title", ""), (p.get("description") or "").strip()
toks = _distinctive(titel) | _distinctive(beschr)
ev = _evidence_pack(folder, p.get("sources") or None, [" ".join(sorted(toks))],
budget=4000) if toks else ""
if not ev:
return # ohne Auszüge kein Anker-Rewrite — Anker-Beleg hat Echtheit schon entschieden
h = _h(titel, beschr, "sanierung")
path = flow.work_dir / f"sanierung-{cid}-{h}.json"
verdict = _sanierung_schema(_json_file(path))
if verdict is None:
status, verdict = await run_single_slot(
ctx, f"Sanierung {cid}", key=f"blocks-{topic}-sanierung-{cid}-{h}",
prompt=_prompt("Blocks-Sanierung", topic=topic, title=titel,
description=beschr or "(leer)", excerpts=ev),
role="judge", capabilities="none",
payload=lambda result, p2=path: _sink_json(result, p2, _sanierung_schema),
timeout=_timeout("selection_mapping", 1))
if status != OK or verdict is None:
return
neu_titel, neu_beschr = verdict
changed = False
if (neu_titel and _norm_title(neu_titel) != _norm_title(titel)
and _hat_anker(neu_titel, ctoks)):
p["title"] = clean_title(neu_titel)
changed = True
if not beschr and neu_beschr:
p["description"] = neu_beschr
changed = True
if changed:
await db.kanban_set_payload(topic, BOARD, cid, p)
async def _proc_naming_check(ctx: GenContext, flow: Flow, cards):
results = await asyncio.gather(*[_namecheck_one(ctx, flow, c) for c in cards], return_exceptions=True)
errs = [r for r in results if isinstance(r, Exception)]

View File

@@ -118,6 +118,13 @@ class Welt:
return j({"keep": keep, "rest": []})
if "-naming-" in key: # deckt auch naming_check (gleicher Key)
return j({"best": 1})
if "-sanierung-" in key: # Titel/Beschreibung unverändert zurück (Fake-Welt ist sauber)
t = re.search(r"^Title: (.+)$", prompt, re.M)
d = re.search(r"^Description: (.+)$", prompt, re.M)
beschr = (d.group(1).strip() if d else "")
if beschr == "(leer)":
beschr = "Beschreibung aus dem Material."
return j({"title": t.group(1).strip() if t else "", "description": beschr})
if "-filter-" in key: # auch filter-recheck
return j({"fragments": {}, "drop": []})
if "-gruppierung-completion-" in key:

View File

@@ -1181,3 +1181,144 @@ async def test_namecheck_ok_behaelt_titel(testdb, tmp_path, monkeypatch):
block = await db.kanban_get_card(TOPIC, B, "b-c9")
assert block["payload"]["title"] == "Eigener Titel"
assert block["payload"]["description"] == "Eigene Beschreibung"
# ── Sanierung: Titel auf Korpus-Form, Beschreibungspflicht (QA: fremd/hygiene) ──────
def test_sanierung_schema_varianten():
assert bi._sanierung_schema({"title": " k-Color ", "description": "d"}) == ("k-Color", "d")
assert bi._sanierung_schema({"title": "", "description": "nur Beschreibung"}) == ("", "nur Beschreibung")
assert bi._sanierung_schema({"title": "x" * 90, "description": "d"}) == ("", "d") # zu lang
assert bi._sanierung_schema({"title": "", "description": ""}) is None
assert bi._sanierung_schema("quatsch") is None
def test_sanierung_noetig():
ctoks = {"color", "graph"}
assert bi._sanierung_noetig({"title": "k-Color", "description": ""}, ctoks) # leere Beschreibung
assert bi._sanierung_noetig({"title": "k-Coloring", "description": "d"}, ctoks) # kein Korpus-Anker
assert not bi._sanierung_noetig({"title": "k-Color", "description": "d"}, ctoks)
assert not bi._sanierung_noetig({"title": "k-Coloring", "description": "d"}, None) # thema: kein Korpus
def _sanierung_env(tmp_path, monkeypatch, antwort):
"""Korpus mit 'k-Color'-Oberflächenform; Judge antwortet mit `antwort`."""
(tmp_path / "korpus.txt").write_text(
"Das k-Color Problem: Kann der Graph mit k Farben gefärbt werden? NP-vollständig.",
encoding="utf-8")
monkeypatch.setattr(bi, "source_folder", lambda t: tmp_path)
seen = {}
async def fake_slot(ctx, label, *, key, prompt, role, capabilities, payload, timeout):
seen[key] = prompt
return "ok", payload((0, json.dumps(antwort["val"]), ""))
monkeypatch.setattr(bi, "run_single_slot", fake_slot)
return GenContext(topic=TOPIC, provider="claude", is_cancelled=lambda: False), seen
async def test_singleton_unverankert_wird_saniert(testdb, tmp_path, monkeypatch):
"""Singleton-Cluster mit Reader-Titel ohne Korpus-Anker: Naming wird nicht mehr
übersprungen — der Titel wird auf die Korpus-Oberflächenform umgeschrieben
(gemessener aak-Fall 'k-Coloring' statt 'k-Color')."""
db = testdb
antwort = {"val": {"title": "k-Color", "description": "Kann der Graph mit k Farben gefärbt werden?"}}
ctx, seen = _sanierung_env(tmp_path, monkeypatch, antwort)
async def fake_members(topic, cid):
return [{"norm": "k-coloring", "title": "k-Coloring",
"description": "Kann der Graph mit k Farben gefärbt werden?",
"readers": ["r1", "r2"], "sources": []}]
monkeypatch.setattr(bi, "_member_rows", fake_members)
payload = {"title": "k-Coloring", "description": "Kann der Graph mit k Farben gefärbt werden?"}
await db.kanban_upsert_card(TOPIC, B, "c1", "cluster", "naming", payload)
await bi._name_one(ctx, _mk_flow(tmp_path), {"card_id": "c1", "payload": payload})
card = await db.kanban_get_card(TOPIC, B, "c1")
assert card["stage"] == "naming_check"
assert card["payload"]["title"] == "k-Color"
assert any("-sanierung-" in k for k in seen)
async def test_leere_beschreibung_wird_gefuellt(testdb, tmp_path, monkeypatch):
"""Verankerter Titel, leere Beschreibung (gemessener aak-Fall der SetCover-Fragmente):
Beschreibung wird belegt nachgefasst, Titel bleibt."""
db = testdb
antwort = {"val": {"title": "k-Color", "description": "Färbbarkeit mit k Farben."}}
ctx, _seen = _sanierung_env(tmp_path, monkeypatch, antwort)
async def fake_members(topic, cid):
return [{"norm": "k-color", "title": "k-Color", "description": "",
"readers": ["r1"], "sources": []}]
monkeypatch.setattr(bi, "_member_rows", fake_members)
payload = {"title": "k-Color", "description": ""}
await db.kanban_upsert_card(TOPIC, B, "c5", "cluster", "naming", payload)
await bi._name_one(ctx, _mk_flow(tmp_path), {"card_id": "c5", "payload": payload})
card = await db.kanban_get_card(TOPIC, B, "c5")
assert card["payload"]["title"] == "k-Color"
assert card["payload"]["description"] == "Färbbarkeit mit k Farben."
async def test_sanierung_unverankerter_vorschlag_verfaellt(testdb, tmp_path, monkeypatch):
"""Judge-Vorschlag ohne Korpus-Anker wird verworfen (dieselbe Messlatte wie QA-fremd);
die Beschreibung wird trotzdem übernommen."""
db = testdb
antwort = {"val": {"title": "Graphfärbungsproblem", "description": "Färbbarkeit mit k Farben."}}
ctx, _seen = _sanierung_env(tmp_path, monkeypatch, antwort)
async def fake_members(topic, cid):
return [{"norm": "k-coloring", "title": "k-Coloring",
"description": "Kann der Graph mit k Farben gefärbt werden?",
"readers": ["r1"], "sources": []}]
monkeypatch.setattr(bi, "_member_rows", fake_members)
payload = {"title": "k-Coloring", "description": "Kann der Graph mit k Farben gefärbt werden?"}
await db.kanban_upsert_card(TOPIC, B, "c2", "cluster", "naming", payload)
await bi._name_one(ctx, _mk_flow(tmp_path), {"card_id": "c2", "payload": payload})
card = await db.kanban_get_card(TOPIC, B, "c2")
assert card["payload"]["title"] == "k-Coloring" # unverankert → verfällt
assert card["stage"] == "naming_check"
async def test_sanierung_fail_open(testdb, tmp_path, monkeypatch):
"""Judge-Ausfall → Karte läuft unverändert weiter (kein Deadletter am Naming)."""
db = testdb
(tmp_path / "korpus.txt").write_text("Der Graph ist endlich.", encoding="utf-8")
monkeypatch.setattr(bi, "source_folder", lambda t: tmp_path)
async def broken_slot(*a, **kw):
return "failed", None
async def fake_members(topic, cid):
return [{"norm": "x", "title": "Graph", "description": "", "readers": ["r1"], "sources": []}]
monkeypatch.setattr(bi, "run_single_slot", broken_slot)
monkeypatch.setattr(bi, "_member_rows", fake_members)
payload = {"title": "Graph", "description": ""}
await db.kanban_upsert_card(TOPIC, B, "c3", "cluster", "naming", payload)
ctx = GenContext(topic=TOPIC, provider="claude", is_cancelled=lambda: False)
await bi._name_one(ctx, _mk_flow(tmp_path), {"card_id": "c3", "payload": payload})
card = await db.kanban_get_card(TOPIC, B, "c3")
assert card["stage"] == "naming_check" and card["payload"]["title"] == "Graph"
async def test_sanierung_anker_und_beschreibung_ok_kein_judge(testdb, tmp_path, monkeypatch):
"""Verankerter Titel + Beschreibung vorhanden → kein Sanierungs-Call."""
db = testdb
(tmp_path / "korpus.txt").write_text("Das k-Color Problem im Graph.", encoding="utf-8")
monkeypatch.setattr(bi, "source_folder", lambda t: tmp_path)
async def never_slot(*a, **kw):
raise AssertionError("Sanierung darf ohne Befund-Form nicht laufen")
async def fake_members(topic, cid):
return [{"norm": "k-color", "title": "k-Color", "description": "d", "readers": ["r1"], "sources": []}]
monkeypatch.setattr(bi, "run_single_slot", never_slot)
monkeypatch.setattr(bi, "_member_rows", fake_members)
payload = {"title": "k-Color", "description": "d"}
await db.kanban_upsert_card(TOPIC, B, "c4", "cluster", "naming", payload)
ctx = GenContext(topic=TOPIC, provider="claude", is_cancelled=lambda: False)
await bi._name_one(ctx, _mk_flow(tmp_path), {"card_id": "c4", "payload": payload})
assert (await db.kanban_get_card(TOPIC, B, "c4"))["stage"] == "naming_check"

View File

@@ -75,11 +75,9 @@ def _mk_race(finder_by_agent):
if "-r1-" in key:
agent = int(key.rsplit("-", 1)[1])
subs = finder_by_agent.get(agent) or []
if subs and (m := _MD_PATH.search(prompt)):
if subs: # Finder antworten als TEXT (Marker-Format), kein out_path mehr
text = "<!-- block: Alpha -->\n" + "\n".join(f"- {s}" for s in subs)
with open(m.group(1), "w", encoding="utf-8") as f:
f.write(text)
outs.append(slot["payload"](None))
outs.append(slot["payload"]((0, text, "")))
outs = [o for o in outs if o]
return outs or None
fake_race.slots_seen = []
@@ -177,11 +175,8 @@ async def test_catchup_adds_and_stops(sub_env, monkeypatch, tmp_path):
if any("-subblock-x" in s["key"] for s in slots):
hit["n"] += 1
if hit["n"] == 1: # first catch-up round: both agents agree on one new sub
for slot in slots[:2]:
m = _MD_PATH.search(slot["prompt"])
with open(m.group(1), "w", encoding="utf-8") as f:
f.write("<!-- block: Alpha -->\n- Vertiefung der Konzepte")
return [slot["payload"](None) for slot in slots[:2]]
text = "<!-- block: Alpha -->\n- Vertiefung der Konzepte"
return [slot["payload"]((0, text, "")) for slot in slots[:2]]
return None
return await base(topic, label, slots, *a, **k)
@@ -251,13 +246,8 @@ async def test_paraphrase_saturation_stops_early(sub_env, monkeypatch):
async def with_r2(topic, label, slots, *a, **k):
if any("-r2-" in s["key"] for s in slots):
outs = []
for slot in slots[:2]:
m = _MD_PATH.search(slot["prompt"])
with open(m.group(1), "w", encoding="utf-8") as f:
f.write("<!-- block: Alpha -->\n- Umbruch erfordert explizite Marker!")
outs.append(slot["payload"](None))
return outs
text = "<!-- block: Alpha -->\n- Umbruch erfordert explizite Marker!"
return [slot["payload"]((0, text, "")) for slot in slots[:2]]
return await base_fake(topic, label, slots, *a, **k)
monkeypatch.setattr(blx, "_race", with_r2)
@@ -282,11 +272,8 @@ async def test_round_cap_stops_endless_finders(sub_env, monkeypatch):
rn = _re.search(r"-r(\d+)-", slots[0]["key"])
n = rn.group(1) if rn else "x"
for slot in slots[:2]:
m = _MD_PATH.search(slot["prompt"])
with open(m.group(1), "w", encoding="utf-8") as f:
f.write(f"<!-- block: Alpha -->\n- Konzept{n} ist eigenständig")
prompts.append((slot["key"], slot["prompt"]))
outs.append(slot["payload"](None))
outs.append(slot["payload"]((0, f"<!-- block: Alpha -->\n- Konzept{n} ist eigenständig", "")))
return outs
monkeypatch.setattr(blx, "_race", endless)
@@ -362,8 +349,9 @@ def test_sink_json_writes_only_valid(tmp_path):
async def test_clarify_inline_evidence_no_tools(sub_env, monkeypatch, tmp_path):
"""Mit Korpus: Judges bekommen Auszüge inline und laufen ohne Tools (Text-Antwort);
die j-Datei schreibt die Engine. Finder bleiben unverändert bei capabilities=files."""
"""Mit Korpus: Judges UND Finder bekommen Auszüge inline und laufen ohne Tools
(Text-Antwort); die Dateien schreibt die Engine. Der Finder verlor vorher 313
Tool-Runden pro Call mit der Material-Suche via glob/grep/bash."""
db, ctx, files = sub_env
d = _corpus(tmp_path)
monkeypatch.setattr(blx, "source_folder", lambda t: d)
@@ -378,7 +366,9 @@ async def test_clarify_inline_evidence_no_tools(sub_env, monkeypatch, tmp_path):
assert judges and all(s["capabilities"] == "none" for s in judges)
assert "── Skript.txt" in judges[0]["prompt"]
assert "ls/find" not in judges[0]["prompt"] # keine Selbst-Recherche-Anweisung mehr
assert finders and all(s["capabilities"] == "files" for s in finders)
assert finders and all(s["capabilities"] == "none" for s in finders)
assert "── Skript.txt" in finders[0]["prompt"] # Auszüge inline statt Dateisystem-Suche
assert "web search" not in finders[0]["prompt"]
assert list(tmp_path.glob("subblock-final-*-j*.md")) # Engine persistiert die Judge-Antwort