update

backend/blocks.py

@@ -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: 3–13 tool rounds, 52–141 s vs 15–35 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)