From 58a4f9b52bac041106ff5121ffaf4e40a07745f8 Mon Sep 17 00:00:00 2001
From: digital-garbage <133010426+digital-garbage@users.noreply.github.com>
Date: Mon, 15 Jun 2026 09:41:09 +0300
Subject: [PATCH] Stage 1a: canonical split_scenes (additive, not yet
 repointed)
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

One rich scene model {anchor, scenes:[{index,raw,keys,effect}]} that every
consumer will derive from — editing (raw), preview (raw->expanded),
attribution (keys), reset (pool = union of keys), generation. Built on the
existing shortcut-aware verbatim splitter + refinement_keys_in_text, so it
can't disagree with itself. Nothing repointed yet; consumers still on their
own splitters until the staged repointing.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---
 conditioning.py               | 57 +++++++++++++++++++++++++++++++++++
 tests/test_refinement_keys.py | 28 +++++++++++++++++
 2 files changed, 85 insertions(+)

diff --git a/conditioning.py b/conditioning.py
index c0f79c96..63852091 100644
--- a/conditioning.py
+++ b/conditioning.py
@@ -903,6 +903,63 @@ def refinement_keys_in_text(text):
             if pc.get("is_shortcut") and str(pc.get("refinement_key") or "")}
 
 
+def split_scenes(prompt):
+    """THE canonical scene split — one source of truth for editing, preview, key attribution,
+    session reset and generation.
+
+    Splits `prompt` into an anchor + scenes at shortcut-aware transition boundaries (detected on
+    BOTH the expanded and the raw text, then projected onto the raw words so nothing is expanded
+    or dropped), and for each scene reports everything any consumer needs:
+
+        {
+          "anchor": <raw anchor text, prepended to every scene>,
+          "scenes": [
+            {"index": i,
+             "raw":    <verbatim scene text>,          # editing / lossless mapping
+             "keys":   <set of refinement keys firing in this scene + anchor>,  # attribution / reset
+             "effect": <transition visual effect before this scene | None>},    # generation
+            ...
+          ],
+        }
+
+    Derive, never re-split: per-scene keys = entry["keys"]; the reset pool = union of all keys;
+    the preview/generation scene texts come from these raw chunks (expanded by the caller). Because
+    every consumer reads ONE split, scene counts and boundaries can never disagree."""
+    verbatim = split_timeline_verbatim(prompt)
+    anchor = verbatim.get("anchor", "") or ""
+    anchor_keys = refinement_keys_in_text(anchor)
+    # transitions[i] = {"after_scene": k, "visual_effect": e} means effect e precedes scene k+1.
+    effect_before = {}
+    for t in verbatim.get("transitions", []) or []:
+        try:
+            effect_before[int(t.get("after_scene", -1)) + 1] = t.get("visual_effect")
+        except (TypeError, ValueError):
+            continue
+    scenes = []
+    for s in verbatim.get("scenes", []) or []:
+        try:
+            idx = int(s.get("index", len(scenes)))
+        except (TypeError, ValueError):
+            idx = len(scenes)
+        raw = s.get("text", "") or ""
+        scenes.append({
+            "index": idx,
+            "raw": raw,
+            "keys": anchor_keys | refinement_keys_in_text(raw),
+            "effect": effect_before.get(idx),
+        })
+    return {"anchor": anchor, "scenes": scenes}
+
+
+def refinement_key_pool_for(prompt):
+    """Every non-default refinement key the prompt activates (anchor + all scenes) — the exact
+    set a session reset wipes. Derived from the one canonical split."""
+    pool = set()
+    for scene in split_scenes(prompt).get("scenes", []) or []:
+        pool |= scene.get("keys") or set()
+    return pool
+
+
 def resolve_scene_refinement_keys(raw_positive, scene_count):
     """Per-scene non-default refinement keys, read straight off each scene's text.
 
diff --git a/tests/test_refinement_keys.py b/tests/test_refinement_keys.py
index f6ed1745..a2b2db94 100644
--- a/tests/test_refinement_keys.py
+++ b/tests/test_refinement_keys.py
@@ -233,6 +233,34 @@ def test_reset_prompt_keys_skips_when_primary_only(monkeypatch):
     assert studio._v2_reset_prompt_keys("default", "plain prompt") == []
 
 
+def test_split_scenes_canonical_model(monkeypatch):
+    """The one canonical split reports per-scene raw text + keys; anchor keys fold into every
+    scene; the reset pool is the union."""
+    db = _db({"anc": (["zeta"], ["ZZZ"], "anchorKey"),
+              "a": (["alpha"], ["AAA"], "keyA"),
+              "b": (["beta"], ["BBB"], "keyB")})
+    monkeypatch.setattr(templates, "load_shortcut_db", lambda: db)
+    monkeypatch.setattr(templates, "load_custom_transition_triggers", lambda: {})
+    out = conditioning.split_scenes("zeta intro scene 1 alpha here scene 2 beta here")
+    scenes = out["scenes"]
+    assert len(scenes) == 2
+    assert scenes[0]["keys"] == {"anchorKey", "keyA"}
+    assert scenes[1]["keys"] == {"anchorKey", "keyB"}
+    assert "alpha" in scenes[0]["raw"]  # verbatim text preserved
+    assert conditioning.refinement_key_pool_for(
+        "zeta intro scene 1 alpha here scene 2 beta here") == {"anchorKey", "keyA", "keyB"}
+
+
+def test_split_scenes_single_scene(monkeypatch):
+    db = _db({"hero": (["hero"], ["a knight"], "mykey")})
+    monkeypatch.setattr(templates, "load_shortcut_db", lambda: db)
+    monkeypatch.setattr(templates, "load_custom_transition_triggers", lambda: {})
+    out = conditioning.split_scenes("hero stands in a field")
+    assert len(out["scenes"]) == 1
+    assert out["scenes"][0]["keys"] == {"mykey"}
+    assert conditioning.refinement_key_pool_for("hero stands in a field") == {"mykey"}
+
+
 def test_single_scene_with_custom_key_is_detected(monkeypatch):
     """Regression: a custom key fired in a single-logical-scene prompt MUST be detected (was
     silently dropped to empty when the old code compared scene counts and fell back)."""