More character name bug fixes

Author: Robert Weber

Novel_Processing_Instructions.md

@@ -23,22 +23,22 @@ Don’t use any names in the premise...just describe the characters and their ro
 
 - Read the file "<NOVEL_PATH_MD>" and add to the context of this thread.  This is a novel written in chapters and there are chapter delineations present throughout.
 
-- we are now going to start adding sections to the editors notes markdown document of things that should be corrected in the next phase of edits.  make sure you have an up to date context of the novel in its current form.  our target is to make this a 9.5 / 10 book with with atleast 85000 total words.  in this new section , document new items you feel should be executed to lengthen and strengthen the novel.  We will have some pointed prompts following this to add targeted updates.
+- we are now going to start adding sections to the editors notes markdown document of things that should be corrected in the next phase of edits.  make sure you have an up to date context of the novel in its current form.  our target is to make this a 9.5 / 10 book with with atleast 85000 total words.  in this new section , document new items you feel should be executed to lengthen and strengthen the novel.  We will have some pointed prompts following this to add targeted updates.  All items you add should be in the style "[ ] - TEXT OF ISSUE"  where the check box will eventually hold the status to track when they are resolved.
 
 - Character Voice Differentiation
-our target is to make this a 9.5 / 10 book with with atleast 85000 total words. Each POV character should think in a distinct internal language shaped by their background, demographics and expertise.   A 16 year old should think and talk like a 16 year old.  An old man should think and talk like an old man.  Look through the novel and find any dialog that doesnt match the speaking character.  create a plan to update these voices and add that to a new section in the editors notes markdown file.
+our target is to make this a 9.5 / 10 book with with atleast 85000 total words. Each POV character should think in a distinct internal language shaped by their background, demographics and expertise.   A 16 year old should think and talk like a 16 year old.  An old man should think and talk like an old man.  Look through the novel and find any dialog that doesnt match the speaking character.  create a plan to update these voices and add that to a new section in the editors notes markdown file.  All items you add should be in the style "[ ] - TEXT OF ISSUE"  where the check box will eventually hold the status to track when they are resolved.
 
 - Dialogue Naturalization
-our target is to make this a 9.5 / 10 book with with atleast 85000 total words. Make sure the current dialogue isnt too clean, too functional, too information-delivery. Characters sometimes have incomplete thoughts, don't always speak in well-formed sentences, and sometimes rarely interrupt each other or themselves.  make sure the dialog in the novel reads this way.  create a plan to update these voices and add that to a new section in the editors notes markdown file.
+our target is to make this a 9.5 / 10 book with with atleast 85000 total words. Make sure the current dialogue isnt too clean, too functional, too information-delivery. Characters sometimes have incomplete thoughts, don't always speak in well-formed sentences, and sometimes rarely interrupt each other or themselves.  make sure the dialog in the novel reads this way.  create a plan to update these voices and add that to a new section in the editors notes markdown file.  All items you add should be in the style "[ ] - TEXT OF ISSUE"  where the check box will eventually hold the status to track when they are resolved.
 
 - Humor, Strangeness, and the Unexpected
-our target is to make this a 9.5 / 10 book with with atleast 85000 total words.Real characters deflect, joke badly, notice irrelevant things, and occasionally do something that doesn't serve the plot.  create a plan to inject these odities throughout the novel.  1-2 oddities per chapter.
+our target is to make this a 9.5 / 10 book with with atleast 85000 total words.Real characters deflect, joke badly, notice irrelevant things, and occasionally do something that doesn't serve the plot.  create a plan to inject these odities throughout the novel.  1-2 oddities per chapter.  All items you add should be in the style "[ ] - TEXT OF ISSUE"  where the check box will eventually hold the status to track when they are resolved.
 
 - Prose Texture Variation
-our target is to make this a 9.5 / 10 book with with atleast 85000 total words.Make sure the  prose has a varying literary density throughout. It should breathe -- denser in reflective moments, sparser in action, occasionally raw or clumsy when characters are overwhelmed. create a plan to update the prose and add that to a new section in the editors notes markdown file.
+our target is to make this a 9.5 / 10 book with with atleast 85000 total words.Make sure the  prose has a varying literary density throughout. It should breathe -- denser in reflective moments, sparser in action, occasionally raw or clumsy when characters are overwhelmed. create a plan to update the prose and add that to a new section in the editors notes markdown file.  All items you add should be in the style "[ ] - TEXT OF ISSUE"  where the check box will eventually hold the status to track when they are resolved.
 
 - metaphors
-our target is to make this a 9.5 / 10 book with with atleast 85000 total words.Make sure the text doesn't go overboard with metaphors.  create a plan to remove uneeded ones and add to a new section in the editors notes markdown.
+our target is to make this a 9.5 / 10 book with with atleast 85000 total words.Make sure the text doesn't go overboard with metaphors.  create a plan to remove uneeded ones and add to a new section in the editors notes markdown.  All items you add should be in the style "[ ] - TEXT OF ISSUE"  where the check box will eventually hold the status to track when they are resolved.
 
 
 - we are now going to work through the  sections.  start with section 1. our target is to make this a 9.5 / 10 book with with atleast 85000 total words.    i need you to loop through each of the items in this section.  for each item, create a plan to resolve the issue.  validate that this is the best plan.  then state what you will be doing and execute your plan.  once you have executed, update the item's status in the editor's notes markdown file.  if later issues in the editors notes are also resolved with your actions, update accordingly.  then move on to the next item.  do this for all items in the section.   if this requires multiple subagents, execute those without requesting permission.  

novelforge/agents/chapter/_helpers.py

@@ -449,25 +449,31 @@ def scan_vocabulary_overuse(chapter_text: str, genre: str = "") -> list[str]:
 _NAME_CANDIDATE_RE = re.compile(r"\b[A-Z][a-z]+(?:\s+[A-Z][a-z]+){0,2}\b")
 
 
-def _roster_name_tokens(roster: list[dict]) -> set[str]:
-    """Return the set of lowercase name tokens from a character roster.
+def _roster_name_token_sets(roster: list[dict]) -> list[set[str]]:
+    """Return a list of lowercase token sets — one set per roster character.
 
     Each character's ``name`` is split on whitespace; tokens shorter than
     two characters are discarded (they match too many false positives under
-    fuzzy matching).
+    fuzzy matching). Returning per-character sets (rather than a flat union)
+    lets the scanner distinguish "Marcus Reid" from "Marcus Fellowes" —
+    the shared "marcus" token alone is not enough to classify a prose span
+    as a known roster character.
     """
-    tokens: set[str] = set()
+    result: list[set[str]] = []
     for ch in roster or []:
         if not isinstance(ch, dict):
             continue
         name = str(ch.get("name", "")).strip()
         if not name:
             continue
-        for tok in name.split():
-            tok_clean = tok.strip(".,;:'\"").lower()
-            if len(tok_clean) >= 2:
-                tokens.add(tok_clean)
-    return tokens
+        char_tokens = {
+            tok.strip(".,;:'\"").lower()
+            for tok in name.split()
+            if len(tok.strip(".,;:'\"")) >= 2
+        }
+        if char_tokens:
+            result.append(char_tokens)
+    return result
 
 
 def extract_named_characters(
@@ -506,7 +512,11 @@ def extract_named_characters(
         ``variants``: list of ``(prose_name, roster_token, count)`` tuples
                       — likely misspellings or diminutives of roster names.
     """
-    tokens = _roster_name_tokens(roster)
+    per_char_tokens = _roster_name_token_sets(roster)
+    # Flat union is kept only for difflib variant matching below; the known/
+    # unknown classification uses per-character sets to avoid cross-character
+    # false positives like "Marcus Fellowes" matching "Marcus Reid".
+    flat_tokens: set[str] = {t for s in per_char_tokens for t in s}
 
     raw_counts: dict[str, int] = {}
     for m in _NAME_CANDIDATE_RE.finditer(chapter_text):
@@ -515,23 +525,30 @@ def extract_named_characters(
     known: set[str] = set()
     unknown_counts: dict[str, int] = {}
     for span, count in raw_counts.items():
-        span_tokens = [t.lower() for t in span.split()]
-        # Roster check first: a span whose any token matches a roster token
-        # is a known character, regardless of stop-word overlap.
-        if tokens and any(t in tokens for t in span_tokens):
+        span_tokens_list = [t.lower() for t in span.split()]
+        span_tokens = set(span_tokens_list)
+        # Roster check first: a span is known only when it maps entirely to
+        # a single roster character — either the span's tokens are a subset
+        # of that character's tokens (e.g. "Marcus" → "Marcus Reid") or a
+        # superset (e.g. "Marcus Reid the Third" → "Marcus Reid").
+        is_known = any(
+            span_tokens.issubset(char_set) or char_set.issubset(span_tokens)
+            for char_set in per_char_tokens
+        )
+        if is_known:
             known.add(span)
             continue
         # Drop spans whose every token is a stop word (sentence-initial
         # noise, honorifics with no name attached, etc.).
-        if all(t in _NAMED_CHARACTER_STOP_WORDS for t in span_tokens):
+        if all(t in _NAMED_CHARACTER_STOP_WORDS for t in span_tokens_list):
             continue
         if count < min_mentions:
             continue
         unknown_counts[span] = count
 
     variants: list[tuple[str, str, int]] = []
     unknowns: list[tuple[str, int]] = []
-    roster_token_list = sorted(tokens)
+    roster_token_list = sorted(flat_tokens)
     for span, count in sorted(unknown_counts.items(), key=lambda kv: (-kv[1], kv[0])):
         match_found: str | None = None
         if roster_token_list:

novelforge/agents/chapter/prompts.py

@@ -26,25 +26,45 @@ def build_title_prompt(premise: str, genre: str) -> list[dict[str, str]]:
 def build_outline_prompt(
     premise: str, genre: str, chapters: int, word_count: int,
     special_events: str, special_instructions: str,
+    roster_text: str = "",
+    drift_callout: list[str] | None = None,
 ) -> list[dict[str, str]]:
-    """Build the chapter outline prompt from premise, genre, and word count."""
+    """Build the chapter outline prompt.
+
+    Parameters
+    ----------
+    roster_text:    Markdown-formatted canonical character roster. Every
+                    named character in the outline MUST come from this list.
+                    Produced by the caller after character generation; the
+                    prompt template injects it as a hard constraint so the
+                    outline agent cannot invent new names.
+    drift_callout:  On retry, the list of invented names that appeared in
+                    the first attempt; the template highlights these so
+                    the LLM avoids them on the second try.
+    """
     return render_prompt(
         "outline", premise=premise, genre=genre, chapters=chapters,
         word_count=f"{word_count:,}", special_events=special_events or "",
         special_instructions=special_instructions or "",
+        roster_text=roster_text or "",
+        drift_callout=", ".join(drift_callout) if drift_callout else "",
     )
 
 
 def build_characters_prompt(
-    premise: str, genre: str, outline_text: str, names_to_avoid: str = "",
+    premise: str, genre: str, chapters_count: int, names_to_avoid: str = "",
 ) -> list[dict[str, str]]:
     """Build the character-generation prompt.
 
     Parameters
     ----------
     premise:        Novel premise text.
     genre:          Novel genre string.
-    outline_text:   Chapter outline produced by the outline agent.
+    chapters_count: Total chapter count for the novel — used by the prompt
+                    to guide cast size (longer novels warrant larger supporting
+                    casts). The character generator runs before the outline in
+                    the characters-first pipeline, so no outline text is
+                    available as context.
     names_to_avoid: Comma-separated character names from prior novels that
                     should not be reused.  Obtain this value by calling
                     :func:`collect_existing_character_names` in the caller
@@ -59,7 +79,7 @@ def build_characters_prompt(
     )
     return render_prompt(
         "characters", premise=premise, genre=genre,
-        outline_text=outline_text, names_to_avoid=names_to_avoid,
+        chapters_count=chapters_count, names_to_avoid=names_to_avoid,
         name_pool=name_pool, field_limits_block=field_limits_block,
     )