feat: add output truncation and offset paging to all large-output tools

- coding_tools: list_files (max_entries=200, sorted), run_command
  (max_chars=8000, output_offset), grep_search (max_lines=100, offset);
  added _truncate_output helper for run_command paging
- pr_tools: get_pull_request_diff / get_pull_request_patch now truncate
  at 15k chars with hunk-boundary awareness and char_offset paging;
  get_pull_request_comments gains limit/offset paging
- issue_tools: get_issue_comments gains limit/offset paging
- file_tools: get_file_content / get_file_content_with_ref now paginate
  by line (offset=1, limit=250), matching existing read_file convention;
  added _paginate_lines helper
- research_tools: fetch_url gains char_offset parameter; _smart_truncate
  now slices from an offset and embeds next char_offset in the footer
- gitea_tools facade: all new params threaded through

gitea/tools/research_tools.py

@@ -42,20 +42,34 @@ class ResearchTools:
     # Internal helpers
     # ------------------------------------------------------------------ #
 
-    def _smart_truncate(self, text: str, max_chars: int = _MAX_CONTENT_CHARS) -> str:
-        """Truncate at a paragraph boundary to preserve coherence.
+    def _smart_truncate(
+        self,
+        text: str,
+        max_chars: int = _MAX_CONTENT_CHARS,
+        char_offset: int = 0,
+    ) -> str:
+        """Slice [char_offset : char_offset+max_chars] and truncate at a paragraph boundary.
 
-        Prefers cutting at a blank-line paragraph boundary rather than
-        mid-sentence so the LLM receives a coherent chunk.
+        Prefers cutting at a blank-line paragraph boundary rather than mid-sentence
+        so the LLM receives a coherent chunk.
         """
-        if len(text) <= max_chars:
-            return text
-        truncated = text[:max_chars]
-        last_para: int = truncated.rfind("\n\n")
-        tail = "\n\n[Content truncated — use fetch_url with a more specific URL or anchor]"
-        if last_para > int(max_chars * 0.7):
-            return truncated[:last_para] + tail
-        return truncated + tail
+        total: int = len(text)
+        chunk: str = text[char_offset : char_offset + max_chars]
+        if char_offset == 0 and len(chunk) <= max_chars and total <= max_chars:
+            return text  # Common fast-path: content fits entirely
+        if len(chunk) >= max_chars:
+            last_para: int = chunk.rfind("\n\n")
+            if last_para > int(max_chars * 0.7):
+                chunk = chunk[:last_para]
+        next_offset: int = char_offset + len(chunk)
+        if next_offset < total:
+            tail = (
+                f"\n\n[Content truncated — {total} chars total. "
+                f"Showing chars {char_offset}–{next_offset}. "
+                f"Re-call fetch_url with char_offset={next_offset} to read more.]"
+            )
+            return chunk + tail
+        return chunk
 
     def _html_to_markdown(self, html: str) -> str:
         """Convert HTML to clean Markdown.
@@ -352,6 +366,7 @@ class ResearchTools:
         url: str,
         extract_text: bool = True,
         max_chars: int = _MAX_CONTENT_CHARS,
+        char_offset: int = 0,
     ) -> str:
         """Fetch the content of a URL and return it as clean, readable Markdown.
 
@@ -371,8 +386,11 @@ class ResearchTools:
             extract_text: If True (default), extract and clean the main content
                           as Markdown, stripping navigation, ads, and boilerplate.
                           Set to False to get raw HTML/JSON (useful for schemas).
-            max_chars: Maximum characters to return (default 20,000).
+            max_chars: Maximum characters to return per call (default 20,000).
                        The tool cuts at a paragraph boundary when truncating.
+            char_offset: Character offset into the extracted content to start
+                         reading from (default 0). Increment by max_chars to
+                         page through content larger than max_chars.
 
         Returns:
             Clean Markdown text of the main content (HTML pages), pretty-printed
@@ -425,7 +443,7 @@ class ResearchTools:
             else:
                 text = raw
 
-            return self._smart_truncate(text, max_chars)
+            return self._smart_truncate(text, max_chars, char_offset)
 
         except httpx.HTTPStatusError as exc:
             return (