fix(tools): PDFの会話間継続をトークン単独で再開可能にする

長いPDFを複数会話に分けて取得する際、モデルがurl_or_pathを省略したり、過去の添付IDや古い継続トークンを組み合わせたりして、先頭または以前のページへ戻る問題を改善する。PDFカーソル自体は正しく進んでいたため、汎用LLM層やセッション形式を変更せず、pdf_to_textの入出力契約だけを明確化する。

- continue_fromを自己完結した再開情報として扱い、PDFソース、ページ番号、ページ内文字位置をトークンから復元

- 継続時はurl_or_pathを省略可能にし、競合する値が渡されてもトークン内のソースを優先

- Ollama生成スキーマでもurl_or_pathを任意にして、continue_from単独のツール呼び出しを許可

- 未完了結果へ最新トークン、旧トークンの置換、単独再開、上限時の保持を表すPDF専用メタデータを追加

- 継続メタデータを固定キーと真偽値で表現し、会話言語に依存しない形式へ統一

- 無効トークン時は先頭再試行を禁止し、直前の最新トークンを再利用する機械可読情報を返却

- メタデータをpage-info側へ配置して本文連結を汚さず、結果全体を6000文字以内に再調整

- llm.py、チャット処理、セッションJSON、scrape_url、search_and_fetch、MCPツールは変更しない

検証:

- pytest: 132 passed

- ruff check: passed

- mypy src: passed

- 1463891.1463912.pdf: 12ページ、52577文字をcontinue_from単独の10回で完全再構成

- 指定PDFの最大ツール結果長はメタデータ込み5992文字

3 files changed, 125 insertions, 27 deletions

diff --git a/docs/tools-schema.md b/docs/tools-schema.md
index 874640a..5b920a4 100644
--- a/docs/tools-schema.md
+++ b/docs/tools-schema.md
@@ -120,13 +120,13 @@ next_token: {token}
 
 | パラメータ | 型 | 必須 | デフォルト | 説明 |
 |------------|----|------|------------|------|
-| `url_or_path` | `str` | ✅ | — | HTTP(S) PDF URL。チャット内では現在のセッションの `attachment:<file_id>` も利用可能。任意のローカルパスは拒否 |
+| `url_or_path` | `str \| None` | 初回のみ | `None` | 初回取得用のHTTP(S) PDF URL。チャット内では現在のセッションの `attachment:<file_id>` も利用可能。`continue_from` 指定時はトークンから復元されるため省略可能。任意のローカルパスは拒否 |
 | `max_chars` | `int \| None` | ❌ | [`settings.fetch_max_chars`](src/ollama_web/config.py:52) (=8000) | 1ページあたりの返却文字数上限 |
-| `continue_from` | `str \| None` | ❌ | `None` | 前回戻り値の `next_token`。指定時は続きページを取得 |
+| `continue_from` | `str \| None` | ❌ | `None` | 同じPDFの直前の戻り値にある `next_token`。PDFソースと次の開始位置を復元して続きページを取得 |
 
 ### 戻り値
 
-`str` — 1ページ分のテキスト、またはエラーメッセージ。続きがある場合は `scrape_url` と同様のページ情報フッターが付与される。
+`str` — 1ページ分のテキスト、またはエラーメッセージ。続きがある場合はページ情報フッターに加え、最新トークンが以前の同一PDF用トークンを置き換えること、上限到達時は最終回答へそのまま残すこと、次回は `continue_from` 単独で再開できることを示すPDF専用メタデータが付与される。このメタデータは会話言語に依存しない固定キーと真偽値で表現し、本文の連結結果には含まれない。
 
 ### 制約
 
@@ -139,10 +139,10 @@ next_token: {token}
 
 ### 処理フロー
 
-1. `continue_from` があればトークンから `(url, page_index, page_char_offset)` を復元
+1. `continue_from` があれば、`url_or_path` の有無や値にかかわらずトークンから `(url, page_index, page_char_offset)` を復元
 2. PDF を全件取得・解析
 3. カーソル位置から、結果全体の上限内で実際に返せる文字数だけ抽出
-4. 未返却文字または残りページがあれば、その直後を指す `next_token` を付与
+4. 未返却文字または残りページがあれば、その直後を指す `next_token` とPDF専用の再開指示を付与
 
 ---
 
diff --git a/src/ollama_web/tools/pdf.py b/src/ollama_web/tools/pdf.py
index 4faf8ac..657edf9 100644
--- a/src/ollama_web/tools/pdf.py
+++ b/src/ollama_web/tools/pdf.py
@@ -12,6 +12,7 @@ from .helper.pagination import (
     PageSizeError,
     format_page_to_limit,
     make_token,
+    parse_paged_result,
     parse_token,
 )
 from .helper.pdf import extract_pdf_page
@@ -23,6 +24,16 @@ logger = logging.getLogger("ollama_web.tools.pdf")
 
 AttachmentLoader = Callable[[str], bytes | None]
 
+_RESUME_RULE = (
+    "\nresume_tool: pdf_to_text"
+    "\nresume_page_index: {page_index}"
+    "\nresume_page_char_offset: {page_char_offset}"
+    "\nnext_token_is_latest: true"
+    "\nsupersedes_previous_tokens: true"
+    "\nresume_with_continue_from_only: true"
+    "\npreserve_next_token_on_tool_limit: true"
+)
+
 
 def _cursor_value(payload: dict[str, Any], key: str, fallback: int = 0) -> int:
     raw = payload.get(key, fallback)
@@ -38,26 +49,32 @@ def _cursor_value(payload: dict[str, Any], key: str, fallback: int = 0) -> int:
 
 
 def _pdf_to_text(
-    url_or_path: str,
+    url_or_path: str | None,
     max_chars: int | None,
     continue_from: str | None,
     attachment_loader: AttachmentLoader | None,
 ) -> str:
-    if not url_or_path or not url_or_path.strip():
-        return "Invalid PDF source."
-
-    src = url_or_path.strip()
+    src = (url_or_path or "").strip()
     page_index = 0
     page_char_offset = 0
     if continue_from:
         try:
             payload = parse_token(continue_from, _TOOL_NAME)
-            src = str(payload.get("url") or src)
+            # A continuation token is self-contained. Its source is
+            # authoritative so a later conversation cannot accidentally pair
+            # the cursor with a stale attachment ID or URL.
+            src = str(payload["url"])
             # Version 1 used offset as the PDF page index.
             page_index = _cursor_value(payload, "page_index", int(payload["offset"]))
             page_char_offset = _cursor_value(payload, "page_char_offset")
         except ValueError as exc:
-            return f"Invalid continuation token: {exc}"
+            return (
+                f"Invalid continuation token: {exc}\n"
+                "restart_from_beginning: false\n"
+                "required_action: reuse_latest_next_token"
+            )
+    elif not src:
+        return "Invalid PDF source."
 
     if src.startswith(_ATTACHMENT_PREFIX):
         file_id = src.removeprefix(_ATTACHMENT_PREFIX)
@@ -98,12 +115,33 @@ def _pdf_to_text(
         )
 
     try:
-        result = format_page_to_limit(
-            page_factory,
-            token_factory,
-            requested_chars=requested,
-            result_limit=settings.max_tool_result_chars,
-        )
+        result_limit = settings.max_tool_result_chars
+        render_limit = result_limit
+        result = ""
+        for _ in range(4):
+            result = format_page_to_limit(
+                page_factory,
+                token_factory,
+                requested_chars=requested,
+                result_limit=render_limit,
+            )
+            _, has_more, next_token = parse_paged_result(result)
+            if not has_more or next_token is None:
+                break
+            next_payload = parse_token(next_token, _TOOL_NAME)
+            resume_metadata = _RESUME_RULE.format(
+                page_index=_cursor_value(next_payload, "page_index", next_payload["offset"]),
+                page_char_offset=_cursor_value(next_payload, "page_char_offset"),
+            )
+            overflow = len(result) + len(resume_metadata) - result_limit
+            if overflow <= 0:
+                result += resume_metadata
+                break
+            render_limit -= overflow + 8
+            if render_limit <= 0:
+                raise PageSizeError("PDF continuation metadata exceeds the result limit.")
+        else:
+            raise PageSizeError("PDF continuation metadata could not fit the result limit.")
     except PageSizeError as exc:
         return f"PDF pagination failed: {exc}"
 
@@ -119,17 +157,21 @@ def _pdf_to_text(
 
 
 def pdf_to_text(
-    url_or_path: str, max_chars: int | None = None, continue_from: str | None = None
+    url_or_path: str | None = None,
+    max_chars: int | None = None,
+    continue_from: str | None = None,
 ) -> str:
     """Extract a losslessly paged PDF from an HTTP(S) URL.
 
     Args:
-      url_or_path: HTTP(S) PDF URL. In chat, a session attachment URI may also
-        be supplied by the attachment context. Arbitrary filesystem paths are
-        never accepted.
+      url_or_path: HTTP(S) PDF URL for the first call. In chat, a session
+        attachment URI may also be supplied by the attachment context.
+        Omit this argument when continue_from is supplied. Arbitrary
+        filesystem paths are never accepted.
       max_chars: Requested maximum source characters for this page. The tool
         may return fewer characters so metadata and its token remain intact.
-      continue_from: Opaque continuation token from the preceding result.
+      continue_from: Opaque, self-contained continuation token from the most
+        recent result for this PDF. It restores both the source and cursor.
     """
     return _pdf_to_text(url_or_path, max_chars, continue_from, None)
 
@@ -138,7 +180,7 @@ def make_session_pdf_tool(attachment_loader: AttachmentLoader) -> Callable[...,
     """Create a session-scoped ``pdf_to_text`` supporting attachment IDs."""
 
     def session_pdf_to_text(
-        url_or_path: str,
+        url_or_path: str | None = None,
         max_chars: int | None = None,
         continue_from: str | None = None,
     ) -> str:
diff --git a/tests/test_tools.py b/tests/test_tools.py
index 95f6514..4960d1d 100644
--- a/tests/test_tools.py
+++ b/tests/test_tools.py
@@ -312,6 +312,8 @@ def test_pdf_to_text_continue_from_invalid_token():
 
     result = pdf_to_text("https://example.com/doc.pdf", continue_from="bad-token")
     assert "Invalid continuation token" in result
+    assert "restart_from_beginning: false" in result
+    assert "required_action: reuse_latest_next_token" in result
 
 
 def test_pdf_to_text_rejects_negative_page_character_cursor():
@@ -351,16 +353,67 @@ def test_pdf_to_text_pagination_roundtrip(monkeypatch):
     token = None
     for _ in range(50):
         result = pdf_mod.pdf_to_text(
-            "https://example.com/doc.pdf", max_chars=8000, continue_from=token
+            "https://example.com/doc.pdf" if token is None else "",
+            max_chars=8000,
+            continue_from=token,
         )
         assert len(result) <= 6000
         body_text, has_more, token = parse_paged_result(result)
+        assert "next_token_is_latest:" not in body_text
+        if has_more:
+            assert "resume_tool: pdf_to_text" in result
+            assert "next_token_is_latest: true" in result
+            assert "supersedes_previous_tokens: true" in result
+            assert "resume_with_continue_from_only: true" in result
+            assert "preserve_next_token_on_tool_limit: true" in result
+        else:
+            assert "resume_tool:" not in result
         collected += body_text
         if not has_more:
             break
     assert collected == expected
 
 
+def test_pdf_to_text_continuation_token_source_is_authoritative(monkeypatch):
+    """A continuation restores its URL and does not require url_or_path."""
+    from ollama_web.tools import pdf as pdf_mod
+
+    data = _make_multipage_pdf(["SOURCE-" + "S" * 9000])
+    fetched: list[str] = []
+
+    def fake_fetch(url: str, timeout=None):
+        fetched.append(url)
+        return data, url, "application/pdf"
+
+    monkeypatch.setattr(pdf_mod, "fetch_public_bytes", fake_fetch)
+    first = pdf_mod.pdf_to_text("https://example.com/source.pdf")
+    _, has_more, token = parse_paged_result(first)
+    assert has_more is True
+    assert token is not None
+
+    resumed = pdf_mod.pdf_to_text(continue_from=token)
+    body, _, _ = parse_paged_result(resumed)
+    assert body
+    assert fetched == [
+        "https://example.com/source.pdf",
+        "https://example.com/source.pdf",
+    ]
+
+    # Even a conflicting argument cannot redirect a valid continuation.
+    pdf_mod.pdf_to_text("https://attacker.example/wrong.pdf", continue_from=token)
+    assert fetched[-1] == "https://example.com/source.pdf"
+
+
+def test_pdf_to_text_ollama_schema_allows_token_only_continuation():
+    from ollama._utils import convert_function_to_tool
+
+    from ollama_web.tools.pdf import pdf_to_text
+
+    schema = convert_function_to_tool(pdf_to_text).function.parameters
+    assert "url_or_path" not in (schema.required or [])
+    assert "continue_from" not in (schema.required or [])
+
+
 def test_session_pdf_tool_reads_original_bytes_and_is_session_scoped(tmp_path):
     from pypdf import PdfReader
 
@@ -386,7 +439,10 @@ def test_session_pdf_tool_reads_original_bytes_and_is_session_scoped(tmp_path):
     token = None
     collected = ""
     for _ in range(10):
-        result = tool(f"attachment:{chat_file.id}", continue_from=token)
+        result = tool(
+            f"attachment:{chat_file.id}" if token is None else "",
+            continue_from=token,
+        )
         body, has_more, token = parse_paged_result(result)
         collected += body
         if not has_more:
@@ -491,7 +547,7 @@ def test_attachment_context_keeps_pdf_continuation_and_registry_can_resume(tmp_p
     token = first_result.split("next_token:", 1)[1].splitlines()[0].strip()
     resumed = registry.execute(
         "pdf_to_text",
-        {"url_or_path": f"attachment:{chat_file.id}", "continue_from": token},
+        {"continue_from": token},
     )
     assert "unavailable" not in resumed.lower()