fix(s08): recognize Anthropic prompt overflow errors

s08_context_compact/README.ja.md

@@ -20,7 +20,7 @@ Agent が作業を続けると、読み込んだファイル、コマンド結
 
 コンテキストウィンドウは、モデルが現在使っている下書き用紙と考えられます。ユーザーメッセージ、モデルの応答、`tool_use`、`tool_result` が順番に書き込まれます。モデルはタスクを続けるたびに、その内容を読み直します。
 
-下書き用紙の大きさは固定です。上限を超えると API はリクエストを拒否し、`prompt_too_long` を返します。コーディングタスクでは、ツール結果が多くの領域を占めます。
+下書き用紙の大きさは固定です。上限を超えると API はリクエストを拒否し、`prompt is too long` を返します。コーディングタスクでは、ツール結果が多くの領域を占めます。
 
 - 長いファイルを読むと、その内容がコンテキストに入ります。
 - テストやビルドのログは、一度に数十 KB 追加されることがあります。
@@ -202,7 +202,7 @@ if self.estimate_chars(messages) > self.CONTEXT_CHAR_LIMIT:
 
 ## API に拒否された後の回復
 
-文字数はモデルが使う token 数の推定値です。そのため API が `prompt_too_long` を返す可能性は残ります。`reactive_compact` は transcript を保存し、古い履歴を要約して、最新 5 メッセージを保持します。
+文字数はモデルが使う token 数の推定値です。そのため API が `prompt is too long` を返す可能性は残ります。`reactive_compact` は transcript を保存し、古い履歴を要約して、最新 5 メッセージを保持します。
 
 ```python
 tail_start = max(0, len(messages) - self.KEEP_RECENT_MESSAGES)
@@ -236,6 +236,7 @@ def agent_loop(messages, active_request):
         except Exception as error:
             message = str(error).lower()
             too_long = ("prompt_too_long" in message
+                        or "prompt is too long" in message
                         or "too many tokens" in message)
             if too_long and reactive_retries < MAX_REACTIVE_RETRIES:
                 messages[:] = COMPACTOR.reactive_compact(

s08_context_compact/README.md

@@ -20,7 +20,7 @@ This lesson adds a four-step compaction pipeline. It first reduces recoverable t
 
 Think of the context window as the model's current scratchpad. User messages, model responses, `tool_use`, and `tool_result` blocks are written onto it in order. The model reads that material again whenever it continues the task.
 
-The scratchpad has a fixed size. When a request exceeds it, the API rejects the call with `prompt_too_long`. Tool results usually consume most of the space in coding tasks:
+The scratchpad has a fixed size. When a request exceeds it, the API rejects the call with `prompt is too long`. Tool results usually consume most of the space in coding tasks:
 
 - Reading a long file puts its contents into the context.
 - Test and build logs can add tens of kilobytes at once.
@@ -202,7 +202,7 @@ Each round therefore starts with the lowest-cost operation whose information is
 
 ## Recovering From an API Rejection
 
-A character count can only estimate the tokens used by a model. The API may still return `prompt_too_long`. `reactive_compact` saves a transcript, summarizes older history, and retains the latest 5 messages:
+A character count can only estimate the tokens used by a model. The API may still return `prompt is too long`. `reactive_compact` saves a transcript, summarizes older history, and retains the latest 5 messages:
 
 ```python
 tail_start = max(0, len(messages) - self.KEEP_RECENT_MESSAGES)
@@ -236,6 +236,7 @@ def agent_loop(messages, active_request):
         except Exception as error:
             message = str(error).lower()
             too_long = ("prompt_too_long" in message
+                        or "prompt is too long" in message
                         or "too many tokens" in message)
             if too_long and reactive_retries < MAX_REACTIVE_RETRIES:
                 messages[:] = COMPACTOR.reactive_compact(

s08_context_compact/README.zh.md

@@ -20,7 +20,7 @@ Agent 持续工作时，读过的文件、执行过的命令和模型回复都
 
 可以把上下文窗口看作模型当前使用的一张草稿纸。用户消息、模型回复、`tool_use` 和 `tool_result` 都会按顺序写在这张纸上。模型每次继续工作时，都要重新读取这些内容。
 
-草稿纸的大小固定。内容超过上限后，API 会拒绝请求并返回 `prompt_too_long`。在代码任务里，工具结果通常占据最多空间：
+草稿纸的大小固定。内容超过上限后，API 会拒绝请求并返回 `prompt is too long`。在代码任务里，工具结果通常占据最多空间：
 
 - 读取一个长文件会把文件内容放进上下文；
 - 测试和构建日志可能一次产生几十 KB 文本；
@@ -202,7 +202,7 @@ if self.estimate_chars(messages) > self.CONTEXT_CHAR_LIMIT:
 
 ## API 拒绝后的补救
 
-字符数只能估算模型实际使用的 token。API 仍可能返回 `prompt_too_long`。`reactive_compact` 会保存 transcript，总结较早历史，并保留最近 5 条消息：
+字符数只能估算模型实际使用的 token。API 仍可能返回 `prompt is too long`。`reactive_compact` 会保存 transcript，总结较早历史，并保留最近 5 条消息：
 
 ```python
 tail_start = max(0, len(messages) - self.KEEP_RECENT_MESSAGES)
@@ -236,6 +236,7 @@ def agent_loop(messages, active_request):
         except Exception as error:
             message = str(error).lower()
             too_long = ("prompt_too_long" in message
+                        or "prompt is too long" in message
                         or "too many tokens" in message)
             if too_long and reactive_retries < MAX_REACTIVE_RETRIES:
                 messages[:] = COMPACTOR.reactive_compact(

s08_context_compact/code.py

@@ -536,7 +536,8 @@ def agent_loop(messages: list, active_request: str):
             reactive_retries = 0
         except Exception as error:
             too_long = any(text in str(error).lower()
-                           for text in ("prompt_too_long", "too many tokens"))
+                           for text in ("prompt_too_long", "prompt is too long",
+                                        "too many tokens"))
             if too_long and reactive_retries < MAX_REACTIVE_RETRIES:
                 print("[reactive compact]")
                 messages[:] = COMPACTOR.reactive_compact(messages, active_request)

tests/test_s08_context_compact.py

@@ -2,6 +2,9 @@
 import sys
 import types
 from pathlib import Path
+from unittest.mock import Mock
+
+import pytest
 
 
 ROOT = Path(__file__).resolve().parents[1]
@@ -136,3 +139,71 @@ def test_prepare_persists_oversized_unseen_result_before_full_compact(
     saved_line = next(line for line in content.splitlines()
                       if line.startswith("Full output: "))
     assert Path(saved_line.removeprefix("Full output: ")).read_text() == output
+
+
+@pytest.mark.parametrize("error_message", [
+    "prompt is too long: 210445 tokens > 200000 maximum",
+    "prompt_too_long",
+    "too many tokens",
+])
+def test_agent_loop_compacts_and_retries_context_overflow(
+        tmp_path, monkeypatch, error_message):
+    lesson = load_lesson(monkeypatch, tmp_path)
+    messages = [{"role": "user", "content": "continue"}]
+    compacted = [{"role": "user", "content": "compacted history"}]
+    requests = []
+    response = types.SimpleNamespace(
+        content=[types.SimpleNamespace(type="text", text="Recovered")])
+
+    def create(**kwargs):
+        requests.append(list(kwargs["messages"]))
+        if len(requests) == 1:
+            raise RuntimeError(error_message)
+        return response
+
+    compact = Mock(return_value=compacted)
+    monkeypatch.setattr(lesson["client"].messages, "create", create)
+    monkeypatch.setattr(lesson["COMPACTOR"], "reactive_compact", compact)
+
+    lesson["agent_loop"](messages, "continue")
+
+    assert len(requests) == 2
+    assert requests[1] == compacted
+    compact.assert_called_once()
+    assert compact.call_args.args[1] == "continue"
+    assert messages == [*compacted,
+                        {"role": "assistant", "content": response.content}]
+
+
+def test_agent_loop_stops_after_one_reactive_retry(tmp_path, monkeypatch):
+    lesson = load_lesson(monkeypatch, tmp_path)
+    messages = [{"role": "user", "content": "continue"}]
+    first_error = RuntimeError("prompt is too long: first request")
+    retry_error = RuntimeError("prompt is too long: retry")
+    create = Mock(side_effect=[first_error, retry_error])
+    compact = Mock(return_value=list(messages))
+    monkeypatch.setattr(lesson["client"].messages, "create", create)
+    monkeypatch.setattr(lesson["COMPACTOR"], "reactive_compact", compact)
+
+    with pytest.raises(RuntimeError) as caught:
+        lesson["agent_loop"](messages, "continue")
+
+    assert caught.value is retry_error
+    assert create.call_count == 2
+    compact.assert_called_once()
+
+
+def test_agent_loop_propagates_unrelated_errors(tmp_path, monkeypatch):
+    lesson = load_lesson(monkeypatch, tmp_path)
+    error = RuntimeError("invalid API key")
+    create = Mock(side_effect=error)
+    compact = Mock()
+    monkeypatch.setattr(lesson["client"].messages, "create", create)
+    monkeypatch.setattr(lesson["COMPACTOR"], "reactive_compact", compact)
+
+    with pytest.raises(RuntimeError) as caught:
+        lesson["agent_loop"]([{"role": "user", "content": "continue"}], "continue")
+
+    assert caught.value is error
+    create.assert_called_once()
+    compact.assert_not_called()