fix: preserve read_file results during micro_compact

agents/s06_context_compact.py

@@ -11,8 +11,9 @@
             |
             v
     [Layer 1: micro_compact]        (silent, every turn)
-      Replace tool_result content older than last 3
+      Replace non-read_file tool_result content older than last 3
       with "[Previous: used {tool_name}]"
+      Keep read_file results intact as reference material
             |
             v
     [Check: tokens > 50000?]
@@ -56,6 +57,7 @@
 THRESHOLD = 50000
 TRANSCRIPT_DIR = WORKDIR / ".transcripts"
 KEEP_RECENT = 3
+PRESERVE_RESULT_TOOLS = {"read_file"}
 
 
 def estimate_tokens(messages: list) -> int:
@@ -83,13 +85,18 @@ def micro_compact(messages: list) -> list:
                 for block in content:
                     if hasattr(block, "type") and block.type == "tool_use":
                         tool_name_map[block.id] = block.name
-    # Clear old results (keep last KEEP_RECENT)
+    # Clear old results (keep last KEEP_RECENT). Preserve read_file outputs because
+    # they are reference material; compacting them into placeholders can make the
+    # agent reread the same files and loop.
     to_clear = tool_results[:-KEEP_RECENT]
     for _, _, result in to_clear:
-        if isinstance(result.get("content"), str) and len(result["content"]) > 100:
-            tool_id = result.get("tool_use_id", "")
-            tool_name = tool_name_map.get(tool_id, "unknown")
-            result["content"] = f"[Previous: used {tool_name}]"
+        if not isinstance(result.get("content"), str) or len(result["content"]) <= 100:
+            continue
+        tool_id = result.get("tool_use_id", "")
+        tool_name = tool_name_map.get(tool_id, "unknown")
+        if tool_name in PRESERVE_RESULT_TOOLS:
+            continue
+        result["content"] = f"[Previous: used {tool_name}]"
     return messages
 
 

docs/en/s06-context-compact.md

@@ -20,8 +20,9 @@ Every turn:
         |
         v
 [Layer 1: micro_compact]        (silent, every turn)
-  Replace tool_result > 3 turns old
+  Replace older non-read_file tool results
   with "[Previous: used {tool_name}]"
+  Keep read_file results intact as reference material
         |
         v
 [Check: tokens > 50000?]
@@ -42,19 +43,18 @@ continue    [Layer 2: auto_compact]
 
 ## How It Works
 
-1. **Layer 1 -- micro_compact**: Before each LLM call, replace old tool results with placeholders.
+1. **Layer 1 -- micro_compact**: Before each LLM call, replace old non-`read_file` tool results with placeholders while preserving prior `read_file` output as reference material.
 
 ```python
 def micro_compact(messages: list) -> list:
     tool_results = []
-    for i, msg in enumerate(messages):
-        if msg["role"] == "user" and isinstance(msg.get("content"), list):
-            for j, part in enumerate(msg["content"]):
-                if isinstance(part, dict) and part.get("type") == "tool_result":
-                    tool_results.append((i, j, part))
+    # ... collect tool_results and build tool_name_map ...
     if len(tool_results) <= KEEP_RECENT:
         return messages
     for _, _, part in tool_results[:-KEEP_RECENT]:
+        tool_name = tool_name_map.get(part.get("tool_use_id", ""), "unknown")
+        if tool_name == "read_file":
+            continue
         if len(part.get("content", "")) > 100:
             part["content"] = f"[Previous: used {tool_name}]"
     return messages
@@ -107,7 +107,7 @@ Transcripts preserve full history on disk. Nothing is truly lost -- just moved o
 |----------------|------------------|----------------------------|
 | Tools          | 5                | 5 (base + compact)         |
 | Context mgmt   | None             | Three-layer compression    |
-| Micro-compact  | None             | Old results -> placeholders|
+| Micro-compact  | None             | Old non-read results -> placeholders; keep file context |
 | Auto-compact   | None             | Token threshold trigger    |
 | Transcripts    | None             | Saved to .transcripts/     |
 
@@ -118,6 +118,6 @@ cd learn-claude-code
 python agents/s06_context_compact.py
 ```
 
-1. `Read every Python file in the agents/ directory one by one` (watch micro-compact replace old results)
+1. `Read every Python file in the agents/ directory one by one` (watch micro-compact keep older `read_file` context while shrinking older command output)
 2. `Keep reading files until compression triggers automatically`
 3. `Use the compact tool to manually compress the conversation`

docs/ja/s06-context-compact.md

@@ -20,8 +20,9 @@ Every turn:
         |
         v
 [Layer 1: micro_compact]        (silent, every turn)
-  Replace tool_result > 3 turns old
+  Replace older non-read_file tool results
   with "[Previous: used {tool_name}]"
+  Keep read_file results intact as reference material
         |
         v
 [Check: tokens > 50000?]
@@ -42,19 +43,18 @@ continue    [Layer 2: auto_compact]
 
 ## 仕組み
 
-1. **第1層 -- micro_compact**: 各LLM呼び出しの前に、古いツール結果をプレースホルダーに置換する。
+1. **第1層 -- micro_compact**: 各LLM呼び出しの前に、古い非 `read_file` ツール結果をプレースホルダーに置換しつつ、以前の `read_file` 出力は参照材料として保持する。
 
 ```python
 def micro_compact(messages: list) -> list:
     tool_results = []
-    for i, msg in enumerate(messages):
-        if msg["role"] == "user" and isinstance(msg.get("content"), list):
-            for j, part in enumerate(msg["content"]):
-                if isinstance(part, dict) and part.get("type") == "tool_result":
-                    tool_results.append((i, j, part))
+    # ... tool_results を集め、tool_name_map を構築する ...
     if len(tool_results) <= KEEP_RECENT:
         return messages
     for _, _, part in tool_results[:-KEEP_RECENT]:
+        tool_name = tool_name_map.get(part.get("tool_use_id", ""), "unknown")
+        if tool_name == "read_file":
+            continue
         if len(part.get("content", "")) > 100:
             part["content"] = f"[Previous: used {tool_name}]"
     return messages
@@ -107,7 +107,7 @@ def agent_loop(messages: list):
 |----------------|------------------|----------------------------|
 | Tools          | 5                | 5 (base + compact)         |
 | Context mgmt   | None             | Three-layer compression    |
-| Micro-compact  | None             | Old results -> placeholders|
+| Micro-compact  | None             | 古い非 read_file 結果 -> プレースホルダー; ファイル文脈は保持 |
 | Auto-compact   | None             | Token threshold trigger    |
 | Transcripts    | None             | Saved to .transcripts/     |
 
@@ -118,6 +118,6 @@ cd learn-claude-code
 python agents/s06_context_compact.py
 ```
 
-1. `Read every Python file in the agents/ directory one by one` (micro-compactが古い結果を置換するのを観察する)
+1. `Read every Python file in the agents/ directory one by one` (micro-compact が古い `read_file` 文脈を保持しつつ、より古いコマンド出力を縮める様子を観察する)
 2. `Keep reading files until compression triggers automatically`
 3. `Use the compact tool to manually compress the conversation`

docs/zh/s06-context-compact.md

@@ -20,8 +20,9 @@ Every turn:
         |
         v
 [Layer 1: micro_compact]        (silent, every turn)
-  Replace tool_result > 3 turns old
+  Replace older non-read_file tool results
   with "[Previous: used {tool_name}]"
+  Keep read_file results intact as reference material
         |
         v
 [Check: tokens > 50000?]
@@ -42,19 +43,18 @@ continue    [Layer 2: auto_compact]
 
 ## 工作原理
 
-1. **第一层 -- micro_compact**: 每次 LLM 调用前, 将旧的 tool result 替换为占位符。
+1. **第一层 -- micro_compact**: 每次 LLM 调用前, 将旧的非 `read_file` tool result 替换为占位符, 同时保留先前的 `read_file` 输出作为参考材料。
 
 ```python
 def micro_compact(messages: list) -> list:
     tool_results = []
-    for i, msg in enumerate(messages):
-        if msg["role"] == "user" and isinstance(msg.get("content"), list):
-            for j, part in enumerate(msg["content"]):
-                if isinstance(part, dict) and part.get("type") == "tool_result":
-                    tool_results.append((i, j, part))
+    # ... 收集 tool_results 并构建 tool_name_map ...
     if len(tool_results) <= KEEP_RECENT:
         return messages
     for _, _, part in tool_results[:-KEEP_RECENT]:
+        tool_name = tool_name_map.get(part.get("tool_use_id", ""), "unknown")
+        if tool_name == "read_file":
+            continue
         if len(part.get("content", "")) > 100:
             part["content"] = f"[Previous: used {tool_name}]"
     return messages
@@ -107,7 +107,7 @@ def agent_loop(messages: list):
 |----------------|------------------|--------------------------------|
 | Tools          | 5                | 5 (基础 + compact)             |
 | 上下文管理     | 无               | 三层压缩                       |
-| Micro-compact  | 无               | 旧结果 -> 占位符               |
+| Micro-compact  | 无               | 旧的非 read_file 结果 -> 占位符；保留文件上下文 |
 | Auto-compact   | 无               | token 阈值触发                 |
 | Transcripts    | 无               | 保存到 .transcripts/           |
 
@@ -120,6 +120,6 @@ python agents/s06_context_compact.py
 
 试试这些 prompt (英文 prompt 对 LLM 效果更好, 也可以用中文):
 
-1. `Read every Python file in the agents/ directory one by one` (观察 micro-compact 替换旧结果)
+1. `Read every Python file in the agents/ directory one by one` (观察 micro-compact 保留旧的 `read_file` 上下文, 同时压缩更旧的命令输出)
 2. `Keep reading files until compression triggers automatically`
 3. `Use the compact tool to manually compress the conversation`

tests/test_s06_micro_compact.py

@@ -0,0 +1,77 @@
+import ast
+import types
+import unittest
+from pathlib import Path
+
+
+def load_micro_compact():
+    source = Path("agents/s06_context_compact.py").read_text()
+    tree = ast.parse(source)
+    selected = []
+    for node in tree.body:
+        if isinstance(node, ast.Assign):
+            for target in node.targets:
+                if isinstance(target, ast.Name) and target.id in {"KEEP_RECENT", "PRESERVE_RESULT_TOOLS"}:
+                    selected.append(node)
+                    break
+        elif isinstance(node, ast.FunctionDef) and node.name == "micro_compact":
+            selected.append(node)
+    module = ast.Module(body=selected, type_ignores=[])
+    ast.fix_missing_locations(module)
+    namespace = {}
+    exec(compile(module, "micro_compact_subset", "exec"), namespace)
+    return namespace["micro_compact"], namespace["KEEP_RECENT"]
+
+
+def make_tool_use(tool_use_id: str, name: str):
+    return types.SimpleNamespace(type="tool_use", id=tool_use_id, name=name)
+
+
+class MicroCompactTests(unittest.TestCase):
+    def setUp(self):
+        self.micro_compact, self.keep_recent = load_micro_compact()
+
+    def test_preserves_old_read_file_results_but_compacts_old_command_output(self):
+        long_read = "A" * 200
+        long_bash = "B" * 200
+        recent_1 = "C" * 200
+        recent_2 = "D" * 200
+        recent_3 = "E" * 200
+        messages = [
+            {"role": "assistant", "content": [make_tool_use("u1", "read_file")]},
+            {"role": "user", "content": [{"type": "tool_result", "tool_use_id": "u1", "content": long_read}]},
+            {"role": "assistant", "content": [make_tool_use("u2", "bash")]},
+            {"role": "user", "content": [{"type": "tool_result", "tool_use_id": "u2", "content": long_bash}]},
+            {"role": "assistant", "content": [make_tool_use("u3", "bash")]},
+            {"role": "user", "content": [{"type": "tool_result", "tool_use_id": "u3", "content": recent_1}]},
+            {"role": "assistant", "content": [make_tool_use("u4", "read_file")]},
+            {"role": "user", "content": [{"type": "tool_result", "tool_use_id": "u4", "content": recent_2}]},
+            {"role": "assistant", "content": [make_tool_use("u5", "edit_file")]},
+            {"role": "user", "content": [{"type": "tool_result", "tool_use_id": "u5", "content": recent_3}]},
+        ]
+
+        self.micro_compact(messages)
+
+        self.assertEqual(messages[1]["content"][0]["content"], long_read)
+        self.assertEqual(messages[3]["content"][0]["content"], "[Previous: used bash]")
+        self.assertEqual(messages[5]["content"][0]["content"], recent_1)
+        self.assertEqual(messages[7]["content"][0]["content"], recent_2)
+        self.assertEqual(messages[9]["content"][0]["content"], recent_3)
+
+    def test_does_nothing_when_tool_results_do_not_exceed_keep_recent(self):
+        messages = [
+            {"role": "assistant", "content": [make_tool_use("u1", "read_file")]},
+            {"role": "user", "content": [{"type": "tool_result", "tool_use_id": "u1", "content": "A" * 200}]},
+            {"role": "assistant", "content": [make_tool_use("u2", "bash")]},
+            {"role": "user", "content": [{"type": "tool_result", "tool_use_id": "u2", "content": "B" * 200}]},
+            {"role": "assistant", "content": [make_tool_use("u3", "edit_file")]},
+            {"role": "user", "content": [{"type": "tool_result", "tool_use_id": "u3", "content": "C" * 200}]},
+        ]
+        before = repr(messages)
+        self.micro_compact(messages)
+        self.assertEqual(repr(messages), before)
+        self.assertEqual(self.keep_recent, 3)
+
+
+if __name__ == "__main__":
+    unittest.main()