fix(autonomous-loop): cap read_file and run_bash output

grep_tool capped its output at 64 KB; read_file_tool returned f.read() whole
and run_bash_tool returned the complete stdout and stderr. A 2 MB log or a
failing test suite therefore entered the transcript in full, and the loop
re-serializes its history into every later prompt, so one oversized result is
paid for on every remaining call of the run.

Lift grep's cap into a shared truncate_tool_output and route all three tools
through it. The marker now reports how much was cut, so the model can narrow
its query rather than guessing that output was complete.

read_file's memory note still reports the file's real length, not the
truncated one.

Closes kyegomez#1971

Author: ayaangazali

swarms/structs/autonomous_loop_utils.py

@@ -45,6 +45,21 @@
 MAX_SUBTASK_LOOPS = 20
 MAX_CONSECUTIVE_THINKS = 2
 
+# The loop re-serializes its history into every later prompt, so one oversized
+# tool result is paid for on every remaining call of the run.
+MAX_TOOL_OUTPUT_CHARS = 65536
+
+
+def truncate_tool_output(text: str) -> str:
+    """Cap a tool result, telling the model how much it is not seeing."""
+    if len(text) <= MAX_TOOL_OUTPUT_CHARS:
+        return text
+    return (
+        text[:MAX_TOOL_OUTPUT_CHARS]
+        + "\n... (output truncated: showing the first "
+        + f"{MAX_TOOL_OUTPUT_CHARS} of {len(text)} characters)"
+    )
+
 
 # Prompts.
 
@@ -784,12 +799,13 @@ def read_file_tool(agent: Any, file_path: str, **kwargs) -> str:
 
         # Read file
         with open(full_path, "r", encoding="utf-8") as f:
-            content = f.read()
+            raw = f.read()
+        content = truncate_tool_output(raw)
 
         # Add to memory
         agent.short_memory.add(
             role="File Operations",
-            content=f"Read file: {full_path} ({len(content)} characters)",
+            content=f"Read file: {full_path} ({len(raw)} characters)",
         )
 
         if agent.verbose:
@@ -1080,7 +1096,7 @@ def run_bash_tool(
         if agent.verbose:
             logger.info(f"Executed bash command: {command[:80]}...")
 
-        return result_msg.strip()
+        return truncate_tool_output(result_msg.strip())
     except subprocess.TimeoutExpired:
         error_msg = f"Error: Command timed out after {timeout_seconds} seconds"
         logger.error(error_msg)
@@ -1099,9 +1115,6 @@ def run_bash_tool(
         return error_msg
 
 
-_GREP_MAX_BYTES = 65536  # cap output at 64 KB
-
-
 def grep_tool(
     agent: Any,
     pattern: str,
@@ -1174,10 +1187,7 @@ def grep_tool(
         stderr = result.stderr or ""
 
         # Truncate oversized output
-        if len(stdout) > _GREP_MAX_BYTES:
-            stdout = (
-                stdout[:_GREP_MAX_BYTES] + "\n... (output truncated)"
-            )
+        stdout = truncate_tool_output(stdout)
 
         if result.returncode == 0:
             output = stdout.strip() or "(no matches)"

tests/agents/test_autonomous_loop.py

@@ -33,7 +33,11 @@
 
 from swarms import Agent
 from swarms.agents.autonomous_loop import AutonomousAgentLoop
-from swarms.structs.autonomous_loop_utils import MAX_SUBTASK_LOOPS
+from swarms.structs.autonomous_loop_utils import (
+    MAX_SUBTASK_LOOPS,
+    MAX_TOOL_OUTPUT_CHARS,
+    read_file_tool,
+)
 
 
 # --------------------------------------------------------------------------
@@ -224,6 +228,20 @@ def test_skipped_counts_as_terminal(self):
 # --------------------------------------------------------------------------
 
 
+class TestToolOutputIsCapped:
+    """One oversized read is re-sent on every later call of the run."""
+
+    def test_read_file_truncates_a_large_file(self, tmp_path):
+        agent = build_agent()
+        big = tmp_path / "big.txt"
+        big.write_text("x" * (MAX_TOOL_OUTPUT_CHARS + 5000))
+
+        output = read_file_tool(agent, str(big))
+
+        assert len(output) < MAX_TOOL_OUTPUT_CHARS + 200
+        assert "output truncated" in output
+
+
 class TestBatchedToolCalls:
     """Every tool call in a response runs, whatever its position."""