fix(models): bound provider-facing observation output

src/minisweagent/models/utils/actions_text.py

@@ -9,6 +9,7 @@
 from jinja2 import StrictUndefined, Template
 
 from minisweagent.exceptions import FormatError
+from minisweagent.models.utils.observations import bounded_observation_output
 from minisweagent.models.utils.openai_multimodal import expand_multimodal_content
 
 
@@ -50,8 +51,9 @@ def format_observation_messages(
     """Format execution outputs into user observation messages."""
     results = []
     for output in outputs:
+        bounded_output = bounded_observation_output(output)
         content = Template(observation_template, undefined=StrictUndefined).render(
-            output=output, **(template_vars or {})
+            output=bounded_output, **(template_vars or {})
         )
         msg: dict = {
             "role": "user",

src/minisweagent/models/utils/actions_toolcall.py

@@ -6,6 +6,7 @@
 from jinja2 import StrictUndefined, Template
 
 from minisweagent.exceptions import FormatError
+from minisweagent.models.utils.observations import bounded_observation_output
 from minisweagent.models.utils.openai_multimodal import expand_multimodal_content
 
 BASH_TOOL = {
@@ -88,8 +89,9 @@ def format_toolcall_observation_messages(
     padded_outputs = outputs + [not_executed] * (len(actions) - len(outputs))
     results = []
     for action, output in zip(actions, padded_outputs):
+        bounded_output = bounded_observation_output(output)
         content = Template(observation_template, undefined=StrictUndefined).render(
-            output=output, **(template_vars or {})
+            output=bounded_output, **(template_vars or {})
         )
         msg = {
             "content": content,

src/minisweagent/models/utils/actions_toolcall_response.py

@@ -6,6 +6,7 @@
 from jinja2 import StrictUndefined, Template
 
 from minisweagent.exceptions import FormatError
+from minisweagent.models.utils.observations import bounded_observation_output
 
 # OpenRouter/OpenAI Responses API uses a flat structure (no nested "function" key)
 BASH_TOOL_RESPONSE_API = {
@@ -116,8 +117,9 @@ def format_toolcall_observation_messages(
     padded_outputs = outputs + [not_executed] * (len(actions) - len(outputs))
     results = []
     for action, output in zip(actions, padded_outputs):
+        bounded_output = bounded_observation_output(output)
         content = Template(observation_template, undefined=StrictUndefined).render(
-            output=output, **(template_vars or {})
+            output=bounded_output, **(template_vars or {})
         )
         msg: dict = {
             "extra": {

src/minisweagent/models/utils/observations.py

@@ -0,0 +1,39 @@
+"""Utilities for preparing execution observations for model providers."""
+
+MAX_OBSERVATION_OUTPUT_BYTES = 2000
+
+
+def truncate_observation_text(text: object, *, max_bytes: int = MAX_OBSERVATION_OUTPUT_BYTES) -> str:
+    """Bound provider-facing command output while preserving both ends."""
+    text = "" if text is None else str(text)
+    encoded = text.encode("utf-8")
+    if len(encoded) <= max_bytes:
+        return text
+
+    omitted = len(encoded) - max_bytes
+    while True:
+        marker = f"\n...[output truncated, {omitted} bytes omitted]...\n"
+        marker_bytes = len(marker.encode("utf-8"))
+        remaining = max(max_bytes - marker_bytes, 0)
+        new_omitted = len(encoded) - remaining
+        if new_omitted == omitted:
+            break
+        omitted = new_omitted
+
+    if marker_bytes > max_bytes:
+        return marker.encode("utf-8")[:max_bytes].decode("utf-8", errors="ignore")
+
+    remaining = max(max_bytes - marker_bytes, 0)
+    head_bytes = remaining // 2
+    tail_bytes = remaining - head_bytes
+
+    head = encoded[:head_bytes].decode("utf-8", errors="ignore")
+    tail = encoded[-tail_bytes:].decode("utf-8", errors="ignore") if tail_bytes else ""
+    return f"{head}{marker}{tail}"
+
+
+def bounded_observation_output(output: dict) -> dict:
+    """Return a shallow copy with only the provider-facing output truncated."""
+    bounded = dict(output)
+    bounded["output"] = truncate_observation_text(bounded.get("output", ""))
+    return bounded

tests/models/test_observation_truncation.py

@@ -0,0 +1,89 @@
+import json
+from pathlib import Path
+
+import yaml
+
+from minisweagent.models.litellm_model import LitellmModel
+from minisweagent.models.utils.actions_text import format_observation_messages
+from minisweagent.models.utils.actions_toolcall import format_toolcall_observation_messages
+from minisweagent.models.utils.actions_toolcall_response import (
+    format_toolcall_observation_messages as format_response_observation_messages,
+)
+from minisweagent.models.utils.observations import MAX_OBSERVATION_OUTPUT_BYTES
+
+
+def _large_output() -> str:
+    return "A" * 10000 + "B" * 10000
+
+
+def test_toolcall_observation_bounds_provider_facing_output():
+    output = _large_output()
+    messages = format_toolcall_observation_messages(
+        actions=[{"command": "generate output", "tool_call_id": "call_1"}],
+        outputs=[{"output": output, "returncode": 0}],
+        observation_template="{{ output.output }}",
+    )
+
+    content = messages[0]["content"]
+    assert len(content.encode("utf-8")) <= MAX_OBSERVATION_OUTPUT_BYTES
+    assert "output truncated" in content
+    assert content.startswith("A")
+    assert content.endswith("B")
+    assert messages[0]["extra"]["raw_output"] == output
+
+
+def test_toolcall_observation_stays_bounded_across_format_error_retries():
+    output = _large_output()
+    messages = format_toolcall_observation_messages(
+        actions=[{"command": "generate output", "tool_call_id": "call_1"}],
+        outputs=[{"output": output, "returncode": 0}],
+        observation_template="{{ output.output }}",
+    )
+
+    assert len(messages[0]["content"].encode("utf-8")) * 3 < 8192
+
+
+def test_mini_template_observation_stays_bounded_for_provider_retries():
+    output = _large_output()
+    config_path = Path(__file__).parents[2] / "src" / "minisweagent" / "config" / "mini.yaml"
+    template = yaml.safe_load(config_path.read_text())["model"]["observation_template"]
+
+    model = LitellmModel(model_name="openai/gpt-4o", observation_template=template)
+    messages = model.format_observation_messages(
+        {"extra": {"actions": [{"command": "generate output", "tool_call_id": "call_1"}]}},
+        [{"output": output, "returncode": 0, "exception_info": ""}],
+    )
+    provider_messages = model._prepare_messages_for_api(messages)
+
+    content = provider_messages[0]["content"]
+    assert len(content.encode("utf-8")) * 3 < 8192
+    assert "output truncated" in content
+    assert "raw_output" not in json.dumps(provider_messages)
+    assert output not in json.dumps(provider_messages)
+
+
+def test_response_api_observation_bounds_provider_facing_output():
+    output = _large_output()
+    messages = format_response_observation_messages(
+        actions=[{"command": "generate output", "tool_call_id": "call_1"}],
+        outputs=[{"output": output, "returncode": 0}],
+        observation_template="{{ output.output }}",
+    )
+
+    content = messages[0]["output"]
+    assert len(content.encode("utf-8")) <= MAX_OBSERVATION_OUTPUT_BYTES
+    assert "output truncated" in content
+    assert messages[0]["extra"]["raw_output"] == output
+
+
+def test_text_observation_bounds_provider_facing_output():
+    output = _large_output()
+    messages = format_observation_messages(
+        [{"output": output, "returncode": 0}],
+        observation_template="{{ output.output }}",
+    )
+
+    content = messages[0]["content"]
+    assert len(content.encode("utf-8")) <= MAX_OBSERVATION_OUTPUT_BYTES
+    assert "output truncated" in content
+    assert messages[0]["extra"]["raw_output"] == output