feat(context): bound per-tool output before it enters agent history

Cap the size of every tool result so a single verbose command (recursive
find, noisy scanner, full page dump) can't pin the conversation near the
model's context window for the rest of a scan.

- New ContextSettings config group with env-tunable caps.
- Default the SDK shell tools' max_output_tokens so exec_command /
  write_stdin truncate head+tail instead of returning unbounded output.
- Bound Strix's own FunctionTool/CustomTool results (line + UTF-8 byte
  head+tail preview with a truncation notice) and cap error strings.

strix/agents/factory.py

@@ -16,6 +16,7 @@ from agents.tool import CustomTool, FunctionTool, Tool
 from pydantic import ValidationError
 
 from strix.agents.prompt import render_system_prompt
+from strix.config import load_settings
 from strix.tools.agents_graph.tools import (
     agent_finish,
     create_agent,
@@ -33,6 +34,7 @@ from strix.tools.notes.tools import (
     list_notes,
     update_note,
 )
+from strix.tools.output_store import bound_text
 from strix.tools.proxy.tools import (
     list_requests,
     list_sitemap,
@@ -108,8 +110,41 @@ def _extract_custom_input(tool: CustomTool, raw_input: str | dict[str, Any]) ->
     return value if isinstance(value, str) else ""
 
 
+def _tool_output_limits() -> tuple[int, int]:
+    context = load_settings().context
+    return context.tool_output_max_lines, context.tool_output_max_bytes
+
+
+def _bound_result(result: Any) -> Any:
+    if not isinstance(result, str):
+        return result
+    max_lines, max_bytes = _tool_output_limits()
+    return bound_text(result, max_lines=max_lines, max_bytes=max_bytes)
+
+
 def _format_tool_error(exc: Exception) -> str:
-    return str(exc) or exc.__class__.__name__
+    message = str(exc) or exc.__class__.__name__
+    max_lines, max_bytes = _tool_output_limits()
+    return bound_text(message, max_lines=max_lines, max_bytes=max_bytes)
+
+
+def _with_bounded_result(tool: FunctionTool) -> FunctionTool:
+    """Cap the size of a tool's result before it enters agent history.
+
+    Idempotent: base tools are shared singletons reused across every agent, so
+    the guard prevents stacking the wrapper on repeated ``build_strix_agent``
+    calls.
+    """
+    if getattr(tool, "_strix_bounded", False):
+        return tool
+    invoke_tool = tool.on_invoke_tool
+
+    async def invoke(ctx: Any, raw_input: str) -> Any:
+        return _bound_result(await invoke_tool(ctx, raw_input))
+
+    tool.on_invoke_tool = invoke
+    tool._strix_bounded = True  # type: ignore[attr-defined]
+    return tool
 
 
 def _function_tool_with_error_result(tool: FunctionTool) -> FunctionTool:
@@ -117,7 +152,7 @@ def _function_tool_with_error_result(tool: FunctionTool) -> FunctionTool:
 
     async def invoke(ctx: Any, raw_input: str) -> Any:
         try:
-            return await invoke_tool(ctx, raw_input)
+            return _bound_result(await invoke_tool(ctx, raw_input))
         except Exception as exc:  # noqa: BLE001 - tool errors should be model-visible results.
             logger.debug("Tool %s failed; returning error as result", tool.name, exc_info=True)
             return _format_tool_error(exc)
@@ -132,7 +167,7 @@ def _custom_tool_as_function_tool(tool: CustomTool) -> FunctionTool:
         if not custom_input:
             return f"`{_custom_tool_input_field(tool)}` must be a non-empty string."
         try:
-            return await tool.on_invoke_tool(ctx, custom_input)
+            return _bound_result(await tool.on_invoke_tool(ctx, custom_input))
         except Exception as exc:  # noqa: BLE001 - matches SDK CustomTool error-as-result behavior.
             logger.debug("Tool %s failed; returning error as result", tool.name, exc_info=True)
             return _format_tool_error(exc)
@@ -210,6 +245,15 @@ def _format_validation_error(tool_name: str, exc: ValidationError) -> str:
     return f"{tool_name}: invalid arguments — " + "; ".join(parts)
 
 
+def _apply_shell_output_cap(parsed: dict[str, Any]) -> None:
+    """Default the SDK shell tools' own token cap so a single command can't
+    dump unbounded output into history. The SDK truncates head+tail when the
+    model omits the field; respect an explicit model-supplied value.
+    """
+    if parsed.get("max_output_tokens") is None:
+        parsed["max_output_tokens"] = load_settings().context.tool_output_max_tokens
+
+
 def _wrap_exec_command(tool: FunctionTool) -> FunctionTool:
     invoke_tool = tool.on_invoke_tool
 
@@ -218,8 +262,10 @@ def _wrap_exec_command(tool: FunctionTool) -> FunctionTool:
             parsed = json.loads(raw_input)
         except (json.JSONDecodeError, TypeError):
             parsed = None
-        if isinstance(parsed, dict) and "shell" not in parsed:
+        if isinstance(parsed, dict):
+            if "shell" not in parsed:
                 parsed["shell"] = "bash"
+            _apply_shell_output_cap(parsed)
             raw_input = json.dumps(parsed)
         try:
             return await invoke_tool(ctx, raw_input)
@@ -245,8 +291,10 @@ def _wrap_write_stdin(tool: FunctionTool) -> FunctionTool:
             parsed = json.loads(raw_input)
         except json.JSONDecodeError:
             parsed = None
-        if isinstance(parsed, dict) and isinstance(parsed.get("chars"), str):
+        if isinstance(parsed, dict):
+            if isinstance(parsed.get("chars"), str):
                 parsed["chars"] = _decode_chars_escape(parsed["chars"])
+            _apply_shell_output_cap(parsed)
             raw_input = json.dumps(parsed)
         try:
             return await invoke_tool(ctx, raw_input)
@@ -447,6 +495,9 @@ def build_strix_agent(
     else:
         tools = [*_BASE_TOOLS, *agent_tools, agent_finish]
     _ensure_unique_tool_names(tools)
+    tools = [
+        _with_bounded_result(tool) if isinstance(tool, FunctionTool) else tool for tool in tools
+    ]
 
     logger.info(
         "Built %s agent '%s' (skills=%d, tools=%d, scan_mode=%s, whitebox=%s)",

strix/config/__init__.py

@@ -17,6 +17,7 @@ from strix.config.loader import (
     persist_current,
 )
 from strix.config.settings import (
+    ContextSettings,
     DedupeSettings,
     IntegrationSettings,
     LlmSettings,
@@ -27,6 +28,7 @@ from strix.config.settings import (
 
 
 __all__ = [
+    "ContextSettings",
     "DedupeSettings",
     "IntegrationSettings",
     "LlmSettings",

strix/config/settings.py

@@ -55,6 +55,23 @@ class DedupeSettings(BaseSettings):
     api_base: str | None = Field(default=None, alias="DEDUPE_LLM_API_BASE")
 
 
+class ContextSettings(BaseSettings):
+    """Context-window management: per-tool-output caps and history compaction."""
+
+    model_config = _BASE_CONFIG
+
+    auto_compact: bool = Field(default=True, alias="STRIX_CONTEXT_AUTO_COMPACT")
+    compact_buffer_tokens: int = Field(default=20_000, gt=0, alias="STRIX_CONTEXT_BUFFER_TOKENS")
+    keep_tokens: int = Field(default=8_000, gt=0, alias="STRIX_CONTEXT_KEEP_TOKENS")
+    fallback_context_tokens: int = Field(
+        default=200_000, gt=0, alias="STRIX_CONTEXT_FALLBACK_TOKENS"
+    )
+    summary_max_tokens: int = Field(default=4_096, gt=0, alias="STRIX_CONTEXT_SUMMARY_TOKENS")
+    tool_output_max_tokens: int = Field(default=8_000, gt=0, alias="STRIX_TOOL_OUTPUT_MAX_TOKENS")
+    tool_output_max_lines: int = Field(default=2_000, gt=0, alias="STRIX_TOOL_OUTPUT_MAX_LINES")
+    tool_output_max_bytes: int = Field(default=50 * 1024, gt=0, alias="STRIX_TOOL_OUTPUT_MAX_BYTES")
+
+
 class RuntimeSettings(BaseSettings):
     model_config = _BASE_CONFIG
 
@@ -99,6 +116,7 @@ class Settings(BaseSettings):
     llm: LlmSettings = Field(default_factory=LlmSettings)
     dedupe: DedupeSettings = Field(default_factory=DedupeSettings)
     runtime: RuntimeSettings = Field(default_factory=RuntimeSettings)
+    context: ContextSettings = Field(default_factory=ContextSettings)
     telemetry: TelemetrySettings = Field(default_factory=TelemetrySettings)
     integrations: IntegrationSettings = Field(default_factory=IntegrationSettings)
     viewer: ViewerSettings = Field(default_factory=ViewerSettings)

strix/tools/output_store.py

@@ -0,0 +1,72 @@
+"""Bound oversized tool results before they enter agent history.
+
+A single verbose tool result (a recursive ``find``, a noisy scanner, a full
+page dump) can otherwise pin the whole conversation near the model's context
+limit for the rest of the scan. This keeps a head + tail slice of the output
+and drops the middle, mirroring how the shell capability truncates its own
+output — the agent still sees the start and end plus how much was removed.
+"""
+
+from __future__ import annotations
+
+
+_TRUNCATION_NOTICE = "[... {lines} lines ({bytes} bytes) truncated ...]"
+
+
+def _byte_len(text: str) -> int:
+    return len(text.encode("utf-8"))
+
+
+def _take_prefix(text: str, max_bytes: int) -> str:
+    budget = 0
+    out: list[str] = []
+    for char in text:
+        size = len(char.encode("utf-8"))
+        if budget + size > max_bytes:
+            break
+        out.append(char)
+        budget += size
+    return "".join(out)
+
+
+def _take_suffix(text: str, max_bytes: int) -> str:
+    budget = 0
+    out: list[str] = []
+    for char in reversed(text):
+        size = len(char.encode("utf-8"))
+        if budget + size > max_bytes:
+            break
+        out.append(char)
+        budget += size
+    out.reverse()
+    return "".join(out)
+
+
+def bound_text(text: str, *, max_lines: int, max_bytes: int) -> str:
+    """Return ``text`` unchanged when small, else a head+tail preview.
+
+    Truncation happens on whichever limit is hit first (line count or UTF-8
+    byte size). The removed middle is replaced with a notice recording how
+    many lines and bytes were dropped so the agent knows output was elided.
+    """
+    lines = text.split("\n")
+    total_bytes = _byte_len(text)
+    if len(lines) <= max_lines and total_bytes <= max_bytes:
+        return text
+
+    head_lines = max(1, max_lines // 2)
+    tail_lines = max_lines - head_lines
+    head = "\n".join(lines[:head_lines])
+    tail = "\n".join(lines[len(lines) - tail_lines :]) if tail_lines > 0 else ""
+
+    # Enforce the byte budget even when the line count alone was fine.
+    half_bytes = max(1, max_bytes // 2)
+    if _byte_len(head) > half_bytes:
+        head = _take_prefix(head, half_bytes)
+    if tail and _byte_len(tail) > half_bytes:
+        tail = _take_suffix(tail, half_bytes)
+
+    dropped_lines = max(0, len(lines) - head_lines - tail_lines)
+    dropped_bytes = max(0, total_bytes - _byte_len(head) - _byte_len(tail))
+    notice = _TRUNCATION_NOTICE.format(lines=dropped_lines, bytes=dropped_bytes)
+    return f"{head}\n\n{notice}\n\n{tail}" if tail else f"{head}\n\n{notice}"

tests/test_agent_factory_shell.py

@@ -9,6 +9,7 @@ import pytest
 from agents.tool import FunctionTool
 
 from strix.agents import factory
+from strix.config import load_settings
 
 
 def _capturing_exec_tool(captured: dict[str, str]) -> FunctionTool:
@@ -32,10 +33,23 @@ async def test_wrap_exec_command_defaults_shell_to_bash() -> None:
     result = await wrapped.on_invoke_tool(cast("Any", None), json.dumps({"cmd": "source /tmp/env"}))
 
     assert result == "ok"
-    assert json.loads(captured["raw_input"]) == {
-        "cmd": "source /tmp/env",
-        "shell": "bash",
-    }
+    parsed = json.loads(captured["raw_input"])
+    assert parsed["cmd"] == "source /tmp/env"
+    assert parsed["shell"] == "bash"
+    expected_cap = load_settings().context.tool_output_max_tokens
+    assert parsed["max_output_tokens"] == expected_cap
+
+
+@pytest.mark.asyncio
+async def test_wrap_exec_command_preserves_explicit_output_cap() -> None:
+    captured: dict[str, str] = {}
+    wrapped = factory._wrap_exec_command(_capturing_exec_tool(captured))
+
+    await wrapped.on_invoke_tool(
+        cast("Any", None), json.dumps({"cmd": "echo hi", "max_output_tokens": 42})
+    )
+
+    assert json.loads(captured["raw_input"])["max_output_tokens"] == 42
 
 
 @pytest.mark.asyncio

tests/test_output_store.py

@@ -0,0 +1,46 @@
+"""Tests for per-tool-output bounding."""
+
+from __future__ import annotations
+
+from strix.tools.output_store import bound_text
+
+
+def test_small_output_passes_through_unchanged() -> None:
+    text = "line 1\nline 2\nline 3"
+    assert bound_text(text, max_lines=100, max_bytes=10_000) == text
+
+
+def test_line_limit_keeps_head_and_tail() -> None:
+    text = "\n".join(str(i) for i in range(1000))
+    bounded = bound_text(text, max_lines=10, max_bytes=1_000_000)
+
+    assert bounded.startswith("0\n1\n2\n3\n4")
+    assert bounded.rstrip().endswith("999")
+    assert "truncated" in bounded
+    # Head + tail only, far fewer than the original 1000 lines.
+    assert len(bounded.splitlines()) < 30
+
+
+def test_byte_limit_enforced_on_single_long_line() -> None:
+    text = "x" * 100_000
+    bounded = bound_text(text, max_lines=2_000, max_bytes=1_000)
+
+    assert "truncated" in bounded
+    assert len(bounded.encode("utf-8")) < 3_000
+
+
+def test_multibyte_characters_not_split() -> None:
+    text = "😀" * 50_000
+    bounded = bound_text(text, max_lines=2_000, max_bytes=1_000)
+
+    # Must remain valid UTF-8 (no broken surrogate halves from a mid-char cut).
+    assert bounded == bounded.encode("utf-8").decode("utf-8")
+    assert "truncated" in bounded
+
+
+def test_notice_reports_dropped_counts() -> None:
+    text = "\n".join("y" * 10 for _ in range(500))
+    bounded = bound_text(text, max_lines=10, max_bytes=1_000_000)
+
+    assert "lines" in bounded
+    assert "bytes" in bounded