Open-source AI penetration testing tool to find and fix your app’s vulnerabilities.
Python agentsai-hackingai-penetration-testingai-pentestingai-securityartificial-intelligencebug-bountycode-qualityctf-toolscybersecuritycybersecurity-toolsethical-hackinghackingllm-securityoffensive-securitypenetration-testingpentesting-toolsred-teamingsecuritysecurity-automation
refactor(context): trim verbose comments
Ahmed Allam74f334cb935c
diff --git a/strix/agents/factory.py b/strix/agents/factory.pyindex d6b36722..872e3564 100644--- a/strix/agents/factory.py+++ b/strix/agents/factory.py@@ -129,12 +129,7 @@ def _format_tool_error(exc: Exception) -> str: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.- """+ """Cap a tool's result size before it enters history (idempotent)."""if getattr(tool, "_strix_bounded", False):return toolinvoke_tool = tool.on_invoke_tool@@ -200,13 +195,7 @@ def _custom_tool_as_function_tool(tool: CustomTool) -> FunctionTool:def _bound_custom_tool(tool: CustomTool) -> CustomTool:- """Bound a native ``CustomTool`` result in place.-- Chat-completions mode converts filesystem ``CustomTool``s to ``FunctionTool``s- (which bounds the result), but the Responses path keeps them native, so a- large ``read_file``/directory listing would otherwise append unbounded text- to history. Wrap ``on_invoke_tool`` so the same head+tail bound applies.- """+ """Bound a native ``CustomTool`` result in place (Responses path)."""invoke_tool = tool.on_invoke_toolasync def invoke(ctx: Any, raw_input: str) -> Any:@@ -223,8 +212,6 @@ def _configure_filesystem_tools(toolset: Any, *, chat_completions: bool) -> Nonesetattr(toolset, name, _custom_tool_as_function_tool(tool))elif isinstance(tool, FunctionTool):setattr(toolset, name, _function_tool_with_error_result(tool))- # Responses-API path: keep tools native but still bound their output so- # filesystem reads can't exhaust the context window on later turns.elif isinstance(tool, CustomTool):setattr(toolset, name, _bound_custom_tool(tool))elif isinstance(tool, FunctionTool):@@ -277,13 +264,8 @@ def _format_validation_error(tool_name: str, exc: ValidationError) -> str:def _apply_shell_output_cap(parsed: dict[str, Any]) -> None:- """Bound the SDK shell tools' own token cap so a single command can't dump- unbounded output into history. The SDK truncates head+tail from this value.-- The configured cap is a ceiling: a missing value defaults to it, and a- larger model-supplied value is clamped down to it. A smaller explicit value- is respected, so the model can still ask for less.- """+ """Clamp the SDK shell tools' ``max_output_tokens`` to the configured+ ceiling; a smaller explicit value is respected."""ceiling = load_settings().context.tool_output_max_tokensrequested = parsed.get("max_output_tokens")parsed["max_output_tokens"] = (diff --git a/strix/config/settings.py b/strix/config/settings.pyindex d7c795b6..0d910e60 100644--- a/strix/config/settings.py+++ b/strix/config/settings.py@@ -69,8 +69,7 @@ class ContextSettings(BaseSettings):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")- # Floor comfortably above the truncation-notice size so a preview- # (head+tail+notice) always fits within the configured ceiling.+ # Floor above the truncation-notice size so a preview always fits.tool_output_max_bytes: int = Field(default=50 * 1024, ge=1024, alias="STRIX_TOOL_OUTPUT_MAX_BYTES")diff --git a/strix/tools/output_store.py b/strix/tools/output_store.pyindex 6e6e6202..ed65dfa8 100644--- a/strix/tools/output_store.py+++ b/strix/tools/output_store.py@@ -1,10 +1,7 @@"""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.+Keeps a head + tail slice and drops the middle, replacing it with a notice of+how much was removed."""from __future__ import annotations@@ -45,22 +42,16 @@ def _take_suffix(text: str, max_bytes: int) -> str: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.- ``max_bytes`` bounds the *entire* joined result, notice and separators- included, and must be large enough to hold the notice itself (guaranteed by- the ``tool_output_max_bytes`` config floor).+ Truncates on whichever limit is hit first (line count or UTF-8 byte size).+ ``max_bytes`` bounds the entire joined result, notice and separators+ included."""lines = text.split("\n")total_bytes = _byte_len(text)if len(lines) <= max_lines and total_bytes <= max_bytes:return text- # Reserve room for the notice and its two blank-line separators so the- # head+tail slices can't consume the whole budget and push the persisted- # value over max_bytes. Upper-bound the notice with the largest possible- # counts; the real notice is never longer. ``+ 4`` covers the separators.+ # Reserve notice + separator bytes up front; ``+ 4`` covers the two "\n\n".notice_overhead = _byte_len(_TRUNCATION_NOTICE.format(lines=len(lines), bytes=total_bytes)) + 4byte_budget = max(2, max_bytes - notice_overhead)@@ -69,16 +60,13 @@ def bound_text(text: str, *, max_lines: int, max_bytes: int) -> str: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, byte_budget // 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)- # Count kept lines from the final slices: the byte pass above may have- # dropped whole lines from head/tail, so deriving this from the original- # head_lines/tail_lines would undercount what was actually removed.+ # Count from the final slices; the byte pass may have dropped whole lines.kept_lines = len(head.split("\n")) + (len(tail.split("\n")) if tail else 0)dropped_lines = max(0, len(lines) - kept_lines)dropped_bytes = max(0, total_bytes - _byte_len(head) - _byte_len(tail))diff --git a/tests/test_agent_factory_shell.py b/tests/test_agent_factory_shell.pyindex 22b662c7..814da60c 100644--- a/tests/test_agent_factory_shell.py+++ b/tests/test_agent_factory_shell.py@@ -82,8 +82,6 @@ async def test_wrap_exec_command_preserves_explicit_shell(shell: str) -> None:@pytest.mark.asyncioasync def test_responses_filesystem_custom_tool_output_is_bounded() -> None:- # In Responses-API mode filesystem tools stay native CustomTools; a large- # read must still be head+tail bounded before it enters history.async def invoke(_ctx: Any, _inp: str) -> str:return "line\n" * 50_000diff --git a/tests/test_config_loader.py b/tests/test_config_loader.pyindex 0d37e1e8..7e14662a 100644--- a/tests/test_config_loader.py+++ b/tests/test_config_loader.py@@ -121,14 +121,7 @@ def test_read_json_overrides_uses_json_when_no_alias_in_environ(tmp_path: Path)assert loader._read_json_overrides(path) == {"llm": {"api_key": "sk-file"}}-# --------------------------------------------------------------------------- #-# ContextSettings validation-# --------------------------------------------------------------------------- #--def test_tool_output_max_bytes_rejects_sub_notice_values() -> None:- # A ceiling below the truncation notice can't fit a bounded preview, so it- # is rejected at load time rather than producing over-cap persisted output.with pytest.raises(ValidationError):ContextSettings(STRIX_TOOL_OUTPUT_MAX_BYTES=64)diff --git a/tests/test_output_store.py b/tests/test_output_store.pyindex 14bff19c..4db8ccf7 100644--- a/tests/test_output_store.py+++ b/tests/test_output_store.py@@ -19,7 +19,6 @@ def test_line_limit_keeps_head_and_tail() -> None: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@@ -28,8 +27,6 @@ def test_byte_limit_enforced_on_single_long_line() -> None:bounded = bound_text(text, max_lines=2_000, max_bytes=1_000)assert "truncated" in bounded- # The whole joined result (head + tail + notice + separators) honours the- # configured maximum, not just the head/tail slices.assert len(bounded.encode("utf-8")) <= 1_000@@ -37,7 +34,7 @@ def test_multibyte_characters_not_split() -> None:text = "😀" * 50_000bounded = 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).+ # Must remain valid UTF-8 (no mid-character cut).assert bounded == bounded.encode("utf-8").decode("utf-8")assert "truncated" in bounded@@ -51,8 +48,7 @@ def test_notice_reports_dropped_counts() -> None:def test_dropped_line_count_accounts_for_byte_trimming() -> None:- # A tight byte budget forces the byte pass to drop whole lines from the- # head/tail slices; the notice must count those, not just the middle.+ # Tight byte budget drops whole lines from head/tail; the notice must count them.text = "\n".join(f"line-{i}" for i in range(200))bounded = bound_text(text, max_lines=20, max_bytes=40)@@ -61,5 +57,4 @@ def test_dropped_line_count_accounts_for_byte_trimming() -> None:dropped = int(match.group(1))kept = [ln for ln in bounded.splitlines() if ln and "truncated" not in ln]assert dropped == 200 - len(kept)- # The naive middle-only count (max_lines split evenly) would under-report.assert dropped > 200 - 20