Files
Teknium c6436783d1 fix(display): browser_exec raw-text masking ends literals at \r and steps over comments
The fail-closed raw-text scan (unparsed snippets, and the whole-pass fallback when the AST
walk gives up) ended a single-quoted literal only at "\n". With bare-\r line endings an
apostrophe in a comment (`# it's here`) opened a "literal" that ran across lines, the
real literals fell out of step and the typed value stayed visible:
`fill_input('#p',\r  # it's here\r  'pw'))\r` showed `'pw'`.

- _literal_end breaks at \r as well as \n (the parser's line terminators).
- _call_end skips `#` comments, so a bracket or quote in one never ends a typing-call
  region early.
- _literal_regions masks quoted text in comments too (fail closed), lets a lone quote
  in a comment end with it, and masks the rest of the region when an unterminated
  literal outside a comment leaves the scan out of step.

Rows for the unparsed and walk-fallback shapes added to
test_browser_exec_typed_password_never_reaches_display (2 red on 8dd38f4f). A
24k-snippet fuzz (clean and malformed) shows 0 snippets leaking that head masked.
2026-11-13 16:48:04 -08:00

393 lines
14 KiB
Python

"""Tests for agent/display.py — build_tool_preview() and inline diff previews."""
import pytest
from unittest.mock import MagicMock
import agent.display as display_module
from agent.display import (
build_tool_preview,
capture_local_edit_snapshot,
extract_edit_diff,
get_cute_tool_message,
prepare_tool_preview,
redact_tool_args_for_display,
set_tool_preview_max_len,
_summarize_rendered_diff_sections,
render_edit_diff_with_delta,
)
@pytest.fixture(autouse=True)
def reset_tool_preview_max_len():
set_tool_preview_max_len(0)
yield
set_tool_preview_max_len(0)
def test_cute_tool_message_falls_back_when_renderer_raises(monkeypatch):
def _boom(*_args, **_kwargs):
raise RuntimeError("cosmetic failure")
monkeypatch.setattr(display_module, "_get_cute_tool_message", _boom)
result = get_cute_tool_message("web_extract", {"urls": []}, 0.25)
assert isinstance(result, str) and result
class TestBuildToolPreview:
"""Tests for build_tool_preview defensive handling and normal operation."""
def test_none_args_returns_none(self):
"""PR #453: None args should not crash, should return None."""
assert build_tool_preview("terminal", None) is None
def test_empty_dict_returns_none(self):
"""Empty dict has no keys to preview."""
assert build_tool_preview("terminal", {}) is None
def test_memory_patch_preview_shows_the_span_being_edited(self):
args = {"action": "patch", "target": "memory", "old_text": "port 8080", "content": "port 9090"}
assert build_tool_preview("memory", args) == '~memory: "port 8080"'
assert '~memory: "port 8080"' in get_cute_tool_message("memory", args, 0.1)
def test_memory_operations_batch_previews_its_first_op_and_the_rest_count(self):
# operations[] is the memory tool's only advertised shape, so progress lines read it.
one = {"target": "user", "operations": [{"action": "add", "content": "Name: Alice"}]}
assert build_tool_preview("memory", one) == '+user: "Name: Alice"'
assert '+user: "Name: Alice"' in get_cute_tool_message("memory", one, 0.1)
three = {"target": "memory", "operations": [
{"action": "remove", "old_text": "port 8080"}, {"action": "add", "content": "a"},
{"action": "patch", "old_text": "b", "content": "c"}]}
assert build_tool_preview("memory", three) == '-memory: "port 8080" (2 more)'
assert '-memory: "port 8080" (2 more)' in get_cute_tool_message("memory", three, 0.1)
def test_browser_type_preview_redacts_api_key(self):
secret = "sk-proj-ABCD1234567890EFGH"
result = build_tool_preview("browser_type", {"ref": "@e3", "text": secret})
assert result is not None
assert secret not in result
assert "sk-pro" in result and "..." in result
def test_browser_type_preview_keeps_normal_text(self):
text = "hello world search query"
result = build_tool_preview("browser_type", {"ref": "@e3", "text": text})
assert result is not None
assert text in result
@pytest.mark.parametrize("code", [
"fill_input('input[type=password]', 'Hunter2Correct!')",
"js(\"document.querySelector('#pwd').value='Hunter2Correct!'\")",
"cdp('Input.insertText', text='Hunter2Correct!')",
"fill_input('#p', 'abc\\'Hunter2Correct!')",
"fill_input('#p', \"\"\"Hunter2Correct!\"\"\")",
"fill_input('#p', 'Hunter2Correct!",
"type_text('Hunter2Correct!')",
"fill_input('input[type=password]:not(.x)', 'Hunter2Correct!')",
"fill_input('#a, #pass', 'Hunter2Correct!')",
"fill_input('#p', f'Hunter2Correct!')",
"fill_input('#p', r'Hunter2Correct!')",
"pw = 'Hunter2Correct!'\nfill_input('#p', pw)",
"params = {'text': 'Hunter2Correct!'}\ncdp('Input.insertText', **params)",
"params = {\"text\": \"Hunter2Correct!\"}\ncdp('Input.insertText', params)",
"params = {'text': 'Hunter2Correct!'}\ncdp('Input.insertText', {**params})",
"params = {'text': 'Hunter2Correct!'}\nparams = {**params}\ncdp('Input.insertText', **params)",
"fill_input('#p', ('Hunter2Correct!'))",
"fill_input('#p', '' 'Hunter2Correct!')",
"fill_input(text='Hunter2Correct!', selector='#p')",
"fill_input('#p', f'{1}Hunter2Correct!')",
"pw = ('Hunter2Correct!')\nfill_input('#p', pw)",
"fill_input('#p', ('Hunter2Correct!'",
# the parser ends a line at \r and \r\n as well as \n: spans must count the same lines
"x = 1\rfill_input('#p', 'Hunter2Correct!')",
"x = 1\r\nfill_input('#p', 'Hunter2Correct!')",
"x = 1\rfill_input('#p', 'Hunter2Correct!')\nprint('hello world, a longer line here')",
"s = '''a\rb'''\r\nfill_input('#p', 'Hunter2Correct!')\rprint('kept')\n",
# names bound by unpacking, loops, splats and calls to a function the snippet defines
"sel, pw = '#p', 'Hunter2Correct!'\nfill_input(sel, pw)",
"fill_input(*('#p', 'Hunter2Correct!'))",
"fill_input(**{'selector': '#p', 'text': 'Hunter2Correct!'})",
"for pw in ['Hunter2Correct!']:\n fill_input('#p', pw)",
"def go(pw):\n fill_input('#p', pw)\ngo('Hunter2Correct!')",
# the raw-text fallback (unparsed: extra ')') ends a literal at \r and skips comments, so the
# quote in "# it's" cannot swallow the typed value
"fill_input('#p',\r # it's here\r 'Hunter2Correct!'))\r",
# a long concatenation makes the walk give up: the whole-pass fallback must hold too
pytest.param("# don't echo the password\rcss = " + " + ".join(f"'.c{i}{{}}'" for i in range(900))
+ "\rfill_input('#password', 'Hunter2Correct!')\r", id="cr-comment-quote-walk-fallback"),
])
def test_browser_exec_typed_password_never_reaches_display(self, code):
"""#72298: an ordinary password typed by a browser_exec snippet matches no secret shape,
yet every display surface (preview, verbose args, cute row) must mask it; the selector stays."""
args = {"code": code}
shown = (build_tool_preview("browser_exec", args, max_len=0),
str(redact_tool_args_for_display("browser_exec", args)),
get_cute_tool_message("browser_exec", args, 0.1))
assert all("Hunter2Correct!" not in s and "***" in s for s in shown)
assert args["code"] == code
def test_browser_exec_display_keeps_selectors_urls_and_untyped_strings(self):
"""Only typed values are masked: a selector, a URL, a text= kwarg outside insertText and a
.value comparison all read as written."""
code = ("new_tab(url='https://example.com/login', text='Search page')\n"
"fill_input('input[type=password]:not(.x)', 'Hunter2Correct!')\n"
"cdp('Input.insertText', text='Hunter2Correct!')\n"
"js(\"document.querySelector('#u').value == 'admin'\")")
shown = redact_tool_args_for_display("browser_exec", {"code": code})["code"]
assert "Hunter2Correct!" not in shown
for kept in ("https://example.com/login", "text='Search page'",
"input[type=password]:not(.x)", ".value == 'admin'"):
assert kept in shown
def test_browser_type_display_args_redact_api_key(self):
secret = "ghp_ABCDEFGHIJ1234567890"
safe_args = redact_tool_args_for_display(
"browser_type", {"ref": "@e3", "text": secret}
)
assert secret not in str(safe_args)
assert safe_args["ref"] == "@e3"
assert safe_args["text"].startswith("ghp_AB")
def test_delegate_task_batch_preview_respects_max_len(self):
result = build_tool_preview(
"delegate_task",
{"tasks": [{"goal": "A" * 80}, {"goal": "B" * 80}]},
max_len=30,
)
assert result is not None and len(result) <= 30
def test_false_like_args_zero(self):
"""Non-dict falsy values should return None, not crash."""
assert build_tool_preview("terminal", 0) is None
assert build_tool_preview("terminal", "") is None
assert build_tool_preview("terminal", []) is None
@pytest.mark.parametrize("max_len", [1, 2, 3, 4])
def test_tiny_max_len_never_exceeded(self, max_len):
"""max_len is a hard cap on every preview path — dedicated builder (terminal), generic
fallback key (web_search), and the cute head-truncated path (#9439)."""
from agent.display import _cute_path, set_tool_preview_max_len
long = "abcdefghijklmnopqrstuvwxyz"
for tool, args in (("terminal", {"command": long}), ("web_search", {"query": long})):
preview = build_tool_preview(tool, args, max_len=max_len)
assert preview and len(preview) <= max_len, (tool, preview)
set_tool_preview_max_len(max_len)
try:
assert len(_cute_path("/" + long + "/file.py")) <= max_len
finally:
set_tool_preview_max_len(0)
class TestPrepareToolPreview:
def test_recovers_and_describes_truncated_url(self):
url = "https://example.com/a/very/long/path/to/a/page"
set_tool_preview_max_len(20)
preview = prepare_tool_preview(
"web_extract",
{"urls": [url]},
fallback=url[:17] + "...",
max_len=20,
)
assert preview.text == url[:17] + "..."
assert preview.truncated is True
assert preview.url == url
def test_untruncated_url_has_no_link_target(self):
url = "https://example.com/page"
preview = prepare_tool_preview(
"browser_navigate", None, fallback=url, max_len=40
)
assert preview.text == url
assert preview.truncated is False
assert preview.url is None
def test_truncated_non_url_has_no_link_target(self):
preview = prepare_tool_preview(
"web_search",
{"query": "how to parse a URL"},
fallback="how to parse a URL",
max_len=12,
)
assert preview.truncated is True
assert preview.url is None
class TestCuteToolMessagePreviewLength:
def test_search_files_preview_uses_positive_configured_limit_not_default(self):
set_tool_preview_max_len(80)
pattern = "function.formatToolCall.context.preview.compactPreview.maxLength.truncate"
line = get_cute_tool_message("search_files", {"pattern": pattern}, 0.1)
assert pattern in line
assert "..." not in line
def test_browser_type_cute_message_redacts_api_key(self):
secret = "sk-proj-ABCD1234567890EFGH"
line = get_cute_tool_message(
"browser_type",
{"ref": "@password", "text": secret},
0.1,
result='{"success": true, "typed": "sk-pro...EFGH"}',
)
assert secret not in line
assert "sk-pro" in line
def test_browser_type_cute_message_keeps_normal_text(self):
text = "hello world"
line = get_cute_tool_message(
"browser_type",
{"ref": "@search", "text": text},
0.1,
result='{"success": true, "typed": "hello world"}',
)
assert text in line
class TestEditDiffPreview:
def test_extract_edit_diff_uses_local_snapshot_for_write_file(self, tmp_path):
target = tmp_path / "note.txt"
target.write_text("old\n", encoding="utf-8")
snapshot = capture_local_edit_snapshot("write_file", {"path": str(target)})
target.write_text("new\n", encoding="utf-8")
diff = extract_edit_diff(
"write_file",
'{"bytes_written": 4}',
function_args={"path": str(target)},
snapshot=snapshot,
)
assert diff is not None
assert "--- a/" in diff
assert "+++ b/" in diff
assert "-old" in diff
assert "+new" in diff
def test_render_edit_diff_with_delta_handles_renderer_errors(self, monkeypatch):
printer = MagicMock()
monkeypatch.setattr("agent.display._summarize_rendered_diff_sections", MagicMock(side_effect=RuntimeError("boom")))
rendered = render_edit_diff_with_delta(
"patch",
'{"diff": "--- a/x\\n+++ b/x\\n"}',
print_fn=printer,
)
assert rendered is False
assert printer.call_count == 0
def test_summarize_rendered_diff_sections_limits_file_count(self):
diff = "".join(
f"--- a/file{i}.py\n+++ b/file{i}.py\n+line{i}\n"
for i in range(8)
)
rendered = _summarize_rendered_diff_sections(diff, max_files=3, max_lines=50)
assert any("a/file0.py" in line for line in rendered)
assert any("a/file1.py" in line for line in rendered)
assert any("a/file2.py" in line for line in rendered)
assert not any("a/file7.py" in line for line in rendered)
assert "additional file" in rendered[-1]
class TestBuildToolLabel:
"""Friendly human-phrased tool labels for built-in tools."""
@pytest.fixture(autouse=True)
def _enable_friendly(self):
from agent.display import set_friendly_tool_labels
set_friendly_tool_labels(True)
yield
set_friendly_tool_labels(True)
def test_disabled_falls_back_to_preview(self):
from agent.display import (
build_tool_label,
build_tool_preview,
set_friendly_tool_labels,
)
set_friendly_tool_labels(False)
args = {"query": "weather in NYC"}
label = build_tool_label("web_search", args)
# With the feature off, must match the raw preview exactly
assert label == build_tool_preview("web_search", args)
assert "Searching the web" not in (label or "")
class TestBuildStatusPhrase:
"""build_status_phrase — live working-state text for Slack's status line."""
def test_caps_length_for_slack_status_line(self):
from agent.display import build_status_phrase
phrase = build_status_phrase(
"terminal", {"command": "x" * 300}, max_len=49
)
assert phrase is not None and len(phrase) <= 49
assert phrase.endswith("…")
def test_respects_friendly_labels_toggle(self):
from agent.display import build_status_phrase, set_friendly_tool_labels
set_friendly_tool_labels(False)
try:
assert build_status_phrase("terminal", {"command": "ls"}) is None
finally:
set_friendly_tool_labels(True)