d6031d2881
AGENTS.md bans tests that read .py source. Replace the tokenize scan with a test that drives the real renderers (clarify summary, fallback turn, verbatim user section, oversized record, active-task line with both quote kinds) and asserts the guard regex matches and no bare idiom appears, plus the small-budget skip. The LSP test asserts via _COMPRESSION_MARKER_RE rather than a literal that a count-less copy would also satisfy. Test count unchanged (2 new vs base). Co-authored-by: ahisblessed <ahisblessed@users.noreply.github.com>
56 lines
2.5 KiB
Python
56 lines
2.5 KiB
Python
"""#121548: model-visible elision mints the non-imitable compression marker.
|
|
|
|
The bare bracketed truncation idiom those renderers used to compose was imitated
|
|
from replayed context into new durable writes (see #83435/#83714). All elision now
|
|
routes through ``agent.compression_marker.elide`` / ``elide_middle``; this file pins
|
|
guard parity and that the real renderers emit a guard-visible marker, never the idiom.
|
|
"""
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
import re
|
|
|
|
from agent.compression_marker import (
|
|
_COMPRESSION_MARKER_RE,
|
|
elide,
|
|
elide_middle,
|
|
)
|
|
from agent.context_compressor import (
|
|
ContextCompressor,
|
|
_build_verbatim_user_section,
|
|
_compact_fallback_turn,
|
|
_summarize_tool_result,
|
|
)
|
|
|
|
# Any "...[<words> truncated]" variant, not just the bare one (e.g. "...[fallback summary truncated]").
|
|
IMITABLE_MARKER_RE = re.compile(r"(?:\.\.\.|…)\s?\[[^\]\n]*truncat")
|
|
|
|
|
|
def test_minted_marker_is_caught_by_the_dispatch_boundary_guard():
|
|
"""A copied marker must refuse a durable write regardless of which renderer leaked it."""
|
|
for out in (elide("y" * 4000, 199), elide_middle("y" * 4000, 100, 100)):
|
|
assert _COMPRESSION_MARKER_RE.search(out)
|
|
|
|
|
|
def test_real_renderers_emit_a_guard_visible_marker_never_the_imitable_idiom():
|
|
"""Oversized input through the real renderers yields the guarded marker, not the idiom.
|
|
|
|
The active-task line quotes both apostrophes and double quotes so repr() cannot escape
|
|
the marker's own apostrophe out of the guard's reach.
|
|
"""
|
|
quoted = "a'b\"c " * 400
|
|
clarify = json.dumps({"user_response": "A" * 5000})
|
|
outputs = {
|
|
"clarify": _summarize_tool_result("clarify", "{}", clarify),
|
|
"fallback_turn": _compact_fallback_turn("z " * 5000),
|
|
"verbatim_user": _build_verbatim_user_section([{"role": "user", "content": "q" * 30000}]),
|
|
"record": ContextCompressor._bound_oversized_record("r" * 50000, 4000),
|
|
"active_task": ContextCompressor._latest_user_task_snapshot([{"role": "user", "content": quoted}]),
|
|
}
|
|
for name, out in outputs.items():
|
|
assert out and _COMPRESSION_MARKER_RE.search(out), (name, out[-300:] if out else out)
|
|
assert not IMITABLE_MARKER_RE.search(out), (name, out[-300:])
|
|
# A budget too small for marker + content skips the straddler instead of a marker-only quote.
|
|
fills_budget = [{"role": "user", "content": "u" * 3_998}] * 6 # 23,988 of the 24,000 budget
|
|
assert "chars omitted" not in _build_verbatim_user_section([{"role": "user", "content": "v" * 500}, *fills_budget])
|