From 356427d43c44bdaa285b3ad1a7af3b4bfe2ba1af Mon Sep 17 00:00:00 2001 From: Florian Brand <46377542+xeophon@users.noreply.github.com> Date: Wed, 18 Nov 2026 17:05:11 -0800 Subject: [PATCH] Match Claude replays that drop standalone blank text blocks --- verifiers/v1/dialects/anthropic.py | 11 +++++++++-- verifiers/v1/harnesses/claude_code/harness.py | 7 +++---- verifiers/v1/harnesses/utils/replay.py | 19 +++++++++++++++++++ verifiers/v1/types.py | 4 ++++ 4 files changed, 35 insertions(+), 6 deletions(-) diff --git a/verifiers/v1/dialects/anthropic.py b/verifiers/v1/dialects/anthropic.py index 0939f00736..520a339e7c 100644 --- a/verifiers/v1/dialects/anthropic.py +++ b/verifiers/v1/dialects/anthropic.py @@ -10,6 +10,7 @@ import re from collections.abc import Callable, Mapping from dataclasses import dataclass, field +from itertools import accumulate from anthropic.types import Message as AnthropicMessage from anthropic.types import Usage as AnthropicUsage @@ -241,6 +242,10 @@ def content_to_wire(content) -> str | list[dict]: return blocks +def _text_block_offsets(blocks: list[str]) -> list[int] | None: + return list(accumulate(map(len, blocks[:-1]))) or None + + def parse_messages(body: dict) -> Messages: """The request's top-level `system` + `messages` -> typed messages. Assistant turns fold their blocks into one message (thinking -> reasoning, tool_use -> tool calls); a user turn's @@ -260,7 +265,7 @@ def parse_messages(body: dict) -> Messages: block for block in blocks if block["type"] not in ("text", "tool_use") ] state.sort(key=lambda block: THINKING_ORDER.get(block["type"], 2)) - text = "".join(b.get("text", "") for b in blocks if b.get("type") == "text") + text = [b.get("text", "") for b in blocks if b.get("type") == "text"] reasoning = "".join( b.get("thinking", "") for b in blocks if b.get("type") == "thinking" ) @@ -276,7 +281,8 @@ def parse_messages(body: dict) -> Messages: ] prompt.append( AssistantMessage( - content=text or None, + content="".join(text) or None, + text_block_offsets=_text_block_offsets(text), reasoning_content=reasoning or None, tool_calls=calls or None, provider_state=state or None, @@ -351,6 +357,7 @@ def response_from_wire(message: AnthropicMessage) -> Response: model=message.model, message=AssistantMessage( content="".join(content) or None, + text_block_offsets=_text_block_offsets(content), reasoning_content="".join(reasoning) or None, tool_calls=calls or None, provider_state=state or None, diff --git a/verifiers/v1/harnesses/claude_code/harness.py b/verifiers/v1/harnesses/claude_code/harness.py index 35451a0449..5bb2dd5807 100644 --- a/verifiers/v1/harnesses/claude_code/harness.py +++ b/verifiers/v1/harnesses/claude_code/harness.py @@ -8,7 +8,7 @@ from verifiers.v1.configs.harness import DialectName, HarnessConfig, PinnedVersion from verifiers.v1.harnesses.node import NODE_BIN_DIR, ensure_node from verifiers.v1.harnesses.utils.install import ensure_installed, remove_dir -from verifiers.v1.harnesses.utils.replay import drop_blank_text +from verifiers.v1.harnesses.utils.replay import drop_blank_text_blocks from verifiers.v1.runtimes import Runtime from verifiers.v1.task import TaskData from verifiers.v1.trace import Trace @@ -44,10 +44,9 @@ class ClaudeCodeHarness(ACPHarness[ClaudeCodeHarnessConfig]): SUPPORTS_SKILLS = True SUPPORTS_TOOL_INTERCEPTION = True - # Claude Code drops whitespace-only text blocks on replay. Closed source: observed - # with 2.1.314, recheck on version bumps. + # Claude Code 2.1.327 filters each whitespace-only text block on replay. def replay_view(self, message: Message, dialect: DialectName | None) -> Message: - return drop_blank_text(message) + return drop_blank_text_blocks(message) async def setup(self, runtime: Runtime) -> None: await ensure_node(runtime) diff --git a/verifiers/v1/harnesses/utils/replay.py b/verifiers/v1/harnesses/utils/replay.py index 8f1886f4f0..7810b75850 100644 --- a/verifiers/v1/harnesses/utils/replay.py +++ b/verifiers/v1/harnesses/utils/replay.py @@ -2,6 +2,8 @@ `Harness.replay_view` overrides. Each function is idempotent and only normalizes whitespace, so the replayed and the sampled form of a turn map to the same view.""" +from itertools import pairwise + from verifiers.v1.types import AssistantMessage, Message @@ -17,6 +19,23 @@ def drop_blank_text(message: Message) -> Message: return message +def drop_blank_text_blocks(message: Message) -> Message: + """Drop independently blank blocks while preserving whitespace inside other blocks.""" + if not isinstance(message, AssistantMessage) or not message.text_block_offsets: + return drop_blank_text(message) + text = message.content or "" + offsets = (0, *message.text_block_offsets, len(text)) + if any(start > end for start, end in pairwise(offsets)): + return drop_blank_text(message) + blocks = (text[start:end] for start, end in pairwise(offsets)) + return message.model_copy( + update={ + "content": "".join(block for block in blocks if block.strip()) or None, + "text_block_offsets": None, + } + ) + + def strip_text(message: Message) -> Message: """Clients that store assistant text stripped, and pad a missing reasoning field (or Anthropic thinking block) with a single space on replay.""" diff --git a/verifiers/v1/types.py b/verifiers/v1/types.py index 9961ef4ea1..d67782c39a 100644 --- a/verifiers/v1/types.py +++ b/verifiers/v1/types.py @@ -136,6 +136,10 @@ class ToolCall(BaseModel): class AssistantMessage(BaseModel): role: Literal["assistant"] = "assistant" content: str | None = None + text_block_offsets: list[int] | None = Field( + default=None, exclude_if=lambda value: value is None + ) + """Character offsets between native text blocks concatenated into content.""" reasoning_content: str | None = None tool_calls: list[ToolCall] | None = None provider_state: list[dict[str, Any]] | None = None