September 27, 2026 20:30 · September 27, 2026 20:21 · September 23, 2026 16:50 · September 22, 2026 22:51 · September 22, 2026 22:47 · September 21, 2026 22:46
 """browser-use 0.13.10 — где живёт содержимое файла: <read_state> (1 шаг) vs history (навсегда).

 Harness fix: history items are only appended when model_output is not None
 (message_manager/service.py "Build the history item"), so we pass a real AgentOutput.
 """
 import asyncio, csv, os, tempfile
 from browser_use import Tools
 from browser_use.agent.message_manager.service import MessageManager
 from browser_use.agent.views import ActionResult, AgentBrain, AgentOutput, AgentStepInfo
 from browser_use.filesystem.file_system import FileSystem
 """Repro for headroom #1099 — "negative savings" under technique=preserve.

 Root cause (headroom 0.39.1, commit d13e196), headroom/image/compressor.py:

  line 254  _extract_image_data()   -> docstring: "Extract FIRST image data from messages"
  line 530  _apply_compression()    -> for "preserve", returns messages UNCHANGED
  line 730  original_tokens   = _estimate_tokens(<FIRST image>) + tile_saved
  line 738  compressed_tokens = _count_result_tokens(<ALL images in ALL messages>)

 The two counters measure different sets, so with N images in the transcript the
 """Where does pydantic-ai lose the "provider sent no usage object" signal?

 Context: pydantic/pydantic-ai#8551. Measured on pydantic-ai-slim 2.48.0,
 openai 3.19.0, opentelemetry-sdk 1.44.0, Python 3.12, macOS.

 Part 1 hits the real production mapper (pydantic_ai.models.openai._map_usage)
 with three responses; C is the positive control and must differ from A.
 Part 2 asks whether RequestUsage.details -- a field that already exists --
 survives accumulation into RunUsage.
 Part 3 asks whether it reaches the OpenTelemetry span attributes unaided.
 """Run the guard from mem0 PR #6859 against edge-shaped indexing results.

 The two methods are lifted verbatim out of the PR's own file (fetched from the branch),
 not re-typed, so nothing about this harness can flatter or punish the result.
 """
 import ast, subprocess, sys, textwrap

 RAW = ("https://raw.githubusercontent.com/iroiro147/mem0/"
       "PR_BRANCH/mem0/vector_stores/azure_ai_search.py")
 """Probe: three shapes of usage that a run span may or may not credit.

 A (the issue): an UNSPANNED nested RUN -> parent span over-credits.
 B (adjacent case raised in the thread): a model request made with NO run at all
  (a tool calling the provider layer directly) -> who gets it?
 C: a provider that returns no usage object -> what survives as a signal?
 """

 import asyncio
 """Cross-language parity vector for RFC 8785 (JCS) as used as a payment claim key.

 Context: stripe/ai#402 proposes deriving a duplicate-charge guard's identity from
 the JCS canonicalization of the request, implemented independently in Python and
 Rust and pinned to parity by cross-language tests.

 RFC 8785 section 3.2.3 requires object members to be sorted on their **UTF-16
 code units**. Python's `sorted()` on `str` sorts by **code point**. Those two
 orders are identical for the whole BMP and then invert above it, because a
 supplementary character encodes as a surrogate pair starting at U+D800, which is
 """Parity probe: fork_session() (disk) vs fork_session_via_store() (store).

 ``_derive_title_from_entries`` documents itself as
    "Precedence matches ``_extract_last_json_string_field`` semantics"
 but it iterates PARSED dicts, so it is spelling-blind, while the disk path
 scans RAW TEXT with a two-pattern loop that prefers the spaced spelling.

 If the documented parity holds, both fork entry points must return the same
 title for the same transcript. Same entries, both paths, one comparison.
 """Probe 2 for openai/openai-agents-python#5085.

 Adds what probe 1 was missing: prompt caching hits on the shared PREFIX
 (system instructions + leading input items), not on the key alone. So measure,
 for consecutive calls to the same sub-agent:

  - is system_instructions byte-identical?
  - how many leading input items are byte-identical?

 across four shapes:
 """GIT-S28 bench: mem0ai/mem0#7211 - compare candidate fixes for Azure AI Search
 indexing-result validation on one shared battery. Run once per variant."""
 import json, sys, traceback
 from unittest.mock import MagicMock, Mock, patch

 VARIANT = sys.argv[1]

 from azure.search.documents._generated.models import IndexingResult as _IR

 def make_obj(key, succeeded, status_code, error_message=None):
 diff --git a/src/claude_agent_sdk/_internal/query.py b/src/claude_agent_sdk/_internal/query.py
 index 2841adf..84e619a 100644
 --- a/src/claude_agent_sdk/_internal/query.py
 +++ b/src/claude_agent_sdk/_internal/query.py
 @@ -171,6 +171,7 @@ class Query:
         self._request_counter = 0
 
         # Message stream
 +        self._prompt_error: Exception | None = None
         self._message_send, self._message_receive = anyio.create_memory_object_stream[
	"""browser-use 0.13.10 — где живёт содержимое файла: <read_state> (1 шаг) vs history (навсегда).

	Harness fix: history items are only appended when model_output is not None
	(message_manager/service.py "Build the history item"), so we pass a real AgentOutput.
	"""
	import asyncio, csv, os, tempfile
	from browser_use import Tools
	from browser_use.agent.message_manager.service import MessageManager
	from browser_use.agent.views import ActionResult, AgentBrain, AgentOutput, AgentStepInfo
	from browser_use.filesystem.file_system import FileSystem
	"""Repro for headroom #1099 — "negative savings" under technique=preserve.

	Root cause (headroom 0.39.1, commit d13e196), headroom/image/compressor.py:

	line 254 _extract_image_data() -> docstring: "Extract FIRST image data from messages"
	line 530 _apply_compression() -> for "preserve", returns messages UNCHANGED
	line 730 original_tokens = _estimate_tokens(<FIRST image>) + tile_saved
	line 738 compressed_tokens = _count_result_tokens(<ALL images in ALL messages>)

	The two counters measure different sets, so with N images in the transcript the
	"""Where does pydantic-ai lose the "provider sent no usage object" signal?

	Context: pydantic/pydantic-ai#8551. Measured on pydantic-ai-slim 2.48.0,
	openai 3.19.0, opentelemetry-sdk 1.44.0, Python 3.12, macOS.

	Part 1 hits the real production mapper (pydantic_ai.models.openai._map_usage)
	with three responses; C is the positive control and must differ from A.
	Part 2 asks whether RequestUsage.details -- a field that already exists --
	survives accumulation into RunUsage.
	Part 3 asks whether it reaches the OpenTelemetry span attributes unaided.
	"""Run the guard from mem0 PR #6859 against edge-shaped indexing results.

	The two methods are lifted verbatim out of the PR's own file (fetched from the branch),
	not re-typed, so nothing about this harness can flatter or punish the result.
	"""
	import ast, subprocess, sys, textwrap

	RAW = ("https://raw.githubusercontent.com/iroiro147/mem0/"
	"PR_BRANCH/mem0/vector_stores/azure_ai_search.py")
	"""Probe: three shapes of usage that a run span may or may not credit.

	A (the issue): an UNSPANNED nested RUN -> parent span over-credits.
	B (adjacent case raised in the thread): a model request made with NO run at all
	(a tool calling the provider layer directly) -> who gets it?
	C: a provider that returns no usage object -> what survives as a signal?
	"""

	import asyncio
	"""Cross-language parity vector for RFC 8785 (JCS) as used as a payment claim key.

	Context: stripe/ai#402 proposes deriving a duplicate-charge guard's identity from
	the JCS canonicalization of the request, implemented independently in Python and
	Rust and pinned to parity by cross-language tests.

	RFC 8785 section 3.2.3 requires object members to be sorted on their **UTF-16
	code units. Python's `sorted()` on `str` sorts by code point**. Those two
	orders are identical for the whole BMP and then invert above it, because a
	supplementary character encodes as a surrogate pair starting at U+D800, which is
	"""Parity probe: fork_session() (disk) vs fork_session_via_store() (store).

	``_derive_title_from_entries`` documents itself as
	"Precedence matches ``_extract_last_json_string_field`` semantics"
	but it iterates PARSED dicts, so it is spelling-blind, while the disk path
	scans RAW TEXT with a two-pattern loop that prefers the spaced spelling.

	If the documented parity holds, both fork entry points must return the same
	title for the same transcript. Same entries, both paths, one comparison.
	"""Probe 2 for openai/openai-agents-python#5085.

	Adds what probe 1 was missing: prompt caching hits on the shared PREFIX
	(system instructions + leading input items), not on the key alone. So measure,
	for consecutive calls to the same sub-agent:

	- is system_instructions byte-identical?
	- how many leading input items are byte-identical?

	across four shapes:
	"""GIT-S28 bench: mem0ai/mem0#7211 - compare candidate fixes for Azure AI Search
	indexing-result validation on one shared battery. Run once per variant."""
	import json, sys, traceback
	from unittest.mock import MagicMock, Mock, patch

	VARIANT = sys.argv[1]

	from azure.search.documents._generated.models import IndexingResult as _IR

	def make_obj(key, succeeded, status_code, error_message=None):
	diff --git a/src/claude_agent_sdk/_internal/query.py b/src/claude_agent_sdk/_internal/query.py
	index 2841adf..84e619a 100644
	--- a/src/claude_agent_sdk/_internal/query.py
	+++ b/src/claude_agent_sdk/_internal/query.py
	@@ -171,6 +171,7 @@ class Query:
	self._request_counter = 0

	# Message stream
	+ self._prompt_error: Exception \| None = None
	self._message_send, self._message_receive = anyio.create_memory_object_stream[