(cherry picked from commit 3d08e03)

Co-authored-by: 123liuziming <[email protected]>

…ments

DashScope's OpenAI-compatible streaming response can emit tool-call argument
deltas with `arguments=None`, which made `"".join(tool_call.arguments)` raise
`TypeError: sequence item N: expected str instance, NoneType found` during
span finalization and aborted every bfclv4 benchmark run. Filter out None
parts at both legacy and current stream-wrapper join sites.

Change-Id: I76b8e0104dacac1a1ecebd41be74283700d46f2c
Co-developed-by: Claude <[email protected]>
Co-Authored-By: Claude Opus 4.7 <[email protected]>

…hecker

multi_turn_checker re-exports execute_multi_turn_func_call and invokes it
twice per entry during evaluation: once to replay the model's tool-call
trace, once to replay the ground-truth trace. Both calls run *after*
inference, outside any ENTRY/AGENT/STEP context, so each one produced a
trace-rooted orphan TOOL span. Drop checker from the wrap targets; the
two inference-side bindings (multi_turn_utils source module + base_handler
re-export) still cover every TOOL span we actually want.

Change-Id: Ife24d8ba2595fc2c10a0dcfc47de5521ad67d3c7
Co-developed-by: Claude <[email protected]>
Co-Authored-By: Claude Opus 4.7 <[email protected]>

Wrapping multi_turn_utils.execute_multi_turn_func_call replaces the
source module attribute, which is fine at instrument time but causes
orphan TOOL spans later: bfcl_eval/__main__.py lazily imports
multi_turn_checker.py during `bfcl evaluate`, and that import resolves
the wrapper into checker's local binding. Ground-truth replay then
emits TOOL spans outside any ENTRY/AGENT/STEP context.

Wrap only base_handler's local binding, which is set during
BaseHandler.inference wrap in step 2 and is the sole inference-time
caller of execute_multi_turn_func_call.

Change-Id: I9145484b1fa9b8bf9cc3899b6a551aec62b856ac
Co-developed-by: Claude <[email protected]>
Co-Authored-By: Claude Opus 4.7 <[email protected]>

* ENTRY span no longer mirrors OpenInference input.value / output.value;
  the same payload is still on gen_ai.input.messages / gen_ai.output.messages.
* Bump _to_jsonable max_depth from 3 to 8 so AGENT message parts (and
  their tool_call arguments / tool_call_response results) serialize as
  real JSON objects instead of Python repr strings with single quotes.

Change-Id: I248356773980d8688e4d76336d92f3935c2e15b2
Co-developed-by: Claude <[email protected]>
Co-Authored-By: Claude Opus 4.7 <[email protected]>

… to AGENT span

The AGENT span now records the system prompt template and tool schema,
enabling full observability of agent configuration in trace data.

Change-Id: I02509ebd7e6a2ff1bdaa64d257fdb231df7f36b5
Co-developed-by: Claude <[email protected]>
Co-Authored-By: Claude Opus 4.6 <[email protected]>

… AGENT span

ENTRY span now captures problem_names/model from CLI kwargs. Runner ENTRY
span falls back to run_spec.template when problem fields are absent. AGENT
span extracts system_prompt from instance.system_template, instance.system_prompt,
or the first system message in _messages/_steps.

Change-Id: I222c187fc780aefe96d64f452a13a22f0bef41ae
Co-developed-by: Claude <[email protected]>
Co-Authored-By: Claude Opus 4.6 <[email protected]>

… STEP span

When using the Python API (DefaultAgent.run() directly), no ENTRY span
was produced because the existing ENTRY was only on the CLI Typer path.
Now DefaultAgentRunWrapper auto-creates an ENTRY span (with input
messages) when one is not already active, using a ContextVar guard to
avoid double ENTRY when invoked via the `mini` CLI.

Also suppress the empty STEP span that occurs when step/cost limits are
already exceeded — the step immediately raises LimitsExceeded with no
LLM or TOOL work, producing noise in traces.

Change-Id: Iae01f70a3c0f7b3a0daa8c03e35e9d482b9fa8d8
Co-developed-by: Claude <[email protected]>
Co-Authored-By: Claude Opus 4.6 <[email protected]>

…ions to AGENT span

The AGENT span now records the system prompt (Current Date from env_info)
and available tool schemas from the test entry, enabling full observability
of agent configuration in trace data.

Change-Id: I64ed458018f5f138a06787b1d8cd938e662a8284
Co-developed-by: Claude <[email protected]>
Co-Authored-By: Claude Opus 4.6 <[email protected]>

WebArena's agent/__init__.py re-exports construct_agent via
`from .agent import construct_agent`. When wrapt wraps the function on
the agent.agent module, the cached reference in agent's namespace remains
unwrapped. This causes `from agent import construct_agent` (the common
import pattern in run.py) to bypass instrumentation.

Add ("agent", "construct_agent") to _PATCH_TARGETS so both references
are wrapped and the AGENT(create_agent) span fires regardless of import
style.

Change-Id: I443c244fa9bb630491852849bea9a75f66d33d58
Co-developed-by: Claude <[email protected]>
Co-Authored-By: Claude Opus 4.6 <[email protected]>

…, tool.definitions per ARMS spec

- Fix capture_message_content() to accept SPAN_ONLY as a valid truthy
  value (matching algotune/wildtool behavior)
- Add gen_ai.input.messages (user intent + previous action) to AGENT span
- Add gen_ai.output.messages (LLM raw prediction) to AGENT span
- Add gen_ai.system_instructions (from PromptConstructor intro) to AGENT span
- Add gen_ai.tool.definitions (browser action types) to AGENT spans
- Add input.mime_type / output.mime_type to TASK span

Implements the field requirements from:
https://help.aliyun.com/zh/arms/application-monitoring/developer-reference/llm-trace-field-definition-description

Change-Id: If233e99bb7354fd5f8ff6e23891d9fe4e95a2503
Co-developed-by: Claude <[email protected]>
Co-Authored-By: Claude Opus 4.6 <[email protected]>

Per ARMS spec, AGENT spans should only carry gen_ai.input.messages,
gen_ai.output.messages, gen_ai.system_instructions, and
gen_ai.tool.definitions — not the generic input.value/output.value.

Change-Id: I7861f4f00a41e2955097343ba032e22a9fa42c1f
Co-developed-by: Claude <[email protected]>
Co-Authored-By: Claude Opus 4.6 <[email protected]>

…CHAIN

1. ENTRY span now records gen_ai.input.messages (task intent) and
   gen_ai.output.messages (initial observation after env.reset)
2. CHAIN span now records output.value with step/tool summary on close
3. Config file reader handles list format (webarena uses [config] arrays)

Change-Id: I12e1d4e2022a0fd1882f11edfb87bf3e25c2a84c
Co-developed-by: Claude <[email protected]>
Co-Authored-By: Claude Opus 4.6 <[email protected]>

ENTRY span should only use gen_ai.input.messages, not input.value.

Change-Id: I76d32679c40107adec84986ef6d8de01bfb06a0a
Co-developed-by: Claude <[email protected]>
Co-Authored-By: Claude Opus 4.6 <[email protected]>

Change-Id: I6a71a6150bb92d336ef19088ef521b6dff5be2ab

Change-Id: I604506992bda330e7ff62705e50ca79173bf8f3b