Pull request overview

Adds a new loongsuite-instrumentation-deepagents package to instrument the deepagents framework with ENTRY span creation, span enrichment, and GenAI metrics emission.

Changes:

• Introduces a DeepAgentsInstrumentor that wraps deepagents.graph.create_deep_agent, installs a LangChain callback enricher, and registers a metrics SpanProcessor.
• Adds internal helpers for extracting metadata/messages and for generating ENTRY spans + metrics.
• Adds a dedicated test suite plus packaging/docs (pyproject, README) and registers the package in the instrumentation index.

Reviewed changes

Copilot reviewed 16 out of 16 changed files in this pull request and generated 7 comments.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

Review by github-manager-bot

Summary

New instrumentation for deepagents (built on LangGraph): wraps create_deep_agent and the resulting graph's invoke/ainvoke/stream/astream to produce ENTRY spans, propagates a subagent registry via contextvars, patches the subagent task tool, and adds a langchain-tracer callback + subagent task-context resolution. Substantial and well-tested. Since it touches the shared langchain tracer, please give those changes a careful eye.

Note: PR is marked [WIP]; focusing on design-level feedback that should survive further iteration.

Findings

• [Warning] _entry_patch.py — _SubagentRunnableProxy overrides only invoke and ainvoke, while __getattr__ delegates everything else (incl. stream/astream) to the raw runnable. The top-level graph wraps all four methods, so streaming subagent graphs silently miss the ReAct-metadata injection that sync/async subagents get — inconsistent telemetry coverage depending on how the subagent is invoked. Either wrap stream/astream in the proxy too, or drop the proxy in favour of the same _wrap_graph_methods used at the top level.
• [Warning] _entry_patch.py — streaming ENTRY span lifecycle. The stream wrappers yield in a try with _finish_entry in except Exception, but no finally. If a consumer abandons a stream early (partial consume, no close()/with), the generator is closed by GeneratorExit, which is a BaseException and is not caught by except Exception → _finish_entry is skipped → the ENTRY span is never stop_entry/fail_entry'd (leak). Add a finally (with a guard to avoid double-finish) so the entry span always ends.
• [Info] _entry_patch.py uses module-level mutable flags (_is_entry_patched, _top_level_patched, _is_subagent_task_patched, _handler). Concurrent instrument()/uninstrument() would race; document single-threaded setup at startup or guard with a lock.
• [Info] _metrics_processor.py / langchain _tracer.py — _get_deepagents_subagent_task_context iterates list(self._runs.values()) per subagent chain start (O(n) per chain). Bounded and fine for typical traces, but note the cost for very large multi-subagent graphs.
• [Info] Truncated (2-line) Apache headers across these files; align with the full header used in e.g. the anthropic plugin for lint consistency.

Suggestions

# in the stream generator wrappers
try:
    yield from inner
finally:
    _finish_entry(invocation, token, result=None, exc=current_exc, finished=bool(...))

Cross-repo Note

No shared API surface with loongsuite-pilot; the langchain-tracer changes are internal to this plugin, no cross-repo change required.

Automated review by github-manager-bot