LangChain Ecosystem Coverage

AgentBridge tracks the LangChain ecosystem from the official recursive documentation index at docs.langchain.com/llms.txt. The current URL snapshot is docs/upstream/langchain-pages.json. The page-level decision ledger is docs/upstream/langchain-coverage.json. It is an audit trail: every page has an explicit status, owner, and next action. Executable native conformance is tracked in docs/upstream/langchain-conformance.json.

Coverage Contract

Every inventoried capability must have one explicit outcome:

Status	Meaning
core	Expressed through framework-neutral AgentBridge types and behavior.
extension	Exposed through a LangChain or LangGraph AgentBridge extension, with adapter tests.
native_only	Available through the preserved raw backend object or native options, but not normalized yet.
planned	A documented gap with an implementation slice and acceptance tests still to add.
unsupported	Deliberately outside the current product boundary, with a reason recorded.

An adapter must not silently discard an upstream option. If a feature cannot be normalized, the adapter should preserve it through an extension configuration, raw backend escape hatch, or an explicit diagnostic.

Current Surface

LangChain area	AgentBridge surface	Current status
LangChain create_agent, tools, retriever tools, structured output, middleware, streaming, runtime context	AgentSpec, ToolSpec, explicit LangChainExtension.config(retriever_tools=...), native stream_options/stream_events_version/stream_protocol, LangChainExtension.context_schema, native context= invocation, plugin configuration, and normalized events	extension
LangChain human-in-the-loop middleware and checkpointed pauses	Native HumanInTheLoopMiddleware, checkpointer pass-through, normalized interrupted workflow event	extension
LangChain MCP tools	Native MCP tool objects supplied through LangChainExtension(mcp_tools=...); optional MCP adapter packages remain external	extension
Newly added LangChain create_agent options	Guarded LangChainExtension.native_options pass-through with diagnostics and offline conformance	native_only until normalized
LangGraph state graphs, validated conditional routing, checkpoints, interrupts, resume, state history/time travel, custom checkpointers, stores, caches, retrievers, native model invocation, sync/async tools, per-run runtime config, batch execution, streaming controls, and native graph escape hatches	Built-in LangGraph adapter, LangGraphExtension.conditional_routing(), get_state/get_state_history/update_state/replay, native batch()/abatch() dispatch, async ToolSpec execution in generated graphs, typed stream_options plus guarded native stream controls, runtime_config forwarding for native RunnableConfig controls, explicit persistence fields, LangGraphConfig.model, LangGraphConfig.native_graph, and guarded native_options compile pass-through	extension
LangSmith tracing, metadata, tags, run names, sessions, trace context	Shared observability helpers and LangChain/LangGraph config; runtime-config conformance is executable offline	extension
Cross-provider trace correlation	Core observability_metadata() maps session, user, project, and trace identifiers consistently for LangChain and LangGraph	extension
OpenAI-compatible, NVIDIA NIM, OpenRouter, local Ollama, Anthropic, Google GenAI/Vertex AI, Mistral, Groq, Cohere, Bedrock, Fireworks, Hugging Face, xAI, Azure OpenAI, and arbitrary LangChain model routes	Lazy provider factories forward model options, or accept any native model object, without moving credentials or provider packages into core	extension
LangChain model retries and fallbacks	LangChainExtension.config(model_retry=..., model_fallbacks=[...]) applies native runnable resilience wrappers to supplied model objects	extension
Langfuse LangChain callback integration	Lazy callback integration in the optional LangChain plugin	extension
LangSmith assistants, graph/schema/subgraph/version introspection, threads, state/checkpoints/pruning, runs, stateless/background/batch execution, crons, interrupts, long-term store, Fleet/Managed Agents, Agent Auth connections, platform tools, A2A JSON-RPC, MCP transport, system health/info/docs/metrics, and SSE APIs	Typed sync/async lifecycle/search/patch/copy/events/join/cancellation/checkpoint/pruning/stateless-run/cron/store/agent/connection/tool-registry/protocol/system helpers over dependency-free JSON/SSE transports plus RemoteGraphClient lifecycle/state helpers	extension
Langfuse traces, OTLP ingestion, observations, scores, metrics, prompts, experiments, datasets, dataset items, score configs, annotation queues, and evaluation reports	Typed sync/async query and lifecycle helpers, native numbered-page dataset/dataset-item/prompt/config/queue iteration, cursor-based experiment/observation/score iteration, reviewer assignments and queue-item lifecycle, config-linked report publication, LangfusePrompt, plus langfuse_evaluation bridge over the dependency-free JSON/SSE/raw transport	extension
Remote LangGraph/Agent Server runs	RemoteGraphClient maps native thread/run SSE events to AgentEvent and RunResult, and exposes thread search/history/copy/pruning, checkpoint snapshots, assistant graph/schema/version discovery, run inspection/events/join/cancel/delete, interrupt resolution, and resume helpers	extension
AgentCore Memory/Gateway bindings	LangChainExtension.agentcore plus credential-free conformance and AgentCore plugin contracts	extension
Async LangChain execution	Native ainvoke path normalized through arun_agent	extension
LangChain batch and async-batch execution	Public batch_agent()/abatch_agent() with native batch()/abatch() dispatch and ordered fallback	extension
Async normalized streaming	Public astream_agent() contract, LangChain native astream/astream_events, and LangGraph native astream with fallback	extension
LangSmith datasets, examples, evaluators, prompts, experiments, monitoring, feedback, REST API, governance, and deployment control plane	Evaluation contract, typed dataset/example CRUD and cursor iteration, bulk example deletion, OpenAI fine-tuning export, dataset version reads/diffs, split updates, public share/unshare, regression deltas, shared-dataset examples with runs, annotation queues and human-review run assignment, feedback configuration CRUD and presigned feedback tokens, native feedback correlation/governance fields, SDK or dependency-free API-client dataset publishing/evaluation bridge, prompt pull/push helpers, typed feedback/thread/MCP/deployment clients, and generic API transport; remaining hosted APIs stay native	extension
Deep Agents, sandboxes, filesystem backends, permissions, skills, interpreters	External agentbridge-deepagents plugin maps create_deep_agent, native backends, skills, memory, subagents, permissions, HITL, persistence, structured output, and native options	extension
LangChain deployment, Agent Server, Studio, Fleet, Managed Deep Agents, and listeners	Typed LangSmith deployment/revision control-plane operations including redeploy, interruption, bulk deletion, logs, and tier updates; v2 listener lifecycle; remote Agent Server thread/run client, Fleet/Managed Deep Agent CRUD, and generic transport; Studio UI and hosted publishing remain native	extension or native_only depending on page
LangChain LLM Gateway and provider administration	Model routing remains delegated to LiteLLM-style strings; no LangSmith gateway control plane	native_only
LangChain TypeScript documentation	Python SDK scope	unsupported for the current SDK

Refresh And Review

Refresh the snapshot intentionally after reviewing new pages:

python scripts/crawl_langchain_docs.py \
  --output docs/upstream/langchain-pages.json
python scripts/build_langchain_coverage.py

Check for upstream drift without changing the repository:

python scripts/crawl_langchain_docs.py \
  --check --fail-on-drift
python scripts/build_langchain_coverage.py --check

When drift is found, review each added or removed page, update the capability ledger and adapter tests, then refresh the snapshot in the same change. The weekly Upstream Compatibility workflow performs this check and uploads the inventory when it fails.

Next LangChain Slices

Run the credential-free native conformance lane:

uv run pytest -q plugins/agentbridge-langchain/tests/test_native_conformance.py

This lane proves translation and normalization for the current native-only surface. It does not claim provider, hosted deployment, or remote LangSmith behavior; those require separate credentialed smoke lanes. It now also checks native async execution/streaming contracts, AgentCore binding preservation, and LangSmith/Langfuse provider option preservation without loading credentials.

Run the explicitly credentialed observability lane:

AGENTBRIDGE_RUN_CREDENTIAL_SMOKE=1 \
  LANGSMITH_API_KEY=... LANGFUSE_PUBLIC_KEY=... LANGFUSE_SECRET_KEY=... \
  python examples/observability_smoke.py

The lane reports LangSmith API authentication and Langfuse callback/runtime wiring separately from the credential-free native conformance report.

The first pass is intentionally ordered by value to application developers:

1. Complete LangChain/LangGraph model, tool, middleware, streaming, structured-output, memory, retrieval, and human-in-the-loop option forwarding.
2. Use the new core EvaluationExample/EvaluationReport contract for provider-neutral offline evaluation, then add LangSmith dataset/evaluator publishing.
3. Expand LangSmith prompt/version and experiment integration while preserving raw clients; Langfuse prompt retrieval/version compilation is now executable.
4. Expand OpenTelemetry-compatible trace export and first-class Langfuse/LangSmith trace correlation.
5. Add typed remote LangGraph/Agent Server operations beyond thread/run streaming, including assistants, checkpoints, store access, interrupts, and deployment lifecycle APIs.
6. Expand the Deep Agents plugin with credentialed sandbox/interpreter lanes and hosted deployment checks; the local translation contract is now executable.

This order keeps the core dependency-free while still giving every upstream feature a visible place in the roadmap and compatibility review.