feat(templates): add the agent-python-langchain template

[tejaskash]

Problem

The refactor branch has no LangChain template. The legacy LangGraph template used the deprecated create_react_agent, unbounded >= pins, and Handlebars branches (filesystem mounts, config bundle, gateway MCP client) that the refactor never renders.

Solution

agent-python-langchain: a Bedrock-backed LangChain agent on AgentCore Runtime, selectable with --template and in the create wizard, on both project create and project add runtime.

• Built with create_agent from langchain.agents (the langgraph source marks create_react_agent deprecated in favour of it). Model from init_chat_model(..., model_provider="bedrock_converse") in model/load.py, same Claude Sonnet 4.5 profile as the Strands template.
• HTTP, CodeZip, Bedrock only. No AgentCore Memory, no MCP client, no container preset. The only Handlebars variable is name.
• InMemorySaver checkpointer keyed on the Runtime session id gives multi-turn history within a session. No eviction, since each Runtime session runs in its own microVM.
• Async generator entrypoint streams each model chunk as {node, content: [blocks]}, tool-call chunks included. A missing or non-string prompt yields one {error} event and stops instead of raising.
• OTEL is always on: opentelemetry-instrumentation-langchain in the dependencies, no explicit instrument call, picked up by ADOT when the L3 construct wraps the CodeZip entrypoint in opentelemetry-instrument.
• Versions were resolved by installing the latest of everything on Python 3.14 and running the agent against Bedrock, then pinned with ~=: langchain 1.4.0, langgraph 1.2.11, langchain-aws 1.7.5, bedrock-agentcore 1.22.0, aws-opentelemetry-distro 0.19.0, opentelemetry-instrumentation-langchain 0.62.3.
• The generated README has an "Extending" section covering tools, middleware, a durable checkpointer, and structured output.

CLI side: one langchain framework value, one preset, one resolver keyed langchain/Python/HTTP, one wizard entry.

Two adjacent changes came out of testing:

• parseAgentEvent in the Agent Inspector proxy only knew the Strands ConverseStream shape, {text}, {error}, and bare strings, so the LangChain stream showed "(empty response)" in the inspector chat. It now joins the text blocks of a {content: [...]} chunk into a text frame and drops tool-call-only chunks.
• The wizard template list is ordered by type, then language, then framework, and the (CodeZip build) / (container build) suffixes are gone since the names carry that. Every row fits on one line at 110 columns.

Recording the manifest snapshot also pruned an orphaned Strands AG-UI snapshot whose test was removed in #2214.

Verification

Unit: bun test (2993 pass), tsc, oxlint, prettier --check.

• Dev and Deploy

[agentcore-devx-automation]

Claude Security Review: no high-confidence findings. (run)

[agentcore-devx-automation]

AgentCore Harness Review

Verdict: Looks good

Nice, tight addition. I ran through the diff and spot-checked the runtime behavior:

• Verified experimentally with langgraph~=1.2.11 / langchain~=1.4.0 that agent.astream(..., stream_mode="messages", version="v2") yields {"type":"messages","data":(AIMessageChunk, metadata)} dicts, so the event["data"] destructure in main.py is correct. Without version="v2" it would yield bare tuples — worth calling out because it's a load-bearing kwarg, but the current code is right.
• Confirmed AIMessageChunk.content_blocks normalizes plain string content into [{type:"text", text:...}], so both string-content Bedrock models and native block outputs land as text through parseAgentEvent.
• init_chat_model("...", model_provider="bedrock_converse") resolves against langchain-aws~=1.7.5 fine (only AWS credential/region errors surface, which is the deploy-time contract).
• The parseAgentEvent extension is properly ordered after error/text, guarded on Array.isArray, and the new test covering a tool-call-only chunk correctly asserts it's dropped.
• The AG-UI snapshot deletion in manager.test.ts.snap looks alarming at first glance but the commit message and prior #2214 confirm it was already an orphan snapshot; the new LangChain snapshot replaces the slot.
• The reordered TEMPLATE_OPTIONS still keys off agent-python-strands as the default in emptyCreateProjectForm, and the wizard tests' up/up navigation matches the new list order.
• Tests use the real HTTP server harness (sseAgent / ServerFarm) rather than mocks — good.

Non-blocking observations, take or leave:

• context.session_id or "default-session" will AttributeError if session_id is ever missing from the context; agent-python-strands/main.py uses the defensive getattr(context, "session_id", "default-session"). Not a real risk given BedrockAgentCoreApp always populates it, just an inconsistency between the two templates.
• pyproject.toml pins aws-opentelemetry-distro ~= 0.19.0 and bedrock-agentcore ~= 1.22.0, both newer than what agent-python-strands ships (~= 0.18.0 / ~= 1.9.1). Presumably intentional, but worth keeping the templates in sync eventually.

LGTM to merge.

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.97%. Comparing base (23df65f) to head (e1df24a).

Additional details and impacted files

@@            Coverage Diff            @@
##           refactor    #2230   +/-   ##
=========================================
  Coverage     96.97%   96.97%           
=========================================
  Files           559      559           
  Lines         38609    38637   +28     
=========================================
+ Hits          37440    37468   +28     
  Misses         1169     1169           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.

[agentcore-devx-automation]

Claude Security Review: no high-confidence findings. (run)

[agentcore-devx-automation]

Claude Security Review: no high-confidence findings. (run)

[agentcore-devx-automation]

Claude Security Review: no high-confidence findings. (run)

[agentcore-devx-automation]

Claude Security Review: no high-confidence findings. (run)

[agentcore-devx-automation]

Claude Security Review: no high-confidence findings. (run)