[jeffdupont]

Python half of AIC-3598. The spec change is launchdarkly/ai-sdks-monorepo#91.

Packaging

• Classifiers. All 14 packages say Development Status :: 5 - Production/Stable. The two vercel-* packages had no classifiers at all.

• Third-party floors move to the version CI tests wherever the old floor crossed a breaking line (a major for 1.x+ libraries, a minor for 0.x):

  • langchain-core>=1.6, langgraph>=1.2, openai>=2.44, openai-agents>=0.17
  • anthropic>=0.113, httpx>=0.28, typesafe-sdk>=0.7

  The rule, as the shared spec states it: a floor sits on the line CI tests, the same major as the tested version for a 1.x+ library and the same minor for a 0.x one. The README says so.

  Floors already on the tested major stay put (opentelemetry-api>=1.25, litellm>=1.83).

• [ld] extra on launchdarkly-ai-server and launchdarkly-ai-python installs launchdarkly-server-sdk>=9.18.3,<10. It's an extra rather than a dependency so bring-your-own-client installs stay lean, as client/AGENTS.md requires. The READMEs and the init_client error message point to it.

• Undeclared LangChain provider packages, found by the new check. Both langchain packages imported langchain_openai and langchain_anthropic without declaring them, the Python side of the JS @langchain/anthropic gap:

  • langchain-agents now depends on langchain-openai, its default model, and offers [anthropic] and [aws] extras.
  • langchain-messages now depends on langchain-openai, its default model, as langchain-agents does, and offers [anthropic] and [aws].
  • The Anthropic and OpenAI imports are now guarded with an install hint, as Bedrock's already was.

• Internal ranges: every package pins launchdarkly-ai-server>=0.6,<2. The native graph adapters import NativeGraphRunner, which published server 0.5.0 lacks, so >=0.4.1 would let a handler resolve to a core it can't import from. >=0.6 resolves whichever lands first, this PR's 0.6.0 or the 1.0.0 cut; chore(AIC-3672): release every Python package as 1.0.0 #242 (AIC-3672) tightens it to >=1.0.0,<2.

Public types

• LDClientInterface becomes a Protocol. It was a plain class with async methods, which matched neither the synchronous ldclient.LDClient nor the "maybe awaitable" way core calls it. ldclient.LDClient now satisfies it structurally; checked with mypy in both directions (a client missing track is rejected). init_client and get_client return it. variation, flush and close may be async; track must be synchronous (-> None), because no call site awaits it.
• tool_handlers on the handler protocol and in every handler is typed Mapping[str, Callable[..., Any] | NativeTool] instead of dict[str, Any], so a caller's precisely typed dict is accepted. Offline evaluations pass a NativeTool to the handler unwrapped, so langchain-agents now checks callable() before calling. It used to raise TypeError: 'NativeTool' object is not callable; there's a test, and it fails without the fix.
• Native graph adapters (to_lang_graph, to_openai_agents, to_claude_agents, to_litellm_agents, to_vercel_agents):
  • take Awaitable[GraphDefinition]
  • take a per-package options TypedDict exported from the root
  • return the new core NativeGraphRunner protocol. They return a small class rather than a SimpleNamespace, because mypy accepts any SimpleNamespace as any protocol.

Packaging check (make check-packaging, new CI job packaging)

scripts/check_packaging.py installs each wheel alone, with no extras, into an empty venv, which proves core loads without launchdarkly-server-sdk. It imports every public module and __all__ name. Then it checks that every module the source imports unconditionally, lazy importlib.import_module calls included, comes from a distribution the package declares; installed through some other dependency is not enough. That found undeclared pydantic in four packages and openai in litellm-agents, now declared. Each declared extra gets its own venv and must import what EXTRA_IMPORTS lists for it; an extra with no entry fails. Internal dependencies resolve to this commit's wheels and third-party ones from PyPI. Each import is resolved by its full dotted name to the distribution that owns it, so a namespace package such as opentelemetry.sdk needs opentelemetry-sdk declared; that found langgraph.prebuilt undeclared in langchain-agents. The 22 runs go through a thread pool.

Checks

• make lint, make format-check, make typecheck (no issues in 104 files), make api-report-check (15 reports current)
• make test: 3,651 passed, 18 skipped
• make check-packaging: all 14 packages and 8 extras pass; removing pydantic from vercel-agents makes it fail with pydantic (from pydantic, not declared)

Versioning

The README documents the 1.0 rule: from 1.0.0, every package is released at the same version.

🤖 Generated with Claude Code

---

Note

Overview
Prepares the Python AI SDK monorepo for GA by tightening packaging, install docs, and public types, plus a CI gate that catches undeclared dependencies workspace tests miss.

Packaging and installs. All packages move to Production/Stable classifiers; third-party floors are raised to versions CI exercises (LangChain 1.x, OpenAI 2.x, etc.). launchdarkly-ai-server[ld] and launchdarkly-ai-python[ld] now install bounded launchdarkly-server-sdk; READMEs drop separate launchdarkly-server-sdk lines. Handlers pin launchdarkly-ai-server>=0.6,<2. LangChain packages declare langchain-openai and add [anthropic] / [aws] extras with guarded import errors; missing deps like pydantic and langgraph-prebuilt are added where the new checker found gaps.

make check-packaging + CI packaging job. New scripts/check_packaging.py builds wheels, installs each package (and each extra) in an isolated venv, validates public __all__ imports, and asserts every unconditional/lazy import is owned by a declared distribution—not a transitive install.

Public API typing. LDClientInterface is a Protocol aligned with real ldclient usage; get_client / init_client return it. tool_handlers is Mapping[str, Callable[..., Any] | NativeTool] across core and handlers (LangChain agents reject unwrapped NativeTool). Native graph to_* adapters take Awaitable[GraphDefinition], export per-package options TypedDicts, and return NativeGraphRunner instead of untyped namespaces.

README adds 1.0 synchronized versioning policy for the SDK.

^{Reviewed by Cursor Bugbot for commit bdd635b. Bugbot is set up for automated code reviews on this repo. Configure here.}

[XieX]

A few comments from a review pass, each checked against the code at this head:

• The biggest one is the LDClientInterface.track() contract (types.py): the new docstring and return type allow an async track, but the SDK never awaits it.
• The rest are about packaging-check coverage, tool_handlers typing, dependency floors, install hints and CI speed.

[jeffdupont]

Re-review of 1304692. Four of the six earlier points are resolved: track() is synchronous again, tool_handlers is a Mapping, the extras are installed and import-checked, and the install hints name the extras. A mypy probe against the head types accepts a real ldclient.LDClient, rejects a client with async def track, and accepts a dict[str, Callable[[dict[str, Any]], str]] passed as tool_handlers. I found no call site that awaits track(), and none that mutates tool_handlers.

Three things are still open:

1. The packaging check can't fail on an undeclared import that another dependency installs. Today pydantic is imported unconditionally by openai-agents, litellm-agents, vercel-agents and vercel-messages, and openai by litellm-agents, and none of them declares it. TESTING.md section 5.3 in ai-sdks-monorepo#91 requires each one to be declared.
2. Core's import scan only runs in a venv that has [ld] installed. An unguarded import ldclient in core would pass, even though section 5.3 says core must load without the base SDK.
3. The internal floor is still launchdarkly-ai-server>=0.4.1. No 0.4.1 release exists, and the handlers now import NativeGraphRunner, which the published 0.5.0 does not have. #242 raises the floor to >=1.0.0,<2, so this PR is only safe if #242 merges before the next release PR.

The floors policy (your point 4) is now written down, but its wording disagrees with the spec text and with some of the floors. See the inline comment on README.md.

[cursor]

Stale Bugbot comment from a previous run.

[XieX]

Thanks for working through all of this. I checked each fix at this head: track() is synchronous and documented that way, tool_handlers is a Mapping with the copies gone, every extra is installed and import-checked, the install hints name the extras, the checks run in parallel, and the floor rule is written down. The declared-distribution check, the core-without-extras run and the >=0.6,<2 internal floor all look right too. Release PR #231 already takes core to 0.6.0, so the floor resolves whichever release goes out first.

Approving. A few non-blocking nits, none of which need to hold this up:

1. Private-spec reference. The module docstring of tests/test_check_packaging.py (line 1) cites the private spec document and a section number. #234, merged to main today but not yet in this branch, removed every such reference from the .py files, so merging this would add one back. Stating the rule in plain words, as #234 did elsewhere, would fix it. The PR description cites section numbers as well.
2. Namespace packages in the declared-import check (latent). The scan reduces each import to its top-level name, so a module like opentelemetry maps to several distributions. A future unguarded lazy from opentelemetry.sdk ... in core would pass, because opentelemetry-api is declared, even though opentelemetry-sdk is only in [otel]. Every such import in core is guarded today, so nothing is broken. Mapping the full dotted module to its distribution (or special-casing namespace packages) would close the gap.
3. LangChain default-model consistency. Both LangChain packages default to OpenAI, but langchain-agents depends on langchain-openai directly while langchain-messages offers it as [openai]. A bare pip install launchdarkly-ai-langchain-messages therefore fails on its default path, though with a clear hint. If that difference is intentional, ignore this.

[jeffdupont]

Thanks @XieX. All three nits are in f3fbe8a:

1. Private-spec reference. The test module docstring no longer cites the spec, and the PR description no longer cites section numbers. I also merged main, so fix(AIC-3568)!: give the barrel core's __all__, and drop private-spec references #234's cleanup is in this branch; a search of the .py files and scripts/ for the spec name or section signs now finds nothing.
2. Namespace packages. The scan keeps each import's full dotted name, and inside the venv each module resolves to the distribution that owns its file, falling back to the top-level lookup only for a namespace leaf with no file. So opentelemetry.sdk needs opentelemetry-sdk declared. It caught a real one straight away: langchain-agents imports langgraph.prebuilt, which comes from langgraph-prebuilt, a dependency of langgraph that we didn't declare. It's now declared at >=1.1. test_a_namespace_package_needs_its_own_distribution_declared covers the case.
3. LangChain default model. Not intentional. langchain-messages now depends on langchain-openai, as langchain-agents does, so a bare install works on the default path. The [openai] extra is gone (it was never released), and the README says so.

Local: make test 3,711 passed, make check-packaging 22/22 (14 packages, 8 extras).

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, have a team admin enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit f3fbe8a. Configure here.}