[andrewklatzke]

Summary

• Add launchdarkly-ai-vercel-messages and launchdarkly-ai-vercel-agents packages that map LaunchDarkly AI Configs onto the official Python ai runtime and AI Gateway creator/model ids.
• Execute tool loops, structured output, usage accumulation, and native graph handoffs instead of reading nonexistent stream fields or stopping at the first tool call.
• Include runnable examples for gateway, injected-OpenAI, and native-graph paths, plus release-please and workspace wiring.

Test plan

• make format-check && make lint && make typecheck && make test in python/
• Run uv run python main.py vercel-agents launch-darkly-documentation-summarizer "..." and confirm a non-empty response, non-zero usage, and a judge score
• Run vercel-messages and vercel-direct success/failure keys from integration-config.json
• Confirm native-graph-vercel fails cleanly if the Gateway free tier still blocks the graph model

Made with Cursor

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Stale Bugbot comment from a previous run.

[cursor]

Stale Bugbot comment from a previous run.

[jeffdupont]

Reviewed with the 1.0 freeze in mind. Despite the name, this isn't a wire format or a JS bridge. It wraps Vercel's own Python runtime, the ai package on PyPI (vercel-labs/ai-python, 0.7.0, classifier "Development Status :: 4 - Beta"), and it's the Python half of launchdarkly/js-ai-sdk#83, which merged on 29 Sep. The public names match JS one for one (vercel_messages, create_vercel_messages_handler, vercel_agents, create_vercel_agents_handler, vercel_graph, to_vercel_agents), and model_id.py is a line-for-line port of JS model-id.ts. Tests pass at 89c52d5: make test 1523 passed, 15 skipped; make typecheck and make lint clean, all exit 0. The branch conflicts with main only in .release-please-manifest.json, because main has since released 0.2.4.

Three things I'd like settled before GA:

1. An AI Config can replace the Gateway API key and the model. _request_params forwards extra_headers and extra_query from model.parameters and puts every other unknown key into extra_body. The Gateway client applies those headers after its own (ai/providers/ai_gateway/client/_client.py, request_headers.update(headers)). I reproduced it at 89c52d5 with the HTTP call stubbed. A config with extra_headers: {Authorization: "Bearer attacker-key", ai-language-model-id: "openai/gpt-5-pro"} sends that Authorization header in place of the customer's AI_GATEWAY_API_KEY, and the ai-language-model-id header switches the model even though model is on the owned list. providerOptions.gateway.only also reaches the request body, so a config can change which upstream provider serves the call. This is the same class of problem as #107: anyone who can edit an AI Config can redirect the customer's traffic. The agents handler (build_request_params) and the native graph share the code. Inline comment below.
2. The native graph doesn't send the per-node events the spec now requires. Since monorepo #26 (AIC-3211, merged into this branch through #114), TESTING.md §2.2 has every native adapter send $ld:ai:graph:node with nodeKey / index when it enters a node. The other three Python adapters do. to_vercel_agents doesn't, and it sends only generation:success per node. JS toVercelAgents sends per-node duration:total and tokens:*, plus $ld:ai:graph:path, which §2.2 now says not to send. So one graph run gives Monitoring three different event sets: Python Vercel, JS Vercel, and the other adapters. Event names are a data contract, so this should be fixed in both languages before 1.0. Inline comment below.
3. Should these packages ship as 1.0? The ai dependency is a 0.x public beta, and the bound is ai>=0.7 with no upper limit, so any 0.8 release can break it. The experimental lifecycle draft (TESTING.md §0) covers experimental names inside launchdarkly_ai_server. It doesn't say what to do with a whole provider package. I'd suggest keeping both Vercel packages on 0.x in both languages when the rest go to 1.0, writing that rule into §0 (a provider package built on a pre-1.0 runtime stays 0.x), and capping ai<0.8 until then. The GA plan's package count doesn't include these two Python packages yet.

Smaller notes:

• The spec PR, launchdarkly/ai-sdks-monorepo#19, has no reviews, hasn't been updated since 22 Sep, and still points the python submodule at e83fa11. Its A.14 says experimental_evaluate is "Not available in official ai 0.7", and it skips vercel_evaluate for Python on that basis. But ai 0.7.0 does export ai.ops.experimental_evaluate(model, state, questions, *, params=None). So either Python gets a vercel_evaluate, or the spec should give a different reason for leaving it out. Since it wraps an upstream experimental_ API, JS vercelEvaluate probably belongs on an experimental entry point and not the package root. #19 also needs the $ld:ai:graph:node line in §2.x.3.
• The last Bugbot comment (step limit, handler.py:422) has no reply yet. With tools and outputFormat both set, a final answer that arrives on step 10 hits continue and then the RuntimeError. The structured follow-up also costs an extra model call and discards the first answer.
• Graph-level events use make_track_data(definition.root, ...), the same as the other adapters. Whatever #121 decides about which configKey graph events carry applies here too.
• JS #83 spreads the whole model.parameters bag into generateText / streamText (modelSettings), so headers and providerOptions probably reach the request there as well. I haven't checked this at runtime in JS. Since #83 has merged, it likely needs a follow-up there.

[andrewklatzke]

Addressed the open review notes in 1389d04:

• AI Config model.parameters no longer forwards extra_headers, extra_query, or leftover keys as extra_body, so a config cannot replace the Gateway Authorization header, ai-language-model-id, or providerOptions. Sampling, max_tokens, reasoning_effort, metadata, and safety_identifier are unchanged. Covered in both Vercel handlers.
• When tools and outputFormat are both set, the structured follow-up no longer counts against the 10-step tool budget. An unconstrained answer on the last step still gets the schema-constrained turn.
• to_vercel_agents now tracks $ld:ai:graph:node on entry for each visited node (nodeKey, 0-based index, value 1), including a node that throws. $ld:ai:graph:path is not sent.
• Both packages pin ai>=0.7,<0.8.

[jeffdupont]

Re-checked at 1389d04. The fix commit covers my 2 Oct points: the parameter allowlist with attacker-key tests, $ld:ai:graph:node with no path event, 0.x with ai>=0.7,<0.8, and the extra structured turn for the step limit. Tests, ruff and mypy pass locally.

Still blocking:

1. Merge conflicts in 6 files: .release-please-manifest.json, .sdk_metadata.json, AGENTS.md, pyproject.toml, release-please-config.json and tests/test_cross_handler_parity.py. CI hasn't run on this head.
2. Spec says Vercel is JS-only. ai-sdks-monorepo TESTING.md A.14 (from #51) says vercel-messages and vercel-agents are not Python packages, and api-surface/python/ has no Vercel lists. I'll settle that line in the monorepo. Either the spec gains Python rows or this doesn't ship.
3. Adopt main's usage-reporting contract (#145). Other handlers keep PACKAGE_NAME in _version.py, call report_usage(...) in each public helper, and point release-please extra-files at _version.py. This PR keeps __version__ in __init__.py and has no report_usage, so Python Vercel would send no $ld:ai:sdk:usage events while JS does.

Non-blocking: other Python handlers now emit AG-UI text and tool events (#161-#165). JS Vercel doesn't either, and the spec doesn't require it.

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 2 potential issues.

^{❌ 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 1ab101a. Configure here.}

[jeffdupont]

Code looks good and lines up with the JS surface. One blocker: main's TESTING.md still says Vercel is JS-only (~L3013), so this needs ai-sdks-monorepo#19 merged first. Can you also add the Python api-surface lists for vercel-agents/messages and the python tags in usage-ids.txt? #19 doesn't include those, so api-surface-check will catch it later otherwise. Happy to approve once the spec side is in.

[andrewklatzke]

Merged main in to resolve the CONFLICTING state (main had just picked up the LiteLLM packages from #102, which collided with this branch's new package entries in .release-please-manifest.json, .sdk_metadata.json, release-please-config.json, AGENTS.md, README.md, main.py, and uv.lock). Kept both sides everywhere — no code/logic conflicts, just two PRs adding new packages to the same lists. Full suite re-run after the merge: 2771 passed, lint/format/typecheck clean. Pushed as c04f87f.

On @jeffdupont's spec blocker: the "Vercel is JS-only" line at main's TESTING.md (~L3013) and the missing Python api-surface lists for vercel-agents/vercel-messages are already fixed in launchdarkly/ai-sdks-monorepo#78, which branches straight off current main (doesn't depend on #19). #78's checks are green; it just needs a review/merge. #19 is a separate, broader PR (adds the full §1.x/§2.x Vercel handler-contract spec sections) that's still open too, but isn't what's gating this specific line.