[apucacao]

The JS and Python SDKs have to forward the same model.parameters keys before 1.0. Jeff's review of launchdarkly/python-ai-sdk#107 found that the lists from #73 differ from Python's. The shared lists are now in the cross-SDK spec (launchdarkly/ai-sdks-monorepo#42, TESTING.md §1.12), and Python is moving to them too. This PR makes each JS handler forward exactly those lists, minus the keys the spec marks as not supported in TypeScript.

What changed:

• Claude Messages: stops forwarding container, inference_geo and metadata. output_config keeps only effort (a top-level effort still moves into it, and an explicit output_config.effort wins). output_config.format is dropped, because structured output belongs to the config's outputFormat: forwarding it let a config ask for one schema in outputFormat while the API enforced another, and made stream return JSON.
• OpenAI Messages: stops forwarding context_management, include, instructions, metadata, moderation, prompt_cache_retention, safety_identifier, store, truncation and user, on both invoke and stream. max_tool_calls is not forwarded: the spec marks it not supported in TypeScript, because the openai package's responses.create has no such parameter.
• OpenAI Agents (handler and native graph): stops forwarding context_management, prompt_cache_retention, store and truncation. tool_choice is forwarded only as a string.
• LiteLLM Agents (handler and native graph): stops forwarding store. reasoning keeps only effort, because the Chat Completions model sends it as reasoning_effort and has no field for summary. tool_choice is forwarded only as a string.
• Claude Agents (handler and native graph): stops forwarding output_format, for the same reason as output_config.format: the handler applies the config's outputFormat itself. thinking is kept only with a type the Claude Agent SDK knows, an enabled thinking needs a numeric budget, and display is kept only as summarized or omitted.
• Vercel (messages and agents): a reasoning object is read as its effort, and any other non-string reasoning is dropped. tool_choice is forwarded only as a string, since the AI SDK reads any other value as a call to a named tool.
• LangChain ChatOpenAI (both LangChain packages and the LangGraph native graph's default model): stops forwarding audio, disable_streaming, metadata, modalities, output_version, prompt_cache_retention, reasoning, stream_usage, streaming, tags, use_responses_api, user, verbose and zdr_enabled. prompt_cache_key is still forwarded: the spec marks it not supported in Python only, where ChatOpenAI has no such field.
• LangChain ChatAnthropic: stops forwarding context_management, disable_streaming, inference_geo, metadata, output_version, stream_usage, streaming, tags and verbose. It now takes stop as stopSequences and max_tokens_to_sample as maxTokens, because ChatAnthropic has neither field under those names. outputConfig keeps only effort, as in Claude Messages.
• LangChain ChatBedrockConverse: stops forwarding disable_streaming, guardrail_config, metadata, output_version, stream_usage, streaming, supports_tool_choice_values, tags and verbose.
• LiteLLM Messages and TypeSafe: no change. LiteLLM Messages already forwarded the spec's list, and TypeSafe forwards no model.parameters key.

Following the spec, nested values with the wrong shape are dropped instead of sent. Examples are a thinking without a type, a reasoning that is not an object, a tool_choice of the wrong type, or an output_config without a string effort. OpenAI Messages reasoning keeps only effort and summary, as OpenAI Agents already did.

Testing:

• Each handler and native graph has a test that the forwarded keys are exactly the spec's list, and, where the handler has both, that invoke and stream forward the same keys and values. This now includes LiteLLM Messages and LiteLLM Agents. TypeSafe has a test that it forwards no model.parameters key at all.
• Every package's never-forward bag is built from one shared key list in packages/client/src/__tests__/never-forwarded-keys.ts, which covers every never-forward category in the spec, including user, metadata, inference_geo and container. Each package's __tests__/never-forwarded.ts adds only the keys its own table leaves out, so the bags can't drift apart.
• New tests cover dropping malformed nested values, output_config.format on invoke and stream, and output_format in Claude Agents.
• Each changed package's tests and tsc --noEmit, and biome check ., pass locally.

---

Note

Overview
Aligns every JS provider handler with TESTING.md §1.12 so model.parameters forwards the same generation-related keys as the Python SDK, and drops everything else (credentials, retention, metadata, wiring, etc.).

Allowlists are tightened across packages: OpenAI Messages/Agents stop forwarding retention and attribution fields (store, user, context_management, …); Claude Messages keeps output_config to { effort } only and strips format; Claude Agents no longer forwards output_format (structured output stays on config outputFormat); LangChain constructors lose base-model wiring keys and narrow ChatAnthropic outputConfig; LiteLLM/Vercel agents drop store and non-string tool_choice; Vercel maps object reasoning to a string effort. Malformed nested values (thinking, tool_choice, reasoning, typed Anthropic objects) are dropped instead of passed through.

Tests: Shared createNeverForwarded() in packages/client replaces duplicated never-forward lists; each handler/native graph gets §1.12 canonical-key tests and invoke/stream parity checks where applicable.

^{Reviewed by Cursor Bugbot for commit 0dd95ab. Bugbot is set up for automated code reviews on this repo. Configure here.}

[apucacao]

bugbot run

[cursor]

Stale Bugbot comment from a previous run.

[apucacao]

bugbot run

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 0e963d3. Configure here.}

[jeffdupont]

The allowlists look right. Every list matches its §1.12 table, minus the keys marked not supported in TypeScript, and credentials, hosts and headers can't reach any provider call. A few things before merging:

1. output_config.format gets through (claude-messages/src/handler.ts:461-467, and the LangChain ChatAnthropic model-parameters.ts). output_config is checked only for being an object, so Anthropic's structured-output format reaches the API. If a config sets outputFormat to schema A and output_config.format to schema B, invoke asks for A and the API enforces B. On stream, which normally returns free text, it returns JSON shaped like B. I've asked in mono#42 for output_config to keep only effort. Once that lands, rebuild it as {effort} and flip the tests that assert it passes through. py#107 has the same issue.
2. The §1.12 required tests are missing for LiteLLM Messages, LiteLLM Agents (handler and native graph), and TypeSafe. They have no list-equals-table probe, no full never-forward bag and no invoke/stream agreement test. TypeSafe has no test that it forwards nothing. The never-forwarded.ts and "cross-SDK allowlist" pattern from openai-agents would carry over.
3. Several never-forward bags leave out keys the spec names, for example user, metadata, inference_geo and container in claude-agents, openai-agents and vercel. The allowlists are fine today, so this is only coverage. One shared key list would keep the bags from drifting.
4. The PR description is out of date. prompt_cache_key is still forwarded for ChatOpenAI, max_tool_calls is marked not supported in TypeScript rather than cut for both SDKs, Claude Agents and Vercel did change, and the LiteLLM Agents changes aren't mentioned.

[apucacao]

@jeffdupont all four addressed:

1. output_config (92804ff): Claude Messages builds it with toOutputConfig and both LangChain packages with toAnthropicOutputConfig, as {effort} only, dropped when there's no string effort. The tests that asserted pass-through are flipped, and new ones check format is dropped on invoke and stream and that the default ChatAnthropic constructor gets outputConfig: {effort} only. Claude Agents outputFormat had the same problem, so it's off the query() allowlist too (4a866f3), handler and native graph, matching the feat: Report AI SDK package identity to LaunchDarkly #42 change.
2. Missing §1.12 tests: LiteLLM Messages and LiteLLM Agents (handler and native graph) now have the list-equals-table probe, the full never-forward bag and invoke/stream agreement where there's a stream (148c3c8). TypeSafe has a test that it never reads model.parameters: the Jev request carries only model, questions and state with the full bag set (0dd95ab).
3. Never-forward bags (76b1c53): one shared list in packages/client/src/__tests__/never-forwarded-keys.ts covering every §1.12 category in both spellings, and each package's never-forwarded.ts builds from it. Packages import it by relative path, like the sdk-usage tests already read client/src/version.ts, so nothing ships. The fuller bag caught keys a handler sets itself (system, instructions), which those call sites now pass as handler-owned, while the deep check still proves the config's value never lands.
4. The PR description is updated against the actual diff: prompt_cache_key still forwarded for ChatOpenAI, max_tool_calls not supported in TypeScript, the real Claude Agents, Vercel and LiteLLM Agents changes, and this round's output_config and outputFormat changes.