feat: added ts-sdk with openAI sdk support

[Adarsh9977]

Summary

Adds @secureagentics/adrian-openai, a TypeScript provider package that instruments the official OpenAI Node SDK for Adrian security monitoring.

• adrian.openai(client) — proxy wrapper around an existing OpenAI client; instruments chat.completions.create and responses.create without changing call sites
• Chat Completions — captures paired LLM events for non-stream and streaming responses, including streamed tool-call aggregation and usage tokens
• Responses API — captures text output, function/tool calls, instructions, and array-shaped input (including prior tool calls and outputs) for both stream and non-stream paths
• adrian.captureTool() — wraps manual tool execution so tool start/end events are emitted and BLOCK/HITL policy can halt execution before the handler runs
• Error capture — failed OpenAI requests and tool executions are recorded as LLM/tool events with error metadata
• Shared capture helpers — reuses @secureagentics/adrian/capture from the core package for consistent event pairing and policy gating
• Depends on @secureagentics/adrian (core TypeScript SDK). feat: added ts-core sdk #50

Test plan

• cd sdk/typescript and run npm test -w @secureagentics/adrian-openai

Checklist

• CLA signed (see CLA.md)
• Tests pass locally
• Docs updated where needed
• British English; no em-dashes; no marketing fluff

[shahtajkhalid]

Thanks for PR @Adarsh9977, needs to fix the following before we can merge:

1. invocation_id is being randomly regenerated per call. Both captureLlmCall (common.ts:27) and captureTool (index.ts:66) is calling runWithInvocationId(randomUUID(), …), hence each LLM call and each tool call gets its own invocation, which degrades our classifier's context, since our backend groups context by the
   (session, invocation, agent_id) tuple.

[Adarsh9977]

Thanks for PR @Adarsh9977, needs to fix the following before we can merge:

1. invocation_id is being randomly regenerated per call. Both captureLlmCall (common.ts:27) and captureTool (index.ts:66) is calling runWithInvocationId(randomUUID(), …), hence each LLM call and each tool call gets its own invocation, which degrades our classifier's context, since our backend groups context by the
   (session, invocation, agent_id) tuple.

I'm looking on it thanks @shahtajkhalid

[Adarsh9977]

@shahtajkhalid Done with the changes. Now it's taking the active invocation Id and if not present then it passes null. Just like we're doing in python.

Please take a look and let me know