Some Python AI integrations drop Anthropic cache-write TTL details

[gouveags]

Version

7.25.0

Steps to Reproduce

PostHog/posthog#71323 adds separate token fields for Anthropic's 5-minute and 1-hour cache writes. I checked each Python integration to see whether that breakdown reaches ingestion.

Two paths currently lose it:

• The LangChain callback reads the aggregate cache-write count but ignores ephemeral_5m_input_tokens and ephemeral_1h_input_tokens from usage_metadata.input_token_details.
• The Claude Agent SDK processor copies aggregate usage fields but discards the nested cache_creation object.

For example, pass this synthetic usage payload through the LangChain callback:

{
    "input_tokens": 800,
    "output_tokens": 50,
    "input_token_details": {
        "cache_read": 200,
        "cache_creation": 0,
        "ephemeral_5m_input_tokens": 300,
        "ephemeral_1h_input_tokens": 200,
    },
}

Expected Result

The resulting $ai_generation event keeps the aggregate and the complete TTL pair:

$ai_cache_creation_input_tokens = 500
$ai_cache_creation_5m_input_tokens = 300
$ai_cache_creation_1h_input_tokens = 200

For this LangChain payload, normalized uncached input is 100: 800 - 200 cache reads - 500 cache writes.

Native PostHog Anthropic clients, including the Bedrock and Vertex wrappers, already preserve the official nested breakdown in $ai_usage; the server PR can extract it. LangGraph uses the same PostHog LangChain callback, so it is covered by the callback fix.

Actual Result

The LangChain path emits an aggregate value of zero, omits both TTL fields, and leaves the 500 cache-write tokens in normalized input. The Claude Agent SDK path emits only its aggregate cache-write value.

The SDK follow-up should fix those loss points, retain aggregate-only behavior, and keep malformed or unavailable TTL data on the existing fallback. Direct boto3/OTel traffic is not included: PostHog Python does not wrap that client, and current instrumentation does not expose the TTL split.

This issue is the Python SDK follow-up to PostHog/posthog#71306 and PostHog/posthog#71323. The companion PR should remain draft until the server-side interface is merged.

[gouveags]

A quick note on why I changed the title and description a few times: I mostly maintain a Python codebase built on LangChain and LangGraph, so the first version naturally described the path I had actually encountered and support day to day.

As I traced the behavior further, I realized the scope was broader, but not identical across every integration. The native PostHog Anthropic wrappers already preserve the raw usage breakdown, and the Bedrock and Vertex clients reuse those adapters. The LangChain callback was dropping the TTL-specific fields, while the Claude Agent SDK kept the aggregate but not the raw breakdown. Direct boto3 and OTel also have different boundaries and cannot be described as the same SDK bug.

I revised the title and body as that evidence became clearer so the issue would describe the real problem without implying that the entire Python SDK fails in the same way. Sorry for the churn here, the investigation improved, and I preferred correcting the framing over keeping the first version just for consistency.

I've also made some manual changes on top of the initial AI version (we need to keep corrrecting the slop, right?)

[turnipdabeets]

@PostHog/team-ai-observability