First /add can fail when tiktoken downloads o200k_base at runtime

[rendigua2025-gif]

Summary

The first /add request can fail with HTTP 500 if tokenizer initialization tries to download o200k_base.tiktoken at runtime and that network request fails.

In an isolated restart/cold-start test, the failure happened before any business write was committed. After prewarming the tokenizer cache with tiktoken.get_encoding("o200k_base"), the same test shape worked with a new memory root/session.

Why this matters

This makes a fresh EverOS service depend on runtime network access to an external tokenizer asset during the first memory write. If that download fails, users see a generic /add 500 even though the memory pipeline itself has not started writing data yet.

This is especially surprising for Docker/Linux isolated tests and server deployments where outbound network/TLS behavior may vary.

Observed evidence

First cold /add:

/add -> HTTP 500

Failure occurred before writes:

unprocessed_buffer: 0
memcell: 0
conversation_status: 0
md_change_state: 0
Markdown marker: none
search result count: 0

Traceback shape:

tiktoken_ext/openai_public.py: o200k_base
tiktoken/load.py: read_file_cached
requests.get("https://openaipublic.blob.core.windows.net/encodings/o200k_base.tiktoken")
SSLError: UNEXPECTED_EOF_WHILE_READING

After tokenizer prewarm:

tiktoken.get_encoding("o200k_base")

A new memory root/session completed successfully, and restart recovery worked as expected.

Expected behavior

EverOS could make this cold-start dependency more explicit and reliable. Possible options:

• Prewarm tokenizer assets during everos init or server startup.
• Document that first use may download tokenizer assets and how to pre-cache them.
• Fail startup or preflight checks with a clear tokenizer/cache error instead of returning a generic /add 500.
• Avoid runtime network fetches in the request path where possible.
• Provide an offline/cache configuration path for Docker/server deployments.

Environment

EverOS: 1.0.0 source checkout
Runtime: Docker Linux runtime
Python: 3.12
Data: synthetic sample only

No real user memory or secrets were used.