Editing Openai/694057b6-101c-8007-9a65-c40578c7252d (section)

==== ### ====

A secondary, low-level system prompt owned by the worker class (not the caller) that injects runtime/control-plane facts the model should always see, such as:
* tool-call budget remaining (and max)
* current date/time (+ timezone)
* worker identity / model identity (optional)
* response constraints that are “platform rules” (optional)
* tool usage rules (e.g., “only call tools from the provided list”)

===== Lock this down because ordering matters. =====

Recommended invariant order:
# BIOS system prompt (worker-generated, dynamic per turn)
# Caller’s system prompt (the user-facing “persona/task” system message, static per request)
# Conversation messages (user/assistant/tool messages)

This gives your platform-level constraints the highest precedence, while still letting the caller define task behavior.

===== For maximum compatibility (since different servers vary in how they treat multiple system messages), the worker should support both: =====
* Two system messages (BIOS + caller system), if the backend supports it cleanly
* Single combined system message, where BIOS is prepended with a clear delimiter (safe fallback)

Example delimiter idea (spec-level, not exact text):

<syntaxhighlight>[BIOS]
...dynamic control-plane info...
[/BIOS]

...caller system prompt...

</syntaxhighlight>

===== The BIOS prompt should be regenerated: =====
* at request start
* before each tool-loop continuation (so “tool iterations remaining” stays accurate)
* optionally on long streams (if you ever add “keepalive” logic)

===== - No secrets in BIOS (API keys, hostnames if sensitive, etc.). Treat it as model-visible. =====
* Hard size cap on BIOS text (it’s easy to accidentally bloat it with debug info).
* Version it: include a short BIOS schema/version string so you can change formatting later without confusion.
* Deterministic format: stable keys so logs/monitoring can parse it.

===== Because you’re using internal streaming, the BIOS layer should also update worker/request status with: =====
* tool budget remaining
* last BIOS timestamp injected (useful when debugging tool-loops)
* BIOS version/hash (handy for correlation)