Problem

When Skyvern runs through an OpenAI-compatible endpoint (ENABLE_OPENAI_COMPATIBLE), every LLM call for every task and workflow run looks the same to the endpoint. A gateway or proxy in front of the model can't tell which run a call belongs to, so it can't attribute cost per run or stop a single run that's looping on a page it can't get past. Skyvern already has task_id / workflow_run_id in context at each call in api_handler_factory.py, but the only headers it adds are static ones (HTTP-Referer, X-Title, the Copilot ones).

Suggestion

An opt-in setting that sends the current run's id as a request header on LLM calls, for example:

OPENAI_COMPATIBLE_RUN_ID_HEADER=X-Run-Id   # unset by default: no change

When it's set, Skyvern adds {<header>: workflow_run_id or task_id} to extra_headers on each LLM call it makes for that run. Nothing changes for anyone who doesn't set it.

Why it helps

• Per-run cost attribution in whatever sits in front of the model (LiteLLM, Helicone, a company's own proxy).
• Per-run spend limits. For example, with Inferrail (an open-source OpenAI-compatible gateway) a self-hosted Skyvern could set OPENAI_COMPATIBLE_RUN_ID_HEADER=X-Inferrail-Attribute-Work-Id, and the gateway's budgets.per_work_default_usd would give each workflow run its own dollar ceiling, refusing calls before they reach the provider once a run is out of budget.

Alternatives considered

A static header through extra_headers can't carry the run id, and a key per run isn't practical for a self-hosted server that starts runs on demand.

Happy to send a PR if this direction works for you. Disclosure: I maintain Inferrail.