docs(proxy): document the provider and operational env vars missing from the reference table

[yassin-berriai]

142 environment variables read under litellm/ have no row in the environment variables reference table. Every one of them is reached through a bare get_secret or get_secret_str call, and the CI gate that enforces this table only recognises those calls when they carry a litellm. prefix, so none of these keys has ever been required to have a row. BerriAI/litellm#TBD widens that gate; this lands the rows it will require, so it has to merge first.

The rows split into two groups, and they deserve different amounts of reviewer attention.

32 are operational settings that nobody could have discovered without reading the source, and those are written individually from their call sites: WORKER_CONFIG, DISABLE_KEY_NAME, PROMETHEUS_SELECTED_INSTANCE, AWS_BEDROCK_RUNTIME_ENDPOINT, FIREWORKS_ACCOUNT_ID, the xAI OAuth token directory and file, and the watsonx alias family among them. REDIS_AZURE_AD_TOKEN is the clearest argument for sourcing descriptions from code rather than from names: it is a boolean flag that switches on Azure AD authentication for Redis, so a description guessed from the name would have told readers to put a token in it.

The remaining 110 are provider API keys and base URLs. Those are uniform, which is right for them, with the default base URL given wherever the call site names one. What is not uniform is the alias handling: every alias row says which name wins. A reader looking at TOGETHER_API_KEY, TOGETHER_AI_API_KEY, TOGETHERAI_API_KEY and TOGETHER_AI_TOKEN cannot tell from the names that they are read in exactly that order, and that ordering means TOGETHERAI_API_KEY, the only one previously documented, is third rather than canonical. Fireworks AI has the same four-name shape.

Rows are inserted alphabetically rather than appended, matching #767.

AI211_API_KEY is deliberately absent. It is a misspelling of AI21_API_KEY read at a single call site in a function that has no callers, and giving it a row would make the typo public API. It is being corrected in the source instead.

Verification

Coverage is asserted rather than eyeballed. The row set is generated against the widened gate's own output, failing if any key it reports lacks a row or if a row is added for a key that is already documented. Table row count goes 793 to 935, exactly +142.

Running the widened gate against this branch leaves only the deliberately excluded key:

$ python tests/documentation_tests/test_env_keys.py
Exception: Environment variables read under ./litellm but mentioned nowhere in the docs: ['AI211_API_KEY']

The gate as it exists on litellm main today also passes against this branch, so this PR is safe to merge on its own and cannot turn code-quality or documentation red for any open litellm PR.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview	Aug 5, 2026 7:34pm