docs(proxy): correct tag_routing_prefix's unprefixed-tag behavior

[deepanshululla]

Documents the corrected behavior from BerriAI/litellm#39903, fixing BerriAI/litellm#39901.

Adds to docs/proxy/tag_routing.md and docs/proxy/config_settings.md:

• Once tag_routing_prefix is configured, an unprefixed request tag is a complete no-op: never matched, required, or excluded, and never a cause of no_deployments_with_tag_routing. The prior text ("keeps going through today's existing handling unchanged") described the bug this PR fixes, not the actual corrected behavior.
• Key/team/project policy tags in metadata.inherited_tags are exempt from this scoping and keep applying regardless of prefix, since they are not caller-controlled.
• A migration caution for operators already routing on plain unprefixed tags, since configuring tag_routing_prefix now stops that routing immediately for any caller not yet migrated to the prefix.

Verified with npm run build locally; no new broken links or anchors on the edited pages.

[vercel]

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

Project	Deployment	Actions	Updated
litellm	Ready	Preview	Sep 5, 2026 4:09pm UTC

[deepanshululla]

Live verification against the fix in litellm PR #39903, real Anthropic and OpenAI API calls, no mocks.

Config used

model_list: chat -> anthropic/claude-haiku-4-5 (tags: default, provider:anthropic, allow_fail_open: true), chat -> openai/gpt-5.4-mini (tags: provider:openai). router_settings: enable_tag_filtering: true.

Quick example (tag_routing_prefix: "route:")

Command: curl http://localhost:4000/v1/chat/completions -d '{"model": "chat", ..., "metadata": {"tags": ["feature:demo", "route:!provider:openai"]}}'

Result: 200, x-litellm-model-name: anthropic/claude-haiku-4-5. feature:demo is inert, route:!provider:openai correctly excludes the OpenAI deployment.

Prefixed tags and the fail-open guard

Command: same, tags ["route:&provider:anthropic", "route:&custom-routing-key"]

Result: 200, x-litellm-model-name: anthropic/claude-haiku-4-5. custom-routing-key matches no deployment's tags but is prefix-marked, so allow_fail_open still proceeds instead of raising.

Migration caution, the new claim in this PR

With tag_routing_prefix: "route:" configured, tags: ["provider:openai"] (no prefix): 200, x-litellm-model-name: anthropic/claude-haiku-4-5, not OpenAI. The plain tag is now a no-op and falls through to the default-tagged deployment.

Same request with tag_routing_prefix: "" (baseline, prefix not configured): 200, x-litellm-model-name: openai/gpt-5.4-mini. The identical plain tag correctly selects the OpenAI deployment.

Conclusion: confirms the corrected unprefixed-tag no-op behavior, the inherited-tag exemption's sibling fail-open behavior, and the migration caution this PR adds are all accurate against the real code.