feat(pricing): infer upstream billing provider from model name for proxied endpoints

[vanhoof]

Problem

When models are served through a local proxy (vertex-proxy, LiteLLM, custom gateway), the Hermes provider name is a user-defined config alias (e.g. vertex-opus46, litellm) that does not match any known billing route. resolve_billing_route() hits the localhost catch-all and returns billing_mode="unknown", so the dashboard and session cost tracker show "unknown" cost for every proxied session.

Observed behavior: every API call through a proxy produces amount_usd=None from estimate_usage_cost(), regardless of the upstream model.

Affected setups: any user routing through a local proxy with a custom provider name in providers: config. This includes vertex-proxy (GCP Vertex AI), LiteLLM, and similar OpenAI-compatible gateways on localhost.

Root Cause

resolve_billing_route() checks provider names against a fixed set of known providers (anthropic, openai, openrouter, etc.). User-defined provider names like vertex-opus46 fall through every named check and land in the custom/local/localhost branch, which returns billing_mode="unknown". The official-docs pricing table has the correct entries keyed by (anthropic, claude-opus-4-6), but the billing route never maps the custom provider name to anthropic.

Secondary issue: the localhost detection only checked for the string "localhost" in the base URL, missing 127.0.0.1 which is the more common form in practice.

Fix

Adds _infer_upstream_provider(), a ~15-line helper that maps well-known model name prefixes to their upstream billing provider:

Prefix	Provider
claude-	anthropic
gpt-	openai
o1-, o3-, o4-	openai
gemini-	google
deepseek-	deepseek

The inference runs as a fallback in two places within resolve_billing_route():

1. Inside the custom/local/localhost branch (before returning unknown)
2. In the final catch-all (for unrecognized provider names that are not on localhost)

When a model name matches, the route resolves to official_docs_snapshot pricing. When it does not match, behavior is unchanged (billing_mode="unknown").

Also adds "127.0.0.1" to the localhost detection alongside "localhost".

Per-file summary

• agent/usage_pricing.py: adds _MODEL_PREFIX_TO_PROVIDER table, _infer_upstream_provider() function, and two inference call sites in resolve_billing_route(). Adds "127.0.0.1" to localhost detection.
• tests/agent/test_usage_pricing.py: 12 new tests covering prefix inference, billing route resolution for proxied models, end-to-end pricing lookup, and end-to-end cost estimation.

Design Notes

This is the simplest approach that solves the problem for the vast majority of proxy users without requiring any config changes. A future billing_provider config field on custom providers could override the inference for edge cases, but was deliberately not included to keep the diff minimal and avoid a config schema change.

The pattern is consistent with existing inference in agent_init.py (which infers api_mode from URL patterns) and _normalize_anthropic_model_name() / _normalize_bedrock_model_name() (which normalize model name variants for pricing lookup).

Related: #18886 attempted to solve this for LiteLLM specifically by probing /model/info. That approach does not generalize to non-LiteLLM proxies (vertex-proxy, custom gateways) and was closed.

Testing

$ python3 -m pytest tests/agent/test_usage_pricing.py -v
27 passed in 0.57s

$ python3 -m pytest tests/agent/test_usage_pricing.py tests/agent/test_insights.py \
    tests/hermes_cli/test_model_cost_guard.py tests/gateway/test_usage_command.py -q
95 passed in 1.46s

Manually verified against a live vertex-proxy (podman, localhost:8788):

Model	Provider	Before	After
claude-opus-4.6	vertex-opus46	None (unknown)	$5.00/$25.00 per M (estimated)
claude-sonnet-4.6	vertex-sonnet46	None (unknown)	$3.00/$15.00 per M (estimated)
gemini-pro	vertex-gemini-pro	None (unknown)	routes to google (official_docs_snapshot)

Signed-off-by: Chris van Hoof vanhoof@ouwish.com

[teknium1]

Thanks for targeting a verified pricing gap: current main still returns an unknown route for custom/local and unrecognized provider aliases at agent/usage_pricing.py:675-677.

Problems

• The added inference checks only raw model prefixes. With a configured proxy alias, current routing strips vendor/ only when no provider was supplied (agent/usage_pricing.py:648-652), so anthropic/claude-* and google/gemini-* do not match the proposed claude-/gemini- prefixes. Vendor-prefixed IDs are supported; the bundled Vertex profile uses google/gemini-3-flash-preview at plugins/model-providers/vertex/__init__.py:72.

Suggested changes

• Normalize a recognized vendor prefix before inference (while retaining the original model for pricing lookup), and add proxy-alias end-to-end tests for vendor-prefixed Anthropic and Gemini IDs.
• Salvage will need conflict-aware integration: resolve_billing_route() changed after this PR's base in db117af4785f79d0adfafcea4d75ee556f4006dd.

Automated hermes-sweeper review.