Skip to content

Hermes attribution

Hermes contributes request-level observations: timestamp, session, profile, provider identifier, model, token classes, requests, and optional cost metadata. The provider contributes account-level usage, balance, quota, or billing data.

When both describe the same workload:

text
provider total = authoritative account activity
Hermes observed = the subset seen flowing through Hermes
unattributed = provider total minus the compatible Hermes subset

They are never added. If a provider lacks detailed native usage, the dashboard can still show Hermes-only workload, explicitly labelled as observed rather than complete account usage.

Mapping and attribution

Hermes provider identifiers such as auto or openai-codex may not match a configured adapter ID. Provider mappings translate those identifiers without rewriting stored raw observations. Profile and model fields power the Breakdown views; session IDs provide session counts.

An unresolved alias remains visible in diagnostics. When multiple dashboard configurations share one provider, Hermes data cannot be assigned safely to a specific connection, so it is presented once at provider level.

Pricing and reconciliation

Model, timestamp, and token class can be priced using the versioned catalogue to produce API-equivalent value. Unknown models and unsupported token classes are unpriced and reduce coverage. Hermes-provided or reconstructed cost does not replace higher-priority provider spend and is never added to it.

See Data sources for setup and Pricing & coverage for completeness rules.

Self-hosted API usage dashboard.