Core Concepts
Observability
What the gateway captures on every request, how it is grouped, and where you see it.
How it works
Spineforge sits in the data path between your agents and the LLM providers, so it sees every request and response without any instrumentation in your agent. There is nothing to install and no agent logic to change: observability comes from the gateway itself.
What is captured per request
| Field | Meaning |
|---|---|
| Agent | The gateway key (agent) that made the call |
| Provider & model | Where the request was routed, e.g. openai / gpt-4o |
| Tokens | Prompt and completion token counts reported for the call |
| Latency | End-to-end duration of the request |
| Cost (USD) | Computed once at write time from the rate card in effect, stored with its rate_card_id |
| Status | Success, error, or blocked by policy (revoked, scope, budget, rate limit) |
Prompt and response content
Whether prompts and responses are stored is a data-handling decision. Spineforge is designed around three modes: Metadata only, Redacted content and Full content. In progress See the Security page for the current status.
Where you see it
- Cost & usage: spend by agent, provider and model.
- Agent graph: every agent and how it connects.
- Lifecycle kanban: agents by lifecycle stage.
- Per-agent timeline: a chronological trail of what each agent did.
A tour of each view is in the Dashboard Guide.
Streaming
Streamed responses are piped through the gateway to your agent as they arrive. A streamed response is recorded as a single request.
In the data path: what that means
The gateway is a stateless, horizontally scaled proxy that adds roughly 15–35 ms of overhead per request. Because it is in the data path, enforcement (revocation, budgets, scopes, rate limits) is inline and immediate. If the gateway were unreachable, agents couldn't reach the model through it. See Service Reliability for measured numbers and our uptime target.
Next steps
- Metrics Reference: how cost and latency are calculated
- Dashboard Guide: the views in detail
- Security Model: enforcement on every request