Naagmani offers discrete hook interception points along the complete request/response lifecycle.

Rendering diagram...

flowchart TD
    A[Client Request] --> B[pre_auth]
    B --> C[pre_route]
    C --> D[pre_model_call]
    D --> E[Upstream AI Execution]
    E --> F[post_model_call]
    F --> G[post_response]
    G --> H[Client Response]
    
    E -.->|Error Encountered| I[on_error]

Supported Hook Points #

Hook Name	When It Executes	Can Mutate Payload?	Common Use Cases
`pre_auth`	Before token validation and project context resolution.	Headers only	Custom JWT decryption, IP whitelisting.
`pre_route`	After project auth, before selecting the upstream provider/model.	Yes (Full)	Semantic routing, RAG context enrichment, PII masking.
`pre_model_call`	Immediately before dispatching payload to the selected provider adapter.	Yes (Provider payload)	Model-specific prompt formatting, token clipping.
`post_model_call`	Immediately after provider response or first stream chunk.	Yes	Output moderation, toxic content filtering.
`post_response`	Final stage before sending bytes to client.	Yes	Watermarking, telemetry calculation.
`on_error`	Triggered when upstream provider returns 4xx/5xx or timeouts.	Error payload	Custom error formatting, alert webhooks.

Hook Priority & Ordering #

Multiple plugins can attach to the same hook point. They execute strictly ordered by their priority integer in descending order:

Plugin A (priority: 100)
Plugin B (priority: 50)
Plugin C (priority: 10)

Each plugin receives the output resulting from the preceding plugin in the pipeline.

Execution Hooks

Supported Hook Points #

Hook Priority & Ordering #

Next Steps #