llm-d Integration Testing

Praxis AI includes an AI-owned ext_proc compatibility layer for llm-d integration.

Praxis AI documentation version

On this page

Overview

Praxis AI includes an AI-owned ext_proc compatibility layer for llm-d integration. This replaces the abandoned Praxis-side crate publication approach (Praxis PR #780). The compatibility layer lives in the AI repository, scoped exclusively to llm-d, and is not a general-purpose Envoy ext_proc extension point.

Crate location and feature

Crate path: integrations/llmd/ext-proc/
Package name: praxis-ai-llmd-ext-proc
Feature name: llmd-ext-proc
publish = false - not available on crates.io.

Default praxis-ai-proxy builds do not link, register, or expose the ext_proc filter. The filter is registered only when the llmd-ext-proc feature is enabled on praxis-ai-proxy.

Architecture

The integration test request path is:

OpenAI-compatible request
    -> Praxis AI runtime
    -> ext_proc filter (gRPC full-duplex)
    -> deterministic mock ExternalProcessor
    -> trusted x-gateway-destination-endpoint mutation
    -> endpoint_selector
    -> llm-d-inference-sim container
    -> Chat Completions response

The mock processor is not the real llm-d Go EPP. These tests prove the transport and routing integration shape, not llm-d scheduling policy.

Support boundary

AI-owned, publish = false, llm-d-only.
Not general-purpose Envoy ext_proc support.
No compatibility or support is promised for non-llm-d uses.
Native Praxis filters are preferred for general extension work.
Other use cases require a maintainer discussion on GitHub.
The Grid demo should consume AI with llmd-ext-proc enabled for provider-gateway/llm-d testing. It should not depend on praxis-proxy-ext-proc or Praxis PR #780.

Running the tests

Requires Docker or Podman for the llm-d-inference-sim container (pinned to v0.10.0).

cargo test -p praxis-tests-environment \
    --features llmd-ext-proc \
    -- --test-threads=1

The six environment tests:

Test	Proves
`simulator_chat_completion_routes_through_praxis`	Full request path through `ext_proc` + `endpoint_selector` to simulator
`simulator_spoofed_destination_header_ignored`	Client-supplied routing header cannot override processor decision
`simulator_repeated_requests_no_crosstalk`	Independent `ext_proc` streams per request
`simulator_health_endpoint_reachable`	Simulator `/health` returns 200
`simulator_metrics_endpoint_reachable`	Simulator `/metrics` returns 200 with content
`simulator_processor_failure_returns_status_on_error`	Unreachable processor returns 503, no bypass

Dependency and build notes

The AI workspace depends on the published Praxis 0.4.1 crates from crates.io. These include the support APIs needed by the llm-d ext_proc compatibility layer.