feat(llm-d): move ext_proc compatibility into AI

[nerdalert]

Summary

Moves the llm-d ext_proc compatibility layer into the AI repository under integrations/llmd/ext-proc/.

This replaces the abandoned Praxis-side crate publication approach. See the comments from @shaneutt in praxis-proxy/praxis#780 for details. I am pushing a draft as I need it for llm-d integration work in grid so this may be temporary if Shane has other plans for gating ext_proc exposure.

The crate is AI-owned, publish = false, and scoped to llm-d’s current EPP protocol rather than general-purpose Envoy ext_proc support.

Runtime registration is gated behind the praxis-ai-proxy/llmd-ext-proc feature. Default praxis-ai-proxy builds do not link, register, or expose the ext_proc filter.

What changed

• Added praxis-ai-llmd-ext-proc at integrations/llmd/ext-proc/
• Added llmd-ext-proc feature to praxis-ai-proxy
• Added six llm-d environment tests using llm-d-inference-sim
• Added docs for the support boundary and test flow

Validation

• cargo check --workspace
• cargo check --features praxis-ai-proxy/llmd-ext-proc
• cargo clippy --workspace --all-targets -- -D warnings
• cargo clippy --workspace --all-targets --features praxis-ai-proxy/llmd-ext-proc -- -D warnings
• cargo test --workspace --exclude praxis-tests-environment
• cargo test -p praxis-tests-environment --features llmd-ext-proc -- --test-threads=1
• RUSTDOCFLAGS="-D warnings" cargo doc --workspace --no-deps --document-private-items
• make container
• git diff --check

[Copilot]

Pull request overview

Moves llm-d ext_proc compatibility into the AI repository behind an opt-in feature.

Changes:

• Adds the AI-owned compatibility crate and vendored protobuf definitions.
• Feature-gates runtime registration and container integration.
• Adds simulator-backed tests and integration documentation.

Reviewed changes

Copilot reviewed 27 out of 29 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
Cargo.lock	Locks new gRPC/protobuf dependencies.
Cargo.toml	Adds workspace crates and dependencies.
Containerfile	Includes the optional crate during builds.
apis/src/lib.rs	Updates test context initialization.
docs/developing/llmd-integration-testing.md	Documents support and testing.
filters/src/agentic/a2a/mod.rs	Normalizes comment typography.
filters/src/lib.rs	Updates test context initialization.
integrations/llmd/ext-proc/Cargo.toml	Defines the compatibility crate.
integrations/llmd/ext-proc/NOTICE	Records vendored-code licenses.
integrations/llmd/ext-proc/build.rs	Compiles vendored protobufs.
integrations/llmd/ext-proc/proto/envoy_common.proto	Vendors Envoy common types.
integrations/llmd/ext-proc/proto/ext_proc.proto	Vendors the external-processing protocol.
integrations/llmd/ext-proc/proto/google/protobuf/any.proto	Vendors Any.
integrations/llmd/ext-proc/proto/google/protobuf/duration.proto	Vendors Duration.
integrations/llmd/ext-proc/proto/google/protobuf/struct.proto	Vendors structured values.
integrations/llmd/ext-proc/proto/google/protobuf/wrappers.proto	Vendors wrapper types.
integrations/llmd/ext-proc/src/callout.rs	Implements single-message callouts.
integrations/llmd/ext-proc/src/duplex.rs	Implements duplex exchange state.
integrations/llmd/ext-proc/src/lib.rs	Implements configuration and filter hooks.
integrations/llmd/ext-proc/src/mutations.rs	Converts and applies mutations.
integrations/llmd/ext-proc/src/proto.rs	Exposes generated protocol modules.
integrations/llmd/ext-proc/src/tests.rs	Adds extensive unit coverage.
server/Cargo.toml	Adds the feature and optional dependency.
server/src/lib.rs	Registers the feature-gated filter.
tests/environment/Cargo.toml	Defines environment tests.
tests/environment/tests/suite/main.rs	Declares the test suite.
tests/environment/tests/suite/sim_backend.rs	Tests simulator-backed routing.
tests/utils/src/net/mod.rs	Exports simulator helpers.
tests/utils/src/net/simulator.rs	Manages simulator containers.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[praxis-bot]

Review: feat(llm-d): move ext_proc compatibility into AI

Well-engineered transfer of the llm-d ext_proc compatibility layer into the AI repository. The crate design, state machine, test coverage, and security posture are all strong.

Findings

Severity	Count
Critical	0
Large	0
Medium	2

Architecture

The duplex exchange state machine (duplex.rs) is the heart of this crate and is well-structured: six orthogonal state domains, transactional validation (local copy before commit), and clean separation between computation (compute_send_transition) and mutation (apply_send_transition). The BootstrapState enum with SyncWrapper for the non-Sync Process future is a solid approach to satisfying filter state constraints.

The feature-gated registration pattern in server/src/lib.rs is clean -- default builds pay zero cost.

Correctness

• Config validation is thorough: deny_unknown_fields, range checks on timeouts and status codes, and clear "not yet supported" rejections for unimplemented Envoy features.
• Pseudo-header and Host protection in mutations prevents ext_proc from rewriting authority or method.
• Body coalescing has overflow-checked size accumulation with a 64 MiB cap.
• The full-duplex on_request_body EOS path correctly sends an empty terminal body marker rather than re-sending accumulated data -- this relies on StreamBuffer having already forwarded each chunk incrementally.
• The ImmediateResponse status code handling correctly falls back to 500 for out-of-range values.

Test Coverage

~8,800 lines of unit tests covering config parsing, validation, proto conversions, all four HeaderAppendAction variants, exchange state machine ordering, timeout overrides, and proto-to-JSON conversion. Six environment integration tests prove the full request path through ext_proc + endpoint_selector to the llm-d-inference-sim container, including spoofed-header rejection and processor failure handling. Strong coverage.

Security

• Pseudo-header and Host mutations are correctly blocked in both request and response directions.
• Timeout overrides are clamped to max_message_timeout with strict protobuf duration validation (rejects negative, zero, sub-millisecond).
• The simulator_spoofed_destination_header_ignored test proves client-supplied routing headers cannot override processor decisions.
• Body mutation size is bounded with checked arithmetic.

Minor Observations (not blocking)

• The filters/src/agentic/a2a/mod.rs changes (em-dash to hyphen, O(n^2) notation) are unrelated editorial cleanups. Not harmful but they add noise to the diff -- a separate commit would keep history cleaner.
• The struct field reordering in apis/src/lib.rs and filters/src/lib.rs test utilities appears to track upstream HttpFilterContext field ordering changes. Fine, just noting the coupling.