feat(filter): add /v1/responses model rewrite filter for Codex passthrough

[nerdalert]

Summary

Adds openai_responses_model_rewrite, a Responses API request filter that translates client-facing model names into backend deployment names.

Capabilities

• Rewrites configured model aliases.
• Injects a default when model is missing or null.
• Publishes original and effective model headers and metadata for downstream routing.
• Preserves every other request field, including tools, instructions, input, function outputs, streaming flags, and unknown fields.
• Skips non-Responses traffic without modifying it.

Workflow

An operator deploys Praxis between their AI clients and inference backends, then configures alias rules such as:

codex-mini-latest -> llama-3.3-70b

When a user or tool such as Codex sends a POST /v1/responses request asking for codex-mini-latest, Praxis silently replaces the model name before forwarding the request to the backend.

The client does not need to know the backend's deployment name.

If a request arrives without a model, the operator can configure a default model that Praxis injects automatically.

Praxis also publishes routing headers such as x-praxis-ai-effective-model. Operators can use these headers to route rewritten models to different backend clusters:

Effective Model	Example Destination
llama-3.3-70b	Llama GPU pool
qwen-2.5-72b	Qwen GPU pool

All other request fields pass through untouched. Praxis does not execute client tools or take ownership of the Codex tool loop.

Example Configuration

  - filter: openai_responses_model_rewrite
    default_model: "llama-3.3-70b"
    model_aliases:
      codex-mini-latest: "llama-3.3-70b"
      gpt-4.1-mini: "qwen-2.5-72b"
    headers:
      effective_model: x-praxis-ai-effective-model
      original_model: x-praxis-ai-original-model

Demo And Results

The runnable passthrough demo, captured results, benchmark data, and documented claim boundaries are available here:

Native /v1/responses Passthrough Demo and Results (https://github.com/nerdalert/praxis-research-spikes/tree/main/demo/v1-responses-passthrough)

Demonstrated Behavior

• Native POST /v1/responses passthrough.
• Model alias rewriting and default injection.
• Effective-model-based backend routing.
• Streaming SSE preservation.
• Codex-shaped tools and function_call_output preservation.
• Mixed Responses and Chat Completions routing.
• Request-path benchmark profiles.

[praxis-bot]

Review: openai_responses_model_rewrite filter

Core logic is clean. Config validation thorough. 40+ unit tests and 11 integration tests. Follows established patterns. One compile-breaking bug and a few test gaps.

Findings: 1 Critical, 2 Medium, 3 Small, 4 Nit. See inline comments.

Non-inline: (1) is_responses_create does own path matching instead of reading classifier metadata, contradicting docs. (2) header_guard started early in content_length test. (3) Integration test file has doc comment.

[praxis-bot]

(superseded by later review)

[praxis-bot]

PR Review: feat(filter): add /v1/responses model rewrite filter for Codex passthrough

Summary: Adds openai_responses_model_rewrite, a Responses API request filter that translates client-facing model names into backend deployment names via alias mapping, with default model injection for missing/null values.

Overall: Well-structured filter implementation that follows the project established patterns closely. The code is clean, the test coverage is thorough with both unit and integration tests, and the integration tests exercise real end-to-end behavior including routing by effective model header. A few items need attention -- primarily a missing #[cfg(feature)] gate on a re-export, a documentation accuracy issue, and a note on JSON round-trip key reordering.

Severity	Count
Medium	3
Small	4
Nit	3

Findings without inline placement

• [Small] tests/integration/tests/suite/main.rs: The new mod openai_responses_model_rewrite; declaration should be gated with #[cfg(feature = "ai-inference")] to match the pattern used for other AI-inference test modules. Without the gate, the module will fail to compile when the feature is disabled.

• [Nit] filter/src/builtins/http/ai/openai/responses/model_rewrite/mod.rs: The on_request_body method checks body.is_none() before calling rewrite_body, but rewrite_body immediately re-checks with let Some(raw) = body.as_ref(). The outer guard makes the inner one unreachable -- not a bug, but a minor redundancy.

---

See inline comments for all other findings.

[leseb]

i'm requesting changes because i'd like to discuss some key point of the filter's goals, otherwise the goal makes sense to me!

[praxis-bot-app]

PR too large: 1763 lines added (limit: 750, excludes Cargo files, tests, docs, examples, and benchmarks). Please split into smaller PRs. Add skip/pr-conventions label to override.