kimi-responses-adapter: Instructions file for Codex

Repository Guidelines

Project Structure & Module Organization

This is a Rust (tokio/axum) service: an OpenAI Responses API adapter for the Kimi Code Anthropic API. It is stateless and holds no credentials.

• src/main.rs — entrypoint; loads env config and starts the axum server.
• src/adapter/ — all adapter logic:
  • config.rs env-driven Config; types.rs Responses/Anthropic API types.
  • convert.rs Responses → Anthropic request conversion.
  • stream.rs Anthropic SSE → Responses SSE state machine (core).
  • nonstream.rs non-streaming response conversion.
  • reasoning.rs thinking/signature ↔ encrypted_content codec.
  • models.rs upstream model-metadata cache (lazy, TTL, single-flight).
  • server.rs routing, upstream calls, raw passthrough for non-Responses paths.
• Tests live next to the code in #[cfg(test)] modules; there are no assets or fixtures directories.

Build, Test, and Development Commands

Toolchain and tasks are managed by mise (mise install once):

mise run build   # cargo build
mise run test    # cargo test
mise run lint    # cargo clippy --all-targets -- -D warnings
mise run fmt     # cargo fmt
mise run ci      # fmt --check + clippy + test (what CI runs)
mise run run     # serve on :8787 (LISTEN_ADDR to override)

Coding Style & Naming Conventions

• Standard Rust: cargo fmt-clean, cargo clippy --all-targets -- -D warnings clean (run mise run ci before committing).
• Names mirror the protocol they belong to: Anthropic* types for the Kimi upstream side, Responses*/InputItem for the OpenAI client side.
• Emitted SSE/JSON payloads are built with serde_json::json! maps (sorted keys, matching Go's map[string]any output); parsed input uses serde structs with #[serde(default)] so unknown fields are ignored. Raw JSON that must round-trip byte-for-byte (tool arguments, schemas) is kept as Box<serde_json::value::RawValue>.
• Comment protocol quirks (e.g. web-search status suppression) where they are handled.