Arkie_models

Imported from arkie888/ArkieAgent-main (crates/Arkie_models/AGENTS.md). Install upstream with npx skills add arkie888/ArkieAgent-main --skill Arkie_models. Copyright stays with the author.

Agent Map — Arkie_models

Start Here

• Read CLAUDE.md first; it contains detailed provider and reasoning notes.
• Read Cargo.toml for provider dependencies and feature shape.
• Relevant files by area:
  • provider.rs, registry.rs, runtime.rs, host.rs — provider contracts and host seams.
  • retry.rs, failover.rs, circuit_breaker.rs, response_cache.rs — reliability and caching.
  • tool_schema.rs, rig_adapter.rs, openai_codex_provider.rs — tool-call/schema boundaries.
  • reasoning.rs — legacy reasoning engine and response shaping.
  • *_oauth.rs, *_auth.rs, session.rs, token_refreshing.rs — provider auth/session handling.

What This Crate Owns

• Multi-provider LLM integration with retry, failover, circuit breaker, and response caching.
• LlmProvider trait, provider registry/chain construction, runtime adapters, tracing/recording, model/cost metadata.
• Provider-specific auth for Arkie, Anthropic/Gemini OAuth, GitHub Copilot, OpenAI Codex/ChatGPT, AWS Bedrock.
• Tool schema normalization and provider-specific tool-call compatibility.
• Transcription, image/vision model helpers, smart routing, and test fault-injection support.

Do Not Move In Here

• Agent-loop/thread ownership or product workflow; those live in engine/turns/product crates.
• Tool execution side effects; complete_with_tools() must pass through cache and leave side effects to callers.
• Unredacted tokens, refresh tokens, request bodies, or provider internals in public errors/logs.
• Direct assumptions that one provider's model override/schema/auth behavior applies globally.

Validation

• Fast local check: cargo test -p Arkie_models
• Lint check: cargo clippy -p Arkie_models --all-targets --all-features -- -D warnings
• Run caller tests when changing provider trait, tool schemas, reasoning output, auth refresh, or failover behavior.

Agent Notes

• RigAdapter forwards per-request model overrides through rig-core's typed request model field; do not add a second model entry through flattened provider parameters.
• complete_with_tools() is never cached because tool calls can have side effects.
• OpenAI strict schema normalization rewrites optional object fields into required-nullable strict mode where required.
• 401 handling differs by auth mode; preserve provider-specific semantics from CLAUDE.md.