feat(executor): add openai-compat tool type for HTTP API endpoints

[RyderFreeman4Logos]

Summary

• Add OpenaiCompat as the 5th tool type in CSA, enabling integration with OpenAI-compatible HTTP APIs (litellm, vllm, local proxy servers)
• New TransportMode::OpenaiCompat — pure HTTP transport, no process management, no cgroup sandbox
• Config resolution from env vars at execute time (OPENAI_COMPAT_BASE_URL, OPENAI_COMPAT_API_KEY, OPENAI_COMPAT_MODEL) via [tools.openai-compat.env] in global config

Design

• Closed enum pattern: exhaustive match arms across all tool enums (compiler-enforced completeness)
• ModelFamily::Other for the new tool (not Anthropic/Google/OpenAI-specific)
• Config resolution order: extra_env → system env → executor model_override → error
• HTTP-only: is_tool_binary_available() returns true unconditionally (no binary to check)
• reqwest dependency already existed in workspace; added to csa-executor

Files changed (6 crates)

• csa-core: ToolName::OpenaiCompat variant, ModelFamily::Other
• csa-config: base_url/api_key fields on ToolConfig, all_known_tools, validation
• csa-executor: Executor variant, TransportFactory, new transport_openai_compat.rs
• cli-sub-agent: parse_tool_name, is_tool_binary_available

Test plan

• All 174 csa-executor tests pass (including new transport tests)
• All 46 csa-config validation tests pass
• Full workspace: 789 tests pass
• just pre-commit passes (fmt, clippy, deny, test, monolith check)
• Manual: csa run --tool openai-compat "hello" with local proxy at localhost:8317

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.