dsh-openai

OpenAI-compatible /v1/chat/completions HTTP bridge backed by the DeepSeek Harness Python SDK (deepseek-harness-sdk), so any OpenAI client (OpenWebUI, official SDKs, curl) can drive a full coding agent — tools, sandbox, persistent shell — instead of a plain chat model. The agent works in a per-session directory and delivers its files through signed download links.

This directory is both the PyPI packaging root and the Docker build context. The full deployment documentation (compose, Caddy/TLS, env matrices, security notes) lives in the repository README.

Install & run

pip install dsh-openai                 # or, from a checkout: pip install -e .
set DEEPSEEK_API_KEY=sk-xxxx           # upstream model credential (export on Linux/macOS)
dsh_openai serve                       # -> http://127.0.0.1:8000/v1

python -m dsh_openai serve is identical. serve applies local-development defaults (loopback bind, ~/.dsh-openai/{home,workspaces}, in-process fakeredis, the shipped editor profile patch); precedence is CLI flag > DSH_OPENAI_* env > local default. Run dsh_openai serve --help for the flag list.

Endpoints

• POST /v1/chat/completions — JSON and SSE streaming
• GET /v1/models, GET /healthz (auth-free liveness)
• GET /v1/dsh/info, GET /v1/dsh/sessions/{id}/files, GET /v1/dsh/files/{token} (signed deliverable download)

Pin a conversation with the x-dsh-session-id header or a stable chat_id body field; both map onto one dsh session and one workspace directory.

Configuration

All runtime settings are DSH_OPENAI_* environment variables resolved by dsh_openai/config.py — including DSH_OPENAI_API_KEY (inbound bearer token), DSH_OPENAI_MODEL, DSH_OPENAI_MAX_SESSIONS, DSH_OPENAI_REDIS_URL, DSH_OPENAI_SYSTEM_PROMPT and DSH_OPENAI_PATCHES. Relative patch paths resolve against the installed package, so the shipped dsh_openai/patches/editor.patch.yml works from any cwd.

Development

pip install -e .[dev]
python -m pytest tests

License

Apache-2.0 — see the LICENSE file (mirrored from the repository root).