# Codex Deepseek Subagents > Install, update, test, or remove a cost-optimized Codex multi-agent setup where GPT stays the planner-reviewer and DeepSeek runs through a single Python runtime with visible routing, SSE events, and approved repository tools. Use when this capability is needed. - Skill: `tomevault-io/codex-deepseek-subagents` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds@latest add tomevault-io/codex-deepseek-subagents` - Raw SKILL.md: https://api.skillmd.com/api/skills/tomevault-io/codex-deepseek-subagents/raw - Safety review: pending (external: skill-scanner PASS, skillspector PASS) - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - Author: tomevault-io (https://skillmd.com/u/tomevault-io) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/tomevault-io/codex-deepseek-subagents --- # Codex DeepSeek Subagents Use this skill when the user wants Codex/GPT to remain the main planner-reviewer while DeepSeek handles explicit, confirmable worker tasks through the local runtime. ## Install Flow Install the skill with Codex `skill-installer`, then materialize project-local files inside the target repository. Windows PowerShell for the initial install: ```powershell powershell -ExecutionPolicy Bypass -File skills/codex-deepseek-subagents/scripts/deepseek-codex.ps1 install -ApiKey ``` After install, prefer the generated local wrappers: ```powershell .\.codex\deepseek-codex.cmd start-runtime .\.codex\deepseek-codex.cmd test-runtime .\.codex\deepseek-codex.cmd doctor .\.codex\deepseek-codex.cmd analyze --prompt "Analyze this repository." .\.codex\deepseek-codex.cmd tui ``` Linux/macOS: ```bash bash skills/codex-deepseek-subagents/scripts/deepseek-codex.sh install --api-key ./.codex/deepseek-codex.sh start-runtime ./.codex/deepseek-codex.sh test-runtime ./.codex/deepseek-codex.sh doctor ./.codex/deepseek-codex.sh analyze --prompt "Analyze this repository." ./.codex/deepseek-codex.sh tui ``` Compatibility aliases still exist for `start-proxy`, `stop-proxy`, and `test-proxy`, but the canonical command surface is the runtime naming. ## Managed Files ```text user_config.json .codex/config.toml .codex/agents/deepseek-worker.toml .codex/deepseek.local.env.sh .codex/deepseek.local.env.ps1 .codex/deepseek-codex.cmd .codex/deepseek-codex.sh .codex/runtime/deepseek_scheduler.py .codex/runtime/deepseek_runtime.py .codex/runtime/deepseek_client.py .codex/runtime/events.py .codex/runtime/render.py .codex/runtime/patch_preview.py .codex/runtime/tool_protocol.py .codex/runtime/usage.py .codex/runtime/doctor.py .codex/runtime/tui.py .codex/runtime/task_queue.json .codex/runtime/sessions.json .codex/runtime/events.log.jsonl .codex/runtime/stdout.log .codex/runtime/stderr.log .codex/test-runtime.sh .codex/test-runtime.ps1 ``` `user_config.json` is shareable and non-secret. API keys must stay in `.codex/*.local.*` or environment variables. ## Runtime Responsibilities The runtime provides: - `GET /healthz` - `GET /v1/agents` - `POST /v1/tasks` - `GET /v1/tasks/{task_id}` - `POST /v1/tasks/{task_id}/approve` - `POST /v1/tasks/{task_id}/retry` - `POST /v1/responses` - `POST /v1/sessions` - `GET /v1/sessions/{session_id}` - `GET /v1/sessions/{session_id}/events` `/v1/responses` supports: - text delegation - approved native repository tools for `execution` tasks - `stream=true` through SSE events - official DeepSeek `tools` / `tool_calls` - patch approval `requires_action` responses Shell command execution remains disabled. ## Delegation Rules Before delegating to DeepSeek, the main agent must describe exactly what will be sent: ```text I can send this to DeepSeek worker now. Scope to be sent: ; files/paths: ; exploration allowed: . This may send repository content to DeepSeek or the configured proxy. Confirm before I delegate. ``` Default to `listed paths only`. Prefer `analyze` for read-only repository analysis. There should be no DeepSeek dispatch before user confirmation. ## Agent Registry `user_config.json` keeps two built-in agent kinds: - `codex_main` - `deepseek_worker` Default routing: - `analysis` -> `codex_main` - `review` -> `codex_main` - `execution` -> `deepseek_worker` Native repository reads and writes must stay inside the approved tool policy for the execution task. ## Thinking Mode Use thinking mode only when the task complexity justifies the extra cost: - simple tasks: `pro` or `flash` - complex implementation or debugging: `pro-thinking` or `flash-thinking` Default `--thinking-view` is `hidden`. `summary` only shows local hidden-chars/token metadata. Raw `reasoning_content` is shown only when the user explicitly chooses `--thinking-view raw`, and it must not be persisted or replayed. ## UX Surface - `delegate` and `analyze` must show a route card and scope card before work is sent. - `analyze` is the preferred read-only path and only enables list/read/search repository tools. - native patch tasks must show `patch.preview` before `patch.applied`. - patch application must always wait for explicit approval through `list-patches`, `show-patch`, `approve-patch`, `reject-patch`, and `apply-patch`. - `tui` opens a runtime dashboard only and must not send a DeepSeek request by itself. - if TUI is unavailable in the current terminal, the runtime must print a fallback message and continue with `stream-cli`. ## Maintenance - `update` rewrites `user_config.json` into the current schema and removes stale legacy artifacts. - `doctor` reports runtime readiness, route display, reasoning stream support, native tool readiness, and stale legacy artifacts separately. - `export-shareable` excludes local runtime state, logs, caches, and secrets. --- > Source: [C-NHXT-N/codex-deepseek-subagents-skill](https://github.com/C-NHXT-N/codex-deepseek-subagents-skill) — distributed by [TomeVault](https://tomevault.io).