# Tri Harness Deepseek Ollama Routing

> Route non-frontier, throughput-safe, and routine heartbeat work across all three CLI harnesses (Claude Code, OpenAI Codex, and OpenCode) onto DeepSeek (deepseek-v4-pro, deepseek-flash, deepseek-chat, deepseek-reasoner) and local Ollama (hybrid-coder, ollama-coder, workstation-ai, qwen2.5-coder:14b) via Claw Router (cr) and the LiteLLM Gateway (https://llm.megabyte.space).

- Skill: `heymegabyte/tri-harness-deepseek-ollama-routing` (Agent Skill)
- Install (CLI): `npx get-skill add heymegabyte/tri-harness-deepseek-ollama-routing`
- Raw SKILL.md: https://api.skillmd.com/api/skills/heymegabyte/tri-harness-deepseek-ollama-routing/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: heymegabyte (https://skillmd.com/u/heymegabyte)
- Updated: 2026-11-16
- Page: https://skillmd.com/skills/heymegabyte/tri-harness-deepseek-ollama-routing

---


# Tri-Harness DeepSeek & Ollama Offload (`claude` + `codex` + `opencode` via Claw Router & LiteLLM Gateway)

## Purpose

Every Paperclip (`swarml.ink`) and OpenClaw (`openclaw.swarml.ink`) agent must preserve premium Frontier subscription quotas (Claude Ultra/Max, Codex Pro, Gemini Ultra) for high-leverage work that genuinely requires frontier reasoning, and offload **all throughput-safe, routine, mechanical, and scheduled work** onto **DeepSeek** and **Ollama** across **all three CLI harnesses**:

1. **Claude Code (`claude` via Claw Router `cr`)**
2. **OpenAI Codex (`codex`)**
3. **OpenCode (`opencode` via `opencode-deepseek.sh`)**

All three harnesses connect through **Claw Router (`cr`)** and the **LiteLLM Gateway (`https://llm.megabyte.space`)**, which provides 4-key DeepSeek load-balancing (`deepseek-v4-pro`, `deepseek-flash`, `deepseek-chat`, `deepseek-reasoner`), local GPU Ollama routing (`hybrid-coder`, `ollama-coder`, `workstation-ai`, `local-llm`, `qwen2.5-coder:14b`) with automatic `deepseek-chat` fallback, and Langfuse observability.

---

## 1. When to Use Frontier vs. DeepSeek / Ollama

| Tier | Allowed Harnesses & Models | Mandatory Workloads |
| --- | --- | --- |
| **Tier 1: Frontier Subscription** | `fleet-cli/claude` (`cr --provider claude`), `fleet-cli/codex` (`codex exec`), `fleet-cli/gemini` (`agy --model gemini-4-carbon-high`) | **OpenSpec Authoring (`proposal.md`, `design.md`, `tasks.md`, `specs/<cap>/spec.md` — NEVER `opencode`)**, cross-repo system architecture, STRIDE threat modeling, complex security reviews, critical production incident root-cause analysis, and final release sign-off. |
| **Tier 2: DeepSeek Throughput (Cloud Offload)** | `fleet-cli/deepseek` (`opencode-deepseek.sh -m litellm/deepseek-v4-pro`), `fleet-cli/claude-deepseek` (`cr @litellm`), `fleet-cli/codex-deepseek` (`codex exec -p deepseek`) | Executing pre-written OpenSpec `tasks.md` checklists, routine heartbeats, Summarizer SITREPs, HR fleet audits, FinOps telemetry, Content/SEO/A11y sweeps, Swarm worktree wave dispatch, and test/lint triage. |
| **Tier 3: Hybrid Ollama + DeepSeek Fallback (Zero-Marginal-Cost Local GPU)** | `fleet-cli/ollama` / `fleet-cli/hybrid-coder` (`opencode-deepseek.sh -m litellm/hybrid-coder`), `fleet-cli/claude-ollama` (`cr @ollama`), `fleet-cli/codex-ollama` (`codex exec -p ollama`) | High-frequency code refactoring (`Refactorer`), mechanical dead-code removal, boilerplate generation, documentation formatting (`Technical-Writer`), and local code edits on `ubuntu.megabyte.space`. |

---

## 2. Exact Invocation Commands Across All 3 Harnesses

### A. Claude Code (`claude`) via Claw Router (`cr`)

Claw Router (`cr`) registers three zero-plaintext API/Gateway backends (`@litellm`, `@ollama`, `@deepseek`) alongside the 3 Claude Max/Ultra OAuth accounts (`brian@megabyte.space`, `hey@megabyte.space`, `blzalewski@gmail.com`):

```bash
# 1. Claude Code harness running DeepSeek V4 Pro via LiteLLM Gateway:
cr @litellm --provider claude -p --output-format stream-json --verbose --dangerously-skip-permissions "Execute task..."

# 2. Claude Code harness running Local Ollama (hybrid-coder -> deepseek-chat fallback) via LiteLLM Gateway:
cr @ollama --provider claude -p --output-format stream-json --verbose --dangerously-skip-permissions "Execute refactor..."

# 3. Claude Code harness running direct DeepSeek Anthropic-compatible endpoint:
cr @deepseek --provider claude -p --output-format stream-json --verbose --dangerously-skip-permissions "Execute task..."

# 4. Frontier Claude Code subscription (auto-rotates lowest-usage OAuth account):
cr --provider claude -p --output-format stream-json --verbose --dangerously-skip-permissions "Author OpenSpec proposal/design/tasks..."
```

### B. OpenAI Codex (`codex`) via LiteLLM Gateway (`wire_api = "responses"`)

Codex CLI is configured in `~/.codex/config.toml` and `~/.codex/{deepseek,deepseek-pro,flash,ollama}.config.toml` to route against `https://llm.megabyte.space/v1` using `LITELLM_API_KEY`:

```bash
# 1. Codex harness running DeepSeek via LiteLLM Gateway:
LITELLM_API_KEY="$(get-secret LITELLM_API_KEY)" codex exec --json --skip-git-repo-check --dangerously-bypass-approvals-and-sandbox -p deepseek "Execute task..."

# 2. Codex harness running Local Ollama (hybrid-coder) via LiteLLM Gateway:
LITELLM_API_KEY="$(get-secret LITELLM_API_KEY)" codex exec --json --skip-git-repo-check --dangerously-bypass-approvals-and-sandbox -p ollama "Execute refactor..."

# 3. Frontier Codex Pro subscription:
codex exec --json --skip-git-repo-check --dangerously-bypass-approvals-and-sandbox "Author OpenSpec or complex multi-file implementation..."
```

### C. OpenCode (`opencode`) via `opencode-deepseek.sh` & LiteLLM Gateway

`~/.config/opencode/opencode.jsonc` declares both `litellm` (`https://llm.megabyte.space/v1`) and `deepseek` (`https://api.deepseek.com`) providers with all DeepSeek and Ollama models:

```bash
# 1. OpenCode running DeepSeek V4 Pro (4-key rotated via LiteLLM):
~/.agentskills/bin/opencode-deepseek.sh run --format json -m litellm/deepseek-v4-pro "Execute task..."

# 2. OpenCode running DeepSeek V4.1 Flash (low-latency triage & summarization):
~/.agentskills/bin/opencode-deepseek.sh run --format json -m litellm/deepseek-flash "Summarize run..."

# 3. OpenCode running Local Ollama Hybrid Coder (qwen2.5-coder:14b with DeepSeek fallback):
~/.agentskills/bin/opencode-deepseek.sh run --format json -m litellm/hybrid-coder "Refactor dead code..."

# 4. OpenCode running Local Workstation GPU Ollama directly:
~/.agentskills/bin/opencode-deepseek.sh run --format json -m litellm/ollama-coder "Generate unit test..."
```

---

## 3. Automatic Task-Tier Offload in `openclaw-fleet/cli.py`

The OpenClaw Fleet CLI plugin (`control-plane/openclaw-fleet/cli.py`) inspects every incoming Paperclip / OpenClaw prompt and `PAPERCLIP_WAKE_REASON`:

1. **Routine Timer Heartbeats & Scheduled Routines (`PAPERCLIP_WAKE_REASON=heartbeat_timer` / `routine_trigger`)**:
   - Automatically offloads onto the agent's harness backed by DeepSeek/Ollama (`claude-deepseek`, `codex-deepseek`, `deepseek`, `ollama`) before touching Frontier subscription quotas.
2. **Throughput-Safe Tasks (Summarization, SITREPs, Audits, Refactors, Boilerplate)**:
   - Automatically executes on `claude-deepseek` / `codex-deepseek` / `deepseek` / `ollama` (`hybrid-coder`).
3. **Frontier OpenSpec & Architecture Tasks (`openspec/changes/`, `proposal.md`, `design.md`, `tasks.md`, `specs/`, `ADR-`, `STRIDE`)**:
   - Enforces Frontier execution (`claude`, `codex`, `gemini`) first, with full PACE fallback continuity.

