DeepSeek V4 Pro Failover Guide

The Problem

What we were actually solving.

DeepSeek V4 Pro runs would start fine and then stop partway through. The root cause was not a crash: the deepseek-v4-pro model was served only by opencode-go. Our opencode-go subscription has a usage limit, so when a goal exhausted that quota mid-run, there was nothing to fall back to and the work halted.

The fix: let the official DeepSeek API also serve the same deepseek-v4-pro id, so the proxy can automatically spill over to it when opencode-go is quota- or cooldown-blocked. Same model name in, uninterrupted work out.

How It Works Now

One model id, two providers, priority-ordered failover.

Because both providers publish the same model id, the router treats them as one pool. fill-first drains opencode-go (priority 9) first, then automatically routes to the official DeepSeek API (priority 5) when opencode-go can't serve the request.

Before vs After

What Changed in the Config

All edits live in /Users/adeel/.cli-proxy-api/config.yaml.

A backup of the original was saved alongside it as config.yaml.bak.deepseek-opencode-primary-fallback.20260709042903.

1. Set routing.strategy: fill-first so the router prefers the highest-priority ready provider and only spills over when it is blocked.
2. Made opencode-go and official Deepseek both publish the shared id deepseek-v4-pro — this is what forms the single failover pool.
3. Kept direct-pin aliases so each path can be tested in isolation: deepseek-v4-pro-oc (opencode-go) and deepseek-v4-pro-official (official).

Routing block

routing:
  strategy: fill-first

Official DeepSeek provider

- name: Deepseek
  base-url: https://api.deepseek.com
  api-key-entries:
    - api-key: sk-b...e102        # official DeepSeek key
  models:
    - name: deepseek-v4-pro
      alias: ''                   # shared id: deepseek-v4-pro
    - name: deepseek-v4-pro
      alias: deepseek-v4-pro-official   # direct-pin for testing
  priority: 5                     # fallback tier

opencode-go provider

- name: opencode-go
  base-url: https://opencode.ai/zen/go/v1
  api-key-entries:
    - api-key: sk-K...z7un         # opencode-go key
  models:
    - name: deepseek-v4-pro
      alias: deepseek-v4-pro       # shared id: deepseek-v4-pro
    - name: deepseek-v4-pro
      alias: deepseek-v4-pro-oc    # direct-pin for testing
  priority: 9                      # preferred tier

Live State & Verification

Model resolution (/v1/models)

Runbook — Reproduce or Adjust

The commands used to edit, validate, restart, and verify.

View the config

sed -n '1,120p' /Users/adeel/.cli-proxy-api/config.yaml

Validate YAML

ruby -ryaml -e 'YAML.load_file(ARGV.fetch(0)); puts "yaml_ok"' \
  /Users/adeel/.cli-proxy-api/config.yaml

Restart the proxy

launchctl kickstart -k user/$(id -u)/com.cliproxyapi

Check the service

launchctl print user/$(id -u)/com.cliproxyapi \
  | rg 'state =|pid =|runs =|last exit code'

Check exposed DeepSeek models

curl -sS -H 'Authorization: Bearer sk-cliproxy-local' \
  http://127.0.0.1:8317/v1/models \
  | ruby -rjson -e 'd=JSON.parse(STDIN.read)["data"];
      puts d.select{|m| m["id"]=~/deepseek/i}.sort_by{|m| m["id"]}
            .map{|m| "#{m["id"]}=#{m["owned_by"]}"}'

Inspect routing logs

rg -n "deepseek-v4-pro" /Users/adeel/.cli-proxy-api/logs/stdout.log | tail -80

Honest Caveat