feat: Integrate opencode-go provider with deepseek-v4-pro and deepseek-v4-flash models

[carlito-hummingbot]

feat: Integrate OpenCode Go provider — deepseek-v4-pro, deepseek-v4-flash and open-source model access

Problem

Condor users who want to run open-weight models (DeepSeek V4 Pro, DeepSeek V4 Flash,
etc.) currently have to self-host or use pay-as-you-go API providers with unpredictable
costs. There is no integrated, flat-rate subscription option that gives predictable
monthly pricing with a curated set of capable open-source coding models.

What is OpenCode Go?

OpenCode Go is a low-cost subscription service for open-source
AI coding models. It costs $5 for the first month, then $10/month, and provides
~$60/month in token value against the same models priced per-token on other providers.

Key features:

• Predictable flat-rate pricing — no per-token surprise bills
• Curated model selection — production-ready open-weight models
• OpenAI-compatible API — works as a drop-in pydantic-ai provider
• Generous usage — $12/5hr soft limit, $30 weekly, $60 monthly equivalent

Available models (through OpenCode Go / Zen)

Model ID	Description
deepseek-v4-pro	Flagship DeepSeek V4 Pro — strong reasoning, coordinator role
deepseek-v4-flash	DeepSeek V4 Flash — fast, efficient for routine/market-making agents
hy3	High-performance coding model
qwen3-235b	Qwen 3 235B — large general-purpose model
qwq-32b	Qwen 32B reasoning model

All resolve against https://opencode.ai/zen/go/v1, an OpenAI-compatible endpoint.

Changes

Provider registration (condor/acp/pydantic_ai_client.py):

• Added opencode-go and opencode to the pydantic-ai provider allowlist
• Registered https://opencode.ai/zen/go/v1 as the default base URL
• Dedicated error messages for opencode-go model resolution failures
• Follows the same pattern as openai: / openrouter: providers

Model assignment (3 AGENT.md files):

• Coordinator (assistants/condor) → opencode-go:deepseek-v4-pro
• Routine builder → opencode-go:deepseek-v4-flash
• Market making expert → opencode-go:deepseek-v4-flash

Config (utils/config.py, main.py):

• Model family registration in config manager
• Provider import chain

Verification

• is_pydantic_ai_model('opencode-go:deepseek-v4-pro') → True
• is_pydantic_ai_model('opencode-go:deepseek-v4-flash') → True
• End-to-end: coordinator → delegate to MM expert (both on opencode-go),
  bot deployed and ran on Binance spot via manage_bots(deploy)

Notes

• Users set OPENCODE_API_KEY in their environment or via Condor's config manager.
  Sign up at https://opencode.ai/go (first month $5, then $10/month).
• Existing providers (openai, openrouter, etc.) are unaffected — additive only.
• Model assignments are defaults; users can override per session.

[fengtality]

don't change default models for condor and its agents. you can tell Claude not to commit these changes

[carlito-hummingbot]

Closing this, use PR#175 instead.