feat(agents): add DeepSeek-V4 native support

[gutierrezx7]

DeepSeek-V4 Support Implementation

Summary

This PR adds native support for DeepSeek-V4 models to oh-my-openagent, enabling users to leverage DeepSeek's advanced agent capabilities within the OMO orchestration framework.

Problem Statement

Users of oh-my-openagent (OMO) who wanted to use DeepSeek-V4 faced compatibility issues:

1. No model detection: DeepSeek-V4 models were not recognized, falling back to generic handling
2. Suboptimal prompts: The default Claude-style prompt (~540 LOC) was too verbose for DeepSeek's attention patterns, causing "lost in the middle" issues
3. Missing fallback chains: DeepSeek-V4 wasn't available in agent fallback chains
4. Incorrect configuration: No proper support for V4-Pro's native thinking mode vs V4-Flash's fast inference

Solution

1. Model Detection (src/agents/types.ts)

Added four detection functions to properly identify DeepSeek model variants:

isDeepSeekV4Model()       // Detects any V4 variant
isDeepSeekV4ProModel()    // Detects Pro variant with thinking support
isDeepSeekV4FlashModel()  // Detects Flash variant for fast inference
isDeepSeekR1Model()       // Detects R1 reasoning model (future use)

Design rationale: Separate detection functions allow granular control over model-specific features (e.g., thinking mode for Pro only).

2. Native Prompt (src/agents/sisyphus/deepseek-v4.ts)

Created a purpose-built prompt for DeepSeek-V4 with the following optimizations:

• Compact structure: ~330 LOC vs ~540 LOC default prompt, reducing "lost in the middle" issues
• High-attention placement: Critical instructions at start and end of prompt
• XML-tagged sections: Clear structure for better parsing
• No corrective overlays: Unlike Gemini, V4 doesn't require behavioral corrections

Key sections:

• Phase 0: Intent Classification (mandatory first step)
• Phase 1: Tool Usage Rules with parallel execution guidelines
• Phase 2: Implementation Rules with delegation structure
• Phase 3: Failure Recovery procedures
• Phase 4: Completion criteria

3. Routing Logic (src/agents/sisyphus.ts)

Added V4-specific routing in createSisyphusAgent():

if (isDeepSeekV4Model(model)) {
  // Use native V4 prompt
  // Enable thinking for Pro variant (32k budget)
  // No reasoningEffort (V4 uses thinking, not OpenAI-style reasoning)
}

Configuration:

• V4-Pro: thinking: { type: "enabled", budgetTokens: 32000 }
• V4-Flash: No thinking config (optimized for speed)

4. Fallback Chains (src/shared/model-requirements.ts)

Added DeepSeek-V4-Pro to agent fallback chains:

Sisyphus:

claude-opus-4-7 (max) → kimi-k2.5 → gpt-5.5 (medium) → deepseek-v4-pro (max) → glm-5 → big-pickle

Oracle:

gpt-5.5 (high) → gemini-3.1-pro (high) → claude-opus-4-7 (max) → deepseek-v4-pro (high) → glm-5

5. Test Coverage (src/agents/types.test.ts)

Added comprehensive tests (42 test cases):

• Provider-prefixed models: llmgateway/deepseek-v4-pro
• Providerless models: deepseek-v4-pro
• Negative cases: V3, R1, other models correctly rejected
• Variant separation: Pro vs Flash detection

Technical Decisions

Why Native Prompt Instead of Overlays?

DeepSeek-V4 differs from Gemini in key ways:

• ✅ Follows system prompts correctly (unlike R1)
• ✅ Has native tool calling with strict mode
• ✅ Doesn't hallucinate tool calls (unlike Gemini)
• ✅ Has larger context window (1M tokens)

A native prompt allows better optimization for V4's architecture than injecting overlays into a generic prompt.

Why Not Support R1 in This PR?

DeepSeek-R1 has fundamentally different requirements:

• Official docs recommend avoiding system prompts
• Requires instructions in user messages instead
• Needs different temperature settings (0.5-0.7)
• Has unique reasoning behavior

Supporting R1 requires architectural changes beyond this PR's scope.

Why Thinking Instead of reasoningEffort?

• reasoningEffort is OpenAI-specific
• DeepSeek-V4 uses Claude-style thinking with budget tokens
• Consistent with existing Claude handling in codebase

Usage

Configuration

Add to oh-my-opencode.jsonc:

{
  "agents": {
    "sisyphus": {
      "model": "llmgateway/deepseek-v4-pro"
    }
  }
}

Or use via LLM Gateway:

{
  "agents": {
    "sisyphus": {
      "model": "llmgateway/deepseek-v4-pro"
    }
  }
}

Available Models

Model	Variant	Context	Features
deepseek-v4-pro	High quality	1M tokens	Thinking mode, strict tool calling
deepseek-v4-flash	Fast inference	1M tokens	Optimized for speed

Testing

All tests pass:

$ bun test src/agents/types.test.ts
42 pass
0 fail

Build succeeds:

$ bun run build
✓ Bundled 1295 modules
✓ TypeScript declarations emitted
✓ JSON Schema generated

Checklist

• Added model detection functions with provider support
• Created native V4 prompt optimized for agent orchestration
• Implemented proper routing with thinking configuration
• Added to fallback chains for Sisyphus and Oracle
• Comprehensive test coverage (42 tests)
• Build passes without errors
• TypeScript type checking passes
• Schema JSON regenerated

References

• DeepSeek-V4 Documentation: https://api-docs.deepseek.com/guides/tool_calls
• Issue [Feature]: Enhancing Support For Other Models #2072: Enhancing Support For Other Models
• Related PR feat(ollama-cloud): add :cloud variant parsing, glm-5.1 fallbacks, and GLM Sisyphus overlay #3517: GLM-5.1 overlay pattern (followed same approach)

Related

Closes #2072 (partially - V4 support, R1 deferred to future PR)

---

^{Codesmith can help with this PR, just tag @codesmith or enable autofix. Settings.}

• Autofix CI and bot reviews

---

Summary by cubic

Adds native DeepSeek-V4 support to Sisyphus with a compact, V4-optimized prompt and model-aware routing. Improves orchestration quality and adds deepseek-v4-pro to Sisyphus and Oracle fallback chains.

• New Features

  • Detects deepseek-v4 variants (pro, flash) and adds an r1 detector for future use.
  • Adds a native DeepSeek-V4 Sisyphus prompt (~330 LOC) with clear, XML-tagged phases. No corrective overlays.
  • Routes V4 models to the native prompt; enables thinking for deepseek-v4-pro (32k budget). No reasoningEffort.
  • Updates fallback chains to include deepseek-v4-pro.
  • Adds 42 tests covering provider-prefixed/providerless names, variant separation, and negatives.

• Migration

  • To use: set agents.sisyphus.model to llmgateway/deepseek-v4-pro (quality) or deepseek-v4-flash (speed).

^{Written for commit b12bab0. Summary will update on new commits.}

[github-actions]

All contributors have signed the CLA. Thank you! ✅
_{Posted by the CLA Assistant Lite bot.}

[cubic-dev-ai]

No issues found across 6 files

Confidence score: 5/5

• Automated review surfaced no issues in the provided summaries.
• No files require special attention.

_{Auto-approved: Native support for DeepSeek-V4 is implemented as a new, isolated code path with specific model detection and comprehensive tests. No existing model logic is modified.}

[gutierrezx7]

I have read the CLA Document and I hereby sign the CLA

[gutierrezx7]

recheck

[gutierrezx7]

I have read the CLA Document and I hereby sign the CLA

recheck

[gutierrezx7]

recheck

[jtoronto]

Waiting for this.