feat(agents): add DeepSeek V4 Pro/Flash Sisyphus prompt routing

[EvangelosMoschou]

What

Adds DeepSeek V4 Pro/Flash model detection, Sisyphus prompt routing, a research-backed Sisyphus prompt with V4-specific guardrails, AND a harness-level earlier compaction threshold to prevent tool call drift.

Why

DeepSeek V4 Pro is a strong agentic coding model (80.6% SWE-bench, 73.6 MCPAtlas, 1M context) at 7x lower cost than Claude. Previously, V4 models fell through to the generic GPT fallback prompt. This PR gives them a dedicated prompt optimized for their strengths and aware of their weaknesses, plus a harness improvement that keeps context under the drift threshold.

Research-backed guardrails

Based on community findings (deepseek-ai/DeepSeek-V3#1244, Maestro, OpenSymphony, akitaonrails benchmark, advance-minimax-m3-cursor-rules):

Known issue	Source	Mitigation
Tool call format drift after ~40 tools	Issue #1244	Prompt: "use tool_calls field ONLY" + Harness: compaction at 35% (vs 78%) to stay under drift threshold
High hallucination rate (94%)	TechJacks	VERIFICATION REQUIRED section + subagent verification: "A subagent report is a lead, not evidence"
tool_choice rejection in thinking mode	Issues #1376, #836	Prompt: prefer "auto", avoid forced tool calls
Language bleed (Chinese tokens)	NV forums	Explicit language constraint
Vague spec fabrication	Maestro project	Precise spec delegation: "Give each delegate a PRECISE spec: file paths, acceptance criteria, scope boundaries"
Pro vs Flash coherence	DeepSeek docs	Variant-specific capability descriptions

Changes (9 files, +350 lines)

Prompt + Detection (7 files)

1. Detectors (model-family-detectors.ts): isDeepSeekV4Model, isDeepSeekV4ProModel, isDeepSeekV4FlashModel
2. Types barrel (agents/types.ts): re-export new detectors
3. Routing (sisyphus-agent-factory.ts): V4 models routed via buildGptSisyphusAgentConfig after Kimi, before GPT-5.5
4. Prompt (sisyphus/deepseek-v4.ts, new): 8-block prompt with 7 V4-specific guardrails + 3 research-backed improvements (context management, clarify-first, precise spec delegation)
5. Sisyphus barrel (sisyphus/index.ts): export buildDeepSeekV4SisyphusPrompt
6. Detectors test (model-family-detectors.test.ts): 9 new assertions
7. Routing test (sisyphus-agent-factory.test.ts): 2 new routing test cases

Harness Improvement (2 files)

8. Compaction trigger (preemptive-compaction-trigger.ts): V4 models get 35% compaction threshold (vs 78% default) to prevent tool call drift
9. Compaction test (preemptive-compaction.test.ts): 2 new tests (V4 triggers at 36%, non-V4 does not at 36%)

TDD Evidence

• Detectors: RED → GREEN (9/9 tests pass)
• Routing: 2 new cases assert Pro/Flash route to V4 prompt
• Compaction: RED (V4 test failed at 36%) → GREEN (V4 triggers at 36%, non-V4 preserved at 78%)
• All 18 existing preemptive-compaction tests still pass
• All 20 routing/detector tests still pass
• CI: green on all platforms

Rebase note

Branch was rebased onto upstream/dev. One conflict in model-family-detectors.test.ts (upstream added Claude Fable/Mythos tests at the same location) was resolved by keeping both test blocks.

Usage

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

The plugin will automatically detect the model family, route to the DeepSeek V4 prompt, apply the V4-specific guardrails, and compact context at 35% to prevent tool call drift.