[examples, gateway, tasks] feat: add DeepEyes training recipe

[Fvvs-iu]

Summary

Add DeepEyes as a complete visual-tool GRPO training workflow for multimodal
visual question answering.

The DeepEyes policy can inspect image crops through ImageZoomInTool before
producing a final answer. DeepEyes-specific behavior remains in the owning
layers: the Agent owns the multimodal ReAct loop and crop action; the Task owns
episode lifecycle and reward context; the Task package owns preprocessing and
Judge-based scoring; Gateway and Framework changes preserve compatible
multimodal requests and task metadata during rollout.

The deepeyes recipe also serves as a lightweight, self-contained quickstart
for the agentic RL training path, keeping data preparation, rollout, reward,
and training entry points in one place.

No separate issue is linked because this PR delivers the complete DeepEyes
training recipe and its supporting core integration.

Changes

• Add registered deepeyes Agent and Task implementations, including the
  image-zoom tool, visual-question preprocessing, and Judge-based reward.
• Add a runnable examples/deepeyes GRPO recipe: parquet dataset adapter,
  task/Judge configuration, training launcher, and eight-NPU Qwen3.5-4B preset.
• Bound rollout generation with configurable turn, per-turn-token, and
  total-token limits; retain the configured system prompt and support
  configurable request retries.
• Normalize multimodal image URLs for Qwen-compatible gateway requests and
  preserve task metadata required for reward evaluation.
• Document data preparation, training controls, and a local fixed 48-sample
  validation result: Qwen3.5-4B accuracy improves from 50.0% to 79.2%.

Validation

• python -m pytest -q tests/uni_agent/agents/test_deepeyes_agent.py 5 passed
• python -m pytest -q tests/uni_agent/framework/test_generate_sequences_on_cpu.py 28 passed
• pre-commit run --all-files --show-diff-on-failure --color=always passed

Checklist

• The PR is focused and explains why no issue is needed.
• The title follows the required format and names the owning layers.
• Tests were added for the core DeepEyes behavior.
• User-facing config and workflow changes include runnable examples and documentation.
• No compatibility-breaking API or migration is required.
• Logs, fixtures, and examples contain no credentials or private data.
• pre-commit run --all-files --show-diff-on-failure passes.

[tardis-key]

The README references examples/deepeyes/run_4b_7p1_container.sh and examples/deepeyes/task_config_4b.yaml, but neither file is included in this PR. Could you add them or update the README to use train_deepeyes.sh directly?

[Fvvs-iu]

The README references examples/deepeyes/run_4b_7p1_container.sh and examples/deepeyes/task_config_4b.yaml, but neither file is included in this PR. Could you add them or update the README to use train_deepeyes.sh directly?

Thanks for catching this. I’ve updated the README to use train_deepeyes.sh directly and added the Judge startup instructions.

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

❌ Patch coverage is 82.12560% with 74 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
uni_agent/agents/deepeyes/tool.py	68.06%	38 Missing ⚠️
uni_agent/agents/deepeyes/agent.py	81.17%	32 Missing ⚠️
uni_agent/gateway/session/codec.py	86.66%	4 Missing ⚠️

📢 Thoughts on this report? Let us know!

[zackcxb]

Is this variable needed? It seems enforced nowhere.