Tinker Integration

[ultmaster]

This pull request adds a new example integration for using Tinker as a backend training service with Agent-lightning. It introduces a comprehensive bridge package (agl_tinker) that adapts Agent-lightning rollouts and datasets to Tinker’s reinforcement learning workflow, allowing seamless fine-tuning and evaluation workflows. The changes include documentation, environment setup, and core adapter modules that enable interoperability between the two frameworks.

New Example Integration and Documentation:

• Added a new tinker example to the catalog, with a clear note on its unmaintained status and compatibility with Agent-lightning v0.2.1.
• Provided a detailed README.md in examples/tinker/ explaining the integration, setup instructions, workflow differences, troubleshooting, and included files.
• Added an .env.example template for environment variables required to run the Tinker integration, including keys for Tinker, OpenAI, WANDB, and CrewAI telemetry settings.

Core Bridge Package (agl_tinker/) Implementation:

• Introduced agl_tinker/algo.py, implementing an Algorithm wrapper that plugs Tinker’s training loop into Agent-lightning’s resource management and dataset system.
• Added agl_tinker/env.py, which provides adapters for Tinker’s RL environment and dataset builders, allowing Agent-lightning tasks to be used in Tinker workflows without modifying rollout logic.
• Included license headers in new package files for compliance.

To fix the issue, the code should avoid binding the socket to all interfaces ("" or "0.0.0.0"). In the context of port discovery, it is safer and functionally equivalent to bind the socket to the loopback address (127.0.0.1), which restricts the socket to local traffic only and avoids any potential security risk. In file examples/tinker/hello.py, replace s.bind(("", 0)) with s.bind(("127.0.0.1", 0)) in the _find_available_port() function. No other changes, methods, or imports are needed.

To fix this problem, change the s.bind(("", 0)) line in _find_available_port() so it binds to the loopback interface only, by specifying '127.0.0.1' instead of the empty string. This restricts the socket to accept connections only from the local machine during the port discovery. No additional imports or significant code changes are needed; just one argument needs updating in the identified function in examples/tinker/q20_train.py at line 61.

[Copilot]

Pull Request Overview

This PR adds a comprehensive Tinker integration example to Agent-lightning, enabling reinforcement learning fine-tuning using Tinker as the backend training service. The integration bridges Agent-lightning's rollout architecture with Tinker's training infrastructure.

Key changes:

• Added new Tinker integration example with bridge code (agl_tinker/) to connect Agent-lightning with Tinker's RL training
• Implemented two example agents: a minimal "Hello" agent and a complex 20 Questions game using CrewAI
• Added wandb dependency to the tinker extra in project configuration

Reviewed Changes

Copilot reviewed 17 out of 18 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
uv.lock	Added wandb dependency for tinker extra
pyproject.toml	Added wandb to tinker dependencies
examples/tinker/README.md	Comprehensive documentation for Tinker integration
examples/tinker/hello.py	Minimal training example demonstrating identity repetition
examples/tinker/q20_*.py	20 Questions game implementation with training and evaluation scripts
examples/tinker/agl_tinker/*.py	Bridge package connecting Agent-lightning with Tinker
examples/tinker/.env.example	Environment variable template
examples/README.md	Updated examples catalog with Tinker entry

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Missing space in string concatenation for 'you're' check. Should be (\"you're \" + task) and (\"you are \" + task) to properly match phrases like "you're 42" instead of "you're42".

Suggested change

	elif ("you're" + task) in content_lower or ("you are" + task) in content_lower:
	elif ("you're " + task) in content_lower or ("you are " + task) in content_lower:

[Copilot]

The validation dataset is being created from the wrong source. It should use the original train_dataset variable from line 233, but after line 251, train_dataset has been reassigned to a list. This line should reference the original dataset before reassignment. Store the original dataset in a separate variable to avoid this issue.