Add LLM Post-Training Engineer - Engineering

[kaining-never-stop]

What does this PR do?

Adds an LLM Post-Training Engineer agent for diagnosing post-training failures and making evidence-based experiment and checkpoint-release decisions.

Agent Information

• Agent Name: LLM Post-Training Engineer
• Category: engineering
• Specialty: SFT, DPO, GRPO/RLVR, MoE post-training, evaluation gates, and checkpoint integrity

Motivation

While running post-training experiments, I needed a focused workflow that distinguishes execution success from learning-signal quality, checkpoint integrity, and held-out behavior. The existing AI Engineer has a broader ML-production scope; this agent focuses on the post-training decision boundary.

Testing

Ran:

• ./scripts/lint-agents.sh engineering/engineering-llm-post-training-engineer.md
• ./scripts/check-agent-originality.sh engineering/engineering-llm-post-training-engineer.md
• ./scripts/check-divisions.sh
• ./scripts/check-tools.sh
• ./scripts/check-runbooks.sh
• ./scripts/convert.sh --tool all --out <temporary directory>

Also tested the agent against real post-training incident scenarios covering SFT loss masking, DPO truncation, reward-signal degeneracy, MoE routing drift, checkpoint integrity, and managed-task liveness.

Checklist

• Original — not a near-duplicate
• Follows the agent template structure
• Includes personality and voice
• Has concrete deliverable templates
• Defines measurable success metrics
• Includes a step-by-step workflow
• Proofread and formatted correctly
• Tested in real scenarios

[msitarzewski]

Merged! 🎉 Landed in #742. Your LLM Post-Training Engineer is impressively rigorous — the evidence-based experiment-gate and checkpoint-integrity discipline (SFT/DPO/GRPO/RLVR) is exactly the kind of specialist depth we want. Distinct from the general AI Engineer. Thanks @kaining-never-stop!