[RobotSail]

Summary

Adds 16 new entries to AGENTS.md distilled from recent merged PRs and agent conversation patterns:

Working With This Codebase (8 new entries):

• trl 1.x SFTConfig parameter migration (packing, packing_strategy must be on SFTConfig not SFTTrainer kwargs)
• Unsloth must be imported before ART in subprocess (GuidedDecodingParams shim ordering)
• Non-standard AdamW beta2=0.95 default in both SFT/OSFT backends
• Callback system subprocess serialization via inspect.getsource + base64
• enforce_eager=True workaround for NVIDIA driver 580+/610+ torch.fx crash
• Dual-registered VLM model detection (Gemma 3 pattern)
• trust_remote_code=True consistency across all AutoConfig/AutoTokenizer calls

Past Incidents (8 new entries):

• datasets 5.0.0 pyarrow compatibility break (mini_trainer PR Update AGENTS.md with learnings from June 2026 work #110)
• LoRA SFT packing crash with trl 1.x (training_hub PR Fix LoRA SFT packing config for trl 1.x compatibility #104)
• vLLM torch.fx crash on NVIDIA driver 580+/610+ (training_hub PR Fix ART backend compatibility with newer NVIDIA drivers #98)
• Unsloth import ordering in ART subprocess (training_hub PR Fix ART backend compatibility with newer NVIDIA drivers #98)
• transformers 5.9.0 + kernels 0.15.1 incompatibility
• Duplicate PR flood from agent dispatch (9 PRs for same issue)
• Agent GitHub token missing PR review and commit status permissions

Sources: mini_trainer PRs #109-110, training_hub PRs #98-108, instructlab/training PRs #695-696, Multica agent conversations MT-2 through MT-20.

Test plan

• Review new entries do not duplicate existing AGENTS.md content
• Verify cited PR numbers are correct and merged
• Check entries are actionable and categorized correctly

🤖 Generated with Claude Code

[coderabbitai]

Warning

Review limit reached

@RobotSail, you've reached your PR review limit, so we couldn't start this review.

Next review available in: 52 minutes

Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available.
You're only billed for reviews past your plan's rate limits ($0.25/file).

How can I continue?

After more reviews become available, a review can be triggered using the @coderabbitai review command as a PR comment. Alternatively, push new commits to this PR.

To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews.

How do review limits work?

CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability.

For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window.

Please refer docs for additional details.

Review details

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: 7528c285-d310-45e2-a4f7-521000c916b8

📥 Commits

Reviewing files that changed from the base of the PR and between b2e28dd and 0e61a32.

📒 Files selected for processing (1)

• AGENTS.md

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch agent/mini-trainer-learner/79e8d731

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}