Copilot Artifacts

GitHub Copilot extensibility artifacts provide AI-assisted workflows for dataset analysis, training job management, and coding standards enforcement. These artifacts are configured in .github/ and activate automatically in VS Code.

📋 Artifact Inventory

Type	Name	Description	Path
Agent	Dataviewer Developer	Interactive dataset analysis and tool development	.github/agents/dataviewer-developer.agent.md
Agent	OSMO Training Manager	LeRobot training lifecycle on OSMO with Azure ML	.github/agents/osmo-training-manager.agent.md
Agent	Physical-AI RPI	Cloud-agent RPI orchestrator (umbrella)	.github/agents/physical-ai-rpi.agent.md
Agent	Physical-AI RPI Worker	Hidden generic subagent shell for Physical-AI RPI	.github/agents/physical-ai-rpi-worker.agent.md
Instruction	Commit Messages	Conventional Commits format for all commit messages	.github/instructions/commit-message.instructions.md
Instruction	Dataviewer	Coding standards for dataviewer development	.github/instructions/dataviewer.instructions.md
Instruction	Docs Style and Conventions	Writing standards for all markdown files	.github/instructions/docs-style-and-conventions.instructions.md
Instruction	Shell Scripts	Implementation standards for bash scripts	.github/instructions/shell-scripts.instructions.md
Prompt	/chatlog	Create and maintain conversation logs	.github/prompts/chatlog.prompt.md
Prompt	/check-training-status	Monitor OSMO training job progress	.github/prompts/check-training-status.prompt.md
Prompt	/start-dataviewer	Launch Dataset Analysis Tool	.github/prompts/start-dataviewer.prompt.md
Prompt	/submit-lerobot-training	Submit LeRobot training job to OSMO	.github/prompts/submit-lerobot-training.prompt.md
Skill	dataviewer	Dataset browsing, annotation, and export	.github/skills/dataviewer/SKILL.md
Skill	osmo-lerobot-training	Training submission, monitoring, and analysis	.github/skills/osmo-lerobot-training/SKILL.md

🔗 Quick Reference

Want to...	Use this artifact
Launch the Dataset Analysis Tool	/start-dataviewer prompt → Dataviewer Developer
Browse and annotate training episodes	Dataviewer Developer agent
Submit a LeRobot training job	/submit-lerobot-training prompt → OSMO Training Manager
Check training job status	/check-training-status prompt → OSMO Training Manager
Save a conversation log	/chatlog prompt
Enforce commit message standards	commit-message instruction (auto-applied)
Enforce coding standards in dataviewer	dataviewer instruction (auto-applied)
Enforce markdown writing standards	docs-style-and-conventions instruction (auto-applied)
Enforce shell script standards	shell-scripts instruction (auto-applied)

🤖 Agents

Dataviewer Developer

Interactive agent for launching, browsing, annotating, and improving the Dataset Analysis Tool.

Property	Value
Handoffs	Start Dataviewer, Browse Dataset, Annotate Episodes
Tools	All (no restrictions)
Skill	dataviewer
Prompts	/start-dataviewer

Four-phase workflow: Launch/Configure → Interactive Browsing (Playwright) → Episode Annotation (API+UI) → Feature Development (React+FastAPI).

OSMO Training Manager

Multi-turn agent for managing LeRobot imitation learning training lifecycle on OSMO with Azure ML integration.

Property	Value
Handoffs	Submit Training Job, Check Training Status, Run Inference Evaluation
Tools	11 explicit (run_in_terminal, memory, runSubagent, ...)
Skill	osmo-lerobot-training
Prompts	/submit-lerobot-training, /check-training-status

Five-phase workflow: Submit → Monitor → Analyze → Summarize → Inference Evaluation. Handles VM eviction recovery, CUDA errors, and KeyError failures.

Physical-AI RPI (Umbrella) and Physical-AI RPI Worker

Paired cloud-agent profiles that surface the autonomous Research → Plan → Implement → Review workflow from microsoft/hve-core inside cloud-agent sessions for this repository. The umbrella is picker-visible; the worker is hidden and reachable only via the umbrella's agent tool fan-out.

Property	Umbrella (physical-ai-rpi)	Worker (physical-ai-rpi-worker)
Target	github-copilot (cloud-agent only)	github-copilot (cloud-agent only)
Picker	Visible (user-invocable: true, default)	Hidden (user-invocable: false, disable-model-invocation: true)
Tools	read, edit, search, bash, agent, github/* write tools	read, edit, search, bash, read-only github/*
Role	Orchestrator: bootstrap verification, RPI procedure, persistence	Content-neutral executor: adopts upstream persona by name
Persists	Posts full phase artifacts as PR comments and maintains an "RPI Artifact Index" in the PR description	None (umbrella owns PR comments and PR description)

The two-agent split exists because cloud-agent surface constraints cannot be satisfied with a single profile:

• Picker visibility versus hiding. A profile is either selectable or not. The umbrella must surface in the picker; the worker must stay hidden so only the umbrella can dispatch it.
• Subagent context isolation. The cloud-agent agent tool runs each dispatch target in a fresh context window. The umbrella needs that isolation when fanning out to upstream researcher-subagent and phase-implementor personas — without it, every research finding and implementation diff would pollute the orchestrator context.
• Decoupling from microsoft/hve-core's subagent roster. The worker takes a persona: <stem> parameter and resolves it to .copilot-tracking/upstream/hve-core-rpi/subagents/<persona>.agent.md at dispatch time. New upstream personas auto-onboard via the next bootstrap with no PR in this repo.
• Different tool grants. The umbrella holds github write tools for per-phase PR comments; the worker is read-only on github and never commits. Merging them would leak orchestrator write authority into every research and implementation invocation.

Upstream RPI persona bodies are downloaded into the workspace by .github/workflows/copilot-setup-steps.yml (Bootstrap hve-core RPI persona step). The umbrella body reads _audit.md for the resolved microsoft/hve-core@main SHA and fails fast with a PR comment if the bootstrap is missing.

📝 Instructions

Instructions activate automatically when files matching their applyTo pattern appear in the chat context.

Name	Applies To	Purpose
Commit Messages	**	Conventional Commits format, scopes, line-length limits
Dataviewer	data-management/viewer/**	SOLID principles, test-first, validation commands
Docs Style and Conventions	**/*.md	Document hierarchy, tables, voice/tone, frontmatter
Shell Scripts	**/*.sh	Script template, library functions, deployment patterns

⚡ Prompts

Prompts are slash commands invoked via / in the chat input. Each prompt targets a specific agent.

Command	Agent Target	Required Inputs
/chatlog	Generic	None
/check-training-status	OSMO Training Manager	workflowId (optional)
/start-dataviewer	Dataviewer Developer	datasetPath
/submit-lerobot-training	OSMO Training Manager	dataset (required)

🛠️ Skills

Skills provide multi-file capabilities with progressive 3-level loading: discovery (frontmatter only) → instructions (SKILL.md body) → resources (bundled reference files).

dataviewer

Property	Value
Directory	.github/skills/dataviewer/
Resources	references/PLAYWRIGHT.md (selectors, interaction recipes, API endpoints)
Used by	Dataviewer Developer agent

osmo-lerobot-training

Property	Value
Directory	.github/skills/osmo-lerobot-training/
Resources	references/DEFAULTS.md (env, datasets, GPU profiles), references/REFERENCE.md (CLI, inference, AzureML navigation)
Used by	OSMO Training Manager agent

🔄 Workflow Chains

Agents compose prompts and skills into end-to-end workflows:

OSMO Training Manager (agent)
  ├── /submit-lerobot-training (prompt)
  ├── /check-training-status   (prompt)
  └── osmo-lerobot-training    (skill)
       ├── references/DEFAULTS.md
       └── references/REFERENCE.md

Dataviewer Developer (agent)
  ├── /start-dataviewer        (prompt)
  └── dataviewer               (skill)
       └── references/PLAYWRIGHT.md

Standalone:
  └── /chatlog                 (prompt, generic)

➕ Adding New Artifacts

VS Code provides generator commands for scaffolding new artifacts:

• /create-agent — Create a new custom agent
• /create-instruction — Create a new instruction file
• /create-prompt — Create a new prompt file
• /create-skill — Create a new agent skill

Place new artifacts in the corresponding .github/ subdirectory and update this inventory page.

Related Documentation

For broader project context, see these companion guides:

• Contributing Guide — Development workflow and coding standards
• Architecture — System architecture and agent skills design
• Prerequisites — Required tools and VS Code settings

---

Crafted with precision by Copilot following brilliant human instruction, then carefully refined by our team of discerning human reviewers.