Onboard or Refresh Context

Workflow

1. •Read top-level AGENTS.md and extract hard constraints.
2. •Find or create agent_assets/<conversation_topic>/ with MEMORIES.md, code/, and artifacts/.
3. •Read relevant prior work:
   • •Existing agent_assets/**/MEMORIES.md files relevant to the current goal.
   • •Any pointers/manifests for prior outputs.
4. •Study the repo:
   • •Top-level README/docs.
   • •Core entry points and scripts/modules tied to the task.
   • •Magic numbers, hardcoded paths, and fragile assumptions.
5. •Load supporting skills:
   • •Discover skills in .agent/skills and .agent/repo_skills (if present).
   • •Run the keeping_memories skill.
   • •Read additional skills that are relevant to the task.
6. •Check conversation summary (if available):
   • •Current status.
   • •Blocking issues.
   • •Next actionable step.
7. •Update agent_assets/<conversation_topic>/MEMORIES.md with:
   • •Current goal and scope.
   • •What exists and where (files, scripts, artifacts).
   • •Known issues and open questions.
   • •Next actionable step.

Required Workspace Layout

All agent-only files must live under agent_assets/<conversation_topic>/:

agent_assets/
└── <conversation_topic>/
    ├── MEMORIES.md
    ├── code/
    └── artifacts/

• •code/: one-off scripts or utilities used only by the agent.
• •artifacts/: temporary logs, manifests, and test outputs.

If first-time onboarding

• •Spend extra effort mapping out the repository structure and workflows.
• •Create an initial baseline snapshot of state and goals in MEMORIES.md.

If refreshing context after a gap

• •Use your knowledge of the codebase to ingest relevant files and code.
• •Improve upon your map of the codebase and workflows.

Behavior expectations (align with repo rules)

• •Follow agentic instructions.
• •Follow rules over long workflows.
• •Write and run little temporary test scripts often (placed in agent_assets/code/) for development and debugging.
• •Keep MEMORIES.md updated and high signal-to-noise.
• •Use and refresh skills (e.g., coding_style, conversion tools). Keep track of which skills are available to you and use them when relevant.
• •Refrain from changing settings, installing packages, or otherwise modifying the environment unless explicitly asked.
• •Ask questions when something is unclear or you can spot a potential problem down the line.
• •Store artifacts liberally.
  • •Keep lightweight and scannable artifacts in agent_assets/ (notes, manifests, small logs, plots, tables, tiny checkpoints).
  • •Route larger artifacts (for example large .pkl/.npy/.npz, model checkpoints, full run dumps, cache directories) to <temp_data_dir> when defined in AGENTS.md.
  • •Leave a pointer/manifest in agent_assets/<conversation_topic>/artifacts/ for anything moved out.