TORCH/CLAUDE.md

CLAUDE.md

Operational playbook for agent execution in this repository.

Scope

Applies to all agent tasks unless superseded by direct runtime instructions.

Operating model

• Keep output and changes generic by default.
• Prefer deterministic, auditable steps.
• Record decisions so future agents can continue with minimal rediscovery.

Handoff documentation workflow

When you learn something that can help future agents:

1. Check for existing related notes in docs/agent-handoffs/learnings/ and docs/agent-handoffs/incidents/.
2. If the learning is new, add a small markdown note using the naming standard in AGENTS.md.
3. Link related issue patterns in KNOWN_ISSUES.md when applicable.
4. Keep note content sanitized and reusable (no secrets, no user-identifying details).

What belongs where

• KNOWN_ISSUES.md: currently active/reproducible issues and known workarounds.
• docs/agent-handoffs/learnings/: durable implementation insights and successful patterns.
• docs/agent-handoffs/incidents/: failure reports, root-cause notes, and prevention guidance.
• CONTRIBUTING.md: process standards for proposing and reviewing changes.

Style for agent notes

• Prefer bullets over long prose.
• Include exact commands where useful.
• State confidence and remaining uncertainty.
• End with a “Next agent should …” line.

Anti-patterns

• Avoid repo-local jargon in shared templates.
• Avoid stale TODO dumps without context or owner.
• Avoid project/brand-specific terms unless explicitly required by the active repo.