Instruction file imported from Mao-pcdue/MyAI (.cursor/rules/01-research-first.mdc). Copyright stays with the author.

Research first, code second

Strict mode. Before any non-trivial change you must produce a research note following docs/RESEARCH-PROTOCOL.md. No exceptions.

What counts as non-trivial

Any of these triggers the protocol:

• A new Rust module, command, or domain
• A change to the trust pipeline (policy / permission / workspace / audit)
• A new React feature, panel, or top-level state
• A new dependency in package.json or Cargo.toml
• A schema change in SQLite tables, settings, or RPC types
• Anything that touches another operator's domain
• Anything labeled with a phase milestone in docs/ROADMAP.md

What does NOT trigger the protocol

• Typo fixes, comment tweaks, formatting
• Single-file refactors that change no behavior
• Updating a string literal that has no security or contract implications

If you're unsure, treat it as non-trivial.

The research note (mandatory, before any code)

Produce this in chat before editing any file:

1. Goal. One sentence: what we're changing and why.
2. MyAI docs read. Cite specific docs from docs/ you consulted (file + section).
3. MyAI code searched. Cite specific files/symbols. Note what already exists vs what's missing.
4. Inspiration scanned. Did you check how ../openhuman or ../JarvisOS handles this? Cite the file and the idea (never the code).
5. Trust boundary impact. Which invariants from .cursor/rules/02-trust-boundary-invariants.mdc apply? How does the design preserve them?
6. Tradeoffs. At least 2 options considered. Why the chosen one wins.
7. Plan. Numbered steps. Each step ≤ 1 file or 1 logical unit.

Only after the operator (or the rules) confirm the note do you start editing.

Anti-patterns (auto-fail the note)

• Skipping straight to code with "let me just try this"
• Re-implementing something already in src-tauri/ without citing why the existing version is insufficient
• Copying logic from ../openhuman or ../JarvisOS (read .cursor/rules/04-inspiration-not-vendoring.mdc)
• Vague tradeoffs ("the other approach is worse") without naming what's worse and why