Skip to content

CLAUDE.md

Latest commit

 

History

History
104 lines (81 loc) · 6.05 KB

CLAUDE.md

File metadata and controls

104 lines (81 loc) · 6.05 KB
Raw

FastLED AI Agent Guidelines

Read the Right File for Your Task

By what you're doing:

Task Read
Writing/editing C++ code agents/docs/cpp-standards.md
Creating an API wrapper type agents/docs/cpp-standards.md → "API Object Pattern"
Writing/editing Python code agents/docs/python-standards.md
Editing meson.build files agents/docs/build-system.md
Running tests, Docker, WASM, QEMU agents/docs/testing-commands.md
Test-Driven Development (TDD) Use /tdd or /tdd-implement skills
Hardware autoresearch / bash autoresearch agents/docs/hardware-autoresearch.md
Debugging a C++ crash agents/docs/debugging.md
Creating a new C++ linter agents/docs/linter-architecture.md
Detailed command reference agents/docs/commands-reference.md
Workflow and task management agents/docs/workflow.md

By directory: src//tests/ → agents/docs/cpp-standards.md | ci/ → agents/docs/python-standards.md, agents/ci.md | tests/ → agents/tests.md | examples/ → agents/examples.md | meson.build → agents/docs/build-system.md

Key Commands

CRITICAL: Always use bash wrapper scripts (NOT direct Python invocation):

  • bash test / bash test --cpp / bash test TestName — Run tests
  • bash lint — Code formatting/linting
  • bash compile wasm --examples Blink — Compile example (WASM is default target)
  • bash compile <platform> --examples Blink — Compile for specific hardware (only when explicitly requested)
  • bash autoresearch --parlio — Live device testing (must specify driver)
  • bash profile <function> — Performance profiling

NEVER use: uv run python test.py — use bash test or uv run test.py FORBIDDEN: --no-fingerprint (use bash test --clean), bare pio/platformio, bare meson/ninja/clang++

See agents/docs/commands-reference.md for Docker, fbuild, WASM, profiling, example compilation, and override mechanism. See agents/docs/build-system.md for full command execution rules and forbidden patterns.

Core Rules (ALL AGENTS)

Git and Code Publishing

  • Default mindset: finish the job. Agents should not leave uncommitted changes dangling on master/main. If you made edits on master, the correct end-state is a feature branch + pushed PR — not a dirty working tree.
  • Feature branches — full autonomy, no user consent required. Agents may freely create branches, commit, push, and open PRs against any branch that is NOT master/main. Do this proactively when work is complete.
  • master/main — extra caution required.
    • NEVER commit directly to master/main.
    • NEVER push directly to master/main.
    • NEVER force-push to master/main (or any branch with an open PR others may be reviewing).
    • If changes exist on master, move them: git checkout -b feat/<topic> carries the working-tree changes to a feature branch, then commit + push + open a PR there.
  • Recovery pattern for uncommitted changes on master:
    1. git status — confirm scope
    2. git checkout -b <descriptive-branch> — changes follow to new branch
    3. git add <specific files> + git commit (conventional commit format)
    4. git push -u origin <branch> and gh pr create
    5. git status again to confirm clean tree

Hook Error Policy

  • ALWAYS stop and fix Write/Edit hook errors immediately before writing the next file
  • IWYU errors may be deferred when laying down multiple new files in a batch

Test Failure Debug Policy

  • When ANY test fails in quick mode, you MUST immediately re-run it in debug mode: bash test <TestName> --debug
  • Debug mode enables ASAN/LSAN/UBSAN sanitizers that catch memory errors, undefined behavior, and leaks
  • Quick mode failures without debug re-run are INCOMPLETE — the root cause is often only visible with sanitizers
  • Do NOT attempt to fix the code based only on quick-mode output — always get debug output first

Error Fixing Policy

  • Fix ALL encountered errors immediately, even pre-existing ones unrelated to your current task

Command Execution

  • Always use bash wrapper scripts (bash test, bash compile, bash lint, bash autoresearch)
  • Stay in project root — never cd to subdirectories
  • Python scripts: Always use uv run python script.py (never bare python)
  • Platform compilation timeout: 15 minutes minimum for platform builds
  • Override: FL_AGENT_ALLOW_ALL_CMDS=1 prefix bypasses forbidden command checks
  • See agents/docs/build-system.md for full rules

Code Standards

  • C++: See agents/docs/cpp-standards.md (span convention, DMA patterns, naming, macros)
  • JavaScript: Run bash lint --js after modifying JS files

Code Review Rule

ALL AGENTS: Run /code-review after making code changes.

Memory Refresh Rule

ALL AGENTS: Read the relevant agents doc before concluding work.

Workflow

See agents/docs/workflow.md for full workflow orchestration and task management.

  • Plan first for non-trivial tasks (3+ steps) — use plan mode
  • Subagents for research, exploration, and parallel analysis
  • Self-improvement: Update agents/tasks/lessons.md after corrections
  • Verify before done — prove it works with tests, logs, demonstrations
  • Test simplicity: Keep tests simple, avoid mocks. See agents/tests.md
  • TDD for features/bugs: Use /tdd (guided cycle) or /tdd-implement (full feature). Write tests FIRST.
  • Orchestrated sub-agents skip bash test — when a sub-agent is one step of a multi-step plan, the orchestrator runs bash test --cpp once at the end (the Stop hook covers this for free). Per-step sub-agents run bash lint only. See agents/tests.md → "Orchestrated Sub-Agent Carve-Out". The orchestrator must say so explicitly in the dispatch prompt.

Core Principles

  • Simplicity First: Make every change as simple as possible. Minimal code impact.
  • No Laziness: Find root causes. No temporary fixes. Senior developer standards.
  • Minimal Impact: Changes should only touch what's necessary. Avoid introducing bugs.