fix-issues

name	fix-issues
description	Use when fixing compiler warnings, static analyzer findings (clang-tidy, etc.), or runtime errors/crashes in the libYSE codebase. Enforces the project's performance-first stance: never trade audio-thread speed for stylistic cleanliness, never modify vendored dependencies, never broaden scope beyond the reported issues.

Fixing warnings, analyzer findings, and runtime errors in libYSE

libYSE is a real-time C++ sound engine. Audio-callback code runs at buffer rate on a dedicated thread and must remain lock-free and allocation-free. "Cleaning up" code here is not free: a well-meaning fix can introduce a glitch, a click, or a crash that is far worse than the original warning.

This skill governs how to triage and fix reported issues. Read it fully before making any change.

Issue tracking — GitHub Issues only

All bugs, follow-ups, and known-issue notes live in GitHub Issues on yvanvds/yse-soundengine. Do not add new entries to a local KNOWN_ISSUES.md or any other in-repo issue list — that file has been retired and removed.

• Read context: before starting a fix, check the relevant issue with gh issue view <n>. Look for prior repro steps, fix sketches, and workarounds-in-tests already documented.
• File new findings: if a fix surfaces a separate bug, file it with gh issue create --title "..." --label bug --body-file ... rather than inlining a TODO comment that no one will find later. Workarounds in test code should reference the issue number (e.g. // see #29) instead of a file path.
• Close on landing — always. Every fixed issue must end up closed on GitHub. Two paths:
  • Preferred: write Closes #<n> (or Fixes #<n>) in the PR body so GitHub auto-closes the issue when the PR merges into the default branch. Verify after merge — auto-close does not fire for PRs targeting non-default branches (e.g. this repo's dev integration branch), so a dev PR still needs the manual close below once the merge to master lands.
  • Manual: gh issue close <n> --reason completed after merge, with a short comment if context is non-obvious (gh issue comment <n> -b "fixed in <sha>"). Use --reason not planned when closing as wontfix or duplicate.
  • Partial fix: leave the issue open and edit the body to describe what is still outstanding, rather than closing and filing a follow-up.
• Search before filing: gh issue list --search "<keyword>" — duplicates are easy to make when descriptions live across multiple subsystems.

The gh CLI is authenticated in the project environment; no MCP server or extra setup is required.

Core principles

1. Correctness bugs first, cosmetic warnings second, never both in one pass. A dangling pointer and an unused-parameter warning are not the same problem. Fix the bug properly. Silence the cosmetic warning with the smallest possible change.

2. Performance is non-negotiable on the audio thread. Anything inside or reachable from process() methods on dspObject / dspSourceObject, the PortAudio/OpenSL ES callback path, or the lock-free message queue consumer must not get slower. If a "fix" adds branches, allocations, virtual dispatch, atomic ops, or mutex acquisitions on this path, it is the wrong fix.

3. Stay in scope. Fix only what was reported. Do not refactor, modernize, reformat, rename, add [[nodiscard]] everywhere, swap containers, sprinkle noexcept, or "improve" anything adjacent. Each of those is a separate decision the user has not made.

4. Never modify vendored dependencies. Anything under dependencies/ (doctest, portaudio headers, rtmidi, libsndfile) is off-limits. Suppress warnings from these at the CMake target level or with a localized pragma around the include site, not by editing the header.

Triage: classify every reported issue before touching code

For each warning, analyzer finding, or error, decide which bucket it falls into:

A. Real bug

Examples: returning a pointer into a destroyed temporary, use-after-free, uninitialized read, missing override that hides a base method, data race, mismatched new/delete, signed/unsigned comparison that actually overflows.

These get fixed properly, even if the fix is bigger than a one-liner. A real bug in audio-thread code is the highest priority issue in this codebase because it manifests as glitches/clicks/crashes that are very hard to debug.

B. Cosmetic / pedantic warning

Examples: __COUNTER__ is a C2y extension, const qualifier on a return value type, unused parameters in virtual base methods with empty default implementations, unused-but-set variables in debug-only paths.

These get silenced with the minimum-impact mechanism (see below). Do not restructure code to satisfy a stylistic warning.

C. Maybe-bug-maybe-not

Examples: an overloaded virtual that hides the base, a signed/unsigned comparison in a loop bound, a "may be uninitialized" the compiler isn't sure about.

Investigate. If it is a bug → bucket A. If the compiler is being conservative and the code is correct → silence with [[maybe_unused]], static_cast, explicit initialization, using Base::method;, etc. Do not silence by disabling the warning globally.

D. Vendored / third-party

The reported file lives under dependencies/. Never edit. Suppress at the build-system level (target-scoped -Wno-...) or with a #pragma clang diagnostic push/ignored/pop block around the #include of the vendored header in our own code.

Performance rules

Before applying any fix, identify whether the affected code is on the audio thread. The audio thread is reached through:

• Any process() override on a dspObject or dspSourceObject subclass
• The PortAudio callback (device/portaudioDeviceManager.cpp) and OpenSL ES equivalent
• The consumer side of the lock-free message queue (utils/lfQueue.hpp)
• Anything called from the above transitively

On the audio thread, these are forbidden as part of a warning fix:

• Adding heap allocation (new, malloc, std::vector::push_back on a vector that wasn't pre-sized, std::string construction, etc.)
• Adding mutex/lock acquisition of any kind
• Adding try/catch or anything that could throw
• Replacing a plain field access with a std::atomic operation that wasn't already atomic (the project has aBool/aInt/aFlt wrappers — use those if atomicity is genuinely needed, but warning fixes rarely require it)
• Adding virtual dispatch where there was none
• Adding work inside hot loops (per-sample, per-frame) — even a single extra branch matters at 48000 × N-channel rates

Off the audio thread (application thread, manager singletons running in update(), file loading, demo code, tests), normal C++ rules apply, but still: minimum change to silence the warning.

Preferred silencing mechanisms (when the issue is genuinely cosmetic)

Ranked from least invasive to most:

1. Local annotation in our own code: [[maybe_unused]], (void)param;, /*paramName*/ comment, using Base::method;, explicit initialization, static_cast to the right type.
2. Localized pragma around a vendored #include: #pragma clang diagnostic push / ignored "-Wfoo" / pop in the file that includes the third-party header.
3. Target-scoped CMake suppression: target_compile_options(yse_tests PRIVATE -Wno-foo) for warnings that only matter in one target (tests are the typical case). Never apply globally to the engine target.
4. Disabling the warning globally: essentially never. Reserve for warnings that have no signal value at all in this codebase.

Platform coordination

If the warning is platform-specific (e.g. only appears on MSVC, only on MSYS2/Clang64, only on Android NDK), check that the fix doesn't break the other platforms. The CMake build covers Windows and Linux; Android is built separately. Suppression flags must be guarded by compiler/platform conditions when the warning only exists on one toolchain.

Workflow

1. Inventory. List every distinct warning/error from the build log. Group identical ones (a doctest __COUNTER__ warning hitting 38 times is one issue, not 38). Note the file path of each.
2. Classify. Assign each group to bucket A/B/C/D above.
3. Audio-thread check. For each bucket-A and bucket-C item, determine whether the affected code is on the audio thread.
4. Plan. Write down the intended fix per group before editing. If a fix would change anything on the audio thread other than removing a real bug, stop and surface it for review rather than applying it.
5. Apply. Make the smallest change that resolves each issue. Do not touch unrelated code in the same file.
6. Verify. Rebuild on the platforms in scope. Run ctest --preset tests-debug. Confirm warning count drops as expected and no new warnings appear. Spot-check at least one demo runs (audio-thread regressions typically show up at runtime, not at compile time).
7. Report. Summarize: groups fixed, mechanism used per group, any items deferred and why, any audio-thread-adjacent code touched.
8. Close. Once the fix merges, ensure every referenced GitHub issue is closed (PR keyword auto-close or gh issue close <n> --reason completed — see "Close on landing" above). Partial fixes stay open with an updated body.

Running clang-tidy locally

A baseline check set lives at .clang-tidy. Invoke analysis through the project wrapper rather than calling clang-tidy directly — the wrapper picks the right compile_commands.json and avoids the "doctest/doctest.h not found" failure mode on test files:

python yse.py analyze YseEngine/dsp/lfo.cpp        # one file
python yse.py analyze YseEngine/device/            # one directory
python yse.py analyze                              # whole project (slow)

Per CLAUDE.md item 6, run python yse.py analyze <changed-files> before committing and clear any new findings on the modified code. The baseline (~50 pre-existing findings across YseEngine/) is tracked as backlog — don't fix in passing during unrelated work.

Things that are out of scope for this skill

• Adding new compiler warnings to the build (enabling -Wfoo that wasn't on before) without an accompanying issue
• Migrating to a newer C++ standard
• Replacing project utilities (aFlt, MULTICHANNELBUFFER, the message queue, dspObject::link()) with standard-library equivalents
• Reformatting code that happens to be near a warning

If any of those would actually help, raise them as a separate proposal, not as part of a warning-fix pass.

Runtime errors and crashes

For runtime errors (assertions, segfaults, audio glitches, hangs) the same priorities apply but with different emphasis:

• Reproduce first. A crash you can't reproduce is not a crash you can fix safely.
• Distinguish audio-thread crashes from application-thread crashes. The former are usually races, lock-free queue misuse, allocations on the audio path, or DSP buffers being read past their length.
• Do not "fix" by adding locks on the audio thread. The lock-free contract is structural; a mutex on the callback is a regression even if it eliminates the immediate symptom.
• Audio glitches/clicks without a hard crash are real bugs, not cosmetic issues. Treat them as bucket A.