---

description: "Fetch a GitHub issue, investigate, plan, implement, test, and commit the fix" argument-hint: "<issue-number>"

Fix Issue Pipeline

Phase 1: Fetch

gh issue view $ARGUMENTS --repo matiaszanolli/ByroRedux --json title,body,labels,state

Save details to .claude/issues/$ARGUMENTS/ISSUE.md.

Phase 2: Classify Domain

Classify the issue — the matched crate becomes the -p test target in Phase 6:

• renderer → byroredux-renderer — Vulkan, pipeline, GPU memory, sync, shaders, RT acceleration
• ecs → byroredux-core — Storage, queries, world, systems, resources, components
• nif → byroredux-nif — NIF parsing, block types, version handling, import pipeline
• animation → byroredux-core (+ byroredux-nif for KF import) — interpolation, blending, animation system
• bsa → byroredux-bsa — BSA/BA2 archive reading, extraction, decompression
• esm → byroredux-plugin — ESM record parsing, cell loading, XCLL lighting
• platform → byroredux-platform — Windowing, input, OS abstractions
• cxx → byroredux-cxx-bridge — C++ interop
• binary → byroredux — Game loop, scene setup, demo code

Phase 3: Investigate

Read only the source files on the code path — don't pre-read a whole crate, and don't re-read files already in context. Trace the path inline.

Specialist agents are a last resort, not a default. Each one is a fresh context window — only delegate when the issue genuinely spans 2+ domains and you can't trace it yourself. When you do, ask the agent for a conclusion (file:line + cause), not file dumps.

INVESTIGATION.md is optional. Skip it for single-site fixes — the commit body covers those. Write it only when the investigation uncovered non-obvious findings worth preserving (cross-file interactions, a wrong-looking-but-correct invariant).

Phase 4: Scope Check

If fix touches >5 files, pause and confirm with the user before proceeding.

Phase 5: Implement

Follow project conventions from CLAUDE.md.

• Shaders: recompile SPIR-V if GLSL changed
• No new dependencies without user approval
• Inner loop = cargo check -p <crate> (the Phase 2 crate), not the full workspace. It's the fastest signal while iterating. Do not run cargo test on every edit — save it for Phase 6.

Phase 6: Verify

Scope first, widen only at the end. Keep output quiet so test logs don't flood context — -q prints dots for passes and full detail only for failures.

cargo test -q -p <crate>          # the Phase 2 crate — the only suite that can regress for a scoped fix
cargo test -q                     # full workspace — ONCE, final gate (skip the redundant `cargo check`: test already built everything)

All tests must pass with zero warnings. Run the full suite a single time at the end, not per fix iteration. For a fix confined to one crate with no cross-crate surface, the scoped run is sufficient evidence — note that you scoped it.

Phase 7: Completeness Checks

Before committing, verify:

• [ ] UNSAFE: Any new unsafe blocks have safety comments
• [ ] SIBLING: Same bug pattern checked in related files
• [ ] DROP: Vulkan object lifecycle correct (destroy-before-parent)
• [ ] TESTS: Regression test added if applicable
• [ ] PINS: A HIGH/CRITICAL fix whose proof is an #[ignore]d real-data test appends <lane> <test fn> to .github/real-data-pins.txt (#6224)

Phase 8: Commit & Close

Commit with conventional message referencing the issue:

Fix #<number>: <description>

Close the issue:

gh issue close $ARGUMENTS --repo matiaszanolli/ByroRedux --comment "Fixed in <commit-hash>"