PSP PRX Plugin Maintenance

Building and debugging PSP kernel plugins (.prx) — cheat devices, VSH tools, sysmon overlays — especially when rebuilding legacy codebases with modern toolchains. Homebrew EBOOT ports are psp-development's territory; this skill covers prx plugins and real-hardware validation.

Trigger

• User reports a PSP plugin fork/rebuild crashes on real hardware ("CI passed but my PSP crashes")
• Working on TempAR / CWCheat-style prx plugins or any seplugins/-deployed kernel plugin
• Toolchain upgrade of a legacy PSP plugin (moving pspdev to GCC 14/15)

Core rules

1. CI green ≠ hardware-valid. Rebuilding 2010s-era PSP plugin C with modern pspdev (GCC 14/15) can pass CI yet crash on a real PSP: the old code relies on UB (stack layout, OOB reads like for(i=0;i<3;i++) over a 2-element array) that newer compilers compile differently. The game may even boot and play intros, then crash at asset-load. Field-test the actual prx before claiming success, and tell the user explicitly that CI is not proof.
2. Find the original working release binary FIRST. Upstream repos often still host original release assets (e.g. raing3/tempar v1.6.4 tar.gz = the old-toolchain TempAR binary, 122,450 B). Download and diff: size, md5, psp-readelf -S section sizes (.text/.data/.bss). Near-identical sections → the regression is codegen behavior, not footprint — stop chasing memory-size theories.
3. Keep the ELF for disassembly. make release runs clean and deletes src/*.elf. Use make prep <tgt> so the ELF survives, then psp-objdump -d src/<tgt>.elf | awk '/<fn>:/,/\$/'. On Windows, objdump/readelf must run via the toolchain on the build host (homelab), not local.
4. Sysmem partitions matter at game RAM peaks. sceKernelAllocPartitionMemory from the game's own partitions (1/6) steals RAM the game needs at its asset-load peak → hard crash (GTA LCS pattern: boots, title cards, crash when 3D assets start). The original TempAR alloc list was {9,1,6} (still visible commented out in kmalloc.c). A "correct" loop-count fix that truncates the fallback partitions removes the safety net that made older builds survive. Fix: restore the author's original partition list AND fail closed — if all named partitions are exhausted, return NULL instead of calling sceKernelAllocPartitionMemory(0, ...) (partition 0 may silently draw from the game's pool).
5. CFLAGS overrides drop include paths. A command-line CFLAGS=... makes build.mak's CFLAGS += -I... inert (command-line vars ignore +=). Re-add explicitly: -I$PSPDEV/psp/include -I$PSPSDK/include -Iinclude -Isrc.
6. Old toolchain images are gone from Docker Hub (oldest pspdev/pspdev tag ≈ v20240701, GCC 14-era; pspdev/pspsdk no longer exists). Original codegen cannot be reproduced via Docker — diff against the original release binary instead.
7. Multiple kernel plugins contaminate field tests. If the stick carries other plugins (AutoCat, PSPSysMon, etc.), disable them for a clean control run before blaming the binary under test.

Release workflow (GitHub Actions)

• Repos of this class often auto-release on v* tags via a build workflow (docker run pspdev/pspdev make prep psp lite + make pack + softprops/action-gh-release).
• Version bump: RELVER is hardcoded in the makefile — bump it or the packaged zip keeps the old name despite the new tag.
• Beware cp1252/MSYS mangling of non-ASCII commit messages on Windows git-bash — the pushed subject garbles (cosmetic; do not force-push history to fix it).
• Watch CI with gh run watch <id> --repo <owner>/<repo> --exit-status; verify release assets (prx sizes) after.

Pitfalls

• Patch-tool edits on CRLF C sources mangled indentation this session — use a Python normalize-replace-save (match on \n, write back \r\n) for block edits in this codebase (same lesson as the homelab repair_*.py pattern).
• Windows Defender/vendors may delete freshly downloaded .prx binaries from temp dirs (curl reports success, file vanishes) — download on the Linux build host instead.
• Shell $(cmd) interpolation inside ssh one-liners can trip the terminal hardline parser — use awk two-pass instead of mid-command variable capture.

See references/psp-plugin-toolchain-triage.md for the full TempAR case study (evidence, fixes, tags).