12 sessions on C with a coding agent: the agent failed at least once in 12, and 0 finished without a human stepping in.
Takeaway: Next time I'd ask upfront how this OS's other interactive JS apps handle input, since the busy-poll getKeyNonBlock() loop turned out to be the real cause of the exit crash, not the VGA I/O port theory the agent chased first.
Takeaway: Next time I'd specify the memory-arena ownership model (who owns terms from clauses vs.
Takeaway: Next time I'd tell it upfront to always launch background threads (HTTP listener, per-connection handlers) as self-cleaning/detached instead of calling thread.destroy, since that single join call silently blocked the whole game window from ever opening.
Takeaway: Next time I'd flag up front that any CLI flags removed during a refactor need a compatibility shim, since removing --recover silently broke one test fixture.
Takeaway: Next time I'd tell it upfront that Stage 3's 'parser/bytecode boundary' can't be a real physical separation in a 63K-line monolith like quickjs-ng, so it skips straight to building the JS_WriteObject/JS_ReadObject wrapper instead of discovering that mid-session.
Takeaway: Next time I'd nail down the QEMU/OVMF/WSL boot pipeline first before layering framebuffer and GRUB experiments on top of it — half the pain was Windows/WSL path translation and stale OVMF vars, not the actual kernel or userland code.
Takeaway: Next time I'd specify upfront that I want both UEFI/MB2 and legacy BIOS/MB1 to keep working, since that dual-path requirement is what caused most of the debugging churn around the framebuffer, paging, and GRUB modules.
Takeaway: Next time I'd tell it up front to match the existing input-loop pattern (blocking getChar, like quill and the shell) instead of letting it invent a getKeyNonBlock busy-poll — that's what caused the GPF on quit.
Takeaway: Next time I'd specify the persistent-arena/variable-registry design up front — the query-vs-clause arena confusion cost the most back-and-forth.
Takeaway: Next time I'd ask upfront for an audit of what quickjs-ng already exposes (JSMallocFunctions, JS_WriteObject/JS_ReadObject) before specifying a stage's API, since the agent ended up discovering mid-stage that the 'ideal' architecture didn't match the 63K-line monolith's reality.
Takeaway: Next time I'd start the prompt by stating upfront that I'm on WSL2/openSUSE driving QEMU from Windows, since half the wasted turns came from the agent trying WSL bash/PowerShell bridging that kept mistranslating paths.
Takeaway: Next time I'd specify upfront that MB1 must stay fully working as a real fallback, not just compile — that black-screen/no-console debugging loop cost the most time.