Each one teaches a concept by making you do it, not read about it.

Real bugs from real repos, not tutorial exercises.

Zero-denominator crashes in FCPXML time parsing — Final Cut uses rational time notation (1001/30000s). Malformed files sent 0/0 framedurations and crashed the parser. Traced every division, added guards at the parsing boundary. 5 fixes, 7 new tests.

Silent marker loss from duplicate clip names — markers were placed by clip name, but editors reuse names. Later markers silently overwrote earlier ones — data loss, no error. Fixed by keying on (clip_name, clip_index) tuples so every placement is positionally unambiguous.

Context window management for autonomous agents — tool outputs consumed 83% of context. Added observation masking (compress to structured summaries), sub-agent isolation (scoped context per task), and a 70% utilization compression trigger. Eliminated mid-task truncation.

SSE invalidation starvation under rapid agent cycles — a plain trailing debounce reset on every event. During bursts the timer restarted endlessly and the UI never refreshed, showing stale data for minutes. Replaced with a hybrid throttle-debounce: 300ms trailing edge for coalescing plus a 1.5s max-wait ceiling guaranteeing freshness.

Per-frame memory allocation in an animation loop — Function.prototype.bind() inside requestAnimationFrame allocated ~60 objects/second. Cached the reference, moved timing to CSS custom properties. Cut ~3,600 allocations per minute.

Enum aliasing flake — MarkerType.INCOMPLETE and MarkerType.TODO shared an underlying value. Python aliases duplicate enum values to the first-defined member, so MarkerType(1) always returned TODO. The test asserted by name, got identity. Fixed by asserting on .value.

Monolithic agent config — a single config.json at 92 modules is a merge-conflict magnet. Should have split to per-module config around module 20. Config decisions compound faster than code decisions.

SQLite access patterns — right database, wrong abstraction. Raw SQL in 30+ modules instead of a thin repository layer. The storage choice was correct; the access pattern was the mistake.

Test strategy — hundreds of unit tests in fcpxml-mcp-server, but the bugs that shipped were integration-level: malformed XML that parsed fine alone and broke combined. Should have started with 10 integration tests against real FCPXML files.

Component explosion — 121 components, not enough composition. A <DataCard> with render props beats three single-use wrappers sharing 80% of their markup. Component count should fall as a codebase matures.

Learning Suite shared shell — 6 games with near-identical shells, each built independently. A shared <GameShell> would have cut 40% of the markup. Noticed at game 4 and kept shipping — speed over architecture was right for the first few, doubling down past that wasn't.