5 sessions on shiki with a coding agent: the agent failed at least once in 2, and 1 finished without a human stepping in.
Takeaway: Next time I'd start the first prompt by asking for a full audit of what each bundling layer (parser, languages, themes) actually pulls in, instead of chasing library swaps first.
Takeaway: Next time I'd point directly at the CSS variables involved (--tool-output-bg, --shiki-dark-bg) since the agent had to spend a lot of turns grepping through multiple components to find the source of the color mismatch.
Takeaway: I'd start by pointing out the uncommitted package.json/pnpm-lock.yaml changes myself so the agent doesn't have to pause and ask — that stop-and-ask was reasonable but cost time I could've saved with more context up front.
Takeaway: Next time I'd ask the agent up front to inspect how @pierre/diffs actually themes its shadow DOM (CSS variables vs.
Takeaway: Next time I'd point directly at toolBlock.ts and ChatTranscript.vue upfront since I already knew roughly where tool rendering logic lived, instead of letting the agent grep broadly first.