Next Cache Components Optimizer

next-cache-components-optimizer

Two loops, shared levers and primitives, different diagnostics:

• Page-render loop ([ppr-loop.md](./ppr-loop.md)) — grow the static shell of a single page. Rank Suspense fallback areas on a shell-only render.
• Nav loop ([instant-nav-loop.md](./instant-nav-loop.md)) — when the user clicks a link from A to B, show B's static layout immediately (chrome, structure, content-shaped fallbacks) instead of holding A's UI until B's data resolves. Capture B's suspended boundaries post-pushstate, classify each by suspended_by[].name, drop SSR-only client hooks.

Pick one and run it end-to-end.

requires

• next-dev-loop initiated for this session — it opens the headed browser, exposes the agent-browser CLI, and wires the dev MCP server that provides mcp get_logs.
• cacheComponents: true in next.config.ts. Refuse otherwise.

preflight (shared)

1. Confirm cacheComponents: true.
2. The user must already be at the page each loop needs in the headed browser (from next-dev-loop) — logged in, with any state set up. This skill can't drive auth, SSO, or MFA; it takes the manual setup as the starting point. (Each sub-loop names which page it expects.)
3. agent-browser get url to anchor the current route.

Each loop sets the instant cookie as needed (see the shared instant cookie section below).

instant cookie (shared)

Both loops use the next-instant-navigation-testing cookie to freeze the framework's dynamic-data writes. Once set, visible content on the page is the static shell + Suspense fallbacks — that's what we capture to assess the optimization.

Set it with a pending-lock tuple [0, ""]. The id is any unique string; the convention is a p-prefixed random stamp so concurrent scopes don't collide:

agent-browser cookies set next-instant-navigation-testing '[0,"p"]' \
  --url 

Each loop's preflight specifies when to set it within the flow. Clear it at the end (see teardown below).

decide which loop

• Page-render when the complaint is about one route's initial load. Read [ppr-loop.md](./ppr-loop.md).
• Nav when it's about navigating between two routes. Read [instant-nav-loop.md](./instant-nav-loop.md).

Ambiguous → ask.

shared refactor levers

• Push down — extract I/O into a Suspense-wrapped child so the parent stays static and static siblings lift into the shell.
• Recurse, don't blind-wrap. If a Suspense boundary already wraps a component containing both static content and the I/O, read inside, extract the I/O-dependent JSX into a new leaf, and lift the static siblings up.
• Cache — 'use cache' + cacheLife(). Always ask the user for freshness; map to a preset (seconds / minutes / hours / days / weeks / max / default).

Push-down and cache compose: push-down lifts static structure, cache eliminates the remaining data gap.

propose via plan mode (shared)

Each refactor goes through plan mode before applying. Treat this as a signal: the application work is non-trivial agentic engineering, not a templated edit. This skill provides the framework — which lever to reach for, which candidate to fix, what the expected visible delta is — but the real work (which file to edit, how to cleanly extract the I/O, where to place the new Suspense boundary, which cacheLife profile to ask the user for) is a judgment call you have to think through. Plan mode forces a coherent proposal before touching code, and gives the user a chance to redirect on any of those decisions.

no-shell bailout (shared)

The levers presume a shell exists to grow or cache toward. If the route is fully blocking — HTTP 500 with blocking-route or NEXT_STATIC_GEN_BAILOUT in mcp get_logs, or zero Suspense boundaries on a visibly-rendered page — there's no shell. Surface the structural blocker and stop; the user has to wrap the offending dynamic access in `` before either loop can help.

verify requires a visible delta (shared)

Each loop captures a baseline screenshot of the shell before applying any change, then re-screenshots after. Report both paths in the final summary so the user can see what changed. The two captures must visibly differ — fallback area shrunk, content promoted to the static surface, target fallback gone or content-shaped. Identical-looking captures mean the refactor didn't land; undo. "Compiles cleanly" is not the bar.

Hide the dev overlay before each screenshot. The Next.js dev overlay (`` at the document root) renders instant-nav guidance, build errors, and other dev chrome that pollute the before/after comparison. Hide it, screenshot, restore:

agent-browser eval "document.querySelector('nextjs-portal').style.display='none'"
agent-browser screenshot 
agent-browser eval "document.querySelector('nextjs-portal').style.display=''"

anti-patterns (shared)

Don't replace granular Suspense boundaries with a top-level loading skeleton. A loading.tsx for the whole segment, or a root-level }> (or worse, fallback={null} that blanks the UI), defeats this skill's optimization — which is to extract real static chrome above each granular boundary and use content-shaped fallbacks per region. A coarse "the page is loading" stand-in bypasses the work entirely.

gotchas (shared)

• Dev doesn't prefetch the way production does, and routes compile on first hit — so after a navigation or reload, the DOM keeps updating for noticeably longer than the eventual production experience. Wait patiently for the DOM to stabilize before capturing the React tree or taking a screenshot — e.g., poll document.documentElement.innerHTML.length until it's unchanged across two consecutive reads. A fixed short delay risks sampling mid-render.
• Don't try to verify nav prefetch by inspecting dev network traffic — dev doesn't fire prefetch requests at all, so the network tab, manual router.prefetch() calls, and ` will all look broken regardless of whether your code is correct. The cookie-locked SPA-nav recipe in [instant-nav-loop.md](./instant-nav-loop.md) under verify` is already the canonical recipe for this — it simulates what production would prerender into the prefetched RSC without requiring prefetch to actually fire. Use it; don't invent a network-tab alternative.
• The diagnose pipeline can be flaky — DevTools attachment timing, DOM-settle races, and dev compilation effects can each produce inconsistent captures from one run to the next. When a result feels off (a candidate appears that you don't expect, or one you expect doesn't), re-run the diagnose 2–3 times and cross-check; boundaries that appear consistently are real, one-off appearances are noise.

reference (shared primitives)

agent-browser react suspense          add --only-dynamic to filter
--json                                server-side to actually-
                                      suspended boundaries. Each
                                      entry has jsx_source +
                                      suspended_by[] with raw blocker
                                      names (usePathname, cookies,
                                      fetch, cache, ...); classify by
                                      name for per-loop rules

POST /__nextjs_original-stack-frames  body { frames: StackFrame[],
                                      isServer, isEdgeServer,
                                      isAppDirectory }; returns one
                                      result per frame with
                                      file:line:column

mcp get_logs                          dev MCP tool from
                                      next-dev-loop; surfaces
                                      blocking-route /
                                      NEXT_STATIC_GEN_BAILOUT 500s

cacheLife('')                default | seconds | minutes
                                      | hours | days | weeks | max

Per-loop primitives in [instant-nav-loop.md](./instant-nav-loop.md).

teardown (shared)

Delete the cookie by name — overwrite with an expired stamp:

agent-browser cookies set next-instant-navigation-testing x \
  --url  --expires 1

Never agent-browser cookies clear (no args) — wipes auth.

---

Sibling of next-dev-loop — initiate that first.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: fellipeutaka
• Source: fellipeutaka/leon
• License: MIT

Install and usage instructions live in the source repository linked above.