Refine Live

Refine Live

Two modes

Persistent (recommended — run and forget)
Run npx transitions-refine live from your project. The CLI starts the relay and wires REFINE_AGENT_CMD so the relay spawns your agent CLI per Refine click. No chat loop; idle = zero credit burn. Works hours later as long as the relay process keeps running. Stop with Ctrl-C (or npx transitions-refine stop).

In-chat loop (fallback — this skill)
Run /refine live in Cursor/Claude/Codex when the relay is up but has no REFINE_AGENT_CMD. You become the poller via GET /jobs/next. The Agent tab stays available only while you keep polling — each idle poll cycle consumes chat turns/credits. Say "stop refine" to exit.

Use the in-chat loop only when you cannot wire a persistent agent CLI.

Turn yourself into the LLM behind the Timeline Inspector's Refine button (in-chat fallback mode). While this loop runs, the panel's LLM tab is "available": each click sends one transition here, you reason about it, and your suggestions appear in the panel.

You are the poller. Nothing is installed — you just talk to a small local relay (default http://localhost:7331) that the npx injector already started.

How it works

Browser (Refine, LLM tab) ──POST /jobs──► relay ──GET /jobs/next──► YOU
                          ◄──GET /jobs/:id── relay ◄──POST /jobs/:id/result── YOU

The loop — stay live, but don't burn credits forever

Keep polling so the panel's LLM tab stays "available", but this loop costs chat turns/credits even while idle, so it is not truly run-and-forget — it has three exits, in priority order:

1. Relay stop signal (authoritative). GET /jobs/next may return 200 with {"stop": true}. The relay sends this when the user clicks Stop in the panel, or automatically after ~10 min with no jobs. Always honor it: stop looping immediately, tell the user the LLM tab will go unavailable and how to resume (/refine live), and end your turn. Never re-poll after a stop signal.
2. The user says so — "stop refine", "exit live", etc.
3. Your own idle backoff (safety net). A long stretch of 204s is normal — it just means no one has clicked Refine yet — but to avoid spending credits on a forgotten loop, back off as idle grows instead of hammering immediately: re-poll right away for the first few empty cycles, then pause ~5s between polls, and after ~10 min of unbroken idle stop on your own (same as the relay's auto-stop) and tell the user how to resume. Any real job resets the backoff.

The relay reports the agent as "available" for ~120s after your last poll, so short pauses keep you live. A successful job always resets idle, so an active session never backs off.

0. Announce yourself once, before the first poll. The relay keeps a sticky Stop latch: after a panel Stop (or the idle auto-stop) it answers every GET /jobs/next with {"stop": true} until a new agent explicitly resumes — so a stopped session can't silently come back. Clear the latch a single time at startup, then begin polling:

   curl -s -X POST http://localhost:7331/poller/start
   

   Do not call this again mid-loop (it would defeat a user's Stop). Only on a fresh /refine live.

1. Claim the next job (long-poll). This call blocks up to ~25s, then returns.

   curl -s http://localhost:7331/jobs/next
   

   • HTTP 204 / empty body → no work yet. Poll again, applying the idle backoff above (immediate at first, then ~5s pauses, then stop after ~10 min).

   • HTTP 200 with {"stop": true} → the loop must end. Stop polling, tell the user the LLM tab is now unavailable and that /refine live resumes it, and end your turn. Do not treat it as a job.

   • HTTP 200 with a job JSON → work to do. Shape:

     {
       "id": "uuid",
       "request": {
         "label": "Resize + Color",
         "selector": ".box-resize",
         "mode": "llm",
         "refineType": "small",
         "timings": [
           { "property": "width", "durationMs": 400, "delayMs": 0, "easing": "ease-out" },
           { "property": "background", "durationMs": 400, "delayMs": 0, "easing": "ease-out" }
         ]
       }
     }
     

   • If request.kind === "scan" this is not a suggestion job — the panel is asking you to group the page's transitions by reading the source. Jump to ## Scan jobs and return groups instead of suggestions.

   • If request.kind === "apply" this is not a suggestion job — the user pressed Accept to write changes to their code. Jump to ## Apply jobs and edit the source instead of posting suggestions. Everything below (refineType, steps 3–4) is for the normal Refine flow.

   • refineType chooses what kinds of suggestions to make (it mirrors the panel's two tabs). The tabs scan independently, so answer only the one you were asked for:

     • "small" (or missing) → Small refinements: nudge the existing declarations toward the motion tokens only (step 3a). Do not propose a recipe swap here — that's the Replace tab's separate job.
     • "replace" → Replace transition: suggest a whole-transition recipe swap only (step 3b). Do not propose motion-token tweaks — skip step 3a entirely.

2. (Optional) post progress so the panel shows what you're doing:

   curl -s -X POST http://localhost:7331/jobs/<id>/status \
     -H 'Content-Type: application/json' \
     -d '{"message":"Matching to transitions.dev motion tokens…"}'
   

3. Answer in ONE shot — speed matters. Each click should feel instant, so resolve the job from the data below plus what's already in this skill. Do not spawn subagents or run a broad codebase search, and do not open the transitions-dev SKILL.md — its tokens and decision rules are inlined here.

   • refineType === "small" → step 3a only, with zero file reads.
   • refineType === "replace" → step 3b only; open at most the one recipe reference file you choose.

   First, infer each declaration's usage from label + selector (modal close, dropdown open, tooltip, badge, resize, color/theme change…). Match on intent, not the nearest number.

   3a. Motion-token tweaks (refineType === "small" only — no file reads). Pick the token that fits the usage and propose a change only where the current value actually differs.

   • Durations: 40ms Stagger (per-item offset) · 80ms Micro (tooltip delay, shake segment) · 150ms Quick (modal/dropdown close, text swap, tooltip appear) · 250ms Fast (icon swap, dropdown/modal open, tabs slide, page slide) · 350ms Medium (panel/toast close) · 400ms Slow (panel open, skeleton reveal, input clear) · 500ms Very slow (emphasis, badge appear, text reveal, success check).
   • Default easing — "Smooth ease out": cubic-bezier(0.22, 1, 0.36, 1) (modal/dropdown/panel open+close, page slide, resize, position change).
   • Other on-grid easings — LEAVE UNCHANGED: ease-out (tooltip), ease-in-out (icon/text swap, text reveal, skeleton reveal), linear (shimmer, pulse, spinner), cubic-bezier(0.34, 1.36, 0.64, 1) (badge pop), cubic-bezier(0.34, 3.85, 0.64, 1) (avatar return).
   • Nudge toward Smooth ease out: generic ease, ease-in, or any hand-rolled cubic-bezier()/linear() that isn't a token above.
   • Scale / blur / distance (the non-resting transform pre-scale or translate distance, and the filter pre-blur a lane carries): scale 0.96 Large (modal) · 0.97 Medium (dropdown open) · 0.98 Small (tooltip) · 0.99 Tiny (dropdown close); blur 2px (panel/icon/text swap, skeleton, number) · 3px (page slide, text reveal) · 8px (success check); distance 4px (text swap) · 6px (shake small) · 8px (badge, page slide, shake large) · 12px (text reveal) · 30px (check badge). Propose kind:"scale"/"blur"/ "distance" (patch key scale/blur/translate, pass translateVarName through) only where a lane's value differs from its usage token.
   • Polish add-on (optional): if the project has the transitions-polish skill installed (.agents/skills/transitions-polish/_refine-rules.md, .claude/…, or skills/…), read that ONE file a single time and also apply its rules — open/close asymmetry (closes faster/quieter than opens), hover in fast / out soft, stagger totals under ~300ms, 80ms intent delays, and never delay a close. This is the only file read allowed on a Small job, and only when that skill is present.

   3b. Whole-transition recipe swap (refineType === "replace" only — no file reads). Match the inferred usage to ONE recipe below (this list is the decision rules — no SKILL.md or reference-file read needed). Emit ONE kind: "replace" suggestion whose patch carries the motion-token duration/easing for the recipe's phase (open vs close) on the property that already transitions (or "all"), with a reference field naming the file and the recipe in title + reason. The patch only drives the live preview — exact keyframes/structure come from the user pasting that reference file, so you never need to open it. If no recipe genuinely fits the usage, return an empty suggestions array with a short summary.

   • Card resize — a container changes width/height on a layout change (01-card-resize.md)
   • Number pop-in — a number/digit updates (02-number-pop-in.md)
   • Notification badge — a small dot/badge appears on a trigger (03-notification-badge.md)
   • Text states swap — text content changes in place (04-text-states-swap.md)
   • Menu dropdown — an anchored surface grows from its trigger (05-menu-dropdown.md)
   • Modal open/close — a centered dialog scales up, softer scale-down on close (06-modal.md)
   • Panel reveal — a surface slides into a region with a cross-blur (07-panel-reveal.md)
   • Page side-by-side — slide between list↔detail or step 1↔step 2 (08-page-side-by-side.md)
   • Icon swap — two icons cross-fade in the same slot (09-icon-swap.md)
   • Success check — a checkmark celebration: fade + rotate + bob + stroke-draw (10-success-check.md)
   • Avatar group hover — hover lifts an item in a horizontal stack (11-avatar-group-hover.md)
   • Error state shake — invalid-input shake (12-error-state-shake.md)
   • Input clear with dissolve — clearing a text field (13-input-clear-dissolve.md)
   • Skeleton loader and reveal — placeholder pulses then swaps to real content (14-skeleton-reveal.md)
   • Shimmer text — in-progress / "thinking" text shimmer (15-shimmer-text.md)
   • Tabs sliding — a moving highlight across segmented options (16-tabs-sliding.md)
   • Tooltip open/close — delayed fade+scale in, instant out (17-tooltip.md)
   • Texts reveal — staggered blurred rise of stacked text lines (18-texts-reveal.md)
   • Card hover tilt — 3D tilt toward the pointer (19-card-tilt.md)
   • Plus to menu morph — a circular trigger becomes the surface it opens (20-plus-menu-morph.md)
   • Accordion expand — a collapsible body grows/shrinks in height (21-accordion.md)

   Tie-break: prefer the lower-overhead recipe (card resize over panel reveal, dropdown over modal). Only propose a swap when the current declarations are clearly a hand-rolled version of a recipe or are missing the structure the usage calls for; if the transition already is the right recipe, return empty.

4. Post the result (this completes the job and renders cards in the panel):

   curl -s -X POST http://localhost:7331/jobs/<id>/result \
     -H 'Content-Type: application/json' \
     -d '{
       "summary": "Tightened the resize and softened the color fade.",
       "suggestions": [
         {
           "id": "width-duration",
           "kind": "duration",
           "property": "width",
           "title": "Duration → Snappy (250ms)",
           "from": "400ms",
           "to": "250ms",
           "patch": { "property": "width", "durationMs": 250 },
           "reason": "A size change reads as direct manipulation — snappy is more responsive than 400ms."
         }
       ]
     }'
   

   The example above is a small job (token tweaks only). A replace job instead returns a single kind: "replace" card as its only suggestion:

   {
     "id": "replace-card-resize",
     "kind": "replace",
     "property": "width",
     "title": "Replace with Card resize",
     "from": "hand-rolled width tween",
     "to": "transitions.dev · Card resize",
     "patch": { "property": "width", "durationMs": 250, "easing": "cubic-bezier(0.22, 1, 0.36, 1)" },
     "reference": "transitions-dev/01-card-resize.md",
     "reason": "This is a width tween on layout change — the Card resize recipe handles it properly. Apply nudges the live timing; paste 01-card-resize.md (run `transitions apply card-resize`) for the full recipe."
   }
   

   If nothing should change, post "suggestions": [] with a short summary. If something goes wrong, report it instead:

   curl -s -X POST http://localhost:7331/jobs/<id>/error \
     -H 'Content-Type: application/json' -d '{"message":"…"}'
   

5. Go back to step 1. Keep looping, but honor the three exits from the loop section: a {"stop": true} from the relay, the user telling you to stop, or your own idle backoff/auto-stop after ~10 min quiet. A real job resets idle. Whenever you do stop, tell them the LLM tab will go unavailable and how to restart (/refine live).

Scan jobs (group from source)

When a claimed job has request.kind === "scan", the panel wants you to turn a flat list of DOM-detected transitions into components with phases. A naive DOM scan only sees each element's current computed transition — it can't tell open from close, and lists related elements (panel, backdrop, staggered items) separately. You fix that by reading the source. The request looks like:

{
  "id": "uuid",
  "request": {
    "kind": "scan",
    "url": "http://localhost:5173/",
    "raw": [
      { "label": "div.dropdown-panel", "selector": ".dropdown-panel",
        "properties": ["opacity","transform"],
        "timings": [{ "property": "opacity", "durationMs": 200, "delayMs": 0, "easing": "ease-out" }],
        "cssRules": [
          ".dropdown .dropdown-panel { opacity: 0; transition: opacity 200ms ease-out 0ms, transform 200ms cubic-bezier(0.22, 1, 0.36, 1) 0ms; }",
          ".dropdown.is-open .dropdown-panel { opacity: 1; transform: translateY(0); }",
          ".dropdown.is-closing .dropdown-panel { transition: opacity 150ms ease-in 0ms; opacity: 0; }"
        ] }
    ]
  }
}

Be fast. The raw.timings are already accurate for each element's current on-screen state — treat them as ground truth and reuse them verbatim. Most raw entries also carry cssRules: the CSS rules harvested live from the page (CSSOM) that drive that element across all states (base + open + close), with var() already resolved to concrete values.

Fast path — prefer cssRules over the filesystem. When an entry has cssRules, they are authoritative and contain everything you need: the opposite phase's timings live on a state-variant selector inside them (e.g. .dd.is-closing .dd-panel, .modal[data-closing] .dialog), and the toggled state is visible in those selectors. Derive grouping, phases, toggled state, and opposite-phase timings directly from cssRules + timings — do not glob/grep/read files for any element whose cssRules is non-empty; it only wastes time. Only fall back to reading source for entries with an empty/missing cssRules (CORS-locked sheets, styled-components, Tailwind, etc.), and even then read the minimum.

Do this:

1. Identify each animated component the raw entries belong to (dropdown, modal, tooltip, accordion, drawer, toast…). The selectors/labels usually make this obvious — only read source (plain CSS / CSS Modules, styled-components/emotion, Tailwind, inline styles, Motion/Framer variants) when the grouping is genuinely unclear.

2. Split each component into phases — usually open and close (a hover-only component can be a single phase). The phase matching the current DOM reuses the provided timings; the opposite phase often lives on a different selector (.is-open vs .is-closing) with different timings — take it from the entry's cssRules (or, only if it has none, read source). Report both even though only one is in the DOM right now.

3. List each phase's members — the elements that animate in that phase. Give each a stable id, a human label, a live-resolvable CSS selector, an optional toState hint (the class/attribute that drives the phase, e.g. .is-open), and its propertyTimings. For the current-state phase, copy the provided raw.timings verbatim; for the opposite phase, quote the real timings from the entry's cssRules (already var()-resolved) — or from source if it has none — never invent.

4. Post the groups (this completes the job):

   curl -s -X POST http://localhost:7331/jobs/<id>/result \
     -H 'Content-Type: application/json' \
     -d '{
       "summary": "Grouped Dropdown into Open/Close.",
       "groups": [
         { "id": "dropdown", "label": "Dropdown", "component": "src/Dropdown.tsx",
           "phases": [
             { "id": "dropdown:open", "phase": "open", "label": "Open", "members": [
               { "id": "panel", "label": "Panel", "selector": ".dropdown-panel", "toState": ".is-open",
                 "propertyTimings": [
                   { "property": "opacity", "durationMs": 200, "delayMs": 0, "easing": "ease-out" },
                   { "property": "transform", "durationMs": 200, "delayMs": 0, "easing": "cubic-bezier(0.22, 1, 0.36, 1)" }
                 ] }
             ] },
             { "id": "dropdown:close", "phase": "close", "label": "Close", "members": [
               { "id": "panel", "label": "Panel", "selector": ".dropdown-panel", "toState": ".is-closing",
                 "propertyTimings": [
                   { "property": "opacity", "durationMs": 150, "delayMs": 0, "easing": "ease-in" }
                 ] }
             ] }
           ] }
       ]
     }'
   

   If you can't confidently group anything, post {"groups":[],"summary":"…"} — the panel keeps its flat DOM scan. Reserve /jobs/<id>/error for unexpected failures.

Then go back to step 1 of the loop.

Apply jobs (write to source)

When a claimed job has request.kind === "apply", the user accepted their current timeline values and wants them written to the codebase. The request looks like:

{
  "id": "uuid",
  "request": {
    "kind": "apply",
    "label": "Dropdown · Close",
    "selector": ".dropdown-panel",
    "component": "src/Dropdown.tsx",
    "group": "Dropdown",
    "phase": "close",
    "changes": [
      { "property": "opacity", "member": "Panel", "selector": ".dropdown-panel",
        "from": { "durationMs": 300, "delayMs": 0, "easing": "ease" },
        "to": { "durationMs": 150, "delayMs": 0, "easing": "cubic-bezier(0.4, 0, 1, 1)" } }
    ]
  }
}

Do this:

1. Locate the real declaration in the source. The selector is a DOM-path hint, not necessarily the source selector. Use the component hint and search by the label/class names; handle whatever the project uses: plain CSS / CSS Modules, styled-components or emotion template literals, Tailwind utilities (duration-300, arbitrary [transition-duration:300ms], or the tailwind.config theme), inline style={{ transition: … }} objects, and Motion/Framer variants. Match by the from values to disambiguate.

   • If phase is set (e.g. "open"/"close"), edit only that state's rule (the .is-open rule for open, the .is-closing/base rule for close) — not the other phase. Each change's member + selector says which element.

2. Edit each change's property to its to values (durationMs ms, easing, delayMs ms) on the right member + phase. Keep the file's existing unit/format (0.25s vs 250ms) and touch only that property's timing. If a CSS variable / design token backs the value, update it at the single most sensible place.

3. Minimal edit — no reformatting or unrelated changes.

4. Post the outcome (this completes the job):

   curl -s -X POST http://localhost:7331/jobs/<id>/result \
     -H 'Content-Type: application/json' \
     -d '{"applied":true,"summary":"Set .t-modal transition to 150ms ease-in","files":["src/Modal.css:42"]}'
   

   If you cannot confidently find the declaration, post {"applied":false,"summary":"<what you searched and why not found>"} (still a result, not an error). Reserve /jobs/<id>/error for unexpected failures.

Then go back to step 1 of the loop.

Suggestion shape (must match the panel)

Each suggestion object:

The panel applies patch live in the browser via the property override. Values are not written to source files — the user copies the ones they keep.

Notes

• Relay port: http://localhost:7331 unless REFINE_RELAY_PORT was changed.
• Only LLM-mode jobs reach you; Deterministic-mode jobs are answered by the relay itself (nearest-token snapping) and never appear here. Whole-transition replace suggestions are therefore LLM-only — the deterministic path can't infer usage well enough to pick a recipe, so a Deterministic + "Replace transition" job just returns an empty result pointing the user back to the Agent tab.
• A replace card's Apply only changes the live timing in the patch. The recipe's structural parts (keyframes, extra properties, JS hooks) aren't applied in the browser — that's why the card points the user at the reference file to paste.
• The relay errors a waiting job after ~120s, so answer promptly once you claim one. The long-poll itself returning 204 is normal — just poll again.