@ammduncan/easel 0.3.1 → 0.3.2

package/CHANGELOG.md

@@ -2,6 +2,15 @@
 
 All notable changes to easel. This project adheres to [Semantic Versioning](https://semver.org/).
 
+## 0.3.2 — 2026-05-26
+
+### Fixed
+- **App-fidelity (`kind:"mockup"`/`"app"`) text painted with `light-dark()` now tracks the easel light/dark toggle instead of the OS color scheme.** Authors paint mockup ink with `color: light-dark(#dark-ink, #light-ink)`, which resolves off the document's *computed* `color-scheme` — not the `data-theme` attribute. The normal wrapper binds `color-scheme` to `data-theme` via `PRESET_TOKENS_CSS`, so `light-dark()` follows the toggle there. But the app-fidelity branch deliberately omits the preset tokens (the agent owns every pixel) and nothing else bound `color-scheme`, so it stayed at the author's `color-scheme: light dark` and `light-dark()` followed the **OS** preference. The symptom was intermittent and maddening: text was perfectly readable when the viewer's theme happened to match the OS, then washed out (white ink on a light card, or dark ink on a dark card) the moment they disagreed. Added a `:root[data-theme]{color-scheme}` binding to the shared `STRUCTURAL_PRIMITIVES_CSS` so `light-dark()` tracks the easel toggle in *every* wrapper branch. A new visual-regression fixture (`mockup-lightdark-ink`) audited across the theme × OS-scheme matrix reproduces the washout with the binding removed (contrast 1.0) and passes with it in place.
+- **Session-id resolution no longer drifts for non-Claude-Code MCP clients (opencode, Cursor, Windsurf, …).** The resolver's tier-4 fallback scanned `~/.claude/projects/<cwd>/` for the most-recently-modified transcript. That's a Claude-Code-specific signal, but it fired for *any* client — so a non-CC client running in a cwd that also holds Claude Code transcripts latched onto whichever transcript was touched last, and the resolved session id changed on every tool call (observed in opencode: `open()`, `push()`, and `label()` each landing on a different session). The scan is now gated behind a positive Claude Code signal (`CLAUDECODE` / `CLAUDE_CODE_ENTRYPOINT`); other clients fall straight through to the stable per-process synthetic id (tier 5), so `open`/`push`/`label` all resolve to one session for the life of the chat. Covered by a new unit suite (`tests/unit/session-id.test.mjs`, run via `npm test`).
+
+### Docs
+- **`push` tool description and the using-easel skill now lead with a fidelity bar: ship high-fidelity, production-grade output by default.** Aimed at getting quality output from non-Claude models that don't infer it — every push should read like a screenshot of shipped software (real content, complete regions, exact values when recreating UI, real iconography, deliberate hierarchy), not a wireframe or grey-box. Low-fidelity is opt-in: only when the user explicitly says rough/wireframe/sketch is fine.
+
 ## 0.3.1 — 2026-05-26
 
 ### Fixed

package/dist/client/viewer.js

@@ -120,6 +120,15 @@
      block, so stripping these in fidelity mode left the skill's own guidance
      ("wrap a mockup in .window") producing unstyled output. */
   const STRUCTURAL_PRIMITIVES_CSS = `
+/* Bind the CSS color-scheme to the host theme so any author CSS that uses
+   light-dark() (text ink, surfaces, borders) tracks the easel light/dark
+   TOGGLE rather than the OS preference. The default wrapper already gets this
+   via PRESET_TOKENS_CSS, but app-fidelity (kind:"mockup") pushes omit the
+   preset tokens — without this rule their light-dark() ink follows the OS
+   scheme and washes out whenever the OS disagrees with the easel toggle. */
+:root[data-theme="light"] { color-scheme: light; }
+:root[data-theme="dark"]  { color-scheme: dark; }
+
 /* Skeuomorphic macOS-style window chrome for UI mockups. Usage:
      <div class="window" data-title="App name"> …mockup content… </div>
    Draws a 40px title bar with the three traffic-light dots (red/yellow/green)

package/dist/mcp.js

@@ -102,6 +102,15 @@ export async function main() {
             {
                 name: TOOL_PUSH,
                 description: "Push an HTML card to this session's live browser tab. Renders in a sandboxed iframe over a host-controlled canvas that can be LIGHT or DARK depending on the user's OS theme. Treat each card as a presentation slide — generous whitespace, presentation-scale type, tangible visuals. Your HTML MUST adapt to both light and dark modes.\n\n" +
+                    "═══ FIDELITY BAR — SHIP HIGH-FIDELITY BY DEFAULT ═══\n" +
+                    "Default to polished, production-grade output that looks like a real screenshot of shipped software or a finished design — NOT a rough sketch, wireframe, or grey-box placeholder. This is the default for EVERY push; you do not need to be asked for quality. Only drop to low-fidelity (wireframe, ASCII-ish boxes, lorem-ipsum, unstyled) when the user EXPLICITLY says rough/lo-fi/wireframe/sketch/quick-and-dirty is fine, or asks for a thumbnail/napkin idea. When in doubt, go high-fidelity.\n" +
+                    "What high-fidelity means concretely:\n" +
+                    "• REAL content, not placeholders. Plausible names, realistic numbers/dates/currency, actual copy — never 'Lorem ipsum', 'Label', 'Item 1 / Item 2', 'Title goes here', or '...'.\n" +
+                    "• COMPLETE, not stubbed. Fill every region you draw — no empty cells, half-built tables, or 'etc.' rows. If a screen has 8 nav items, draw 8.\n" +
+                    "• EXACT values when recreating real UI — pull true colors, spacing, radii, type, and layout from the component/theme/Figma/DevTools (see the recreation rules below). A close-but-wrong mock misleads more than none.\n" +
+                    "• Visual craft: deliberate hierarchy, aligned grids, consistent spacing scale, real iconography (inline SVG, not emoji-as-icon), proper empty/hover/active states where they matter. Avoid the generic-AI look (one purple gradient, evenly-sized boxes, centered everything).\n" +
+                    "• Tangible over abstract (see VISUALS): a mock should read as the actual thing, not labeled rectangles.\n" +
+                    "If you genuinely can't reach the bar (missing real values, ambiguous source), say so in ONE line in chat and push your best honest attempt — don't pass a rough draft off as final, and don't silently ship a grey-box.\n\n" +
                     "═══ ADAPTIVE COLOR (gets wrong most often) ═══\n" +
                     "• Do NOT set `background` on `body` or your root wrapper. The host paints the canvas — setting bg fights it and creates a wrong-shade block in the opposite mode.\n" +
                     "• Use `light-dark()` for ALL text colors, card backgrounds, borders, and decorative shades. Add `:root { color-scheme: light dark; }` so the function resolves. Hardcoded `color: #475569` goes invisible in dark mode; hardcoded `border: 1px solid #e5e5e5` becomes a hard white line.\n" +

package/dist/session-id.js

@@ -10,7 +10,11 @@ import { HOOK_DIR } from "./paths.js";
  *   3. Hook file at ~/.easel/hook/cc-session-<ppid>.txt (Claude Code's
  *      SessionStart hook writes this; pitstop-style PPID bridging)
  *   4. Most-recently-modified transcript under ~/.claude/projects/<cwd>/
- *      (Claude Code transcript scan)
+ *      (Claude Code transcript scan) — ONLY when a positive Claude Code env
+ *      signal is present. Other MCP clients (opencode, Cursor, Windsurf, …)
+ *      can share a cwd that already holds CC transcripts; without this guard
+ *      they'd latch onto whichever unrelated transcript was touched last and
+ *      the resolved session would drift on every tool call.
  *   5. Synthetic id derived from this MCP child's PPID — gives every other
  *      MCP client (Cursor, Windsurf, Claude Desktop, etc.) a stable session
  *      per chat without requiring any hook. The MCP child IS the session.
@@ -38,25 +42,33 @@ export function resolveClaudeSessionId(opts = {}) {
     catch {
         /* fall through */
     }
-    try {
-        const encoded = cwd.replace(/\//g, "-");
-        const dir = join(home, ".claude", "projects", encoded);
-        let bestId;
-        let bestMtime = 0;
-        for (const f of readdirSync(dir)) {
-            if (!f.endsWith(".jsonl"))
-                continue;
-            const m = statSync(join(dir, f)).mtimeMs;
-            if (m > bestMtime) {
-                bestMtime = m;
-                bestId = f.slice(0, -".jsonl".length);
+    // Tier 4 (transcript scan) is Claude-Code-specific: only trust it when we
+    // have positive evidence we're actually running inside Claude Code. For any
+    // other MCP client the scan would pick an unrelated, actively-changing CC
+    // transcript in the same cwd and the session id would drift per call — so
+    // skip straight to the stable per-process synthetic id (tier 5).
+    const isClaudeCode = Boolean(env.CLAUDECODE || env.CLAUDE_CODE_ENTRYPOINT);
+    if (isClaudeCode) {
+        try {
+            const encoded = cwd.replace(/\//g, "-");
+            const dir = join(home, ".claude", "projects", encoded);
+            let bestId;
+            let bestMtime = 0;
+            for (const f of readdirSync(dir)) {
+                if (!f.endsWith(".jsonl"))
+                    continue;
+                const m = statSync(join(dir, f)).mtimeMs;
+                if (m > bestMtime) {
+                    bestMtime = m;
+                    bestId = f.slice(0, -".jsonl".length);
+                }
             }
+            if (bestId)
+                return bestId;
+        }
+        catch {
+            /* fall through */
         }
-        if (bestId)
-            return bestId;
-    }
-    catch {
-        /* fall through */
     }
     return syntheticSessionIdFromPpid(ppid);
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ammduncan/easel",
-  "version": "0.3.1",
+  "version": "0.3.2",
   "description": "A live browser tab for every Claude Code (and MCP) session. The push MCP tool appends HTML cards to a scrolling feed you keep open in split-screen.",
   "type": "module",
   "license": "MIT",
@@ -43,6 +43,7 @@
     "build": "tsc && node scripts/copy-client.mjs",
     "start": "node dist/cli.js",
     "dev": "tsc --watch",
+    "test": "npm run build && node --test tests/unit/*.test.mjs",
     "prepublishOnly": "npm run build"
   },
   "publishConfig": {

package/skills/using-easel/SKILL.md

@@ -49,6 +49,20 @@ Don't poll. Just react to the hint when it appears.
 
 Pushed cards are **presentations**, not UI dashboards. Read each rule and apply it; the wrapper gives you good defaults but they only carry so far.
 
+### 0. Fidelity bar — ship high-fidelity by default
+
+Every push should look like a **screenshot of shipped software or a finished design** — polished and production-grade — not a rough sketch, wireframe, or grey-box placeholder. Quality is the default; you don't need to be asked for it. Only drop to low-fidelity when the user **explicitly** says rough / lo-fi / wireframe / sketch / quick-and-dirty is fine, or asks for a napkin-level thumbnail. When unsure, go high-fidelity.
+
+Concretely, high-fidelity means:
+
+- **Real content, never placeholders.** Plausible names, realistic numbers/dates/currency, actual copy — no "Lorem ipsum", "Label", "Item 1 / Item 2", "Title goes here", or "…".
+- **Complete, not stubbed.** Fill every region you draw — no empty cells, half-built tables, or "etc." rows. A nav with 8 items shows 8.
+- **Exact values when recreating real UI.** Pull true colors, spacing, radii, type, and layout from the component / theme / Figma / DevTools (see [Use the actual values](#use-the-actual-values-not-approximations)). A close-but-wrong mock misleads more than none.
+- **Visual craft.** Deliberate hierarchy, aligned grids, a consistent spacing scale, real iconography (inline SVG — not emoji standing in for icons), and the states that matter (empty / hover / active). Avoid the generic-AI look: one purple gradient, evenly-sized boxes, everything centered.
+- **Tangible over abstract** (see [Visualizations](#5-visualizations--tangible-over-abstract)) — the mock should read as the actual thing, not labeled rectangles.
+
+If you genuinely can't clear the bar (missing real values, ambiguous source), say so in one line in chat and push your best honest attempt — don't pass a rough draft off as final, and don't silently ship a grey-box.
+
 ### 1. Typography
 
 - **Page lede**: 40–52 px, weight 500, letter-spacing ≈ -0.025em