gentle-shell-bridge 0.3.0 → 0.3.2

package/test/tool-design-cost.test.ts

@@ -1,473 +0,0 @@
-/**
- * S13/T23 — the WORK a design switch does, counted instead of timed.
- *
- * Wall-clock numbers measured on real pi-tui and pi's `ToolExecutionComponent` (T22) say the freeze
- * came from two calls the host's own card-style switch never makes:
- *
- * - `tui.invalidate()` clears every mounted component's frame cache, so the next frame re-parses
- *   every Markdown message of the transcript (0.06 ms at 50 children, 1.1 ms at 600 for the walk
- *   alone, and +709 ms for the frame that re-parses 300 messages with 45-line code blocks);
- * - `requestRender(force = true)` calls `resetRenderState()`, throwing away the differential state so
- *   the whole buffer is cleared and rewritten (6.9 MB at 600 children) instead of only the rows that
- *   changed.
- *
- * Neither is needed to change a card that is already drawn: the bridge's wrapper renders the HOST's
- * component and applies the live selection on top of whatever rows it answered with — including the
- * rows the host's own `ToolCardRows` cache hands back (`cards-bash-v2` `extensions/quiet-tools.ts`
- * 612-624 keys its frame on width + style + running, and the design is NOT in that key). This file
- * proves that against a host component that caches exactly like that one, and bounds the switch:
- *
- * - the tree is never invalidated, and no prose component re-parses;
- * - exactly one frame is requested, and it is differential;
- * - at most one persisted-selection read per card per frame;
- * - none of those counts grows with the number of Markdown components in the transcript.
- *
- * The bound is call counts, never milliseconds, so it cannot flake on a busy machine.
- */
-
-import assert from "node:assert/strict";
-import { mkdtempSync, rmSync } from "node:fs";
-import { tmpdir } from "node:os";
-import { join } from "node:path";
-import test from "node:test";
-
-import { resetNoticesForTests } from "../src/notices.ts";
-import { resetPluginLinesForTests } from "../src/plugin-lines.ts";
-import {
-  getRepaintDiagnostics,
-  requestSurfaceRepaint,
-  resetRepaintForTests,
-  setRepaintHost,
-} from "../src/repaint.ts";
-import { buildRows, resetRowBuildStateForTests } from "../src/rows.ts";
-import { resetStoreStateForTests } from "../src/store.ts";
-import {
-  TOOL_DESIGN_NATIVE,
-  getToolDesignSelectionReads,
-  registerToolDesign,
-  resetToolDesignSelectionReads,
-  resetToolDesignsForTests,
-  setToolDesignSelection,
-} from "../src/tool-designs.ts";
-import {
-  installToolDesignRows,
-  resetToolDesignRowStateForTests,
-} from "../src/tool-design-rows.ts";
-import { resetToolDesignWiringForTests, toolDesignResolver } from "../src/tool-design-wiring.ts";
-import type { CustomizeRow, ToolDesignContext, ToolDesignRegistration } from "../src/types.ts";
-
-let configHome: string;
-let previousConfigHome: string | undefined;
-
-test.before(() => {
-  configHome = mkdtempSync(join(tmpdir(), "gsb-tool-design-cost-"));
-  previousConfigHome = process.env.GENTLE_PI_CONFIG_HOME;
-  process.env.GENTLE_PI_CONFIG_HOME = configHome;
-});
-
-test.after(() => {
-  if (previousConfigHome === undefined) delete process.env.GENTLE_PI_CONFIG_HOME;
-  else process.env.GENTLE_PI_CONFIG_HOME = previousConfigHome;
-  rmSync(configHome, { recursive: true, force: true });
-});
-
-test.beforeEach(() => {
-  rmSync(join(configHome, "gentle-shell-bridge"), { recursive: true, force: true });
-  resetToolDesignsForTests();
-  resetToolDesignRowStateForTests();
-  resetToolDesignWiringForTests();
-  resetRepaintForTests();
-  resetStoreStateForTests();
-  resetNoticesForTests();
-  resetPluginLinesForTests();
-  resetRowBuildStateForTests();
-});
-
-const WIDTH = 80;
-const THEME = { name: "fake-theme" };
-const BASH_HEADING = "bash heading";
-
-/** The host's own rows for one bash card, as `quiet-tools.ts` draws them. */
-const HOST_CARD = [
-  "╭─ ✿ bash card ───────────── ctrl+o ╮",
-  "│ \u001b[36m$ ls -la\u001b[0m                │",
-  "╰───────────────────────────────────╯",
-];
-
-const DESIGNED_CARD = [
-  HOST_CARD[0],
-  HOST_CARD[1].replace("$ ", "bash "),
-  HOST_CARD[2],
-];
-
-/** How many times the design applied since the last reset (one per card per frame while selected). */
-let applied = 0;
-
-const bashHeadingDesign: ToolDesignRegistration = {
-  tool: "bash",
-  id: "loon.bash-heading",
-  label: BASH_HEADING,
-  apply(lines: readonly string[], ctx: ToolDesignContext): string[] {
-    applied += 1;
-    assert.equal(ctx.tool, "bash");
-    return lines.map((line) => line.replace("$ ", "bash "));
-  },
-};
-
-// ---------------------------------------------------------------------------------------------
-// The instrumented host: a frame counter on every component, exactly like pi's TUI
-// ---------------------------------------------------------------------------------------------
-
-interface Counters {
-  /** `tui.invalidate()`: the call that cleared every cache in the tree. */
-  treeInvalidations: number;
-  /** What each `requestRender()` was called with: `plain` or `forced`. */
-  frameRequests: Array<"plain" | "forced">;
-  /** A prose row re-parsed its source because its cache was dropped. */
-  proseCacheMisses: number;
-  /** A prose row had its cache dropped. */
-  proseInvalidations: number;
-  /** The host card's renderer ran (it may answer from its own cache). */
-  hostRenders: number;
-  /** The host card answered from its own frame cache (the rows the design must post-process). */
-  hostCacheHits: number;
-  /** The host card rebuilt its rows (width/style/running changed). */
-  hostRebuilds: number;
-  /** Cards whose `invalidate()` the bridge called. */
-  cardInvalidations: number;
-  /** Design applications. */
-  designApplies: number;
-}
-
-/** A `Markdown`-shaped row: it caches its lines and re-parses only after `invalidate()`. */
-function mountProse(text: string, counters: Counters): { render(width: number): string[]; invalidate(): void } {
-  let cached: string[] | undefined;
-  return {
-    render(width: number): string[] {
-      if (cached !== undefined) return cached;
-      counters.proseCacheMisses += 1;
-      cached = [`${text} `.repeat(4).trim().padEnd(width, " ")];
-      return cached;
-    },
-    invalidate(): void {
-      counters.proseInvalidations += 1;
-      cached = undefined;
-    },
-  };
-}
-
-/**
- * A host bash card whose renderer caches its frame the way `ToolCardRows` does
- * (`extensions/quiet-tools.ts:612-624`): keyed on width, the host's own style and whether it is
- * running. The DESIGN is not in that key — which is why the rows it answers with are cached under
- * the previous selection, and why the wrapper must post-process them instead of the tree being
- * invalidated.
- */
-function hostCardDefinition(counters: Counters): Record<string, unknown> {
-  const frame = { width: -1, style: "", lines: [...HOST_CARD] };
-  const hostStyle = "float";
-  return {
-    renderShell: "default",
-    renderCall: (..._args: unknown[]) => ({
-      render(width: number): string[] {
-        counters.hostRenders += 1;
-        if (frame.width === width && frame.style === hostStyle) {
-          counters.hostCacheHits += 1;
-          return frame.lines;
-        }
-        counters.hostRebuilds += 1;
-        frame.width = width;
-        frame.style = hostStyle;
-        frame.lines = [...HOST_CARD];
-        return frame.lines;
-      },
-      invalidate(): void {
-        frame.width = -1;
-      },
-    }),
-  };
-}
-
-/** The part of pi's `ToolExecutionComponent` this test needs: it stores and re-invokes the renderer. */
-interface FakeCard {
-  toolName: string;
-  toolDefinition: Record<string, unknown>;
-  renders: number;
-  component?: { render(width: number): string[] };
-  updateDisplay(): void;
-  invalidate(): void;
-  render(width: number): string[];
-}
-
-function mountCard(counters: Counters): FakeCard {
-  const card: FakeCard = {
-    toolName: "bash",
-    toolDefinition: toolDesignResolver("bash", () => hostCardDefinition(counters)) as Record<
-      string,
-      unknown
-    >,
-    renders: 0,
-    updateDisplay(): void {
-      const renderCall = card.toolDefinition.renderCall;
-      card.component =
-        typeof renderCall === "function"
-          ? (renderCall as (...args: unknown[]) => { render(width: number): string[] })(
-              { command: "ls -la" },
-              THEME,
-              { args: {} },
-            )
-          : undefined;
-    },
-    invalidate(): void {
-      counters.cardInvalidations += 1;
-      card.updateDisplay();
-    },
-    render(width: number): string[] {
-      card.renders += 1;
-      return card.component === undefined ? [] : card.component.render(width);
-    },
-  };
-  card.updateDisplay();
-  return card;
-}
-
-interface Transcript {
-  counters: Counters;
-  /** The conversation render pass: every mounted child, in order (pi's `Container.render`). */
-  drawFrame(): string[][];
-  cards: FakeCard[];
-  proseInvalidations: () => number;
-  proseCacheMisses: () => number;
-}
-
-function transcript(options: { children: number; markdown: number }): Transcript {
-  const counters: Counters = {
-    treeInvalidations: 0,
-    frameRequests: [],
-    proseCacheMisses: 0,
-    proseInvalidations: 0,
-    hostRenders: 0,
-    hostCacheHits: 0,
-    hostRebuilds: 0,
-    cardInvalidations: 0,
-    designApplies: 0,
-    // The count of design applications is kept by the registered design itself; see `press`.
-  } as Counters;
-
-  const nodes: Array<{ render(width: number): string[]; invalidate?(): void }> = [];
-  const cards: FakeCard[] = [];
-  const bash = options.children - options.markdown;
-  for (let index = 0; index < options.markdown; index += 1) {
-    nodes.push(mountProse(`assistant message ${index}`, counters));
-  }
-  for (let index = 0; index < bash; index += 1) {
-    const card = mountCard(counters);
-    cards.push(card);
-    nodes.push(card);
-  }
-
-  setRepaintHost({
-    terminal: {},
-    invalidate: (): void => {
-      counters.treeInvalidations += 1;
-      for (const node of nodes) node.invalidate?.();
-    },
-    requestRender: (force?: boolean): void => {
-      counters.frameRequests.push(force === true ? "forced" : "plain");
-    },
-  });
-
-  return {
-    counters,
-    drawFrame: () => nodes.map((node) => node.render(WIDTH)),
-    cards,
-    proseInvalidations: () => counters.proseInvalidations,
-    proseCacheMisses: () => counters.proseCacheMisses,
-  };
-}
-
-/** The menu's live rows, built through the host contract the menu reads. */
-function rows(): CustomizeRow[] {
-  installToolDesignRows();
-  return buildRows({ theme: undefined, configHome }).rows;
-}
-
-/** Press one value of the bash line, exactly as the menu runs a row action. */
-async function press(value: string): Promise<void> {
-  const row = rows().find((candidate) => candidate.choice?.value === value);
-  assert.ok(row?.action !== undefined, `the menu has a ${JSON.stringify(value)} value`);
-  await row.action({
-    input: async () => undefined,
-    confirm: async () => true,
-    capture: async () => undefined,
-  });
-}
-
-/** Everything one switch costs, as counts. */
-interface SwitchCost {
-  treeInvalidations: number;
-  frameRequests: Array<"plain" | "forced">;
-  selectionReadsInFrame: number;
-  designAppliesInFrame: number;
-  proseCacheMissesInFrame: number;
-  hostCacheHitsInFrame: number;
-  cardInvalidations: number;
-}
-
-async function measureSwitch(options: { children: number; markdown: number }): Promise<{
-  cost: SwitchCost;
-  drawn: string[][];
-  transcript: Transcript;
-}> {
-  registerToolDesign(bashHeadingDesign);
-  setToolDesignSelection("bash", TOOL_DESIGN_NATIVE);
-  const world = transcript(options);
-  // Warm every cache with one frame of the host's own rows.
-  world.drawFrame();
-
-  await press(BASH_HEADING);
-
-  resetToolDesignSelectionReads();
-  const appliesBefore = applied;
-  const proseBefore = world.proseCacheMisses();
-  const hitsBefore = world.counters.hostCacheHits;
-  const drawn = world.drawFrame();
-
-  return {
-    cost: {
-      treeInvalidations: world.counters.treeInvalidations,
-      frameRequests: [...world.counters.frameRequests],
-      selectionReadsInFrame: getToolDesignSelectionReads(),
-      designAppliesInFrame: applied - appliesBefore,
-      proseCacheMissesInFrame: world.proseCacheMisses() - proseBefore,
-      hostCacheHitsInFrame: world.counters.hostCacheHits - hitsBefore,
-      cardInvalidations: world.counters.cardInvalidations,
-    },
-    drawn,
-    transcript: world,
-  };
-}
-
-// ---------------------------------------------------------------------------------------------
-// T23 — the bound
-// ---------------------------------------------------------------------------------------------
-
-test("T23/S13: a switch asks for ONE differential frame and invalidates nothing", async () => {
-  registerToolDesign(bashHeadingDesign);
-  const world = transcript({ children: 6, markdown: 3 });
-  world.drawFrame();
-
-  await press(BASH_HEADING);
-
-  assert.equal(
-    world.counters.treeInvalidations,
-    0,
-    "the tree is not invalidated: every Markdown row would re-parse and every card would rebuild",
-  );
-  assert.deepEqual(
-    world.counters.frameRequests,
-    ["plain"],
-    "one frame, differential: no forced full redraw of the whole buffer",
-  );
-  assert.equal(world.counters.cardInvalidations, 0, "no card is invalidated by a selection change");
-
-  const drawn = world.drawFrame();
-  assert.deepEqual(drawn.filter((lines) => lines.length === HOST_CARD.length), [
-    DESIGNED_CARD,
-    DESIGNED_CARD,
-    DESIGNED_CARD,
-  ], "every already-constructed card draws the design");
-  assert.equal(
-    world.counters.hostCacheHits,
-    3,
-    "the host answered from its OWN frame cache; the wrapper post-processed those rows",
-  );
-  assert.equal(world.counters.hostRebuilds, 3, "the host rebuilt its rows once, at warm-up");
-  assert.equal(getRepaintDiagnostics().lastTrigger?.outcome, "repainted");
-
-  // S2's other half: `native` restores the host's exact bytes on the SAME components.
-  await press(TOOL_DESIGN_NATIVE);
-  assert.deepEqual(
-    world.drawFrame().filter((lines) => lines.length === HOST_CARD.length),
-    [HOST_CARD, HOST_CARD, HOST_CARD],
-    "native restores the host's own rows, byte for byte",
-  );
-  assert.equal(world.counters.treeInvalidations, 0, "the way back costs no invalidation either");
-  assert.deepEqual(world.counters.frameRequests, ["plain", "plain"], "one frame per press");
-});
-
-test("T23/S13: the work of a switch does not grow with the number of Markdown rows", async () => {
-  const small = await measureSwitch({ children: 6, markdown: 3 });
-  await press(TOOL_DESIGN_NATIVE);
-  const large = await measureSwitch({ children: 203, markdown: 200 });
-
-  assert.deepEqual(
-    large.cost,
-    {
-      ...small.cost,
-      // The host renders once per card per frame in both; that is the switch's real work, and it is
-      // per CARD, never per Markdown row. Nothing else may differ.
-    },
-    "same cards, same frame, same reads: 200 prose rows cost a switch exactly what 3 do",
-  );
-  assert.equal(large.cost.treeInvalidations, 0);
-  assert.deepEqual(large.cost.frameRequests, ["plain"]);
-});
-
-test("T23/S13: a switch performs at most one selection read per card per frame", async () => {
-  registerToolDesign(bashHeadingDesign);
-  const world = transcript({ children: 10, markdown: 6 });
-  world.drawFrame();
-
-  await press(BASH_HEADING);
-
-  resetToolDesignSelectionReads();
-  world.drawFrame();
-  const firstFrame = getToolDesignSelectionReads();
-  assert.equal(firstFrame, 4, "one read per card, no more");
-
-  resetToolDesignSelectionReads();
-  world.drawFrame();
-  assert.equal(getToolDesignSelectionReads(), 4, "and the same on every later frame");
-
-  // A frame with no design still reads the selection once per card: the bound is about the CARD,
-  // not about which value is selected.
-  await press(TOOL_DESIGN_NATIVE);
-  resetToolDesignSelectionReads();
-  world.drawFrame();
-  assert.equal(getToolDesignSelectionReads(), 4);
-
-  // Reads never grow with the transcript's prose, only with its cards.
-  const wide = await measureSwitch({ children: 104, markdown: 100 });
-  assert.equal(wide.cost.selectionReadsInFrame, 4, "100 prose rows still cost four selection reads");
-  assert.ok(
-    wide.cost.selectionReadsInFrame <= 4,
-    "the bound: at most one read per card per frame",
-  );
-});
-
-test("T23/S13: the prose rows never re-parse for a switch", async () => {
-  const { cost } = await measureSwitch({ children: 30, markdown: 25 });
-  assert.equal(cost.proseCacheMissesInFrame, 0, "no Markdown row re-parses on a design switch");
-  assert.equal(cost.cardInvalidations, 0);
-});
-
-// S19 replaced the PRESERVE test that pinned the heavy path (`tui.invalidate()` + a forced frame) for
-// bridge-owned rows. Measured with real pi components, the rows are byte-identical without them on the
-// conversation, the rail's host panels and the bridge's sections, and the invalidation was the
-// expensive half (697 ms vs 420 ms for `plain`, 99 ms vs 3 ms for a rail section toggle at 600
-// children). test/surface-repaint-cost.test.ts holds the full bound; this pins the same request here.
-test("T36/S19: the bridge-owned surface repaint is one differential frame too", () => {
-  const world = transcript({ children: 2, markdown: 1 });
-  world.counters.treeInvalidations = 0;
-  world.counters.frameRequests.length = 0;
-
-  requestSurfaceRepaint({ cause: "card style → neon", surfaces: ["conversation", "rail"] });
-
-  assert.equal(world.counters.treeInvalidations, 0, "no tree invalidation: the Markdown caches survive");
-  assert.deepEqual(world.counters.frameRequests, ["plain"], "one differential frame, not a forced one");
-});
-
-// Keep the compiler honest about the design type used above.
-const _contextProbe: ToolDesignContext = { width: WIDTH, tool: "bash", theme: THEME, part: "call" };
-void _contextProbe;