diff --git a/.gitignore b/.gitignore
index 044b6ea46..0c22cc6eb 100644
--- a/.gitignore
+++ b/.gitignore
@@ -70,3 +70,6 @@ python/robomp/.cache/
 python/robomp/src/robomp/static/
 python/robomp/web/dist/
 python/robomp/.env
+
+# Local, machine-specific boot perf baseline (see packages/coding-agent/scripts/bench-guard.ts)
+packages/coding-agent/bench/boot-baseline.json
diff --git a/packages/coding-agent/CHANGELOG.md b/packages/coding-agent/CHANGELOG.md
index bef9ae4ac..3003906e0 100644
--- a/packages/coding-agent/CHANGELOG.md
+++ b/packages/coding-agent/CHANGELOG.md
@@ -2,6 +2,14 @@
 
 ## [Unreleased]
 
+### Changed
+
+- Interactive boot no longer blocks first paint on MCP server discovery. For UI sessions, `createAgentSession` returns immediately and connects to configured MCP servers in the background; their tools and slash commands stream in through the existing live-refresh channel once connected. An explicitly requested MCP tool whose server has not finished connecting resolves to a deterministic "still connecting" placeholder instead of an "unknown tool" error, and MCP server instructions join the system prompt on the next session rather than retroactively. Non-UI modes (`print`/`rpc`/`acp`) keep the blocking discovery path. Measured ~290 ms (≈24% of cold boot) off the first-paint critical path with MCP servers configured.
+- Assistant-message streaming is cheaper per token. During a stream the component now reuses its `Markdown` subtree across reveal ticks — only the growing block is re-rendered — instead of tearing down and re-lexing every block on each ~30 fps tick, and grapheme counting in the reveal controller is memoized. A completed thinking block that precedes a still-streaming answer is no longer re-highlighted every frame (~66% less render work on think-then-answer streams in benchmarks; single-block streams are unchanged).
+- Cold boot no longer builds the model catalog's canonical-equivalence index on the first-paint critical path. The `ModelRegistry` constructor built `buildCanonicalModelIndex` over the entire ~3,200-model catalog synchronously (~210 ms); it is now built lazily on first read (`getCanonicalModels`/`getCanonicalVariants`/`getCanonicalId`, reached by the model picker and by `enabledModels`/default-role pattern resolution), which a default interactive launch never touches before paint. Measured ~244 ms (≈16% of cold-boot wall) off first paint; the picker pays the one-time build on first open.
+- Repeat `read` summaries of an unchanged file no longer re-run the tree-sitter parse. The per-session summary is memoized on the content hash of the freshly-read bytes — the file is still read fresh on every call, so results stay correct without a staleness window — dropping a repeated same-file summary read from ~17 ms to ~2.5 ms.
+- Attributed the previously-unlabeled synchronous boot region in the `PI_TIMING` startup table with `modelRegistry:init`, `buildCanonicalModelIndex`, and `initTelemetryExport` spans.
+
 ## [15.10.8] - 2026-06-09
 
 ### Added
diff --git a/packages/coding-agent/bench/rendering.ts b/packages/coding-agent/bench/rendering.ts
index e15d82976..8e32428a6 100644
--- a/packages/coding-agent/bench/rendering.ts
+++ b/packages/coding-agent/bench/rendering.ts
@@ -1,6 +1,18 @@
 import { initTheme } from "../src/modes/theme/theme";
 import { truncateToVisualLines } from "../src/modes/components/visual-truncate";
 import { WelcomeComponent } from "../src/modes/components/welcome";
+import type { AssistantMessage } from "@oh-my-pi/pi-ai";
+import { Editor } from "@oh-my-pi/pi-tui";
+import { AssistantMessageComponent } from "../src/modes/components/assistant-message";
+import { TranscriptContainer } from "../src/modes/components/transcript-container";
+import { Settings } from "../src/config/settings";
+import { getEditorTheme } from "../src/modes/theme/theme";
+import { buildDisplayMessage, nextStep, visibleUnits } from "../src/modes/controllers/streaming-reveal";
+import * as os from "node:os";
+import * as path from "node:path";
+import * as fs from "node:fs";
+import { ReadTool } from "../src/tools/read";
+import type { ToolSession } from "../src/tools";
 
 const ITERATIONS = 500;
 const WIDTH = 100;
@@ -20,6 +32,7 @@ function bench(name: string, fn: () => void): number {
 	return elapsed;
 }
 
+await Settings.init({ inMemory: true });
 await initTheme("dark");
 
 console.log(`Rendering benchmark (${ITERATIONS} iterations)\n`);
@@ -39,3 +52,223 @@ const welcome = new WelcomeComponent("8.12.3", "claude-3.7", "anthropic",	[
 bench("WelcomeComponent.render", () => {
 	welcome.render(WIDTH);
 });
+
+// ── A2: streaming reveal + editor render baselines ──────────────────────────
+//
+// Diagnostic series, not a fixed-iteration micro-op. `streamingReveal` proves
+// or refutes the O(N^2) reveal hypothesis: per-step cost (visibleUnits +
+// buildDisplayMessage, the work every stream delta/30fps tick does) is sampled
+// at growing revealed lengths. Rising per-step ms => O(N) per tick => O(N^2)
+// over the message. Flat per-step => already linear.
+
+function makeMarkdownCorpus(targetGraphemes: number): string {
+	const para =
+		"The quick brown fox jumps over the lazy dog while 🚀 emoji and a `code span` " +
+		"plus **bold** and _italic_ text exercise the markdown lexer and the grapheme segmenter. ";
+	const codeBlock = "\n```ts\nconst x: number = compute(a, b) + delta;\nreturn x.toFixed(2);\n```\n\n";
+	const list = "\n- first bullet item\n- second bullet item with `inline`\n- third\n\n";
+	let out = "";
+	let i = 0;
+	while (out.length < targetGraphemes) {
+		out += `## Section ${++i}\n\n${para}${para}${codeBlock}${list}`;
+	}
+	return out.slice(0, targetGraphemes);
+}
+
+function makeTextMessage(text: string): AssistantMessage {
+	return {
+		role: "assistant",
+		content: [{ type: "text", text }],
+		api: "anthropic-messages",
+		provider: "anthropic",
+		model: "bench",
+		usage: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, totalTokens: 0, cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 } },
+		stopReason: "stop",
+		timestamp: 0,
+	};
+}
+
+/** Average ms for one call of `fn`, over `reps` repeats. */
+function benchStep(reps: number, fn: () => void): number {
+	const start = Bun.nanoseconds();
+	for (let i = 0; i < reps; i++) fn();
+	return (Bun.nanoseconds() - start) / 1e6 / reps;
+}
+
+/** Average ms for one awaited call of `fn`, over `reps` repeats. */
+async function benchStepAsync(reps: number, fn: () => Promise<unknown>): Promise<number> {
+	const start = Bun.nanoseconds();
+	for (let i = 0; i < reps; i++) await fn();
+	return (Bun.nanoseconds() - start) / 1e6 / reps;
+}
+
+const REVEAL_CORPUS = makeMarkdownCorpus(6000);
+const REVEAL_CHECKPOINTS = [1000, 2000, 3000, 4000, 5000, 6000];
+const STEP_REPS = 40;
+
+console.log("\nstreamingReveal (isolated C1: visibleUnits + buildDisplayMessage per delta):");
+for (const n of REVEAL_CHECKPOINTS) {
+	const msg = makeTextMessage(REVEAL_CORPUS.slice(0, n));
+	const revealed = Math.floor(n * 0.9);
+	const ms = benchStep(STEP_REPS, () => {
+		visibleUnits(msg, false);
+		buildDisplayMessage(msg, revealed, false);
+	});
+	console.log(`  len=${n}: ${ms.toFixed(4)}ms/step`);
+}
+
+// Real streaming cost: text GROWS every tick, so Markdown's text-keyed cache
+// misses each step (the actual interactive path). Total ms to fully reveal an
+// N-grapheme message in nextStep increments — the number C1+C2 reduce.
+console.log("\nstreamingRevealFull (C1+C2: full incremental reveal, growing text => cache-miss/tick):");
+try {
+	for (const n of REVEAL_CHECKPOINTS) {
+		const full = makeTextMessage(REVEAL_CORPUS.slice(0, n));
+		const total = visibleUnits(full, false);
+		const component = new AssistantMessageComponent();
+		const start = Bun.nanoseconds();
+		let revealed = 0;
+		let steps = 0;
+		while (revealed < total) {
+			revealed = Math.min(total, revealed + nextStep(total - revealed));
+			component.updateContent(buildDisplayMessage(full, revealed, false));
+			component.render(WIDTH);
+			steps++;
+		}
+		const ms = (Bun.nanoseconds() - start) / 1e6;
+		console.log(`  len=${n}: ${ms.toFixed(2)}ms total over ${steps} steps (${(ms / steps).toFixed(4)}ms/step)`);
+	}
+} catch (err) {
+	console.log(`  (skipped: ${(err as Error).message})`);
+}
+
+// Multi-block variant: a finalized thinking block (stable) precedes the growing
+// text block — the shape C2 targets. Current code re-lexes BOTH every tick;
+// after C2 the finalized thinking block stays L1-cached and only the tail re-lexes.
+function makeThinkingPlusText(thinking: string, text: string): AssistantMessage {
+	return { ...makeTextMessage(text), content: [{ type: "thinking", thinking }, { type: "text", text }] };
+}
+console.log("\nstreamingRevealMultiBlock (C2: finalized thinking block + growing text):");
+try {
+	const thinking = makeMarkdownCorpus(2500);
+	for (const n of [2000, 4000, 6000]) {
+		const full = makeThinkingPlusText(thinking, REVEAL_CORPUS.slice(0, n));
+		const total = visibleUnits(full, false);
+		const component = new AssistantMessageComponent();
+		const start = Bun.nanoseconds();
+		let revealed = 0;
+		let steps = 0;
+		while (revealed < total) {
+			revealed = Math.min(total, revealed + nextStep(total - revealed));
+			component.updateContent(buildDisplayMessage(full, revealed, false));
+			component.render(WIDTH);
+			steps++;
+		}
+		const ms = (Bun.nanoseconds() - start) / 1e6;
+		console.log(`  text=${n} (+2500 thinking): ${ms.toFixed(2)}ms total over ${steps} steps (${(ms / steps).toFixed(4)}ms/step)`);
+	}
+} catch (err) {
+	console.log(`  (skipped: ${(err as Error).message})`);
+}
+
+console.log("\neditorKeystroke (C3: layout recompute vs no-mutation render):");
+try {
+	const buffer = Array.from({ length: 50 })
+		.map((_, i) => `Line ${i + 1}: some editor content with words to wrap at width ${WIDTH} and more text here`)
+		.join("\n");
+
+	const e1 = new Editor(getEditorTheme());
+	e1.setText(buffer);
+	e1.render(WIDTH); // warm
+	const noMutMs = benchStep(200, () => {
+		e1.render(WIDTH);
+	});
+	console.log(`  no-mutation render: ${noMutMs.toFixed(4)}ms/render`);
+
+	const e2 = new Editor(getEditorTheme());
+	e2.setText(buffer);
+	e2.render(WIDTH); // warm
+	const editMs = benchStep(200, () => {
+		e2.insertText("x");
+		e2.render(WIDTH);
+	});
+	console.log(`  edit + render: ${editMs.toFixed(4)}ms/op`);
+} catch (err) {
+	console.log(`  (skipped: ${(err as Error).message})`);
+}
+
+// ── E3: long-transcript frame cost ──────────────────────────────────────────
+//
+// E3 root cause: Container.render walks EVERY child and concatenates their line
+// arrays on every frame. Finalized messages hit their Markdown L1 cache (no
+// re-lex) but still pay the tree walk + line-array rebuild/concat per frame.
+// Build N finalized assistant messages (prose + closed code fences) + 1 growing
+// tail, then time one render(WIDTH) of the whole tree per streaming frame.
+// Rising ms/frame in N => the stable history is re-walked/re-concatenated each
+// frame (the cost E3 culls); flat => the walk is already cheap.
+console.log("\nlongTranscriptFrame (E3: whole-tree render cost vs transcript length N):");
+try {
+	const histText = makeMarkdownCorpus(800);
+	const tailCorpus = makeMarkdownCorpus(1200);
+	for (const n of [50, 100, 200]) {
+		const container = new TranscriptContainer();
+		for (let i = 0; i < n; i++) {
+			const c = new AssistantMessageComponent();
+			c.updateContent(makeTextMessage(histText));
+			container.addChild(c);
+		}
+		const tail = new AssistantMessageComponent();
+		container.addChild(tail);
+		let revealed = Math.floor(tailCorpus.length * 0.5);
+		tail.updateContent(makeTextMessage(tailCorpus.slice(0, revealed)));
+		container.render(WIDTH); // warm finalized history (L1 caches hot)
+		const ms = benchStep(60, () => {
+			revealed += 20;
+			if (revealed > tailCorpus.length) revealed = Math.floor(tailCorpus.length * 0.5);
+			tail.updateContent(makeTextMessage(tailCorpus.slice(0, revealed)));
+			container.render(WIDTH);
+		});
+		console.log(`  N=${n}: ${ms.toFixed(4)}ms/frame`);
+	}
+} catch (err) {
+	console.log(`  (skipped: ${(err as Error).message})`);
+}
+
+// ── E4: tool read/parse redundancy ──────────────────────────────────────────
+//
+// E4 root cause: the read tool re-parses (tree-sitter `summarizeCode`, ~12-18ms
+// for a ~1500-line file) on every summary read of the same unchanged file. E4-ii
+// memoizes the parse per session keyed on the content hash of the freshly-read
+// bytes, so a repeat read of the same file reuses the parse (the file is still
+// read fresh, so the result stays correct). A repeated same-session summary read
+// should drop from ~17ms to a few ms; a fresh session each call stays full cost.
+console.log("\ntoolReadReparse (E4: repeat summary read, memoized parse vs cold):");
+try {
+	const dir = fs.mkdtempSync(path.join(os.tmpdir(), "bench-e4-"));
+	const file = path.join(dir, "big.ts");
+	let src = "";
+	for (let i = 0; i < 375; i++) {
+		src += `export function fn${i}(a: number, b: string): boolean {\n  const x = a + ${i};\n  return x > 0 && b.length === ${i};\n}\n`;
+	}
+	fs.writeFileSync(file, src);
+	const mkSession = (): ToolSession =>
+		({
+			cwd: dir,
+			hasUI: false,
+			getSessionFile: () => path.join(dir, "s.jsonl"),
+			getSessionSpawns: () => "*",
+			getArtifactsDir: () => path.join(dir, "sess"),
+			allocateOutputArtifact: async (t: string) => ({ id: "a", path: path.join(dir, `a.${t}.log`) }),
+			settings: Settings.isolated(),
+		}) as unknown as ToolSession;
+	const sameSession = mkSession();
+	const rt = new ReadTool(sameSession);
+	for (let i = 0; i < 3; i++) await rt.execute("warm", { path: file });
+	const repeatMs = await benchStepAsync(20, () => rt.execute("c", { path: file }));
+	const coldMs = await benchStepAsync(20, () => new ReadTool(mkSession()).execute("c", { path: file }));
+	console.log(`  same-session repeat read: ${repeatMs.toFixed(3)}ms/call (memoized parse)`);
+	console.log(`  fresh-session each read:  ${coldMs.toFixed(3)}ms/call (cold parse)`);
+	fs.rmSync(dir, { recursive: true, force: true });
+} catch (err) {
+	console.log(`  (skipped: ${(err as Error).message})`);
+}
diff --git a/packages/coding-agent/package.json b/packages/coding-agent/package.json
index d252cf5ed..81b33faa1 100644
--- a/packages/coding-agent/package.json
+++ b/packages/coding-agent/package.json
@@ -41,7 +41,8 @@
 		"format-prompts": "bun scripts/format-prompts.ts",
 		"generate-docs-index": "bun scripts/generate-docs-index.ts",
 		"prepack": "bun scripts/generate-docs-index.ts",
-		"generate-template": "bun scripts/generate-template.ts"
+		"generate-template": "bun scripts/generate-template.ts",
+		"bench:guard": "bun scripts/bench-guard.ts"
 	},
 	"dependencies": {
 		"@agentclientprotocol/sdk": "catalog:",
diff --git a/packages/coding-agent/scripts/bench-guard.ts b/packages/coding-agent/scripts/bench-guard.ts
new file mode 100755
index 000000000..2ce365d9f
--- /dev/null
+++ b/packages/coding-agent/scripts/bench-guard.ts
@@ -0,0 +1,71 @@
+#!/usr/bin/env bun
+/**
+ * Boot-time regression guard (Phase A1 of the boot/TUI perf work).
+ *
+ * Re-runs the `PI_TIMING=x` cold-boot benchmark under hyperfine and fails when
+ * the median regresses past `baseline * THRESHOLD`. `PI_TIMING=x` runs the full
+ * pre-paint chain in `runRootCommand` and then `process.exit(0)`, so the
+ * never-exiting interactive launch becomes a terminating, benchmarkable boot.
+ *
+ * Boot wall-clock is MACHINE-RELATIVE: a baseline captured on one machine is
+ * meaningless on another (and on CI). This is a LOCAL guard — regenerate the
+ * baseline on the machine you measure on, then compare on that same machine.
+ * It is intentionally NOT wired into CI for that reason.
+ *
+ *   bun scripts/bench-guard.ts --update   # capture/refresh the baseline
+ *   bun scripts/bench-guard.ts            # measure + compare; exit 1 on regression
+ *
+ * Requires `hyperfine` on PATH.
+ */
+import * as fs from "node:fs";
+import * as path from "node:path";
+
+const THRESHOLD = 1.05; // 5% regression budget
+const BASELINE_PATH = path.join(import.meta.dir, "..", "bench", "boot-baseline.json");
+const BENCH_COMMAND = "PI_TIMING=x bun src/cli.ts";
+const cwd = path.join(import.meta.dir, "..");
+
+function medianOf(hyperfineJson: string): number {
+	const parsed = JSON.parse(hyperfineJson) as { results: Array<{ mean: number; median?: number }> };
+	const result = parsed.results[0];
+	if (!result) throw new Error("hyperfine produced no result");
+	return result.median ?? result.mean;
+}
+
+async function measure(): Promise<{ seconds: number; raw: string }> {
+	const tmp = path.join(import.meta.dir, "..", "bench", `.boot-run-${Date.now()}.json`);
+	const proc = Bun.spawn(["hyperfine", "--warmup", "3", "--min-runs", "10", "--export-json", tmp, BENCH_COMMAND], {
+		cwd,
+		stdout: "inherit",
+		stderr: "inherit",
+	});
+	const code = await proc.exited;
+	if (code !== 0) throw new Error(`hyperfine exited ${code}`);
+	const raw = await Bun.file(tmp).text();
+	fs.rmSync(tmp, { force: true });
+	return { seconds: medianOf(raw), raw };
+}
+
+const update = process.argv.includes("--update");
+const { seconds, raw } = await measure();
+
+if (update) {
+	fs.mkdirSync(path.dirname(BASELINE_PATH), { recursive: true });
+	await Bun.write(BASELINE_PATH, raw);
+	console.log(`Baseline updated: ${(seconds * 1000).toFixed(0)}ms median -> ${BASELINE_PATH}`);
+	process.exit(0);
+}
+
+if (!fs.existsSync(BASELINE_PATH)) {
+	console.error("No baseline found. Run `bun scripts/bench-guard.ts --update` on this machine first.");
+	process.exit(2);
+}
+
+const baseline = medianOf(await Bun.file(BASELINE_PATH).text());
+const ratio = seconds / baseline;
+const verdict = ratio > THRESHOLD ? "REGRESSION" : "ok";
+console.log(
+	`boot median: ${(seconds * 1000).toFixed(0)}ms vs baseline ${(baseline * 1000).toFixed(0)}ms ` +
+		`(${((ratio - 1) * 100).toFixed(1)}%, budget ${((THRESHOLD - 1) * 100).toFixed(0)}%) -> ${verdict}`,
+);
+process.exit(ratio > THRESHOLD ? 1 : 0);
diff --git a/packages/coding-agent/src/config/model-registry.ts b/packages/coding-agent/src/config/model-registry.ts
index 51b62be06..dc83a3171 100644
--- a/packages/coding-agent/src/config/model-registry.ts
+++ b/packages/coding-agent/src/config/model-registry.ts
@@ -899,6 +899,7 @@ function getConfiguredProviderOrderFromSettings(): string[] {
 export class ModelRegistry {
 	#models: Model<Api>[] = [];
 	#canonicalIndex: CanonicalModelIndex = { records: [], byId: new Map(), bySelector: new Map() };
+	#canonicalIndexDirty: boolean = true;
 	#customProviderApiKeys: Map<string, string> = new Map();
 	#keylessProviders: Set<string> = new Set();
 	#discoverableProviders: DiscoveryProviderConfig[] = [];
@@ -2158,10 +2159,25 @@ export class ModelRegistry {
 			this.#rebuildPending = true;
 			return;
 		}
-		this.#canonicalIndex = buildCanonicalModelIndex(this.#models, this.#equivalenceConfig);
+		// Defer the catalog-wide index build to first read. Boot model
+		// resolution reads it only when enabledModels or a default-role pattern
+		// is configured; the empty interactive launch never reads it pre-paint,
+		// so the ~200ms build over the full catalog moves off the first-paint
+		// critical path.
+		this.#canonicalIndexDirty = true;
 		this.#rebuildPending = false;
 	}
 
+	#ensureCanonicalIndex(): CanonicalModelIndex {
+		if (this.#canonicalIndexDirty) {
+			this.#canonicalIndex = logger.time("buildCanonicalModelIndex", () =>
+				buildCanonicalModelIndex(this.#models, this.#equivalenceConfig),
+			);
+			this.#canonicalIndexDirty = false;
+		}
+		return this.#canonicalIndex;
+	}
+
 	#suspendRebuild(): void {
 		this.#rebuildSuspended += 1;
 	}
@@ -2172,7 +2188,7 @@ export class ModelRegistry {
 		}
 		if (this.#rebuildSuspended === 0 && this.#rebuildPending) {
 			this.#rebuildPending = false;
-			this.#canonicalIndex = buildCanonicalModelIndex(this.#models, this.#equivalenceConfig);
+			this.#canonicalIndexDirty = true;
 		}
 	}
 
@@ -2290,7 +2306,7 @@ export class ModelRegistry {
 
 	getCanonicalModels(options?: CanonicalModelQueryOptions): CanonicalModelRecord[] {
 		const records: CanonicalModelRecord[] = [];
-		for (const record of this.#canonicalIndex.records) {
+		for (const record of this.#ensureCanonicalIndex().records) {
 			const variants = this.#filterCanonicalVariants(record, options);
 			if (variants.length === 0) {
 				continue;
@@ -2305,7 +2321,7 @@ export class ModelRegistry {
 	}
 
 	getCanonicalVariants(canonicalId: string, options?: CanonicalModelQueryOptions): CanonicalModelVariant[] {
-		const record = this.#canonicalIndex.byId.get(canonicalId.trim().toLowerCase());
+		const record = this.#ensureCanonicalIndex().byId.get(canonicalId.trim().toLowerCase());
 		if (!record) {
 			return [];
 		}
@@ -2322,7 +2338,7 @@ export class ModelRegistry {
 	}
 
 	getCanonicalId(model: Model<Api>): string | undefined {
-		return this.#canonicalIndex.bySelector.get(formatCanonicalVariantSelector(model).toLowerCase());
+		return this.#ensureCanonicalIndex().bySelector.get(formatCanonicalVariantSelector(model).toLowerCase());
 	}
 
 	/**
diff --git a/packages/coding-agent/src/main.ts b/packages/coding-agent/src/main.ts
index b690978d0..aedf6d396 100644
--- a/packages/coding-agent/src/main.ts
+++ b/packages/coding-agent/src/main.ts
@@ -804,7 +804,7 @@ export async function runRootCommand(
 
 	// Create AuthStorage and ModelRegistry upfront
 	const authStorage = await logger.time("discoverModels", deps.discoverAuthStorage ?? discoverAuthStorage);
-	const modelRegistry = new ModelRegistry(authStorage);
+	const modelRegistry = logger.time("modelRegistry:init", () => new ModelRegistry(authStorage));
 
 	if (parsedArgs.version) {
 		process.stdout.write(`${VERSION}\n`);
@@ -1045,7 +1045,7 @@ export async function runRootCommand(
 	// Both are no-ops when OTEL_EXPORTER_OTLP_ENDPOINT is unset. An empty config
 	// is enough to enable telemetry — content capture is governed by the
 	// standard OTEL_INSTRUMENTATION_GENAI_CAPTURE_MESSAGE_CONTENT env var.
-	await initTelemetryExport();
+	await logger.time("initTelemetryExport", initTelemetryExport);
 	if (isTelemetryExportEnabled()) {
 		sessionOptions.telemetry = {};
 	}
diff --git a/packages/coding-agent/src/modes/components/assistant-message.ts b/packages/coding-agent/src/modes/components/assistant-message.ts
index 61b59804a..35c11968e 100644
--- a/packages/coding-agent/src/modes/components/assistant-message.ts
+++ b/packages/coding-agent/src/modes/components/assistant-message.ts
@@ -36,6 +36,11 @@ export class AssistantMessageComponent extends Container {
 	 * transcript keeps the error in history.
 	 */
 	#errorPinned = false;
+	// Fast-path state: reuse Markdown children when message shape is stable during streaming.
+	#fastPathKey: string | undefined;
+	#fastPathItems:
+		| Array<{ md: Markdown; contentIndex: number; blockType: "text" | "thinking"; lastText: string }>
+		| undefined;
 
 	constructor(
 		message?: AssistantMessage,
@@ -211,12 +216,107 @@ export class AssistantMessageComponent extends Container {
 		}
 	}
 
+	#computeShapeKey(message: AssistantMessage): string {
+		const parts: string[] = [`htb:${this.hideThinkingBlock ? 1 : 0}`];
+		for (const content of message.content) {
+			if (content.type === "text") {
+				parts.push(content.text.trim() ? "T1" : "T0");
+			} else if (content.type === "thinking") {
+				if (!content.thinking.trim()) parts.push("K0");
+				else if (this.hideThinkingBlock) parts.push("KH");
+				else parts.push("KV");
+			} else {
+				// Non-rendered blocks (toolCall, redactedThinking, …) still occupy a
+				// content index. Encode their position so an inserted/removed one shifts
+				// the key and forces the teardown path instead of mis-indexing children.
+				parts.push(`O:${content.type}`);
+			}
+		}
+		if (settings.get("display.showTokenUsage") && this.#usageInfo) {
+			const u = this.#usageInfo;
+			parts.push(`u:${u.input + u.cacheWrite}:${u.output}:${u.cacheRead}`);
+		} else {
+			parts.push("u:");
+		}
+		return parts.join("|");
+	}
+
+	#canFastPath(message: AssistantMessage): boolean {
+		for (const content of message.content) {
+			if (content.type === "toolCall") return false;
+		}
+		if (this.#toolImagesByCallId.size > 0) return false;
+		if (message.stopReason === "aborted" && shouldRenderAbortReason(message.errorMessage)) return false;
+		if (message.stopReason === "error" && !this.#errorPinned) return false;
+		if (
+			message.errorMessage &&
+			shouldRenderAbortReason(message.errorMessage) &&
+			message.stopReason !== "aborted" &&
+			message.stopReason !== "error"
+		)
+			return false;
+		// Extension stability: if thinking renderers exist and any tracked thinking
+		// block's text changed, extensions may produce a different child count.
+		if (this.thinkingRenderers.length > 0 && this.#fastPathItems) {
+			for (const item of this.#fastPathItems) {
+				if (item.blockType === "thinking") {
+					const content = message.content[item.contentIndex];
+					if (content?.type === "thinking" && content.thinking.trim() !== item.lastText) return false;
+				}
+			}
+		}
+		return true;
+	}
+
+	#tryFastPathUpdate(message: AssistantMessage): boolean {
+		if (!this.#fastPathKey || !this.#fastPathItems) return false;
+		if (!this.#canFastPath(message)) {
+			this.#fastPathKey = undefined;
+			this.#fastPathItems = undefined;
+			return false;
+		}
+		if (this.#computeShapeKey(message) !== this.#fastPathKey) {
+			this.#fastPathKey = undefined;
+			this.#fastPathItems = undefined;
+			return false;
+		}
+		// Shape is identical — setText only on Markdown children whose source changed.
+		for (const item of this.#fastPathItems) {
+			const content = message.content[item.contentIndex];
+			let newText: string;
+			if (item.blockType === "text" && content?.type === "text") {
+				newText = content.text.trim();
+			} else if (item.blockType === "thinking" && content?.type === "thinking") {
+				newText = content.thinking.trim();
+			} else {
+				// Block at this index is gone or changed type (index shift) — fail closed.
+				this.#fastPathKey = undefined;
+				this.#fastPathItems = undefined;
+				return false;
+			}
+			if (newText !== item.lastText) {
+				item.md.setText(newText);
+				item.lastText = newText;
+			}
+		}
+		return true;
+	}
+
 	updateContent(message: AssistantMessage): void {
 		this.#lastMessage = message;
 
+		// Fast path: reuse Markdown children when shape is stable during streaming
+		if (this.#tryFastPathUpdate(message)) return;
+
 		// Clear content container
 		this.#contentContainer.clear();
 
+		// Determine if we should capture Markdown instances for next fast path
+		const shouldCapture = this.#canFastPath(message);
+		const captureItems:
+			| Array<{ md: Markdown; contentIndex: number; blockType: "text" | "thinking"; lastText: string }>
+			| undefined = shouldCapture ? [] : undefined;
+
 		const hasVisibleContent = message.content.some(
 			c => (c.type === "text" && c.text.trim()) || (c.type === "thinking" && c.thinking.trim()),
 		);
@@ -228,7 +328,10 @@ export class AssistantMessageComponent extends Container {
 			if (content.type === "text" && content.text.trim()) {
 				// Assistant text messages with no background - trim the text
 				// Set paddingY=0 to avoid extra spacing before tool executions
-				this.#contentContainer.addChild(new Markdown(content.text.trim(), 1, 0, getMarkdownTheme()));
+				const trimmed = content.text.trim();
+				const md = new Markdown(trimmed, 1, 0, getMarkdownTheme());
+				this.#contentContainer.addChild(md);
+				captureItems?.push({ md, contentIndex: i, blockType: "text", lastText: trimmed });
 			} else if (content.type === "thinking" && content.thinking.trim()) {
 				// Add spacing only when another visible assistant content block follows.
 				// This avoids a superfluous blank line before separately-rendered tool execution blocks.
@@ -245,12 +348,12 @@ export class AssistantMessageComponent extends Container {
 				} else {
 					const thinkingText = content.thinking.trim();
 					// Thinking traces in thinkingText color, italic
-					this.#contentContainer.addChild(
-						new Markdown(thinkingText, 1, 0, getMarkdownTheme(), {
-							color: (text: string) => theme.fg("thinkingText", text),
-							italic: true,
-						}),
-					);
+					const md = new Markdown(thinkingText, 1, 0, getMarkdownTheme(), {
+						color: (text: string) => theme.fg("thinkingText", text),
+						italic: true,
+					});
+					this.#contentContainer.addChild(md);
+					captureItems?.push({ md, contentIndex: i, blockType: "thinking", lastText: thinkingText });
 					this.#appendThinkingExtensions(i, thinkingIndex, thinkingText);
 					thinkingIndex += 1;
 					if (hasVisibleContentAfter) {
@@ -299,5 +402,14 @@ export class AssistantMessageComponent extends Container {
 			this.#contentContainer.addChild(new Spacer(1));
 			this.#contentContainer.addChild(new Text(theme.fg("dim", parts.join("  ")), 1, 0));
 		}
+
+		// Store fast-path state for next call
+		if (shouldCapture) {
+			this.#fastPathItems = captureItems;
+			this.#fastPathKey = this.#computeShapeKey(message);
+		} else {
+			this.#fastPathKey = undefined;
+			this.#fastPathItems = undefined;
+		}
 	}
 }
diff --git a/packages/coding-agent/src/modes/controllers/streaming-reveal.ts b/packages/coding-agent/src/modes/controllers/streaming-reveal.ts
index 1e4edccbf..6ed42cc9d 100644
--- a/packages/coding-agent/src/modes/controllers/streaming-reveal.ts
+++ b/packages/coding-agent/src/modes/controllers/streaming-reveal.ts
@@ -1,5 +1,6 @@
 import type { AssistantMessage } from "@oh-my-pi/pi-ai";
 import { getSegmenter } from "@oh-my-pi/pi-tui";
+import { LRUCache } from "lru-cache/raw";
 import type { AssistantMessageComponent } from "../components/assistant-message";
 
 export const STREAMING_REVEAL_FRAME_MS = 1000 / 30;
@@ -15,11 +16,17 @@ type StreamingRevealControllerOptions = {
 	requestRender(): void;
 };
 
+const graphemeCountCache = new LRUCache<string, number>({ max: 128 });
+
 function countGraphemes(text: string): number {
+	if (text.length === 0) return 0;
+	const cached = graphemeCountCache.get(text);
+	if (cached !== undefined) return cached;
 	let count = 0;
 	for (const _segment of getSegmenter().segment(text)) {
 		count += 1;
 	}
+	graphemeCountCache.set(text, count);
 	return count;
 }
 
diff --git a/packages/coding-agent/src/sdk.ts b/packages/coding-agent/src/sdk.ts
index 75fdb3966..a70b400a2 100644
--- a/packages/coding-agent/src/sdk.ts
+++ b/packages/coding-agent/src/sdk.ts
@@ -87,7 +87,14 @@ import { type FileSlashCommand, loadSlashCommands as loadSlashCommandsInternal }
 import type { HindsightSessionState } from "./hindsight/state";
 import { LocalProtocolHandler, type LocalProtocolOptions } from "./internal-urls";
 import { LSP_STARTUP_EVENT_CHANNEL, type LspStartupEvent } from "./lsp/startup-events";
-import { discoverAndLoadMCPTools, MCPManager, type MCPToolsLoadResult } from "./mcp";
+import {
+	discoverAndLoadMCPTools,
+	type MCPLoadResult,
+	MCPManager,
+	MCPToolCache,
+	type MCPToolsLoadResult,
+	parseMCPToolName,
+} from "./mcp";
 import { resolveMemoryBackend } from "./memory-backend";
 import type { MnemopiSessionState } from "./mnemopi/state";
 import asyncResultTemplate from "./prompts/tools/async-result.md" with { type: "text" };
@@ -141,6 +148,7 @@ import {
 	type DiscoverableTool,
 	filterBySource,
 	formatDiscoverableToolServerSummary,
+	isMCPToolName,
 	selectDiscoverableToolNamesByServer,
 	summarizeDiscoverableTools,
 } from "./tool-discovery/tool-index";
@@ -296,6 +304,76 @@ function buildMcpNotificationBatchMessage(entries: McpNotificationEntry[]): Agen
 	};
 }
 
+type DeferredMCPActivation = {
+	mcpDiscoveryEnabled: boolean;
+	explicitlyRequestedMCPToolNames: string[];
+	activateAllMCPTools: boolean;
+};
+
+function formatMCPConnectingMessage(serverNames: string[]): string {
+	return `Connecting to MCP servers: ${serverNames.join(", ")}…`;
+}
+
+function createPendingMCPTool(name: string): Tool {
+	const parsed = parseMCPToolName(name);
+	const serverName = parsed?.serverName;
+	const mcpToolName = parsed?.toolName ?? name;
+	const label = serverName ? `${serverName}/${mcpToolName}` : name;
+	const message = serverName
+		? `MCP server "${serverName}" is still connecting; tool "${name}" is not yet available. Retry after the MCP connection completes.`
+		: `MCP discovery is still in progress; tool "${name}" is not yet available. Retry after MCP connection completes.`;
+	const tool: Tool & { mcpServerName?: string; mcpToolName?: string } = {
+		name,
+		label,
+		description: `Pending MCP tool. ${message}`,
+		parameters: {
+			type: "object",
+			properties: {},
+			additionalProperties: true,
+		},
+		approval: "write",
+		intent: "omit",
+		mcpServerName: serverName,
+		mcpToolName,
+		async execute() {
+			return {
+				content: [{ type: "text", text: message }],
+				details: { serverName, mcpToolName, isError: true },
+				isError: true,
+			};
+		},
+	};
+	return tool;
+}
+
+function collectPendingMCPToolNames(
+	explicitToolNames: readonly string[] | undefined,
+	restoredSelectedToolNames: readonly string[],
+): string[] {
+	const names = new Set<string>();
+	for (const name of explicitToolNames ?? []) {
+		const normalized = name.toLowerCase();
+		if (isMCPToolName(normalized)) names.add(normalized);
+	}
+	for (const name of restoredSelectedToolNames) {
+		const normalized = name.toLowerCase();
+		if (isMCPToolName(normalized)) names.add(normalized);
+	}
+	return [...names];
+}
+
+function logMCPLoadErrors(errors: MCPLoadResult["errors"]): void {
+	for (const [serverName, error] of errors) {
+		logger.error("MCP tool load failed", { path: `mcp:${serverName}`, error });
+	}
+}
+
+function applyMCPEnvironment(result: { exaApiKeys: string[] }): void {
+	if (result.exaApiKeys.length > 0 && !$env.EXA_API_KEY) {
+		Bun.env.EXA_API_KEY = result.exaApiKeys[0];
+	}
+}
+
 // Types
 export interface CreateAgentSessionOptions {
 	/** Working directory for project-local discovery. Default: getProjectDir() */
@@ -1456,46 +1534,89 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 		let mcpManager: MCPManager | undefined = options.mcpManager;
 		toolSession.mcpManager = mcpManager;
 		const enableMCP = options.enableMCP ?? true;
+		const deferMCPDiscoveryForUI = enableMCP && !mcpManager && options.hasUI === true;
+		const includeMCPServerInstructionsInPrompt = !deferMCPDiscoveryForUI;
 		const customTools: CustomTool[] = [];
+		let startDeferredMCPDiscovery:
+			| ((liveSession: AgentSession, activation: DeferredMCPActivation) => void)
+			| undefined;
+		const onMCPConnecting = (serverNames: string[]) => {
+			if (!options.hasUI || serverNames.length === 0) return;
+			process.stderr.write(`${chalk.gray(formatMCPConnectingMessage(serverNames))}\n`);
+		};
+		const mcpDiscoverOptions = {
+			onConnecting: onMCPConnecting,
+			enableProjectConfig: settings.get("mcp.enableProjectConfig") ?? true,
+			// Always filter Exa - we have native integration
+			filterExa: true,
+			// Filter browser MCP servers when builtin browser tool is active
+			filterBrowser: settings.get("browser.enabled") ?? false,
+		};
 		if (enableMCP && !mcpManager) {
-			const mcpResult = await logger.time("discoverAndLoadMCPTools", discoverAndLoadMCPTools, cwd, {
-				onConnecting: serverNames => {
-					if (options.hasUI && serverNames.length > 0) {
-						process.stderr.write(`${chalk.gray(`Connecting to MCP servers: ${serverNames.join(", ")}…`)}\n`);
-					}
-				},
-				enableProjectConfig: settings.get("mcp.enableProjectConfig") ?? true,
-				// Always filter Exa - we have native integration
-				filterExa: true,
-				// Filter browser MCP servers when builtin browser tool is active
-				filterBrowser: settings.get("browser.enabled") ?? false,
-				cacheStorage: settings.getStorage(),
-				authStorage,
-			});
-			mcpManager = mcpResult.manager;
-			toolSession.mcpManager = mcpManager;
+			if (deferMCPDiscoveryForUI) {
+				const cacheStorage = settings.getStorage();
+				mcpManager = new MCPManager(cwd, cacheStorage ? new MCPToolCache(cacheStorage) : null);
+				mcpManager.setAuthStorage(authStorage);
+				toolSession.mcpManager = mcpManager;
 
-			if (settings.get("mcp.notifications")) {
-				mcpManager.setNotificationsEnabled(true);
-			}
-			// If we extracted Exa API keys from MCP configs and EXA_API_KEY isn't set, use the first one
-			if (mcpResult.exaApiKeys.length > 0 && !$env.EXA_API_KEY) {
-				Bun.env.EXA_API_KEY = mcpResult.exaApiKeys[0];
-			}
+				if (settings.get("mcp.notifications")) {
+					mcpManager.setNotificationsEnabled(true);
+				}
 
-			// Log MCP errors
-			for (const { path, error } of mcpResult.errors) {
-				logger.error("MCP tool load failed", { path, error });
-			}
+				const deferredMCPManager = mcpManager;
+				startDeferredMCPDiscovery = (liveSession, activation) => {
+					void (async () => {
+						try {
+							const mcpResult = await logger.time("discoverAndLoadMCPTools", () =>
+								deferredMCPManager.discoverAndConnect(mcpDiscoverOptions),
+							);
+							applyMCPEnvironment(mcpResult);
+							logMCPLoadErrors(mcpResult.errors);
+							await liveSession.refreshMCPTools(mcpResult.tools, {
+								activateAll: activation.activateAllMCPTools,
+							});
+							if (!activation.mcpDiscoveryEnabled && activation.explicitlyRequestedMCPToolNames.length > 0) {
+								await liveSession.setActiveToolsByName([
+									...liveSession.getActiveToolNames(),
+									...activation.explicitlyRequestedMCPToolNames,
+								]);
+							}
+						} catch (error) {
+							logger.error("MCP tool load failed", {
+								path: ".mcp.json",
+								error: error instanceof Error ? error.message : String(error),
+							});
+						}
+					})();
+				};
+			} else {
+				const mcpResult = await logger.time("discoverAndLoadMCPTools", discoverAndLoadMCPTools, cwd, {
+					...mcpDiscoverOptions,
+					cacheStorage: settings.getStorage(),
+					authStorage,
+				});
+				mcpManager = mcpResult.manager;
+				toolSession.mcpManager = mcpManager;
 
-			if (mcpResult.tools.length > 0) {
-				// MCP tools are LoadedCustomTool, extract the tool property
-				customTools.push(...mcpResult.tools.map(loaded => loaded.tool));
+				if (settings.get("mcp.notifications")) {
+					mcpManager.setNotificationsEnabled(true);
+				}
+				applyMCPEnvironment(mcpResult);
+
+				// Log MCP errors
+				for (const { path, error } of mcpResult.errors) {
+					logger.error("MCP tool load failed", { path, error });
+				}
+
+				if (mcpResult.tools.length > 0) {
+					// MCP tools are LoadedCustomTool, extract the tool property
+					customTools.push(...mcpResult.tools.map(loaded => loaded.tool));
+				}
 			}
 		}
 		// Only top-level sessions own the global MCPManager. Subagents already
 		// receive the parent's manager via `options.mcpManager`, and reassigning
-		// the singleton to the same value is a no-op \u2014 keep the gate explicit
+		// the singleton to the same value is a no-op — keep the gate explicit
 		// to mirror the AsyncJobManager ownership rule.
 		if (mcpManager && !options.parentTaskPrefix) MCPManager.setInstance(mcpManager);
 
@@ -1748,6 +1869,14 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 		for (const tool of wrappedExtensionTools) {
 			toolRegistry.set(tool.name, tool);
 		}
+		if (deferMCPDiscoveryForUI && mcpManager) {
+			for (const name of collectPendingMCPToolNames(options.toolNames, existingSession.selectedMCPToolNames)) {
+				if (!toolRegistry.has(name)) {
+					toolRegistry.set(name, createPendingMCPTool(name));
+				}
+			}
+		}
+
 		// Wrap every tool with `ExtensionToolWrapper` so the per-tool approval gate runs on every
 		// call site, regardless of whether any user extensions are loaded. See the runner-construction
 		// comment above for the safety invariant this enforces.
@@ -1838,8 +1967,13 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 			const memoryBackend = await resolveMemoryBackend(settings);
 			const memoryInstructions = await memoryBackend.buildDeveloperInstructions(agentDir, settings, session);
 
-			// Build combined append prompt: memory instructions + MCP server instructions
-			const serverInstructions = mcpManager?.getServerInstructions();
+			// Build combined append prompt: memory instructions + MCP server instructions.
+			// Interactive deferred MCP intentionally omits MCP instructions permanently;
+			// prompt rebuilds caused by later tool refreshes keep cache keys stable by
+			// updating tool metadata only, not server-provided prompt text.
+			const serverInstructions = includeMCPServerInstructionsInPrompt
+				? mcpManager?.getServerInstructions()
+				: undefined;
 			let appendPrompt: string | undefined = memoryInstructions ?? undefined;
 			if (serverInstructions && serverInstructions.size > 0) {
 				const parts: string[] = [];
@@ -2202,20 +2336,23 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 			rebuildSystemPrompt,
 			reloadSshTool,
 			requestedToolNames: requestedToolNameSet,
-			getMcpServerInstructions: mcpManager
-				? () => {
-						const raw = mcpManager.getServerInstructions();
-						if (!raw || raw.size === 0) return raw;
-						const out = new Map<string, string>();
-						for (const [name, text] of raw) {
-							out.set(
-								name,
-								text.length > MAX_MCP_INSTRUCTIONS_LENGTH ? text.slice(0, MAX_MCP_INSTRUCTIONS_LENGTH) : text,
-							);
+			getMcpServerInstructions:
+				includeMCPServerInstructionsInPrompt && mcpManager
+					? () => {
+							const raw = mcpManager.getServerInstructions();
+							if (!raw || raw.size === 0) return raw;
+							const out = new Map<string, string>();
+							for (const [name, text] of raw) {
+								out.set(
+									name,
+									text.length > MAX_MCP_INSTRUCTIONS_LENGTH
+										? text.slice(0, MAX_MCP_INSTRUCTIONS_LENGTH)
+										: text,
+								);
+							}
+							return out;
 						}
-						return out;
-					}
-				: undefined,
+					: undefined,
 			mcpDiscoveryEnabled,
 			initialSelectedMCPToolNames,
 			defaultSelectedMCPToolNames,
@@ -2349,7 +2486,26 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 		// Skip when reusing a parent's manager — the parent owns the callbacks.
 		if (mcpManager && !options.mcpManager) {
 			mcpManager.setOnToolsChanged(tools => {
-				void session.refreshMCPTools(tools);
+				void (async () => {
+					try {
+						await session.refreshMCPTools(
+							tools,
+							deferMCPDiscoveryForUI && !mcpDiscoveryEnabled && options.toolNames === undefined
+								? { activateAll: true }
+								: undefined,
+						);
+						if (deferMCPDiscoveryForUI && !mcpDiscoveryEnabled && explicitlyRequestedMCPToolNames.length > 0) {
+							await session.setActiveToolsByName([
+								...session.getActiveToolNames(),
+								...explicitlyRequestedMCPToolNames,
+							]);
+						}
+					} catch (error) {
+						logger.warn("MCP tool refresh failed", {
+							error: error instanceof Error ? error.message : String(error),
+						});
+					}
+				})();
 			});
 			// Wire prompt refresh → rebuild MCP prompt slash commands
 			mcpManager.setOnPromptsChanged(serverName => {
@@ -2382,6 +2538,12 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 			});
 		}
 
+		startDeferredMCPDiscovery?.(session, {
+			mcpDiscoveryEnabled,
+			explicitlyRequestedMCPToolNames,
+			activateAllMCPTools: !mcpDiscoveryEnabled && options.toolNames === undefined,
+		});
+
 		return {
 			session,
 			extensionsResult,
diff --git a/packages/coding-agent/src/tools/read.ts b/packages/coding-agent/src/tools/read.ts
index 9b39f4b97..300078240 100644
--- a/packages/coding-agent/src/tools/read.ts
+++ b/packages/coding-agent/src/tools/read.ts
@@ -8,6 +8,7 @@ import { glob, type SummaryResult, summarizeCode } from "@oh-my-pi/pi-natives";
 import type { Component } from "@oh-my-pi/pi-tui";
 import { Text } from "@oh-my-pi/pi-tui";
 import { getRemoteDir, logger, prompt, readImageMetadata, untilAborted } from "@oh-my-pi/pi-utils";
+import { LRUCache } from "lru-cache/raw";
 import * as z from "zod/v4";
 import {
 	canonicalSnapshotKey,
@@ -99,6 +100,24 @@ import {
 import { ToolAbortError, ToolError, throwIfAborted } from "./tool-errors";
 import { toolResult } from "./tool-result";
 
+// Per-session memo for tree-sitter summaries. `summarizeCode` is a pure function
+// of (code, path, fold settings) but costs ~12-18ms for a ~1500-line file, and a
+// repeat summary read of the same unchanged file re-parses from scratch. Key on
+// the content hash of the freshly-read bytes (+ path + fold settings): the file
+// is still read fresh on every call, so a hit only reuses the deterministic
+// parse — there is no staleness window and no stat guard is needed. Bounded LRU,
+// aged out with the session via WeakMap.
+const SUMMARY_CACHE_MAX = 48;
+const summaryParseCaches = new WeakMap<object, LRUCache<string, SummaryResult>>();
+function getSummaryParseCache(session: object): LRUCache<string, SummaryResult> {
+	let cache = summaryParseCaches.get(session);
+	if (!cache) {
+		cache = new LRUCache<string, SummaryResult>({ max: SUMMARY_CACHE_MAX });
+		summaryParseCaches.set(session, cache);
+	}
+	return cache;
+}
+
 // Document types converted to markdown via markit.
 const CONVERTIBLE_EXTENSIONS = new Set([".pdf", ".doc", ".docx", ".ppt", ".pptx", ".xls", ".xlsx", ".rtf", ".epub"]);
 
@@ -1512,14 +1531,23 @@ export class ReadTool implements AgentTool<typeof readSchema, ReadToolDetails> {
 			if (lineCount > MAX_SUMMARY_LINES) return null;
 			if (lineCount < this.session.settings.get("read.summarize.minTotalLines")) return null;
 
+			const minBodyLines = this.session.settings.get("read.summarize.minBodyLines");
+			const minCommentLines = this.session.settings.get("read.summarize.minCommentLines");
+			const unfoldUntilLines = this.session.settings.get("read.summarize.unfoldUntil");
+			const unfoldLimitLines = this.session.settings.get("read.summarize.unfoldLimit");
+			const cache = getSummaryParseCache(this.session);
+			const cacheKey = `${absolutePath}\0${Bun.hash(code)}\0${minBodyLines},${minCommentLines},${unfoldUntilLines},${unfoldLimitLines}`;
+			const memoized = cache.get(cacheKey);
+			if (memoized) return memoized;
 			const result = summarizeCode({
 				code,
 				path: absolutePath,
-				minBodyLines: this.session.settings.get("read.summarize.minBodyLines"),
-				minCommentLines: this.session.settings.get("read.summarize.minCommentLines"),
-				unfoldUntilLines: this.session.settings.get("read.summarize.unfoldUntil"),
-				unfoldLimitLines: this.session.settings.get("read.summarize.unfoldLimit"),
+				minBodyLines,
+				minCommentLines,
+				unfoldUntilLines,
+				unfoldLimitLines,
 			});
+			cache.set(cacheKey, result);
 			return result;
 		} catch {
 			return null;
diff --git a/packages/coding-agent/test/modes/components/assistant-message-streaming-fastpath.test.ts b/packages/coding-agent/test/modes/components/assistant-message-streaming-fastpath.test.ts
new file mode 100644
index 000000000..6a640b2ea
--- /dev/null
+++ b/packages/coding-agent/test/modes/components/assistant-message-streaming-fastpath.test.ts
@@ -0,0 +1,141 @@
+import { afterEach, beforeAll, beforeEach, describe, expect, it } from "bun:test";
+import type { AssistantMessage } from "@oh-my-pi/pi-ai";
+import { resetSettingsForTest, Settings } from "@oh-my-pi/pi-coding-agent/config/settings";
+import { AssistantMessageComponent } from "@oh-my-pi/pi-coding-agent/modes/components/assistant-message";
+import { initTheme } from "@oh-my-pi/pi-coding-agent/modes/theme/theme";
+
+const W = 100;
+
+function msg(content: AssistantMessage["content"], extra: Partial<AssistantMessage> = {}): AssistantMessage {
+	return {
+		role: "assistant",
+		content,
+		api: "anthropic-messages",
+		provider: "anthropic",
+		model: "m",
+		usage: {
+			input: 0,
+			output: 0,
+			cacheRead: 0,
+			cacheWrite: 0,
+			totalTokens: 0,
+			cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
+		},
+		stopReason: "stop",
+		timestamp: 0,
+		...extra,
+	};
+}
+
+/** Render `m` on a brand-new component, which always takes the teardown path. */
+function teardownRender(m: AssistantMessage): string {
+	const fresh = new AssistantMessageComponent();
+	fresh.updateContent(m);
+	return fresh.render(W).join("\n");
+}
+
+beforeAll(async () => {
+	await initTheme(false);
+});
+
+beforeEach(async () => {
+	resetSettingsForTest();
+	await Settings.init({ inMemory: true });
+});
+
+afterEach(() => {
+	resetSettingsForTest();
+});
+
+// Contract: the streaming fast path (a component reused across updateContent
+// calls, which reuses Markdown children via setText) MUST render byte-identical
+// output to the teardown path (a fresh component that rebuilds every child) for
+// the same message — at every step. If they ever diverge, the optimization
+// silently corrupts the transcript.
+describe("AssistantMessageComponent streaming fast path", () => {
+	it("matches teardown output across a growing thinking + text stream", () => {
+		const reused = new AssistantMessageComponent();
+		const thinking = "Reasoning about the **problem** with `code` and a list:\n- a\n- b";
+		const steps = [
+			"He",
+			"Hello, ",
+			"Hello, world.",
+			"Hello, world.\n\n## Heading\n\nSome `inline` and **bold** text.",
+			"Hello, world.\n\n## Heading\n\nSome `inline` and **bold** text.\n\n```ts\nconst x = 1;\n```",
+		];
+		for (const text of steps) {
+			const m = msg([
+				{ type: "thinking", thinking },
+				{ type: "text", text },
+			]);
+			reused.updateContent(m);
+			expect(reused.render(W).join("\n")).toBe(teardownRender(m));
+		}
+	});
+
+	it("matches teardown for a single growing text block", () => {
+		const reused = new AssistantMessageComponent();
+		let text = "";
+		for (const chunk of ["The ", "quick ", "brown ", "**fox** ", "jumps."]) {
+			text += chunk;
+			const m = msg([{ type: "text", text }]);
+			reused.updateContent(m);
+			expect(reused.render(W).join("\n")).toBe(teardownRender(m));
+		}
+	});
+
+	// Regression: #fastPathItems are keyed by raw content index, but a
+	// `redactedThinking` block is not rendered. If one appears mid-stream it
+	// shifts the indices of the visible blocks; the shape key must reflect that
+	// (or the fast path must fail closed) so children are not mis-targeted.
+	it("matches teardown when a redactedThinking block shifts indices mid-stream", () => {
+		const reused = new AssistantMessageComponent();
+		const a = msg([
+			{ type: "thinking", thinking: "step one details here" },
+			{ type: "text", text: "answer one" },
+		]);
+		reused.updateContent(a);
+		expect(reused.render(W).join("\n")).toBe(teardownRender(a));
+
+		// A redactedThinking block appears at index 0, pushing thinking->1, text->2.
+		const b = msg([
+			{ type: "redactedThinking", data: "opaque-blob" },
+			{ type: "thinking", thinking: "step two with more detail" },
+			{ type: "text", text: "answer two is longer now" },
+		]);
+		reused.updateContent(b);
+		expect(reused.render(W).join("\n")).toBe(teardownRender(b));
+	});
+
+	it("matches teardown when an error trailer appears after streamed text", () => {
+		const reused = new AssistantMessageComponent();
+		const ok = msg([{ type: "text", text: "partial answer in progress" }]);
+		reused.updateContent(ok);
+		expect(reused.render(W).join("\n")).toBe(teardownRender(ok));
+
+		const errored = msg([{ type: "text", text: "partial answer in progress" }], {
+			stopReason: "error",
+			errorMessage: "upstream 502",
+		});
+		reused.updateContent(errored);
+		expect(reused.render(W).join("\n")).toBe(teardownRender(errored));
+	});
+
+	it("matches teardown when a block visibility toggles (empty -> non-empty)", () => {
+		const reused = new AssistantMessageComponent();
+		// First an empty trailing text block (not rendered), then it gains content.
+		const empty = msg([
+			{ type: "thinking", thinking: "thinking out loud" },
+			{ type: "text", text: "" },
+		]);
+		reused.updateContent(empty);
+		expect(reused.render(W).join("\n")).toBe(teardownRender(empty));
+
+		const filled = msg([
+			{ type: "thinking", thinking: "thinking out loud" },
+			{ type: "text", text: "now there is an answer" },
+		]);
+		reused.updateContent(filled);
+		expect(reused.render(W).join("\n")).toBe(teardownRender(filled));
+	});
+});
diff --git a/packages/coding-agent/test/sdk-mcp-defer.test.ts b/packages/coding-agent/test/sdk-mcp-defer.test.ts
new file mode 100644
index 000000000..2f99b89bd
--- /dev/null
+++ b/packages/coding-agent/test/sdk-mcp-defer.test.ts
@@ -0,0 +1,95 @@
+import { afterAll, afterEach, beforeAll, beforeEach, describe, expect, it } from "bun:test";
+import * as fs from "node:fs";
+import * as os from "node:os";
+import * as path from "node:path";
+import { AuthStorage, getBundledModel } from "@oh-my-pi/pi-ai";
+import { ModelRegistry } from "@oh-my-pi/pi-coding-agent/config/model-registry";
+import { Settings } from "@oh-my-pi/pi-coding-agent/config/settings";
+import { createAgentSession } from "@oh-my-pi/pi-coding-agent/sdk";
+import { SessionManager } from "@oh-my-pi/pi-coding-agent/session/session-manager";
+import { Snowflake } from "@oh-my-pi/pi-utils";
+
+// Contract for B1 (interactive MCP deferral): when `hasUI` is true, MCP
+// discovery is deferred off the first-paint path, so an explicitly requested
+// MCP tool (e.g. via `--tools`) whose server has not yet connected MUST still
+// be a *known* tool — registered as a deterministic "still connecting"
+// placeholder — rather than vanishing and surfacing as "unknown tool" if the
+// model calls it before the background connection completes. With `hasUI`
+// false there is no deferral, so an MCP tool name with no real backing is not
+// registered at all (the non-UI paths keep the blocking discover path).
+describe("createAgentSession MCP deferral (B1)", () => {
+	let registryDir: string;
+	let tempDir: string;
+	let authStorage: AuthStorage;
+	let modelRegistry: ModelRegistry;
+
+	const PENDING_MCP_TOOL = "mcp__pending_connectingtool";
+
+	const baseOptions = () => ({
+		cwd: tempDir,
+		agentDir: tempDir,
+		modelRegistry,
+		sessionManager: SessionManager.inMemory(),
+		settings: Settings.isolated({}),
+		model: getBundledModel("openai", "gpt-4o-mini"),
+		disableExtensionDiscovery: true,
+		skills: [],
+		contextFiles: [],
+		promptTemplates: [],
+		slashCommands: [],
+		enableLsp: false,
+		skipPythonPreflight: true,
+		// No .mcp.json in tempDir, so no real MCP server can ever back this name.
+		enableMCP: true,
+		toolNames: ["read", PENDING_MCP_TOOL],
+	});
+
+	beforeAll(async () => {
+		registryDir = path.join(os.tmpdir(), `pi-sdk-mcp-defer-registry-${Snowflake.next()}`);
+		fs.mkdirSync(registryDir, { recursive: true });
+		authStorage = await AuthStorage.create(path.join(registryDir, "auth.db"));
+		modelRegistry = new ModelRegistry(authStorage);
+	});
+
+	afterAll(() => {
+		authStorage.close();
+		if (registryDir && fs.existsSync(registryDir)) {
+			fs.rmSync(registryDir, { recursive: true, force: true });
+		}
+	});
+
+	beforeEach(() => {
+		tempDir = path.join(os.tmpdir(), `pi-sdk-mcp-defer-${Snowflake.next()}`);
+		fs.mkdirSync(tempDir, { recursive: true });
+	});
+
+	afterEach(() => {
+		if (tempDir && fs.existsSync(tempDir)) {
+			fs.rmSync(tempDir, { recursive: true, force: true });
+		}
+	});
+
+	it("registers a pending placeholder for an explicit MCP tool when hasUI defers discovery", async () => {
+		const { session } = await createAgentSession({ ...baseOptions(), hasUI: true });
+		try {
+			// The explicitly requested MCP tool is a known, resolvable tool even
+			// though no server has connected — deterministic, not "unknown tool".
+			expect(session.getActiveToolNames()).toContain(PENDING_MCP_TOOL);
+		} finally {
+			await session.dispose();
+		}
+	});
+
+	it("does not fabricate the MCP tool in non-UI mode (no deferral, no backing server)", async () => {
+		const { session } = await createAgentSession({ ...baseOptions(), hasUI: false });
+		try {
+			// Without deferral there is no placeholder; the name has no real
+			// server backing, so it is simply not a registered tool.
+			expect(session.getActiveToolNames()).not.toContain(PENDING_MCP_TOOL);
+			// A normal builtin is unaffected.
+			expect(session.getActiveToolNames()).toContain("read");
+		} finally {
+			await session.dispose();
+		}
+	});
+});