fix(coding-agent): reopen oversized sessions and wrap tool output

packages/coding-agent/src/core/changes.md

@@ -1,5 +1,34 @@
 # changes
 
+## 2026-09-05 - Skip live-history budget on same-model resume
+
+### What changed
+
+- `createAgentSession` treats a restored session as same-persisted-model when existing
+  messages are present and the saved provider/modelId equals the resolved model
+  provider/id. That path admits with `liveContextTokens = 0` so only the fixed prompt,
+  tools, and reserves are validated. Different explicit targets and fallback replacements
+  still sum the restored transcript. Resume still omits speculation lead
+  (`includeSpeculationLead: false`, `admission: "resume"`). `_setModel` and prompt
+  pre-compaction guards are unchanged.
+
+### Why
+
+- Same-model resume of an already-saved oversized history was rejected before the TUI
+  opened, so the user could not compact or continue. The history was produced by that
+  model; startup should not re-charge it as a downswitch. Switching to a different model
+  must still refuse a transcript that cannot fit.
+
+### Why an extension could not handle it
+
+- `createAgentSession` throws `ModelUsabilityBudgetError` before extensions are wired;
+  no extension hook can intercept that constructor-time reject.
+
+### Expected merge conflict zones
+
+- LOW: the startup `assertModelUsable` call in `sdk.ts`;
+  `test/suite/model-usability-budget.test.ts`.
+
 ## 2026-09-05 - Require explicit fallback chains
 
 ### What changed

packages/coding-agent/src/core/extensions/builtin/terminal/changes.md

@@ -1,5 +1,23 @@
 # terminal builtin extension — fork surface
 
+## bash_output completed/expanded TUI wrap (2026-09-05)
+
+### What changed
+
+- `tools/render.ts`: completed and expanded `bash_output` results wrap with `wrapTextWithAnsi(..., Math.max(1, width))` instead of raw `split("\n")`. Partial preview still uses `truncateToVisualLines`.
+
+### Why
+
+- A completed or expanded result could emit 102/180-column lines into a 98-column Box, overflowing the TUI width contract. Wrapping is display-only; it is not the process-exit root cause.
+
+### Why an extension could not handle it
+
+- The overflow is in this builtin's own `renderResult` component.
+
+### Expected merge conflict zones
+
+- LOW: `tools/render.ts` completed/expanded `render()` branch and `test/suite/terminal-bash-output-width.test.ts`.
+
 ## `persistent` reads as the standing-watch switch (2026-09-04)
 
 ### What changed

packages/coding-agent/src/core/extensions/builtin/terminal/tools/render.ts

@@ -1,4 +1,4 @@
-import { Text } from "@earendil-works/pi-tui";
+import { Text, wrapTextWithAnsi } from "@earendil-works/pi-tui";
 import { truncateToVisualLines } from "../../../../../modes/interactive/components/visual-truncate.ts";
 import type { AgentToolResult, ToolDefinition } from "../../../types.ts";
 import type { BashOutputInput, bashOutputSchema } from "./bash-output.ts";
@@ -24,7 +24,7 @@ class BashOutputResultComponent {
 	}
 
 	render(width: number): string[] {
-		if (!this.#isPartial || this.#expanded) return this.#text.split("\n");
+		if (!this.#isPartial || this.#expanded) return wrapTextWithAnsi(this.#text, Math.max(1, width));
 		return truncateToVisualLines(this.#text, OUTPUT_PREVIEW_LINES, Math.max(1, width)).visualLines.map((line) =>
 			line.trimEnd(),
 		);

packages/coding-agent/src/core/sdk.ts

@@ -516,9 +516,16 @@ export async function createAgentSession(options: CreateAgentSessionOptions = {}
 		sessionStartEvent,
 		autoTitleSessions: options.autoTitleSessions,
 	});
-	const liveContextTokens = hasExistingSession
-		? existingSession.messages.reduce((total, message) => total + estimateTokens(message), 0)
-		: 0;
+	const samePersistedModel =
+		hasExistingSession &&
+		existingSession.model != null &&
+		model != null &&
+		existingSession.model.provider === model.provider &&
+		existingSession.model.modelId === model.id;
+	const liveContextTokens =
+		hasExistingSession && !samePersistedModel
+			? existingSession.messages.reduce((total, message) => total + estimateTokens(message), 0)
+			: 0;
 	session.assertModelUsable(
 		undefined,
 		liveContextTokens,

packages/coding-agent/test/suite/model-usability-budget.test.ts

@@ -367,6 +367,157 @@ describe("model usability budget", () => {
 		fitting.session.dispose();
 	});
 
+	it("resumes a same-saved-model transcript that exceeds the startup budget", async () => {
+		// given
+		const harness = await createHarness({
+			models: [{ id: "startup", contextWindow: 100_000, maxTokens: 4_000 }],
+		});
+		harnesses.push(harness);
+		const model = harness.getModel();
+		const sessionManager = SessionManager.inMemory(harness.tempDir);
+		sessionManager.appendModelChange(model.provider, model.id);
+		sessionManager.appendMessage({
+			role: "user",
+			content: [{ type: "text", text: "restored transcript ".repeat(200_000) }],
+			timestamp: Date.now(),
+		});
+
+		// when
+		const resumed = await createAgentSession({
+			cwd: harness.tempDir,
+			agentDir: join(harness.tempDir, "same-saved-agent"),
+			model,
+			sessionManager,
+		});
+
+		// then
+		expect(resumed.session.agent.state.messages).toEqual(sessionManager.buildSessionContext().messages);
+		expect(resumed.session.model?.id).toBe(model.id);
+		resumed.session.dispose();
+	});
+
+	it("rejects a same-saved-model resume when the fixed prompt and reserves cannot fit", async () => {
+		// given
+		const harness = await createHarness({
+			models: [{ id: "startup", contextWindow: 16_000, maxTokens: 4_000 }],
+		});
+		harnesses.push(harness);
+		const model = harness.getModel();
+		const sessionManager = SessionManager.inMemory(harness.tempDir);
+		sessionManager.appendModelChange(model.provider, model.id);
+		sessionManager.appendMessage({
+			role: "user",
+			content: [{ type: "text", text: "restored transcript ".repeat(200_000) }],
+			timestamp: Date.now(),
+		});
+
+		// when
+		const error = await createAgentSession({
+			cwd: harness.tempDir,
+			agentDir: join(harness.tempDir, "same-saved-fixed-base"),
+			model,
+			sessionManager,
+		}).then(
+			() => undefined,
+			(reason: unknown) => reason,
+		);
+
+		// then
+		expect(error).toBeInstanceOf(ModelUsabilityBudgetError);
+		if (!(error instanceof ModelUsabilityBudgetError)) throw new Error("expected fixed-base resume rejection");
+		expect(error.projection).toMatchObject({
+			model: `${model.provider}/${model.id}`,
+			usable: false,
+			liveContextTokens: 0,
+			admission: "resume",
+			speculationLeadTokens: 0,
+			contextWindow: 16_000,
+		});
+	});
+
+	it("rejects a resumed transcript when the explicit target differs from the saved model", async () => {
+		// given
+		const harness = await createHarness({
+			models: [
+				{ id: "saved", contextWindow: 100_000, maxTokens: 4_000 },
+				{ id: "other", contextWindow: 100_000, maxTokens: 4_000 },
+			],
+		});
+		harnesses.push(harness);
+		const saved = harness.getModel("saved");
+		const other = harness.getModel("other");
+		if (!saved) throw new Error("missing saved model fixture");
+		if (!other) throw new Error("missing other model fixture");
+		const sessionManager = SessionManager.inMemory(harness.tempDir);
+		sessionManager.appendModelChange(saved.provider, saved.id);
+		sessionManager.appendMessage({
+			role: "user",
+			content: [{ type: "text", text: "restored transcript ".repeat(200_000) }],
+			timestamp: Date.now(),
+		});
+
+		// when
+		const error = await createAgentSession({
+			cwd: harness.tempDir,
+			agentDir: join(harness.tempDir, "different-target-agent"),
+			model: other,
+			sessionManager,
+		}).then(
+			() => undefined,
+			(reason: unknown) => reason,
+		);
+
+		// then
+		expect(error).toBeInstanceOf(ModelUsabilityBudgetError);
+		if (!(error instanceof ModelUsabilityBudgetError)) throw new Error("expected different-target resume rejection");
+		expect(error.projection).toMatchObject({
+			model: `${other.provider}/${other.id}`,
+			usable: false,
+			admission: "resume",
+			speculationLeadTokens: 0,
+		});
+		expect(error.projection.liveContextTokens).toBeGreaterThan(0);
+	});
+
+	it("rejects a resumed transcript when fallback replaces an unrestorable saved model", async () => {
+		// given
+		const harness = await createHarness({
+			models: [{ id: "startup", contextWindow: 100_000, maxTokens: 4_000 }],
+		});
+		harnesses.push(harness);
+		const fallback = harness.getModel();
+		const sessionManager = SessionManager.inMemory(harness.tempDir);
+		sessionManager.appendModelChange("gone", "gone");
+		sessionManager.appendMessage({
+			role: "user",
+			content: [{ type: "text", text: "restored transcript ".repeat(200_000) }],
+			timestamp: Date.now(),
+		});
+
+		// when
+		const error = await createAgentSession({
+			cwd: harness.tempDir,
+			agentDir: join(harness.tempDir, "fallback-target-agent"),
+			sessionManager,
+			modelRegistry: harness.modelRegistry,
+			authStorage: harness.authStorage,
+		}).then(
+			() => undefined,
+			(reason: unknown) => reason,
+		);
+
+		// then
+		expect(error).toBeInstanceOf(ModelUsabilityBudgetError);
+		if (!(error instanceof ModelUsabilityBudgetError)) throw new Error("expected fallback resume rejection");
+		expect(error.projection).toMatchObject({
+			model: `${fallback.provider}/${fallback.id}`,
+			usable: false,
+			admission: "resume",
+			speculationLeadTokens: 0,
+		});
+		expect(error.projection.liveContextTokens).toBeGreaterThan(0);
+	});
+
 	it("accepts the exact minimum and rejects one token below it", async () => {
 		// given
 		const harness = await createHarness();

packages/coding-agent/test/suite/terminal-bash-output-width.test.ts

@@ -0,0 +1,142 @@
+import { Box, stripTerminalSequences, visibleWidth } from "@earendil-works/pi-tui";
+import { beforeAll, describe, expect, it } from "vitest";
+import { renderBashOutputResult } from "../../src/core/extensions/builtin/terminal/tools/render.ts";
+import type { ToolRenderContext, ToolRenderResultOptions } from "../../src/core/extensions/types.ts";
+import { initTheme, theme } from "../../src/modes/interactive/theme/theme.ts";
+
+type BashOutputResult = Parameters<typeof renderBashOutputResult>[0];
+
+const BOX_WIDTH = 98;
+const LINE_102 = "c".repeat(102);
+const LINE_180 = "p".repeat(180);
+const LONG_PATH =
+	"make[1]: Entering directory '/workspace/project/models/share/Library/src/vendor/third_party/fortran/modules'";
+const LONG_COMMAND =
+	"nvfortran -module /workspace/project/models/share/include -c -Mpreprocess -r8 -O3 -Munroll=c:1 -Mlre -Mvect=simd -Mcache_align -r8 ModHdf5Utils.f90";
+const ANSI_LONG_LINE = `\x1b[31m${LONG_PATH}\x1b[39m`;
+const WIDE_UNICODE_LINE = `wide ${"全".repeat(60)} end`;
+
+function bashOutputText(): string {
+	return [
+		"status: exited_2 exit_code: 2",
+		LONG_PATH,
+		LONG_COMMAND,
+		LINE_102,
+		LINE_180,
+		ANSI_LONG_LINE,
+		WIDE_UNICODE_LINE,
+	].join("\n");
+}
+
+function renderContext(options: ToolRenderResultOptions): ToolRenderContext {
+	return {
+		args: { bash_id: "bash_1" },
+		toolCallId: "bash-output-width",
+		invalidate: () => {},
+		lastComponent: undefined,
+		state: undefined,
+		cwd: "/tmp/workspace",
+		executionStarted: true,
+		argsComplete: true,
+		isPartial: options.isPartial,
+		expanded: options.expanded,
+		showImages: false,
+		isError: false,
+	};
+}
+
+function createResult(options: ToolRenderResultOptions, text = bashOutputText()) {
+	const result: BashOutputResult = { content: [{ type: "text", text }], details: undefined };
+	return renderBashOutputResult(result, options, theme, renderContext(options));
+}
+
+function renderBox(width: number, options: ToolRenderResultOptions, text = bashOutputText()): string[] {
+	const box = new Box(1, 1, (value: string) => `\x1b[48;5;22m${value}\x1b[49m`);
+	box.addChild(createResult(options, text));
+	return box.render(width);
+}
+
+function nonWhitespaceCodepoints(text: string): string {
+	return [...text].filter((ch) => !/\s/u.test(ch)).join("");
+}
+
+function expectLinesFit(lines: string[], width: number): void {
+	for (const line of lines) {
+		expect(visibleWidth(line)).toBeLessThanOrEqual(width);
+	}
+}
+
+function expectPreserved(lines: string[], original: string): void {
+	expect(nonWhitespaceCodepoints(lines.map(stripTerminalSequences).join(""))).toBe(
+		nonWhitespaceCodepoints(stripTerminalSequences(original)),
+	);
+}
+
+describe("bash_output completed/expanded width wrap", () => {
+	beforeAll(() => {
+		initTheme("dark");
+	});
+
+	it("fits a 98-column Box for complete output with ANSI, CJK, and long command/path lines", () => {
+		expect(visibleWidth(LINE_102)).toBeGreaterThan(BOX_WIDTH - 2);
+		expect(visibleWidth(LINE_180)).toBeGreaterThan(BOX_WIDTH - 2);
+		expect(visibleWidth(LONG_PATH)).toBeGreaterThan(BOX_WIDTH - 2);
+		expect(visibleWidth(LONG_COMMAND)).toBeGreaterThan(BOX_WIDTH - 2);
+		expect(visibleWidth(WIDE_UNICODE_LINE)).toBeGreaterThan(BOX_WIDTH - 2);
+
+		const text = bashOutputText();
+		const lines = renderBox(BOX_WIDTH, { expanded: false, isPartial: false }, text);
+		expect(lines.length).toBeGreaterThan(0);
+		expectLinesFit(lines, BOX_WIDTH);
+		expectPreserved(lines, text);
+		expect(lines.join("")).toContain("\x1b[31m");
+	});
+
+	it("fits a 98-column Box for expanded partial output", () => {
+		const text = bashOutputText();
+		const lines = renderBox(BOX_WIDTH, { expanded: true, isPartial: true }, text);
+		expectLinesFit(lines, BOX_WIDTH);
+		expectPreserved(lines, text);
+	});
+
+	it("wraps complete component output across resize widths without dropping tokens", () => {
+		const text = bashOutputText();
+		const component = createResult({ expanded: false, isPartial: false }, text);
+		for (const width of [24, 40, 80, BOX_WIDTH, 160]) {
+			const lines = component.render(width);
+			expectLinesFit(lines, width);
+			expectPreserved(lines, text);
+			expect(lines.join("")).toContain("\x1b[31m");
+		}
+	});
+
+	it("wraps expanded ANSI and CJK lines without dropping tokens", () => {
+		const text = bashOutputText();
+		const component = createResult({ expanded: true, isPartial: true }, text);
+		for (const width of [20, 41, 97, 98]) {
+			const lines = component.render(width);
+			expectLinesFit(lines, width);
+			expectPreserved(lines, text);
+			const payload = lines.join("");
+			expect(payload).toContain("\x1b[31m");
+			expect(payload).toContain("全");
+		}
+	});
+
+	it("keeps partial preview truncated while preserving remaining non-whitespace characters", () => {
+		const text = Array.from(
+			{ length: 40 },
+			(_, index) => `preview-line-${String(index).padStart(2, "0")}-${"x".repeat(30)}`,
+		).join("\n");
+		const preview = createResult({ expanded: false, isPartial: true }, text).render(40);
+		const expanded = createResult({ expanded: true, isPartial: true }, text).render(40);
+		expectLinesFit(preview, 40);
+		expectLinesFit(expanded, 40);
+		expect(preview.length).toBeLessThan(expanded.length);
+		const previewChars = nonWhitespaceCodepoints(preview.map(stripTerminalSequences).join(""));
+		const originalChars = nonWhitespaceCodepoints(stripTerminalSequences(text));
+		expect(originalChars.endsWith(previewChars)).toBe(true);
+		expect(previewChars.length).toBeGreaterThan(0);
+		expectPreserved(expanded, text);
+	});
+});