gsd-pi 2.80.0-dev.e146beb20 → 2.80.0-dev.e51d2c88c

package/packages/pi-coding-agent/src/core/agent-session-abort-order.test.ts

@@ -120,6 +120,7 @@ describe("#4243 — abort() must run before _disconnectFromAgent()", () => {
 
 	it("newSession() invokes abort() before _disconnectFromAgent()", async () => {
 		const session = await createSession();
+		(session as any).agent.state.isStreaming = true;
 		const order = recordCallOrder(session as any, ["abort", "_disconnectFromAgent"]);
 
 		const ok = await session.newSession();
@@ -138,6 +139,40 @@ describe("#4243 — abort() must run before _disconnectFromAgent()", () => {
 		);
 	});
 
+	it("newSession() waits instead of aborting when the prior turn is idle but not settled", async () => {
+		const session = await createSession();
+		const order: string[] = [];
+		let releaseIdle!: () => void;
+		const idle = new Promise<void>((resolve) => {
+			releaseIdle = resolve;
+		});
+
+		(session as any).agent.state.isStreaming = false;
+		(session as any).agent.waitForIdle = () => {
+			order.push("waitForIdle");
+			return idle;
+		};
+		(session as any).abort = async () => {
+			order.push("abort");
+		};
+		const originalDisconnect = (session as any)._disconnectFromAgent.bind(session);
+		(session as any)._disconnectFromAgent = () => {
+			order.push("_disconnectFromAgent");
+			originalDisconnect();
+		};
+
+		const pendingNewSession = session.newSession();
+		await Promise.resolve();
+		assert.deepEqual(order, ["waitForIdle"]);
+		assert.equal(order.includes("abort"), false);
+
+		releaseIdle();
+		const ok = await pendingNewSession;
+		assert.equal(ok, true);
+		assert.deepEqual(order, ["waitForIdle", "_disconnectFromAgent"]);
+		assert.equal(order.includes("abort"), false);
+	});
+
 	it("newSession() waits instead of aborting while agent_end processing is still streaming", async () => {
 		const session = await createSession();
 		const order: string[] = [];
@@ -432,6 +467,7 @@ describe("#4243 — abort() must run before _disconnectFromAgent()", () => {
 		const sessionFile = session.sessionFile;
 		assert.ok(typeof sessionFile === "string" && sessionFile.length > 0, "need a session file to switch to");
 
+		(session as any).agent.state.isStreaming = true;
 		const order = recordCallOrder(session as any, ["abort", "_disconnectFromAgent"]);
 
 		const ok = await session.switchSession(sessionFile);

package/packages/pi-coding-agent/src/core/agent-session.ts

@@ -1626,6 +1626,11 @@ export class AgentSession {
 		// message_end/agent_end events fire while listeners are still connected.
 		// During agent_end handling the turn is already ending; aborting there can
 		// convert a successful auto-mode handoff into an aborted provider message.
+		if (!this.agent.state.isStreaming) {
+			this._retryHandler.abortRetry();
+			await this.agent.waitForIdle();
+			return;
+		}
 		await this.abort();
 	}
 
@@ -2211,6 +2216,9 @@ export class AgentSession {
 					})();
 				},
 				getSystemPrompt: () => this.systemPrompt,
+				setCompactionThresholdOverride: (percent) => {
+					this.settingsManager.setCompactionThresholdOverride(percent);
+				},
 			},
 		);
 	}

package/packages/pi-coding-agent/src/core/chat-controller-ordering.test.ts

@@ -1080,7 +1080,9 @@ test("chat-controller rolls up low-signal direct tool execution events on agent_
 		} as any);
 	}
 
-	assert.equal(host.chatContainer.children.length, 3, "direct tool events render as individual rows while running");
+	assert.equal(host.chatContainer.children.length, 1, "direct tool events roll up as they finish");
+	assert.equal(host.chatContainer.children[0]?.constructor?.name, "ToolPhaseSummaryComponent");
+	assert.match(host.chatContainer.children[0].render(120).join("\n"), /Setup \/ shell 3 actions/);
 
 	await handleAgentEvent(host, { type: "agent_end" } as any);
 

package/packages/pi-coding-agent/src/core/compaction/compaction.ts

@@ -86,6 +86,13 @@ export interface CompactionSettings {
 	enabled: boolean;
 	reserveTokens: number;
 	keepRecentTokens: number;
+	/**
+	 * Optional percent-of-context-window threshold (0 < value < 1). When set,
+	 * `shouldCompact()` fires once `contextTokens > contextWindow * thresholdPercent`,
+	 * overriding the absolute `reserveTokens` calculation. Lets host integrations
+	 * (e.g. GSD) express compaction policy as a fraction independent of model size.
+	 */
+	thresholdPercent?: number;
 }
 
 export const DEFAULT_COMPACTION_SETTINGS: CompactionSettings = {
@@ -185,9 +192,20 @@ export function estimateContextTokens(messages: AgentMessage[]): ContextUsageEst
 
 /**
  * Check if compaction should trigger based on context usage.
+ *
+ * When `thresholdPercent` is set (and within (0, 1)), it overrides the absolute
+ * `reserveTokens` calculation: compaction fires at `contextWindow * thresholdPercent`.
+ * Otherwise the legacy `contextWindow - reserveTokens` headroom is used.
  */
 export function shouldCompact(contextTokens: number, contextWindow: number, settings: CompactionSettings): boolean {
 	if (!settings.enabled) return false;
+	if (
+		settings.thresholdPercent !== undefined &&
+		settings.thresholdPercent > 0 &&
+		settings.thresholdPercent < 1
+	) {
+		return contextTokens > contextWindow * settings.thresholdPercent;
+	}
 	return contextTokens > contextWindow - settings.reserveTokens;
 }
 

package/packages/pi-coding-agent/src/core/compaction-threshold.test.ts

@@ -0,0 +1,121 @@
+// pi-coding-agent / Regression tests for compaction threshold percent (#5475)
+
+import assert from "node:assert/strict";
+import { describe, it } from "node:test";
+
+import { shouldCompact, type CompactionSettings } from "./compaction/compaction.js";
+import { SettingsManager } from "./settings-manager.js";
+
+const REGISTRY_DEFAULTS: CompactionSettings = {
+	enabled: true,
+	reserveTokens: 16_384,
+	keepRecentTokens: 20_000,
+};
+
+describe("shouldCompact — thresholdPercent (#5475)", () => {
+	it("uses absolute reserveTokens when thresholdPercent is unset (legacy behavior)", () => {
+		// 200K window, 16384 reserve → fires at 183_617 tokens
+		assert.equal(shouldCompact(183_616, 200_000, REGISTRY_DEFAULTS), false);
+		assert.equal(shouldCompact(183_617, 200_000, REGISTRY_DEFAULTS), true);
+	});
+
+	it("uses thresholdPercent when set, ignoring reserveTokens", () => {
+		const settings: CompactionSettings = { ...REGISTRY_DEFAULTS, thresholdPercent: 0.7 };
+		// 200K * 0.7 = 140_000 → fires above that
+		assert.equal(shouldCompact(140_000, 200_000, settings), false);
+		assert.equal(shouldCompact(140_001, 200_000, settings), true);
+		// reserveTokens-based math would have said false at 183_616 — the percent override changes that
+		assert.equal(shouldCompact(150_000, 200_000, settings), true);
+	});
+
+	it("falls back to reserveTokens when thresholdPercent is out of range", () => {
+		// Defense in depth: reject 0, 1, negative, NaN, Infinity
+		for (const bad of [0, 1, -0.1, 1.5, Number.NaN, Number.POSITIVE_INFINITY]) {
+			const settings: CompactionSettings = { ...REGISTRY_DEFAULTS, thresholdPercent: bad };
+			assert.equal(
+				shouldCompact(183_616, 200_000, settings),
+				false,
+				`bad=${bad} should fall back to reserveTokens math`,
+			);
+			assert.equal(shouldCompact(183_617, 200_000, settings), true, `bad=${bad}`);
+		}
+	});
+
+	it("respects enabled=false regardless of thresholdPercent", () => {
+		const settings: CompactionSettings = {
+			...REGISTRY_DEFAULTS,
+			enabled: false,
+			thresholdPercent: 0.5,
+		};
+		assert.equal(shouldCompact(199_999, 200_000, settings), false);
+	});
+
+	it("scales with contextWindow — same percent, different windows", () => {
+		const settings: CompactionSettings = { ...REGISTRY_DEFAULTS, thresholdPercent: 0.8 };
+		// 100K window: fires above 80_000
+		assert.equal(shouldCompact(80_000, 100_000, settings), false);
+		assert.equal(shouldCompact(80_001, 100_000, settings), true);
+		// 1M window: fires above 800_000
+		assert.equal(shouldCompact(800_000, 1_000_000, settings), false);
+		assert.equal(shouldCompact(800_001, 1_000_000, settings), true);
+	});
+});
+
+describe("SettingsManager — compaction threshold override (#5475)", () => {
+	it("getCompactionThresholdPercent returns undefined by default", () => {
+		const sm = SettingsManager.inMemory({});
+		assert.equal(sm.getCompactionThresholdPercent(), undefined);
+		assert.equal(sm.getCompactionSettings().thresholdPercent, undefined);
+	});
+
+	it("setCompactionThresholdOverride applies in-memory and is exposed via getCompactionSettings", () => {
+		const sm = SettingsManager.inMemory({});
+		sm.setCompactionThresholdOverride(0.7);
+		assert.equal(sm.getCompactionThresholdPercent(), 0.7);
+		assert.equal(sm.getCompactionSettings().thresholdPercent, 0.7);
+	});
+
+	it("setCompactionThresholdOverride(undefined) clears a prior override", () => {
+		const sm = SettingsManager.inMemory({});
+		sm.setCompactionThresholdOverride(0.7);
+		sm.setCompactionThresholdOverride(undefined);
+		assert.equal(sm.getCompactionThresholdPercent(), undefined);
+		assert.equal(sm.getCompactionSettings().thresholdPercent, undefined);
+	});
+
+	it("setCompactionThresholdOverride preserves other compaction fields (enabled, reserveTokens)", () => {
+		const sm = SettingsManager.inMemory({
+			compaction: { enabled: true, reserveTokens: 30_000, keepRecentTokens: 25_000 },
+		});
+		sm.setCompactionThresholdOverride(0.6);
+		const settings = sm.getCompactionSettings();
+		assert.equal(settings.enabled, true);
+		assert.equal(settings.reserveTokens, 30_000);
+		assert.equal(settings.keepRecentTokens, 25_000);
+		assert.equal(settings.thresholdPercent, 0.6);
+	});
+
+	it("setCompactionThresholdOverride works when no compaction config exists yet", () => {
+		const sm = SettingsManager.inMemory({});
+		sm.setCompactionThresholdOverride(0.85);
+		assert.equal(sm.getCompactionThresholdPercent(), 0.85);
+		// Other compaction fields fall back to their defaults
+		const settings = sm.getCompactionSettings();
+		assert.equal(settings.enabled, true);
+		assert.equal(typeof settings.reserveTokens, "number");
+		assert.equal(typeof settings.keepRecentTokens, "number");
+	});
+});
+
+describe("end-to-end — getCompactionSettings + shouldCompact (#5475)", () => {
+	it("70% threshold on a 200K window fires at the documented bug-report value (140_001 not 183_617)", () => {
+		const sm = SettingsManager.inMemory({});
+		sm.setCompactionThresholdOverride(0.7);
+		const settings = sm.getCompactionSettings();
+
+		assert.equal(shouldCompact(140_000, 200_000, settings), false);
+		assert.equal(shouldCompact(140_001, 200_000, settings), true);
+		// Pre-fix behavior would have required 183_617 — verify we no longer wait that long
+		assert.equal(shouldCompact(150_000, 200_000, settings), true);
+	});
+});

package/packages/pi-coding-agent/src/core/extensions/runner.test.ts

@@ -141,6 +141,7 @@ describe("ExtensionRunner.emitToolCall", () => {
 			getContextUsage: () => undefined,
 			compact: () => {},
 			getSystemPrompt: () => "",
+			setCompactionThresholdOverride: () => {},
 		});
 
 		const errors: any[] = [];
@@ -220,6 +221,7 @@ describe("ExtensionRunner.createContext", () => {
 			getContextUsage: () => undefined,
 			compact: () => {},
 			getSystemPrompt: () => "",
+			setCompactionThresholdOverride: () => {},
 		});
 
 		const errors: any[] = [];

package/packages/pi-coding-agent/src/core/extensions/runner.ts

@@ -235,6 +235,7 @@ export class ExtensionRunner {
 	private getContextUsageFn: () => ContextUsage | undefined = () => undefined;
 	private compactFn: (options?: CompactOptions) => void = () => {};
 	private getSystemPromptFn: () => string = () => "";
+	private setCompactionThresholdOverrideFn: (percent: number | undefined) => void = () => {};
 	private newSessionHandler: NewSessionHandler = async () => {
 		throw new Error("Command context not yet bound: newSession is unavailable during early lifecycle");
 	};
@@ -428,6 +429,7 @@ export class ExtensionRunner {
 		this.getContextUsageFn = contextActions.getContextUsage;
 		this.compactFn = contextActions.compact;
 		this.getSystemPromptFn = contextActions.getSystemPrompt;
+		this.setCompactionThresholdOverrideFn = contextActions.setCompactionThresholdOverride;
 
 		// Flush provider registrations queued during extension loading
 		for (const { name, config } of this.runtime.pendingProviderRegistrations) {
@@ -714,6 +716,7 @@ export class ExtensionRunner {
 			getContextUsage: () => this.getContextUsageFn(),
 			compact: (options) => this.compactFn(options),
 			getSystemPrompt: () => this.getSystemPromptFn(),
+			setCompactionThresholdOverride: (percent) => this.setCompactionThresholdOverrideFn(percent),
 		};
 	}
 

package/packages/pi-coding-agent/src/core/extensions/types.ts

@@ -289,6 +289,12 @@ export interface ExtensionContext {
 	compact(options?: CompactOptions): void;
 	/** Get the current effective system prompt. */
 	getSystemPrompt(): string;
+	/**
+	 * Set or clear an in-memory compaction threshold-percent override (0 < value < 1).
+	 * Pass `undefined` to clear. The override is not persisted; host integrations
+	 * are expected to re-apply on each session_start.
+	 */
+	setCompactionThresholdOverride(percent: number | undefined): void;
 }
 
 /**
@@ -1741,6 +1747,7 @@ export interface ExtensionContextActions {
 	getContextUsage: () => ContextUsage | undefined;
 	compact: (options?: CompactOptions) => void;
 	getSystemPrompt: () => string;
+	setCompactionThresholdOverride: (percent: number | undefined) => void;
 }
 
 /**

package/packages/pi-coding-agent/src/core/settings-manager.ts

@@ -15,6 +15,13 @@ export interface CompactionSettings {
 	enabled?: boolean; // default: true
 	reserveTokens?: number; // default: 16384
 	keepRecentTokens?: number; // default: 20000
+	/**
+	 * Optional percent-of-context-window trigger (0 < value < 1). When set,
+	 * compaction fires at `contextWindow * thresholdPercent` and overrides
+	 * `reserveTokens`. Typically set as a runtime override by host integrations
+	 * (see `setCompactionThresholdOverride`) and not persisted by users directly.
+	 */
+	thresholdPercent?: number;
 }
 
 export interface BranchSummarySettings {
@@ -812,11 +819,42 @@ export class SettingsManager {
 		return this.settings.compaction?.keepRecentTokens ?? COMPACTION_KEEP_RECENT_TOKENS;
 	}
 
-	getCompactionSettings(): { enabled: boolean; reserveTokens: number; keepRecentTokens: number } {
+	getCompactionThresholdPercent(): number | undefined {
+		return this.settings.compaction?.thresholdPercent;
+	}
+
+	/**
+	 * Set or clear an in-memory compaction threshold-percent override.
+	 *
+	 * Applied to `this.settings` only; never persisted to disk. Pass `undefined`
+	 * to clear a previously set override (necessary for idempotent re-sync from
+	 * host integrations whose preference may have been removed).
+	 *
+	 * Direct mutation is used instead of `applyOverrides()` because deep-merge
+	 * semantics skip `undefined` values, which would prevent clearing.
+	 */
+	setCompactionThresholdOverride(percent: number | undefined): void {
+		if (!this.settings.compaction) {
+			this.settings.compaction = {};
+		}
+		if (percent === undefined) {
+			delete this.settings.compaction.thresholdPercent;
+		} else {
+			this.settings.compaction.thresholdPercent = percent;
+		}
+	}
+
+	getCompactionSettings(): {
+		enabled: boolean;
+		reserveTokens: number;
+		keepRecentTokens: number;
+		thresholdPercent?: number;
+	} {
 		return {
 			enabled: this.getCompactionEnabled(),
 			reserveTokens: this.getCompactionReserveTokens(),
 			keepRecentTokens: this.getCompactionKeepRecentTokens(),
+			thresholdPercent: this.getCompactionThresholdPercent(),
 		};
 	}
 

package/packages/pi-coding-agent/src/modes/interactive/components/tool-execution.ts

@@ -1287,6 +1287,10 @@ export class ToolPhaseSummaryComponent extends Container {
 		super();
 	}
 
+	getPhases(): ToolExecutionPhase[] {
+		return this.phases.map((phase) => ({ ...phase }));
+	}
+
 	override render(width: number): string[] {
 		const frameWidth = Math.max(20, width);
 		const rows = this.phases.map((phase) => {

package/packages/pi-coding-agent/src/modes/interactive/controllers/chat-controller.test.ts

@@ -1,5 +1,7 @@
 import assert from "node:assert/strict";
 import test from "node:test";
+import { Container } from "@gsd/pi-tui";
+import stripAnsi from "strip-ansi";
 
 import { findLatestPinnableText, handleAgentEvent } from "./chat-controller.js";
 import { initTheme } from "../theme/theme.js";
@@ -111,3 +113,57 @@ test("handleAgentEvent: agent_start clears stale adaptive blocking error", async
 	assert.equal(cleared, true);
 	assert.equal(requestedRender, true);
 });
+
+test("handleAgentEvent: standalone completed tool events roll up incrementally", async () => {
+	initTheme("dark", false);
+	const chatContainer = new Container();
+	let renderCount = 0;
+	const host = {
+		isInitialized: true,
+		footer: { invalidate() {} },
+		settingsManager: {
+			getTimestampFormat() {
+				return "date-time-iso";
+			},
+			getShowImages() {
+				return false;
+			},
+		},
+		getRegisteredToolDefinition() {
+			return undefined;
+		},
+		chatContainer,
+		pendingTools: new Map(),
+		ui: {
+			requestRender() {
+				renderCount++;
+			},
+		},
+	} as any;
+
+	for (const [toolCallId, toolName] of [
+		["read-1", "read"],
+		["read-2", "read"],
+		["edit-1", "edit"],
+	] as const) {
+		await handleAgentEvent(host, {
+			type: "tool_execution_start",
+			toolCallId,
+			toolName,
+			args: { path: `/tmp/${toolCallId}.txt` },
+		} as any);
+		await handleAgentEvent(host, {
+			type: "tool_execution_end",
+			toolCallId,
+			toolName,
+			result: { content: [], isError: false },
+			isError: false,
+		} as any);
+	}
+
+	const rendered = stripAnsi(chatContainer.render(100).join("\n"));
+	assert.match(rendered, /Context reads 2 actions\s+success · \d+(ms|s)/);
+	assert.match(rendered, /File changes 1 action\s+success · \d+(ms|s)/);
+	assert.doesNotMatch(rendered, /\bread\s+success ·/);
+	assert.ok(renderCount > 0);
+});

package/packages/pi-coding-agent/src/modes/interactive/controllers/chat-controller.ts

@@ -28,7 +28,8 @@ type RenderedSegment =
 		contentType: "text" | "thinking";
 		component: AssistantMessageComponent;
 	}
-	| { kind: "tool"; contentIndex: number; component: ToolExecutionComponent };
+	| { kind: "tool"; contentIndex: number; component: ToolExecutionComponent }
+	| { kind: "tool-summary"; component: ToolPhaseSummaryComponent; phases: ToolExecutionPhase[] };
 
 let renderedSegments: RenderedSegment[] = [];
 // When providers reuse one assistant lifecycle across internal sub-turns,
@@ -124,17 +125,25 @@ function replaceCompactToolRowsWithPhaseSummary(
 ): void {
 	let changed = false;
 	const nextRenderedSegments: RenderedSegment[] = [];
-	let rollupRun: Array<{ seg: Extract<RenderedSegment, { kind: "tool" }>; phase: ToolExecutionPhase }> = [];
+	let rollupRun: Array<{
+		seg: Extract<RenderedSegment, { kind: "tool" | "tool-summary" }>;
+		phases: ToolExecutionPhase[];
+	}> = [];
 
 	const flushRollupRun = () => {
-		if (rollupRun.length < 2) {
+		const actionCount = rollupRun.reduce(
+			(total, item) => total + item.phases.reduce((sum, phase) => sum + phase.count, 0),
+			0,
+		);
+		if (actionCount < 2) {
 			nextRenderedSegments.push(...rollupRun.map((item) => item.seg));
 			rollupRun = [];
 			return;
 		}
 
 		const firstIndex = Math.max(0, host.chatContainer.children.indexOf(rollupRun[0].seg.component));
-		const summary = new ToolPhaseSummaryComponent(mergeToolPhases(rollupRun.map((item) => item.phase)));
+		const phases = mergeToolPhases(rollupRun.flatMap((item) => item.phases));
+		const summary = new ToolPhaseSummaryComponent(phases);
 
 		for (const { seg } of rollupRun) {
 			host.chatContainer.removeChild(seg.component);
@@ -149,13 +158,18 @@ function replaceCompactToolRowsWithPhaseSummary(
 		}
 
 		changed = true;
+		nextRenderedSegments.push({ kind: "tool-summary", component: summary, phases });
 		rollupRun = [];
 	};
 
 	for (const seg of renderedSegments) {
 		const phase = seg.kind === "tool" ? seg.component.getRollupPhase() : null;
 		if (seg.kind === "tool" && phase) {
-			rollupRun.push({ seg, phase });
+			rollupRun.push({ seg, phases: [phase] });
+			continue;
+		}
+		if (seg.kind === "tool-summary") {
+			rollupRun.push({ seg, phases: seg.component.getPhases() });
 			continue;
 		}
 
@@ -409,6 +423,7 @@ export async function handleAgentEvent(host: InteractiveModeStateHost & {
 							details: externalToolResult.details,
 							isError: externalToolResult.isError,
 						});
+						replaceCompactToolRowsWithPhaseSummary(host);
 					}
 				}
 
@@ -843,7 +858,6 @@ export async function handleAgentEvent(host: InteractiveModeStateHost & {
 					host.streamingComponent.setShowMetadata(true);
 					host.streamingComponent.updateContent(host.streamingMessage);
 				}
-				replaceCompactToolRowsWithPhaseSummary(host);
 
 				if (host.streamingMessage.stopReason === "aborted" || host.streamingMessage.stopReason === "error") {
 					if (!errorMessage) {
@@ -862,6 +876,7 @@ export async function handleAgentEvent(host: InteractiveModeStateHost & {
 					for (const [, component] of host.pendingTools.entries()) {
 						component.setArgsComplete();
 					}
+					replaceCompactToolRowsWithPhaseSummary(host);
 				}
 				host.streamingComponent = undefined;
 				host.streamingMessage = undefined;
@@ -912,7 +927,7 @@ export async function handleAgentEvent(host: InteractiveModeStateHost & {
 			const component = host.pendingTools.get(event.toolCallId);
 			if (component) {
 				component.updateResult({ ...event.result, isError: event.isError });
-				host.pendingTools.delete(event.toolCallId);
+				replaceCompactToolRowsWithPhaseSummary(host);
 				host.ui.requestRender();
 			}
 			break;

package/packages/pi-coding-agent/src/modes/interactive/interactive-mode.ts

@@ -1378,6 +1378,9 @@ export class InteractiveMode {
 				})();
 			},
 			getSystemPrompt: () => this.session.systemPrompt,
+			setCompactionThresholdOverride: (percent) => {
+				this.session.settingsManager.setCompactionThresholdOverride(percent);
+			},
 		});
 
 		// Set up the extension shortcut handler on the default editor