co-mono/packages/coding-agent/test/compaction.test.ts

import type { AgentMessage } from "@mariozechner/pi-agent-core";
import type { AssistantMessage, Usage } from "@mariozechner/pi-ai";
import { getModel } from "@mariozechner/pi-ai";
import { readFileSync } from "fs";
import { join } from "path";
import { beforeEach, describe, expect, it } from "vitest";
import {
	type CompactionSettings,
	calculateContextTokens,
	compact,
	DEFAULT_COMPACTION_SETTINGS,
	findCutPoint,
	getLastAssistantUsage,
	shouldCompact,
} from "../src/core/compaction.js";
import {
	buildSessionContext,
	type CompactionEntry,
	type ModelChangeEntry,
	migrateSessionEntries,
	parseSessionEntries,
	type SessionEntry,
	type SessionMessageEntry,
	type ThinkingLevelChangeEntry,
} from "../src/core/session-manager.js";

// ============================================================================
// Test fixtures
// ============================================================================

function loadLargeSessionEntries(): SessionEntry[] {
	const sessionPath = join(__dirname, "fixtures/large-session.jsonl");
	const content = readFileSync(sessionPath, "utf-8");
	const entries = parseSessionEntries(content);
	migrateSessionEntries(entries); // Add id/parentId for v1 fixtures
	return entries.filter((e): e is SessionEntry => e.type !== "session");
}

function createMockUsage(input: number, output: number, cacheRead = 0, cacheWrite = 0): Usage {
	return {
		input,
		output,
		cacheRead,
		cacheWrite,
		totalTokens: input + output + cacheRead + cacheWrite,
		cost: { input: 0, output: 0, cacheRead: 0, cacheWrite: 0, total: 0 },
	};
}

function createUserMessage(text: string): AgentMessage {
	return { role: "user", content: text, timestamp: Date.now() };
}

function createAssistantMessage(text: string, usage?: Usage): AssistantMessage {
	return {
		role: "assistant",
		content: [{ type: "text", text }],
		usage: usage || createMockUsage(100, 50),
		stopReason: "stop",
		timestamp: Date.now(),
		api: "anthropic-messages",
		provider: "anthropic",
		model: "claude-sonnet-4-5",
	};
}

let entryCounter = 0;
let lastId: string | null = null;

function resetEntryCounter() {
	entryCounter = 0;
	lastId = null;
}

// Reset counter before each test to get predictable IDs
beforeEach(() => {
	resetEntryCounter();
});

function createMessageEntry(message: AgentMessage): SessionMessageEntry {
	const id = `test-id-${entryCounter++}`;
	const entry: SessionMessageEntry = {
		type: "message",
		id,
		parentId: lastId,
		timestamp: new Date().toISOString(),
		message,
	};
	lastId = id;
	return entry;
}

function createCompactionEntry(summary: string, firstKeptEntryId: string): CompactionEntry {
	const id = `test-id-${entryCounter++}`;
	const entry: CompactionEntry = {
		type: "compaction",
		id,
		parentId: lastId,
		timestamp: new Date().toISOString(),
		summary,
		firstKeptEntryId,
		tokensBefore: 10000,
	};
	lastId = id;
	return entry;
}

function createModelChangeEntry(provider: string, modelId: string): ModelChangeEntry {
	const id = `test-id-${entryCounter++}`;
	const entry: ModelChangeEntry = {
		type: "model_change",
		id,
		parentId: lastId,
		timestamp: new Date().toISOString(),
		provider,
		modelId,
	};
	lastId = id;
	return entry;
}

function createThinkingLevelEntry(thinkingLevel: string): ThinkingLevelChangeEntry {
	const id = `test-id-${entryCounter++}`;
	const entry: ThinkingLevelChangeEntry = {
		type: "thinking_level_change",
		id,
		parentId: lastId,
		timestamp: new Date().toISOString(),
		thinkingLevel,
	};
	lastId = id;
	return entry;
}

// ============================================================================
// Unit tests
// ============================================================================

describe("Token calculation", () => {
	it("should calculate total context tokens from usage", () => {
		const usage = createMockUsage(1000, 500, 200, 100);
		expect(calculateContextTokens(usage)).toBe(1800);
	});

	it("should handle zero values", () => {
		const usage = createMockUsage(0, 0, 0, 0);
		expect(calculateContextTokens(usage)).toBe(0);
	});
});

describe("getLastAssistantUsage", () => {
	it("should find the last non-aborted assistant message usage", () => {
		const entries: SessionEntry[] = [
			createMessageEntry(createUserMessage("Hello")),
			createMessageEntry(createAssistantMessage("Hi", createMockUsage(100, 50))),
			createMessageEntry(createUserMessage("How are you?")),
			createMessageEntry(createAssistantMessage("Good", createMockUsage(200, 100))),
		];

		const usage = getLastAssistantUsage(entries);
		expect(usage).not.toBeNull();
		expect(usage!.input).toBe(200);
	});

	it("should skip aborted messages", () => {
		const abortedMsg: AssistantMessage = {
			...createAssistantMessage("Aborted", createMockUsage(300, 150)),
			stopReason: "aborted",
		};

		const entries: SessionEntry[] = [
			createMessageEntry(createUserMessage("Hello")),
			createMessageEntry(createAssistantMessage("Hi", createMockUsage(100, 50))),
			createMessageEntry(createUserMessage("How are you?")),
			createMessageEntry(abortedMsg),
		];

		const usage = getLastAssistantUsage(entries);
		expect(usage).not.toBeNull();
		expect(usage!.input).toBe(100);
	});

	it("should return undefined if no assistant messages", () => {
		const entries: SessionEntry[] = [createMessageEntry(createUserMessage("Hello"))];
		expect(getLastAssistantUsage(entries)).toBeUndefined();
	});
});

describe("shouldCompact", () => {
	it("should return true when context exceeds threshold", () => {
		const settings: CompactionSettings = {
			enabled: true,
			reserveTokens: 10000,
			keepRecentTokens: 20000,
		};

		expect(shouldCompact(95000, 100000, settings)).toBe(true);
		expect(shouldCompact(89000, 100000, settings)).toBe(false);
	});

	it("should return false when disabled", () => {
		const settings: CompactionSettings = {
			enabled: false,
			reserveTokens: 10000,
			keepRecentTokens: 20000,
		};

		expect(shouldCompact(95000, 100000, settings)).toBe(false);
	});
});

describe("findCutPoint", () => {
	it("should find cut point based on actual token differences", () => {
		// Create entries with cumulative token counts
		const entries: SessionEntry[] = [];
		for (let i = 0; i < 10; i++) {
			entries.push(createMessageEntry(createUserMessage(`User ${i}`)));
			entries.push(
				createMessageEntry(createAssistantMessage(`Assistant ${i}`, createMockUsage(0, 100, (i + 1) * 1000, 0))),
			);
		}

		// 20 entries, last assistant has 10000 tokens
		// keepRecentTokens = 2500: keep entries where diff < 2500
		const result = findCutPoint(entries, 0, entries.length, 2500);

		// Should cut at a valid cut point (user or assistant message)
		expect(entries[result.firstKeptEntryIndex].type).toBe("message");
		const role = (entries[result.firstKeptEntryIndex] as SessionMessageEntry).message.role;
		expect(role === "user" || role === "assistant").toBe(true);
	});

	it("should return startIndex if no valid cut points in range", () => {
		const entries: SessionEntry[] = [createMessageEntry(createAssistantMessage("a"))];
		const result = findCutPoint(entries, 0, entries.length, 1000);
		expect(result.firstKeptEntryIndex).toBe(0);
	});

	it("should keep everything if all messages fit within budget", () => {
		const entries: SessionEntry[] = [
			createMessageEntry(createUserMessage("1")),
			createMessageEntry(createAssistantMessage("a", createMockUsage(0, 50, 500, 0))),
			createMessageEntry(createUserMessage("2")),
			createMessageEntry(createAssistantMessage("b", createMockUsage(0, 50, 1000, 0))),
		];

		const result = findCutPoint(entries, 0, entries.length, 50000);
		expect(result.firstKeptEntryIndex).toBe(0);
	});

	it("should indicate split turn when cutting at assistant message", () => {
		// Create a scenario where we cut at an assistant message mid-turn
		const entries: SessionEntry[] = [
			createMessageEntry(createUserMessage("Turn 1")),
			createMessageEntry(createAssistantMessage("A1", createMockUsage(0, 100, 1000, 0))),
			createMessageEntry(createUserMessage("Turn 2")), // index 2
			createMessageEntry(createAssistantMessage("A2-1", createMockUsage(0, 100, 5000, 0))), // index 3
			createMessageEntry(createAssistantMessage("A2-2", createMockUsage(0, 100, 8000, 0))), // index 4
			createMessageEntry(createAssistantMessage("A2-3", createMockUsage(0, 100, 10000, 0))), // index 5
		];

		// With keepRecentTokens = 3000, should cut somewhere in Turn 2
		const result = findCutPoint(entries, 0, entries.length, 3000);

		// If cut at assistant message (not user), should indicate split turn
		const cutEntry = entries[result.firstKeptEntryIndex] as SessionMessageEntry;
		if (cutEntry.message.role === "assistant") {
			expect(result.isSplitTurn).toBe(true);
			expect(result.turnStartIndex).toBe(2); // Turn 2 starts at index 2
		}
	});
});

describe("buildSessionContext", () => {
	it("should load all messages when no compaction", () => {
		const entries: SessionEntry[] = [
			createMessageEntry(createUserMessage("1")),
			createMessageEntry(createAssistantMessage("a")),
			createMessageEntry(createUserMessage("2")),
			createMessageEntry(createAssistantMessage("b")),
		];

		const loaded = buildSessionContext(entries);
		expect(loaded.messages.length).toBe(4);
		expect(loaded.thinkingLevel).toBe("off");
		expect(loaded.model).toEqual({ provider: "anthropic", modelId: "claude-sonnet-4-5" });
	});

	it("should handle single compaction", () => {
		// IDs: u1=test-id-0, a1=test-id-1, u2=test-id-2, a2=test-id-3, compaction=test-id-4, u3=test-id-5, a3=test-id-6
		const u1 = createMessageEntry(createUserMessage("1"));
		const a1 = createMessageEntry(createAssistantMessage("a"));
		const u2 = createMessageEntry(createUserMessage("2"));
		const a2 = createMessageEntry(createAssistantMessage("b"));
		const compaction = createCompactionEntry("Summary of 1,a,2,b", u2.id); // keep from u2 onwards
		const u3 = createMessageEntry(createUserMessage("3"));
		const a3 = createMessageEntry(createAssistantMessage("c"));

		const entries: SessionEntry[] = [u1, a1, u2, a2, compaction, u3, a3];

		const loaded = buildSessionContext(entries);
		// summary + kept (u2, a2) + after (u3, a3) = 5
		expect(loaded.messages.length).toBe(5);
		expect(loaded.messages[0].role).toBe("compactionSummary");
		expect((loaded.messages[0] as any).summary).toContain("Summary of 1,a,2,b");
	});

	it("should handle multiple compactions (only latest matters)", () => {
		// First batch
		const u1 = createMessageEntry(createUserMessage("1"));
		const a1 = createMessageEntry(createAssistantMessage("a"));
		const compact1 = createCompactionEntry("First summary", u1.id);
		// Second batch
		const u2 = createMessageEntry(createUserMessage("2"));
		const b = createMessageEntry(createAssistantMessage("b"));
		const u3 = createMessageEntry(createUserMessage("3"));
		const c = createMessageEntry(createAssistantMessage("c"));
		const compact2 = createCompactionEntry("Second summary", u3.id); // keep from u3 onwards
		// After second compaction
		const u4 = createMessageEntry(createUserMessage("4"));
		const d = createMessageEntry(createAssistantMessage("d"));

		const entries: SessionEntry[] = [u1, a1, compact1, u2, b, u3, c, compact2, u4, d];

		const loaded = buildSessionContext(entries);
		// summary + kept from u3 (u3, c) + after (u4, d) = 5
		expect(loaded.messages.length).toBe(5);
		expect((loaded.messages[0] as any).summary).toContain("Second summary");
	});

	it("should keep all messages when firstKeptEntryId is first entry", () => {
		const u1 = createMessageEntry(createUserMessage("1"));
		const a1 = createMessageEntry(createAssistantMessage("a"));
		const compact1 = createCompactionEntry("First summary", u1.id); // keep from first entry
		const u2 = createMessageEntry(createUserMessage("2"));
		const b = createMessageEntry(createAssistantMessage("b"));

		const entries: SessionEntry[] = [u1, a1, compact1, u2, b];

		const loaded = buildSessionContext(entries);
		// summary + all messages (u1, a1, u2, b) = 5
		expect(loaded.messages.length).toBe(5);
	});

	it("should track model and thinking level changes", () => {
		const entries: SessionEntry[] = [
			createMessageEntry(createUserMessage("1")),
			createModelChangeEntry("openai", "gpt-4"),
			createMessageEntry(createAssistantMessage("a")),
			createThinkingLevelEntry("high"),
		];

		const loaded = buildSessionContext(entries);
		// model_change is later overwritten by assistant message's model info
		expect(loaded.model).toEqual({ provider: "anthropic", modelId: "claude-sonnet-4-5" });
		expect(loaded.thinkingLevel).toBe("high");
	});
});

// ============================================================================
// Integration tests with real session data
// ============================================================================

describe("Large session fixture", () => {
	it("should parse the large session", () => {
		const entries = loadLargeSessionEntries();
		expect(entries.length).toBeGreaterThan(100);

		const messageCount = entries.filter((e) => e.type === "message").length;
		expect(messageCount).toBeGreaterThan(100);
	});

	it("should find cut point in large session", () => {
		const entries = loadLargeSessionEntries();
		const result = findCutPoint(entries, 0, entries.length, DEFAULT_COMPACTION_SETTINGS.keepRecentTokens);

		// Cut point should be at a message entry (user or assistant)
		expect(entries[result.firstKeptEntryIndex].type).toBe("message");
		const role = (entries[result.firstKeptEntryIndex] as SessionMessageEntry).message.role;
		expect(role === "user" || role === "assistant").toBe(true);
	});

	it("should load session correctly", () => {
		const entries = loadLargeSessionEntries();
		const loaded = buildSessionContext(entries);

		expect(loaded.messages.length).toBeGreaterThan(100);
		expect(loaded.model).not.toBeNull();
	});
});

// ============================================================================
// LLM integration tests (skipped without API key)
// ============================================================================

describe.skipIf(!process.env.ANTHROPIC_OAUTH_TOKEN)("LLM summarization", () => {
	it("should generate a compaction result for the large session", async () => {
		const entries = loadLargeSessionEntries();
		const model = getModel("anthropic", "claude-sonnet-4-5")!;

		const compactionResult = await compact(
			entries,
			model,
			DEFAULT_COMPACTION_SETTINGS,
			process.env.ANTHROPIC_OAUTH_TOKEN!,
		);

		expect(compactionResult.summary.length).toBeGreaterThan(100);
		expect(compactionResult.firstKeptEntryId).toBeTruthy();
		expect(compactionResult.tokensBefore).toBeGreaterThan(0);

		console.log("Summary length:", compactionResult.summary.length);
		console.log("First kept entry ID:", compactionResult.firstKeptEntryId);
		console.log("Tokens before:", compactionResult.tokensBefore);
		console.log("\n--- SUMMARY ---\n");
		console.log(compactionResult.summary);
	}, 60000);

	it("should produce valid session after compaction", async () => {
		const entries = loadLargeSessionEntries();
		const loaded = buildSessionContext(entries);
		const model = getModel("anthropic", "claude-sonnet-4-5")!;

		const compactionResult = await compact(
			entries,
			model,
			DEFAULT_COMPACTION_SETTINGS,
			process.env.ANTHROPIC_OAUTH_TOKEN!,
		);

		// Simulate appending compaction to entries by creating a proper entry
		const lastEntry = entries[entries.length - 1];
		const parentId = lastEntry.id;
		const compactionEntry: CompactionEntry = {
			type: "compaction",
			id: "compaction-test-id",
			parentId,
			timestamp: new Date().toISOString(),
			...compactionResult,
		};
		const newEntries = [...entries, compactionEntry];
		const reloaded = buildSessionContext(newEntries);

		// Should have summary + kept messages
		expect(reloaded.messages.length).toBeLessThan(loaded.messages.length);
		expect(reloaded.messages[0].role).toBe("compactionSummary");
		expect((reloaded.messages[0] as any).summary).toContain(compactionResult.summary);

		console.log("Original messages:", loaded.messages.length);
		console.log("After compaction:", reloaded.messages.length);
	}, 60000);
});