@prestyj/cli 5.13.0 → 5.16.0

package/dist/core/agent-session-compaction.test.js

@@ -1,1156 +0,0 @@
-import fs from "node:fs/promises";
-import os from "node:os";
-import path from "node:path";
-import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
-import { MODELS } from "./model-registry.js";
-import { estimateConversationTokens } from "./compaction/token-estimator.js";
-import { useFakeHome } from "../test-support/fake-home.js";
-const shouldCompactMock = vi.hoisted(() => vi.fn());
-const compactMock = vi.hoisted(() => vi.fn());
-const agentLoopMock = vi.hoisted(() => vi.fn());
-vi.mock("./compaction/compactor.js", async () => {
-    const actual = await vi.importActual("./compaction/compactor.js");
-    return {
-        ...actual,
-        shouldCompact: shouldCompactMock,
-        compact: compactMock,
-    };
-});
-vi.mock("@prestyj/agent", async () => {
-    const actual = await vi.importActual("@prestyj/agent");
-    return {
-        ...actual,
-        agentLoop: agentLoopMock,
-    };
-});
-vi.mock("./mcp/index.js", async () => {
-    const actual = await vi.importActual("./mcp/index.js");
-    return {
-        ...actual,
-        MCPClientManager: vi.fn(function MCPClientManagerMock() {
-            return {
-                connectAll: vi.fn(async () => []),
-                dispose: vi.fn(async () => { }),
-            };
-        }),
-    };
-});
-let restoreHome;
-let tmpHome;
-let tmpProject;
-async function writeJson(filePath, value) {
-    await fs.mkdir(path.dirname(filePath), { recursive: true });
-    await fs.writeFile(filePath, JSON.stringify(value, null, 2), "utf-8");
-}
-function compactionResult(messages, compacted = true, contextSelection) {
-    return {
-        messages,
-        result: {
-            compacted,
-            ...(compacted ? {} : { reason: "too_few_messages" }),
-            originalCount: compacted ? 6 : messages.length,
-            newCount: messages.length,
-            tokensBeforeEstimate: 180_000,
-            tokensAfterEstimate: compacted ? 2_000 : 180_000,
-            ...(contextSelection ? { contextSelection } : {}),
-        },
-    };
-}
-const providerModels = MODELS.filter((model, index, models) => models.findIndex((candidate) => candidate.provider === model.provider) === index).map((model) => ({ provider: model.provider, model: model.id }));
-beforeEach(async () => {
-    tmpHome = await fs.mkdtemp(path.join(os.tmpdir(), "agent-session-home-"));
-    tmpProject = await fs.mkdtemp(path.join(os.tmpdir(), "agent-session-project-"));
-    restoreHome = useFakeHome(tmpHome);
-    shouldCompactMock.mockReset();
-    compactMock.mockReset();
-    agentLoopMock.mockReset();
-    const authProviders = [
-        "anthropic",
-        "openai",
-        "sakana",
-        "xai",
-        "gemini",
-        "moonshot",
-        "glm",
-        "minimax",
-        "xiaomi",
-        "xiaomi-credits",
-        "deepseek",
-        "openrouter",
-    ];
-    await writeJson(path.join(tmpHome, ".ezcoder", "auth.json"), Object.fromEntries(authProviders.map((provider) => [
-        provider,
-        {
-            accessToken: `test-${provider}-token`,
-            refreshToken: `test-${provider}-refresh`,
-            expiresAt: Date.now() + 3_600_000,
-            ...(provider === "openai" ? { accountId: "chatgpt-account" } : {}),
-        },
-    ])));
-    await writeJson(path.join(tmpHome, ".ezcoder", "settings.json"), {
-        autoCompact: true,
-        compactThreshold: 0.1,
-    });
-});
-afterEach(async () => {
-    restoreHome?.();
-    await fs.rm(tmpHome, { recursive: true, force: true });
-    await fs.rm(tmpProject, { recursive: true, force: true });
-    vi.clearAllMocks();
-});
-describe("AgentSession worker auto-compaction", () => {
-    it("auto-compacts transient JSON-mode worker sessions before the agent loop runs", async () => {
-        const compactedMessages = [
-            { role: "system", content: "worker system prompt" },
-            { role: "user", content: "[compacted worker context]\n\nDo worker task" },
-        ];
-        shouldCompactMock.mockReturnValue(true);
-        compactMock.mockResolvedValue({
-            messages: compactedMessages,
-            result: {
-                compacted: true,
-                originalCount: 2,
-                newCount: 2,
-                tokensBeforeEstimate: 100_000,
-                tokensAfterEstimate: 1_000,
-            },
-        });
-        agentLoopMock.mockImplementation(async function* (messages) {
-            messages.push({ role: "assistant", content: "worker done" });
-            yield { type: "agent_done" };
-        });
-        const { AgentSession } = await import("./agent-session.js");
-        const session = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "worker system prompt",
-            transient: true,
-        });
-        await session.initialize();
-        await session.setApprovedPlan("/tmp/approved-plan.md");
-        await session.prompt("Do worker task");
-        await session.dispose();
-        expect(shouldCompactMock).toHaveBeenCalledWith(expect.arrayContaining([
-            expect.objectContaining({ role: "user", content: "Do worker task" }),
-        ]), expect.any(Number), 0.1, undefined);
-        expect(compactMock).toHaveBeenCalledWith(expect.arrayContaining([
-            expect.objectContaining({ role: "user", content: "Do worker task" }),
-        ]), expect.objectContaining({
-            provider: "anthropic",
-            model: "claude-test",
-            apiKey: "test-anthropic-token",
-            targetTokens: 20_000,
-            approvedPlanPath: "/tmp/approved-plan.md",
-        }));
-        expect(agentLoopMock).toHaveBeenCalledWith(expect.arrayContaining([
-            { role: "user", content: "[compacted worker context]\n\nDo worker task" },
-        ]), expect.objectContaining({ provider: "anthropic", model: "claude-test" }));
-    }, 15_000);
-});
-describe("AgentSession compaction persistence", () => {
-    it("persists the compaction summary as a new checkpoint in the session log", async () => {
-        // "Model-visible means logged" — the compacted history that replaces the
-        // conversation is exactly what the next model request will be derived
-        // from, so it must land in a NEW persisted session file (the old file
-        // stays as the parent for history reconstruction).
-        shouldCompactMock.mockReturnValue(false); // manual compact() below, no pre-run auto-compaction
-        compactMock.mockResolvedValue({
-            messages: [
-                { role: "system", content: "worker system prompt" },
-                {
-                    role: "user",
-                    content: "[session compacted] Summary of earlier work: the task was started.",
-                },
-            ],
-            result: {
-                compacted: true,
-                originalCount: 3,
-                newCount: 2,
-                tokensBeforeEstimate: 100_000,
-                tokensAfterEstimate: 500,
-            },
-        });
-        agentLoopMock.mockImplementation(async function* (messages) {
-            messages.push({ role: "assistant", content: "working on it" });
-            yield { type: "agent_done" };
-        });
-        const { AgentSession } = await import("./agent-session.js");
-        // Non-transient: compact() takes the lease branch that persists a
-        // compaction checkpoint (transient sessions compact in memory only).
-        const session = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "worker system prompt",
-        });
-        await session.initialize();
-        await session.prompt("Do the task.");
-        const preCompactPath = session.getState().sessionPath;
-        await session.compact();
-        expect(compactMock).toHaveBeenCalledTimes(1);
-        const postCompactPath = session.getState().sessionPath;
-        await session.dispose();
-        // The compacted branch is a NEW session file...
-        expect(postCompactPath).toBeTruthy();
-        expect(postCompactPath).not.toBe(preCompactPath);
-        // ...whose persisted rows are the replacement history, not the old one.
-        const rows = (await fs.readFile(postCompactPath, "utf-8"))
-            .split("\n")
-            .filter((line) => line.trim())
-            .map((line) => JSON.parse(line));
-        const texts = rows
-            .filter((row) => row.type === "message" && row.message)
-            .map((row) => typeof row.message.content === "string"
-            ? row.message.content
-            : JSON.stringify(row.message.content));
-        expect(texts.some((t) => t.includes("[session compacted] Summary of earlier work"))).toBe(true);
-        expect(texts.some((t) => t.includes("Do the task."))).toBe(false);
-    }, 15_000);
-});
-describe("AgentSession stale tool-output pruning", () => {
-    it("stubs superseded reads in-place during the pre-step transform", async () => {
-        shouldCompactMock.mockReturnValue(false);
-        let capturedMessages = [];
-        agentLoopMock.mockImplementation(async function* (messages, options) {
-            messages.push({
-                role: "assistant",
-                content: [
-                    { type: "tool_call", id: "old-read", name: "read", args: { file_path: "src/a.ts" } },
-                ],
-            }, {
-                role: "tool",
-                content: [{ type: "tool_result", toolCallId: "old-read", content: "x".repeat(150_000) }],
-            }, { role: "user", content: "turn 2" }, {
-                role: "assistant",
-                content: [
-                    { type: "tool_call", id: "new-read", name: "read", args: { file_path: "src/a.ts" } },
-                ],
-            }, {
-                role: "tool",
-                content: [{ type: "tool_result", toolCallId: "new-read", content: "fresh" }],
-            }, {
-                role: "assistant",
-                content: [{ type: "tool_call", id: "latest", name: "grep", args: { pattern: "x" } }],
-            }, {
-                role: "tool",
-                content: [{ type: "tool_result", toolCallId: "latest", content: "latest evidence" }],
-            }, { role: "user", content: "turn 3" });
-            capturedMessages = await options.transformContext(messages, { pendingMessages: [] });
-            yield { type: "agent_done" };
-        });
-        const { AgentSession } = await import("./agent-session.js");
-        const session = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            transient: true,
-        });
-        await session.initialize();
-        await session.prompt("prune me");
-        await session.dispose();
-        const results = capturedMessages
-            .filter((msg) => msg.role === "tool")
-            .flatMap((msg) => (Array.isArray(msg.content) ? msg.content : []));
-        const oldRead = results.find((r) => r.type === "tool_result" && r.toolCallId === "old-read");
-        const newRead = results.find((r) => r.type === "tool_result" && r.toolCallId === "new-read");
-        expect(oldRead?.content).toContain("superseded by a newer read");
-        expect(newRead?.content).toBe("fresh");
-        expect(compactMock).not.toHaveBeenCalled();
-    });
-    it("discards the turn's provider usage after a prune so freed tokens defer compaction", async () => {
-        shouldCompactMock.mockReturnValue(false);
-        agentLoopMock.mockImplementation(async function* (messages, options) {
-            messages.push({
-                role: "assistant",
-                content: [
-                    { type: "tool_call", id: "old-read", name: "read", args: { file_path: "src/a.ts" } },
-                ],
-            }, {
-                role: "tool",
-                content: [{ type: "tool_result", toolCallId: "old-read", content: "x".repeat(150_000) }],
-            }, { role: "user", content: "turn 2" }, {
-                role: "assistant",
-                content: [
-                    { type: "tool_call", id: "new-read", name: "read", args: { file_path: "src/a.ts" } },
-                ],
-            }, {
-                role: "tool",
-                content: [{ type: "tool_result", toolCallId: "new-read", content: "fresh" }],
-            }, {
-                role: "assistant",
-                content: [{ type: "tool_call", id: "latest", name: "grep", args: { pattern: "x" } }],
-            }, {
-                role: "tool",
-                content: [{ type: "tool_result", toolCallId: "latest", content: "latest evidence" }],
-            }, { role: "user", content: "turn 3" });
-            // Usage counted the soon-to-be-pruned 150k-char read. After the prune
-            // it must NOT reach shouldCompact — estimation on the pruned history
-            // takes over instead.
-            await options.transformContext(messages, {
-                usage: { inputTokens: 180_000, outputTokens: 500 },
-                pendingMessages: [],
-            });
-            yield { type: "agent_done" };
-        });
-        const { AgentSession } = await import("./agent-session.js");
-        const session = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            transient: true,
-        });
-        await session.initialize();
-        await session.prompt("prune usage fallback");
-        await session.dispose();
-        // The transform after the in-loop mutation is the last shouldCompact call;
-        // its actualTokens argument must be an estimate of the pruned history,
-        // far below the stale 180k usage figure.
-        const lastCall = shouldCompactMock.mock.calls.at(-1);
-        const actualTokens = lastCall[3];
-        expect(actualTokens).toBeLessThan(50_000);
-    });
-});
-describe("AgentSession overflow recovery", () => {
-    // Regression: the desktop app drives the loop through AgentSession (not the
-    // TUI's useContextCompaction hook), and AgentSession never passed
-    // `transformContext` — so a provider `request_too_large` / context-overflow
-    // (the loop's force-compact path) had NO auto recovery and surfaced straight
-    // to the user. This proves the loop's { force: true } call now triggers a
-    // real compaction and hands the shrunken history back for retry.
-    it("force-compacts and returns the shrunken history when the loop reports overflow", async () => {
-        // Pre-turn compaction disabled so we isolate the overflow force path.
-        shouldCompactMock.mockReturnValue(false);
-        const compactedMessages = [
-            { role: "system", content: "system prompt" },
-            { role: "user", content: "[compacted]" },
-        ];
-        compactMock.mockResolvedValue({
-            messages: compactedMessages,
-            result: {
-                compacted: true,
-                originalCount: 6,
-                newCount: 2,
-                tokensBeforeEstimate: 500_000,
-                tokensAfterEstimate: 2_000,
-            },
-        });
-        let forceResult;
-        agentLoopMock.mockImplementation(async function* (messages, options) {
-            // Non-force pre-call invocation must pass through untouched.
-            const passthrough = await options.transformContext(messages, { pendingMessages: [] });
-            expect(passthrough).toBe(messages);
-            expect(compactMock).not.toHaveBeenCalled();
-            // Force invocation (overflow) must compact and return the smaller array.
-            forceResult = await options.transformContext(messages, {
-                force: true,
-                pendingMessages: [],
-            });
-            yield { type: "agent_done" };
-        });
-        const { AgentSession } = await import("./agent-session.js");
-        const session = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            transient: true,
-        });
-        await session.initialize();
-        await session.prompt("do the thing");
-        await session.dispose();
-        expect(compactMock).toHaveBeenCalledTimes(1);
-        expect(forceResult).toEqual(compactedMessages);
-    });
-});
-describe("AgentSession mid-turn compaction", () => {
-    it("compacts non-forced in-flight history exactly at the configured 80% boundary", async () => {
-        await writeJson(path.join(tmpHome, ".ezcoder", "settings.json"), {
-            autoCompact: true,
-            compactThreshold: 0.8,
-        });
-        const pendingMessage = {
-            role: "tool",
-            content: [
-                {
-                    type: "tool_result",
-                    toolCallId: "t1",
-                    content: "pending tool output ".repeat(8),
-                },
-            ],
-        };
-        const usage = {
-            inputTokens: 159_900,
-            cacheRead: 30,
-            cacheWrite: 20,
-            outputTokens: 40,
-        };
-        shouldCompactMock.mockImplementation((_messages, contextWindow, threshold, actualTokens) => actualTokens !== undefined && actualTokens >= Math.ceil(contextWindow * threshold));
-        const compactedMessages = [
-            { role: "system", content: "system prompt" },
-            { role: "user", content: "[compacted in-flight history]" },
-        ];
-        compactMock.mockResolvedValue(compactionResult(compactedMessages));
-        let transformed;
-        agentLoopMock.mockImplementation(async function* (messages, options) {
-            messages.push({
-                role: "assistant",
-                content: [{ type: "tool_call", id: "t1", name: "read", args: {} }],
-            });
-            messages.push(pendingMessage);
-            transformed = await options.transformContext(messages, {
-                usage,
-                pendingMessages: [pendingMessage],
-            });
-            yield { type: "agent_done" };
-        });
-        const { AgentSession } = await import("./agent-session.js");
-        const session = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            transient: true,
-        });
-        await session.initialize();
-        await session.prompt("run a tool loop");
-        await session.dispose();
-        // Computed after the run: the anchored usage sample calibrates the token
-        // estimator, so the pending-message estimate must use the same calibrated
-        // ratio the transform used when it called shouldCompact.
-        const expectedActiveTokens = usage.inputTokens +
-            usage.cacheRead +
-            usage.cacheWrite +
-            usage.outputTokens +
-            estimateConversationTokens([pendingMessage]);
-        expect(expectedActiveTokens).toBeGreaterThanOrEqual(160_000);
-        expect(transformed).toEqual(compactedMessages);
-        expect(shouldCompactMock).toHaveBeenCalledWith(expect.any(Array), 200_000, 0.8, expectedActiveTokens);
-        expect(compactMock.mock.calls.at(-1)?.[0]).toContainEqual(pendingMessage);
-    });
-    it("reuses authoritative usage for the first context check of the next prompt", async () => {
-        const usage = { inputTokens: 120_000, outputTokens: 100 };
-        shouldCompactMock.mockImplementation((_messages, contextWindow, threshold, actualTokens) => actualTokens !== undefined && actualTokens >= Math.ceil(contextWindow * threshold));
-        compactMock.mockResolvedValue(compactionResult([
-            { role: "system", content: "system prompt" },
-            { role: "user", content: "[compacted across prompts]" },
-        ]));
-        let run = 0;
-        agentLoopMock.mockImplementation(async function* (messages, options) {
-            run += 1;
-            await options.transformContext(messages, { pendingMessages: [] });
-            if (run === 1) {
-                messages.push({ role: "assistant", content: "first response" });
-                yield {
-                    type: "turn_end",
-                    turn: 1,
-                    stopReason: "end_turn",
-                    usage,
-                    timing: {
-                        startedAt: 1,
-                        completedAt: 2,
-                        providerDurationMs: 1,
-                    },
-                };
-            }
-            yield { type: "agent_done" };
-        });
-        const { AgentSession } = await import("./agent-session.js");
-        const session = new AgentSession({
-            provider: "anthropic",
-            model: "claude-fable-5",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            transient: true,
-        });
-        await session.initialize();
-        await session.prompt("first prompt");
-        await session.prompt("second prompt");
-        await session.dispose();
-        expect(compactMock).toHaveBeenCalledTimes(1);
-        expect(shouldCompactMock.mock.calls.some((call) => typeof call[3] === "number" && call[3] >= usage.inputTokens + usage.outputTokens)).toBe(true);
-    });
-    it("clears authoritative usage after compaction and does not recompact unchanged context", async () => {
-        await writeJson(path.join(tmpHome, ".ezcoder", "settings.json"), {
-            autoCompact: true,
-            compactThreshold: 0.85,
-        });
-        const trigger = 170_000;
-        let hasCompacted = false;
-        const activeTokensAfterCompaction = [];
-        shouldCompactMock.mockImplementation((_messages, _contextWindow, _threshold, actualTokens) => {
-            if (hasCompacted && actualTokens !== undefined)
-                activeTokensAfterCompaction.push(actualTokens);
-            return actualTokens !== undefined && actualTokens >= trigger;
-        });
-        compactMock.mockImplementation(async () => {
-            hasCompacted = true;
-            return compactionResult([
-                { role: "system", content: "system prompt" },
-                { role: "user", content: "[compacted below target]" },
-            ]);
-        });
-        let run = 0;
-        agentLoopMock.mockImplementation(async function* (messages, options) {
-            run += 1;
-            await options.transformContext(messages, { pendingMessages: [] });
-            messages.push({ role: "assistant", content: `response ${run}` });
-            if (run === 1) {
-                yield {
-                    type: "turn_end",
-                    turn: 1,
-                    stopReason: "end_turn",
-                    usage: { inputTokens: 180_000, outputTokens: 1_000 },
-                    timing: { startedAt: 1, completedAt: 2, providerDurationMs: 1 },
-                };
-            }
-            yield { type: "agent_done" };
-        });
-        const { AgentSession } = await import("./agent-session.js");
-        const session = new AgentSession({
-            provider: "anthropic",
-            model: "claude-fable-5",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            transient: true,
-        });
-        await session.initialize();
-        await session.prompt("first prompt");
-        await session.prompt("second prompt triggers compaction");
-        await session.prompt("third prompt stays below target");
-        await session.dispose();
-        expect(compactMock).toHaveBeenCalledTimes(1);
-        expect(activeTokensAfterCompaction.length).toBeGreaterThan(0);
-        expect(activeTokensAfterCompaction.every((tokens) => tokens < trigger)).toBe(true);
-    });
-    it.each(providerModels)("$provider uses the same normalized usage formula", async ({ provider, model }) => {
-        await writeJson(path.join(tmpHome, ".ezcoder", "settings.json"), {
-            autoCompact: true,
-            compactThreshold: 0.8,
-        });
-        shouldCompactMock.mockReturnValue(false);
-        const usage = {
-            inputTokens: 100,
-            cacheRead: 30,
-            cacheWrite: 20,
-            outputTokens: 40,
-        };
-        const expectedActiveTokens = 190;
-        agentLoopMock.mockImplementation(async function* (messages, options) {
-            await options.transformContext(messages, {
-                usage,
-                pendingMessages: [],
-            });
-            yield { type: "agent_done" };
-        });
-        const { AgentSession } = await import("./agent-session.js");
-        const session = new AgentSession({
-            provider: provider,
-            model,
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            transient: true,
-        });
-        await session.initialize();
-        await session.prompt("test normalized usage");
-        await session.dispose();
-        expect(shouldCompactMock.mock.calls.some((call) => call[3] === expectedActiveTokens)).toBe(true);
-    }, 15_000);
-    it("honors a custom threshold during a non-forced transform", async () => {
-        await writeJson(path.join(tmpHome, ".ezcoder", "settings.json"), {
-            autoCompact: true,
-            compactThreshold: 0.65,
-        });
-        shouldCompactMock.mockReturnValue(false);
-        const usage = { inputTokens: 1_000, outputTokens: 100 };
-        agentLoopMock.mockImplementation(async function* (messages, options) {
-            await options.transformContext(messages, { usage, pendingMessages: [] });
-            yield { type: "agent_done" };
-        });
-        const { AgentSession } = await import("./agent-session.js");
-        const session = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            transient: true,
-        });
-        await session.initialize();
-        await session.prompt("custom threshold");
-        await session.dispose();
-        expect(shouldCompactMock).toHaveBeenCalledWith(expect.any(Array), 200_000, 0.65, 1_100);
-    });
-    it("honors autoCompact false for non-forced calls but force bypasses settings and cooldown", async () => {
-        await writeJson(path.join(tmpHome, ".ezcoder", "settings.json"), {
-            autoCompact: false,
-            compactThreshold: 0.8,
-        });
-        const compactedMessages = [
-            { role: "system", content: "system prompt" },
-            { role: "user", content: "[forced compaction]" },
-        ];
-        compactMock
-            .mockResolvedValueOnce(compactionResult([
-            { role: "system", content: "system prompt" },
-            { role: "user", content: "too little history" },
-            { role: "assistant", content: "reply" },
-        ], false))
-            .mockResolvedValueOnce(compactionResult(compactedMessages));
-        let nonForcedResult;
-        let forcedResult;
-        agentLoopMock.mockImplementation(async function* (messages, options) {
-            nonForcedResult = await options.transformContext(messages, { pendingMessages: [] });
-            await options.transformContext(messages, { force: true, pendingMessages: [] });
-            forcedResult = await options.transformContext(messages, {
-                force: true,
-                pendingMessages: [],
-            });
-            yield { type: "agent_done" };
-        });
-        const { AgentSession } = await import("./agent-session.js");
-        const session = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            transient: true,
-        });
-        await session.initialize();
-        await session.prompt("force despite settings");
-        await session.dispose();
-        expect(nonForcedResult).toBeDefined();
-        expect(shouldCompactMock).not.toHaveBeenCalled();
-        expect(compactMock).toHaveBeenCalledTimes(2);
-        expect(forcedResult).toEqual(compactedMessages);
-    });
-    it("cools down after a non-forced no-op instead of retrying every tool step", async () => {
-        shouldCompactMock.mockImplementation((_messages, _contextWindow, _threshold, actualTokens) => actualTokens !== undefined);
-        compactMock.mockImplementation(async (messages) => compactionResult([...messages], false));
-        agentLoopMock.mockImplementation(async function* (messages, options) {
-            const transformOptions = {
-                usage: { inputTokens: 180_000, outputTokens: 1_000 },
-                pendingMessages: [],
-            };
-            await options.transformContext(messages, transformOptions);
-            await options.transformContext(messages, transformOptions);
-            yield { type: "agent_done" };
-        });
-        const { AgentSession } = await import("./agent-session.js");
-        const session = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            transient: true,
-        });
-        await session.initialize();
-        await session.prompt("no-op cooldown");
-        await session.dispose();
-        expect(compactMock).toHaveBeenCalledTimes(1);
-    });
-    it("cools down after a failed proactive summary", async () => {
-        shouldCompactMock.mockImplementation((_messages, _contextWindow, _threshold, actualTokens) => actualTokens !== undefined);
-        compactMock.mockRejectedValue(new Error("summary unavailable"));
-        agentLoopMock.mockImplementation(async function* (messages, options) {
-            const transformOptions = {
-                usage: { inputTokens: 180_000, outputTokens: 1_000 },
-                pendingMessages: [],
-            };
-            await options.transformContext(messages, transformOptions);
-            await options.transformContext(messages, transformOptions);
-            yield { type: "agent_done" };
-        });
-        const { AgentSession } = await import("./agent-session.js");
-        const session = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            transient: true,
-        });
-        await session.initialize();
-        await session.prompt("failure cooldown");
-        await session.dispose();
-        expect(compactMock).toHaveBeenCalledTimes(1);
-    });
-    it("continues the run and cools down across prompts after pre-run compaction fails", async () => {
-        shouldCompactMock.mockReturnValue(true);
-        compactMock.mockRejectedValue(new Error("summary unavailable"));
-        agentLoopMock.mockImplementation(async function* () {
-            yield { type: "agent_done" };
-        });
-        const { AgentSession } = await import("./agent-session.js");
-        const session = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            transient: true,
-        });
-        await session.initialize();
-        await session.prompt("first prompt");
-        await session.prompt("second prompt");
-        await session.dispose();
-        expect(compactMock).toHaveBeenCalledTimes(1);
-        expect(agentLoopMock).toHaveBeenCalledTimes(2);
-    });
-    it("cools down across prompts after pre-run compaction is a no-op", async () => {
-        shouldCompactMock.mockReturnValue(true);
-        compactMock.mockImplementation(async (messages) => compactionResult([...messages], false));
-        agentLoopMock.mockImplementation(async function* () {
-            yield { type: "agent_done" };
-        });
-        const { AgentSession } = await import("./agent-session.js");
-        const session = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            transient: true,
-        });
-        await session.initialize();
-        await session.prompt("first prompt");
-        await session.prompt("second prompt");
-        await session.dispose();
-        expect(compactMock).toHaveBeenCalledTimes(1);
-        expect(agentLoopMock).toHaveBeenCalledTimes(2);
-    });
-});
-/** Every .jsonl under the ezcoder session store — must stay empty for
- * transient sessions (Nolan chat/autopilot and subagent spawns). */
-async function listSessionFiles() {
-    const sessionsDir = path.join(tmpHome, ".ezcoder", "sessions");
-    const out = [];
-    const walk = async (dir) => {
-        let entries;
-        try {
-            entries = await fs.readdir(dir, { withFileTypes: true });
-        }
-        catch {
-            return;
-        }
-        for (const e of entries) {
-            const full = path.join(dir, e.name);
-            if (e.isDirectory())
-                await walk(full);
-            else if (e.name.endsWith(".jsonl"))
-                out.push(full);
-        }
-    };
-    await walk(sessionsDir);
-    return out;
-}
-describe("transient sessions never leak to the session store", () => {
-    // Regression: the Nolan autopilot session (transient: true) was leaking one
-    // 3-line "## Who you are … Nolan Grout" session file per review cycle into the
-    // project's session list — via compact() assigning a real sessionPath and
-    // via newSession() (autopilot's per-cycle resetReviewer) creating a file
-    // unconditionally.
-    it("compact() keeps a transient session fully in memory — no session file", async () => {
-        shouldCompactMock.mockReturnValue(true);
-        compactMock.mockResolvedValue({
-            messages: [
-                { role: "system", content: "nolan system prompt" },
-                { role: "user", content: "[compacted]" },
-            ],
-            result: {
-                compacted: true,
-                originalCount: 2,
-                newCount: 2,
-                tokensBeforeEstimate: 100_000,
-                tokensAfterEstimate: 1_000,
-            },
-        });
-        agentLoopMock.mockImplementation(async function* (messages) {
-            messages.push({ role: "assistant", content: "IGNORE" });
-            yield { type: "agent_done" };
-        });
-        const { AgentSession } = await import("./agent-session.js");
-        const session = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "nolan system prompt",
-            transient: true,
-        });
-        await session.initialize();
-        await session.prompt("review this turn");
-        // Post-compaction persistence paths must all stay no-ops.
-        await session.persistNolanTurn("q", "a");
-        await session.persistAutopilotMarker("done");
-        await session.dispose();
-        expect(compactMock).toHaveBeenCalled();
-        expect(await listSessionFiles()).toEqual([]);
-    });
-    it("newSession() on a transient session creates no file and detaches the DAG leaf", async () => {
-        shouldCompactMock.mockReturnValue(false);
-        agentLoopMock.mockImplementation(async function* (messages) {
-            messages.push({ role: "assistant", content: "ALL_CLEAR" });
-            yield { type: "agent_done" };
-        });
-        const { AgentSession } = await import("./agent-session.js");
-        const session = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "nolan system prompt",
-            transient: true,
-        });
-        await session.initialize();
-        await session.prompt("cycle 1");
-        // Autopilot's per-cycle resetReviewer path.
-        await session.newSession();
-        await session.prompt("cycle 2");
-        await session.dispose();
-        expect(await listSessionFiles()).toEqual([]);
-    });
-});
-describe("AgentSession compaction events", () => {
-    it("marks a skipped compaction as unsuccessful", async () => {
-        const { AgentSession } = await import("./agent-session.js");
-        const session = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            transient: true,
-        });
-        await session.initialize();
-        compactMock.mockResolvedValue(compactionResult([...session.getMessages()], false));
-        const events = [];
-        session.eventBus.on("compaction_end", (event) => events.push(event));
-        await session.compact({ accessToken: "token" });
-        expect(events).toEqual([{ compacted: false, originalCount: 1, newCount: 1 }]);
-        await session.dispose();
-    });
-    it("emits query-selection token statistics when the compactor ran retrieval", async () => {
-        const { AgentSession } = await import("./agent-session.js");
-        const session = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            transient: true,
-        });
-        await session.initialize();
-        compactMock.mockResolvedValue(compactionResult([...session.getMessages()], false, {
-            strategy: "query_aware",
-            selectedMessages: 8,
-            selectedTokens: 1200,
-            droppedMessages: 4,
-            queryTerms: 3,
-        }));
-        const events = [];
-        session.eventBus.on("compaction_end", (event) => events.push(event));
-        await session.compact({ accessToken: "token" });
-        expect(events).toEqual([
-            {
-                compacted: false,
-                originalCount: 1,
-                newCount: 1,
-                selectionStrategy: "query_aware",
-                selectedMessages: 8,
-                selectedTokens: 1200,
-                droppedMessages: 4,
-                queryTerms: 3,
-            },
-        ]);
-        await session.dispose();
-    });
-});
-describe("AgentSession.getContextUsage", () => {
-    it("reports the model window and drops after a compaction", async () => {
-        const { AgentSession } = await import("./agent-session.js");
-        const model = MODELS[0];
-        const session = new AgentSession({
-            provider: model.provider,
-            model: model.id,
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            transient: true,
-        });
-        await session.initialize();
-        const messages = session.getMessages();
-        for (let i = 0; i < 40; i += 1) {
-            messages.push({ role: "user", content: "a fairly long user message ".repeat(20) });
-        }
-        const before = session.getContextUsage();
-        expect(before.size).toBeGreaterThan(0);
-        expect(before.used).toBeGreaterThan(0);
-        // No authoritative pricing exists yet, so cost is omitted rather than
-        // reported as zero — a client showing $0.00 would be stating a fact we
-        // do not have.
-        expect(before.costUsd).toBeUndefined();
-        compactMock.mockResolvedValue(compactionResult([messages[0]], true));
-        await session.compact({ accessToken: "test-token" });
-        const after = session.getContextUsage();
-        // Clients detect compaction from exactly this: used falls, size does not.
-        expect(after.used).toBeLessThan(before.used);
-        expect(after.size).toBe(before.size);
-        await session.dispose();
-    });
-});
-describe("load-time auto-compaction deferral (deferLoadCompaction)", () => {
-    // Regression: resuming an over-context session ran a summary LLM call (30s
-    // timeout) inline in loadExistingSession — inside initialize(), which the
-    // ezcoder-app sidecar's readiness (waitForReady → the whole webview) is gated on.
-    // A slow/hanging summary call froze the window for the full timeout. With
-    // deferLoadCompaction the resume returns immediately and runLoop()'s
-    // pre-run auto-compaction handles it on the first prompt.
-    /** Create + persist a real session file, returning its path. */
-    async function persistSession() {
-        shouldCompactMock.mockReturnValue(false);
-        agentLoopMock.mockImplementation(async function* (messages) {
-            messages.push({ role: "assistant", content: "first reply" });
-            yield { type: "agent_done" };
-        });
-        const { AgentSession } = await import("./agent-session.js");
-        const session = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-        });
-        await session.initialize();
-        await session.prompt("hello");
-        await session.dispose();
-        const files = await listSessionFiles();
-        expect(files.length).toBeGreaterThan(0);
-        return files[0];
-    }
-    const compactedResult = {
-        messages: [
-            { role: "system", content: "system prompt" },
-            { role: "user", content: "[compacted]" },
-        ],
-        result: {
-            compacted: true,
-            originalCount: 3,
-            newCount: 2,
-            tokensBeforeEstimate: 500_000,
-            tokensAfterEstimate: 2_000,
-        },
-    };
-    it("defers compaction out of initialize() and runs it on the first prompt", async () => {
-        const sessionPath = await persistSession();
-        shouldCompactMock.mockReturnValue(true);
-        compactMock.mockResolvedValue(compactedResult);
-        agentLoopMock.mockImplementation(async function* (messages) {
-            messages.push({ role: "assistant", content: "resumed reply" });
-            yield { type: "agent_done" };
-        });
-        const { AgentSession } = await import("./agent-session.js");
-        const resumed = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            sessionId: sessionPath,
-            deferLoadCompaction: true,
-        });
-        await resumed.initialize();
-        // Readiness path must NOT have paid for a summary LLM call.
-        expect(compactMock).not.toHaveBeenCalled();
-        // First prompt triggers runLoop()'s existing pre-run auto-compaction.
-        await resumed.prompt("continue");
-        expect(compactMock).toHaveBeenCalledTimes(1);
-        await resumed.dispose();
-    });
-    it("coordinates simultaneous compactions and adopts the single checkpoint", async () => {
-        const sessionPath = await persistSession();
-        shouldCompactMock.mockReturnValue(true);
-        compactMock.mockImplementation(async () => {
-            await new Promise((resolve) => setTimeout(resolve, 25));
-            return compactedResult;
-        });
-        const { AgentSession } = await import("./agent-session.js");
-        const first = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            sessionId: sessionPath,
-            deferLoadCompaction: true,
-        });
-        const second = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            sessionId: sessionPath,
-            deferLoadCompaction: true,
-        });
-        await Promise.all([first.initialize(), second.initialize()]);
-        const credentials = { accessToken: "test-token" };
-        await Promise.all([
-            first.compact(credentials, "automatic"),
-            second.compact(credentials, "automatic"),
-        ]);
-        expect(compactMock).toHaveBeenCalledTimes(1);
-        expect(first.getState().sessionPath).toBe(second.getState().sessionPath);
-        expect(first.getState().sessionPath).not.toBe(sessionPath);
-        await Promise.all([first.dispose(), second.dispose()]);
-    });
-    it("rebases stale compaction onto a divergent canonical checkpoint", async () => {
-        const sessionPath = await persistSession();
-        shouldCompactMock.mockReturnValue(false);
-        agentLoopMock.mockImplementation(async function* (messages) {
-            messages.push({ role: "assistant", content: "canonical reply" });
-            yield { type: "agent_done" };
-        });
-        compactMock.mockImplementation(async (messages) => {
-            const hasCanonicalProgress = messages.some((message) => typeof message.content === "string" && message.content.includes("canonical progress"));
-            return {
-                ...compactedResult,
-                messages: [
-                    { role: "system", content: "system prompt" },
-                    {
-                        role: "user",
-                        content: hasCanonicalProgress ? "[compacted canonical progress]" : "[compacted stale]",
-                    },
-                ],
-            };
-        });
-        const { AgentSession } = await import("./agent-session.js");
-        const creator = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            sessionId: sessionPath,
-            deferLoadCompaction: true,
-        });
-        const stale = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            sessionId: sessionPath,
-            deferLoadCompaction: true,
-        });
-        await Promise.all([creator.initialize(), stale.initialize()]);
-        await creator.prompt("canonical progress");
-        shouldCompactMock.mockReturnValue(true);
-        await creator.compact({ accessToken: "test-anthropic-token" }, "automatic");
-        await stale.compact({ accessToken: "test-anthropic-token" }, "automatic");
-        const staleCompactionInput = compactMock.mock.calls[1]?.[0];
-        expect(staleCompactionInput).toEqual(expect.arrayContaining([
-            expect.objectContaining({ content: "[compacted canonical progress]" }),
-        ]));
-        expect(stale.getMessages()).toEqual(expect.arrayContaining([
-            expect.objectContaining({ content: "[compacted canonical progress]" }),
-        ]));
-        await Promise.all([creator.dispose(), stale.dispose()]);
-    });
-    it("deduplicates unchanged no-op attempts but lets manual compaction retry", async () => {
-        const sessionPath = await persistSession();
-        shouldCompactMock.mockReturnValue(true);
-        compactMock.mockImplementation(async (messages) => compactionResult([...messages], false));
-        const { AgentSession } = await import("./agent-session.js");
-        const session = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            sessionId: sessionPath,
-            deferLoadCompaction: true,
-        });
-        await session.initialize();
-        const credentials = { accessToken: "test-token" };
-        await session.compact(credentials, "automatic");
-        await session.compact(credentials, "automatic");
-        expect(compactMock).toHaveBeenCalledTimes(1);
-        await session.compact(credentials, "manual");
-        expect(compactMock).toHaveBeenCalledTimes(2);
-        expect(session.getState().sessionPath).toBe(sessionPath);
-        await session.dispose();
-    });
-    it("resolves a stale below-threshold checkpoint path before loading history", async () => {
-        const sessionPath = await persistSession();
-        shouldCompactMock.mockReturnValue(true);
-        compactMock.mockResolvedValue(compactedResult);
-        const { AgentSession } = await import("./agent-session.js");
-        const creator = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            sessionId: sessionPath,
-            deferLoadCompaction: true,
-        });
-        await creator.initialize();
-        await creator.compact({ accessToken: "test-anthropic-token" }, "automatic");
-        const checkpointPath = creator.getState().sessionPath;
-        await creator.dispose();
-        shouldCompactMock.mockReturnValue(false);
-        const resumed = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            sessionId: sessionPath,
-            deferLoadCompaction: true,
-        });
-        await resumed.initialize();
-        expect(resumed.getState().sessionPath).toBe(checkpointPath);
-        expect(resumed.getMessages()).toEqual(expect.arrayContaining([expect.objectContaining({ content: "[compacted]" })]));
-        expect(compactMock).toHaveBeenCalledTimes(1);
-        await resumed.dispose();
-    });
-    it("adopts a checkpoint created after deferred load before persisting the prompt", async () => {
-        const sessionPath = await persistSession();
-        shouldCompactMock.mockReturnValue(true);
-        compactMock.mockResolvedValue(compactedResult);
-        agentLoopMock.mockImplementation(async function* () {
-            yield { type: "agent_done" };
-        });
-        const { AgentSession } = await import("./agent-session.js");
-        const creator = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            sessionId: sessionPath,
-            deferLoadCompaction: true,
-        });
-        const follower = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            sessionId: sessionPath,
-            deferLoadCompaction: true,
-        });
-        await Promise.all([creator.initialize(), follower.initialize()]);
-        await creator.compact({ accessToken: "test-token" }, "automatic");
-        const checkpointPath = creator.getState().sessionPath;
-        shouldCompactMock.mockReturnValue(false);
-        await follower.prompt("continue from the shared checkpoint");
-        expect(compactMock).toHaveBeenCalledTimes(1);
-        expect(follower.getState().sessionPath).toBe(checkpointPath);
-        expect(follower.getMessages()).toEqual(expect.arrayContaining([
-            expect.objectContaining({ content: "[compacted]" }),
-            expect.objectContaining({ content: "continue from the shared checkpoint" }),
-        ]));
-        await Promise.all([creator.dispose(), follower.dispose()]);
-    });
-    it("still compacts inline during initialize() without the flag (CLI resume)", async () => {
-        const sessionPath = await persistSession();
-        shouldCompactMock.mockReturnValue(true);
-        compactMock.mockResolvedValue(compactedResult);
-        const { AgentSession } = await import("./agent-session.js");
-        const resumed = new AgentSession({
-            provider: "anthropic",
-            model: "claude-test",
-            cwd: tmpProject,
-            systemPrompt: "system prompt",
-            sessionId: sessionPath,
-        });
-        await resumed.initialize();
-        expect(compactMock).toHaveBeenCalledTimes(1);
-        await resumed.dispose();
-    });
-});
-//# sourceMappingURL=agent-session-compaction.test.js.map