fix(eval): coalesced python kernel replacement

Tracked retained Python kernel generations and shared one replacement promise per dead generation. Reset and disposal now invalidate and drain replacement work before allowing a new session to take ownership.

Added deterministic fake-kernel coverage for concurrent callers, cancellation, reset, owner/global disposal, and independent cwd keys.

Fixes #6367
This commit is contained in:
roboomp
2026-07-23 19:00:08 +00:00
parent 130578aced
commit 03489d1ebe
3 changed files with 326 additions and 34 deletions
@@ -1,5 +1,9 @@
import { afterEach, describe, expect, it } from "bun:test";
import { disposeAllKernelSessions, executePython } from "@oh-my-pi/pi-coding-agent/eval/py/executor";
import {
disposeAllKernelSessions,
disposeKernelSessionsByOwner,
executePython,
} from "@oh-my-pi/pi-coding-agent/eval/py/executor";
import {
type KernelExecuteOptions,
type KernelExecuteResult,
@@ -29,6 +33,10 @@ class FakeKernel {
return this.#alive;
}
markDead(): void {
this.#alive = false;
}
async execute(code: string, options?: KernelExecuteOptions): Promise<KernelExecuteResult> {
this.executeCalls.push(code);
this.#onExecute?.(options);
@@ -53,6 +61,10 @@ const okResult: KernelExecuteResult = {
stdinRequested: false,
};
async function flushMicrotasks(): Promise<void> {
for (let i = 0; i < 5; i += 1) await Promise.resolve();
}
describe("executePython session lifecycle", () => {
const originalStart = PythonKernel.start;
@@ -95,6 +107,215 @@ describe("executePython session lifecycle", () => {
expect(liveKernel.executeCalls).toEqual(["print('restart')"]);
});
it("coalesces concurrent replacement of one dead session generation", async () => {
const deadKernel = new FakeKernel(okResult);
const replacementOne = new FakeKernel(okResult);
const replacementTwo = new FakeKernel(okResult);
const replacementStarted = Promise.withResolvers<void>();
const releaseReplacement = Promise.withResolvers<void>();
const replacements = [replacementOne, replacementTwo];
let startCount = 0;
PythonKernel.start = async () => {
startCount += 1;
if (startCount === 1) return deadKernel as unknown as PythonKernel;
replacementStarted.resolve();
await releaseReplacement.promise;
return replacements.shift() as unknown as PythonKernel;
};
await executePython("print('setup')", { sessionId: "session-concurrent-restart" });
deadKernel.markDead();
const first = executePython("print('first')", { sessionId: "session-concurrent-restart" });
await replacementStarted.promise;
const second = executePython("print('second')", { sessionId: "session-concurrent-restart" });
await Promise.resolve();
await Promise.resolve();
expect(startCount).toBe(2);
releaseReplacement.resolve();
await Promise.all([first, second]);
expect(replacementOne.executeCalls).toEqual(["print('first')", "print('second')"]);
expect(replacementTwo.executeCalls).toEqual([]);
await disposeAllKernelSessions();
expect(replacementOne.shutdownCalls).toBe(1);
expect(replacementTwo.shutdownCalls).toBe(0);
});
it("keeps a shared replacement alive when one caller cancels", async () => {
const deadKernel = new FakeKernel(okResult);
const replacement = new FakeKernel(okResult);
const replacementStarted = Promise.withResolvers<void>();
const releaseReplacement = Promise.withResolvers<void>();
let startCount = 0;
PythonKernel.start = async () => {
startCount += 1;
if (startCount === 1) return deadKernel as unknown as PythonKernel;
replacementStarted.resolve();
await releaseReplacement.promise;
return replacement as unknown as PythonKernel;
};
await executePython("print('setup')", { sessionId: "session-cancelled-restart" });
deadKernel.markDead();
const abortController = new AbortController();
const cancelled = executePython("print('cancelled')", {
sessionId: "session-cancelled-restart",
signal: abortController.signal,
});
await replacementStarted.promise;
const retained = executePython("print('retained')", { sessionId: "session-cancelled-restart" });
await flushMicrotasks();
abortController.abort(Object.assign(new Error("replacement wait cancelled"), { name: "AbortError" }));
expect((await cancelled).cancelled).toBe(true);
expect(startCount).toBe(2);
releaseReplacement.resolve();
expect((await retained).cancelled).toBe(false);
expect(replacement.executeCalls).toEqual(["print('retained')"]);
});
it("invalidates an in-flight replacement before resetting to a fresh generation", async () => {
const deadKernel = new FakeKernel(okResult);
const staleReplacement = new FakeKernel(okResult);
const freshKernel = new FakeKernel(okResult);
const replacementStarted = Promise.withResolvers<void>();
const releaseReplacement = Promise.withResolvers<void>();
let startCount = 0;
PythonKernel.start = async () => {
startCount += 1;
if (startCount === 1) return deadKernel as unknown as PythonKernel;
if (startCount === 2) {
replacementStarted.resolve();
await releaseReplacement.promise;
return staleReplacement as unknown as PythonKernel;
}
return freshKernel as unknown as PythonKernel;
};
await executePython("print('setup')", { sessionId: "session-reset-replacement" });
deadKernel.markDead();
const obsolete = executePython("print('obsolete')", { sessionId: "session-reset-replacement" });
await replacementStarted.promise;
const reset = executePython("print('reset')", {
sessionId: "session-reset-replacement",
reset: true,
});
await flushMicrotasks();
releaseReplacement.resolve();
expect((await obsolete).cancelled).toBe(true);
expect((await reset).cancelled).toBe(false);
expect(staleReplacement.executeCalls).toEqual([]);
expect(staleReplacement.shutdownCalls).toBe(1);
expect(freshKernel.executeCalls).toEqual(["print('reset')"]);
await executePython("print('later')", { sessionId: "session-reset-replacement" });
expect(startCount).toBe(3);
expect(freshKernel.executeCalls).toEqual(["print('reset')", "print('later')"]);
});
it("drains replacements invalidated by owner and global disposal", async () => {
const ownerKernel = new FakeKernel(okResult);
const globalKernel = new FakeKernel(okResult);
const ownerReplacement = new FakeKernel(okResult);
const globalReplacement = new FakeKernel(okResult);
const replacementsStarted = Promise.withResolvers<void>();
const releaseReplacements = Promise.withResolvers<void>();
const initialKernels = [ownerKernel, globalKernel];
const replacementKernels = [ownerReplacement, globalReplacement];
let replacementStartCount = 0;
PythonKernel.start = async () => {
const initial = initialKernels.shift();
if (initial) return initial as unknown as PythonKernel;
replacementStartCount += 1;
if (replacementStartCount === 2) replacementsStarted.resolve();
await releaseReplacements.promise;
return replacementKernels.shift() as unknown as PythonKernel;
};
await executePython("print('owner setup')", {
sessionId: "session-owner-disposal-replacement",
kernelOwnerId: "replacement-owner",
});
await executePython("print('global setup')", { sessionId: "session-global-disposal-replacement" });
ownerKernel.markDead();
globalKernel.markDead();
const ownerExecution = executePython("print('owner obsolete')", {
sessionId: "session-owner-disposal-replacement",
kernelOwnerId: "replacement-owner",
});
const globalExecution = executePython("print('global obsolete')", {
sessionId: "session-global-disposal-replacement",
});
await replacementsStarted.promise;
const ownerDisposal = disposeKernelSessionsByOwner("replacement-owner");
await flushMicrotasks();
const globalDisposal = disposeAllKernelSessions();
await flushMicrotasks();
releaseReplacements.resolve();
expect((await ownerExecution).cancelled).toBe(true);
expect((await globalExecution).cancelled).toBe(true);
await Promise.all([ownerDisposal, globalDisposal]);
expect(ownerReplacement.executeCalls).toEqual([]);
expect(globalReplacement.executeCalls).toEqual([]);
expect(ownerReplacement.shutdownCalls).toBe(1);
expect(globalReplacement.shutdownCalls).toBe(1);
});
it("keeps replacement coordination independent across normalized cwd keys", async () => {
const deadOne = new FakeKernel(okResult);
const deadTwo = new FakeKernel(okResult);
const replacementOne = new FakeKernel(okResult);
const replacementTwo = new FakeKernel(okResult);
const kernels = [deadOne, deadTwo, replacementOne, replacementTwo];
let startCount = 0;
PythonKernel.start = async () => {
startCount += 1;
return kernels.shift() as unknown as PythonKernel;
};
await executePython("print('setup one')", {
cwd: "/tmp/replacement-key-one",
sessionId: "session-independent-replacement",
});
await executePython("print('setup two')", {
cwd: "/tmp/replacement-key-two",
sessionId: "session-independent-replacement",
});
deadOne.markDead();
deadTwo.markDead();
await Promise.all([
executePython("print('one')", {
cwd: "/tmp/replacement-key-one",
sessionId: "session-independent-replacement",
}),
executePython("print('two')", {
cwd: "/tmp/replacement-key-two",
sessionId: "session-independent-replacement",
}),
]);
expect(startCount).toBe(4);
expect(replacementOne.executeCalls).toEqual(["print('one')"]);
expect(replacementTwo.executeCalls).toEqual(["print('two')"]);
});
it("resets the session kernel when requested", async () => {
const firstKernel = new FakeKernel(okResult);
const secondKernel = new FakeKernel(okResult);