fix(plugins): retry errored plugins at boot instead of leaving them dead (#12054)

## Thinking Path

> - Paperclip is the open source app people use to manage AI agents for
work
> - Plugins extend the server with sandbox providers, tools, and jobs; a
loader activates them at boot
> - When activation fails, the loader marks the plugin `error` and skips
it on every later boot
> - Activation failures are often environmental — missing package
dependencies, a stale build output, a module that moved under a pull —
and the fix lands on disk without any write to the plugin row
> - The plugin therefore stays dead forever, and every feature behind it
(sandbox destroys, cleanup sweeps, probes) silently stops working until
an operator flips the row by hand
> - This pull request makes `loadAll` retry errored plugins once per
boot: flip to `ready`, attempt activation, and re-record the error if
the attempt fails
> - The benefit is that a plugin recovers on the next boot after its
environment is fixed, with no manual database or lifecycle intervention

## Linked Issues or Issue Description

**What happened?**

Several sandbox-provider plugins sat in `error` status for weeks after a
transient activation failure (a module resolution error from an older
checkout state). The boot loader only loads plugins in `ready` status,
so it never retried them. Environments backed by those providers lost
sandbox destroys, cleanup sweeps, and probes with no visible signal
other than the stale `last_error`.

**Expected behavior**

A plugin whose activation failure has been fixed on disk recovers on the
next server boot. A plugin that still fails stays in `error` with a
fresh error message.

**Steps to reproduce**

1. Install a plugin whose worker cannot start (for example, delete one
of its dependencies), then boot the server. The plugin lands in `error`
status.
2. Restore the dependency.
3. Restart the server. Before this change, the plugin stays in `error`
forever. After this change, the boot retries it and the plugin
activates.

## What Changed

- `server/src/services/plugin-loader.ts`: `loadAll` also fetches plugins
in `error` status, flips each to `ready`, and activates it with the
normal batch. The flip runs before activation because the `error` status
only legally transitions to `ready` or `uninstalled`; a retry that
failed while still in `error` could not re-mark itself. A failed flip
logs a warning and never aborts the boot load. The stale comment at the
`markError` site now describes the retry.
- `server/src/__tests__/plugin-loader-error-retry.test.ts`: covers the
flip-then-retry flow, the failed-flip isolation, and the empty case.

## Verification

- `npx vitest run server/src/__tests__/plugin-loader-error-retry.test.ts
server/src/__tests__/bundled-plugins.test.ts
server/src/__tests__/plugin-lifecycle-restart.test.ts
server/src/__tests__/cloud-image-bundled-plugins.test.ts`
- Manual: mark an installed plugin's status to `error`, restart the
server, and observe the loader log line `retrying plugins that failed
activation on a previous boot` followed by a successful activation (or a
fresh `last_error` if the plugin is genuinely broken).

## Risks

- A genuinely broken plugin now costs one bounded activation attempt per
boot (the attempts run in parallel with the ready batch under
`Promise.allSettled`). It cannot crash-loop within a running process,
and it returns to `error` with a fresh message.
- The flip clears `last_error` before the attempt. If the process dies
between the flip and the activation, the row is `ready` with no error
text; the next boot simply loads it as a ready plugin.
- Operators who relied on `error` as a manual "keep this off" latch
should use the `disabled` status, which this change does not touch.

## Model Used

- Claude (Anthropic) — Fable 5, model id `claude-fable-5`, extended
thinking, agentic tool use via Claude Code CLI.

## Checklist

- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have not referenced internal/instance-local Paperclip issues or
links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip`
URLs)
- [x] My branch name describes the change (e.g. `docs/...`, `fix/...`)
and contains no internal Paperclip ticket id or instance-derived details
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [ ] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [ ] All Paperclip CI gates are green
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge
This commit is contained in:
Devin Foley 2026-08-23 16:52:42 -07:00 committed by GitHub
parent ae6761e2b0
commit 627eef7cbd
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 196 additions and 8 deletions

View File

@ -0,0 +1,151 @@
/**
* loadAll boot retry for plugins stranded in error status.
*
* An activation failure marks a plugin `error`, and the loader used to skip
* those rows on every later boot the row stayed dead until an operator
* flipped it back to `ready` by hand, even when the underlying cause (missing
* package dependencies, a stale build output) had long been fixed on disk.
* loadAll now queues errored plugins for one retry per boot: it flips each row
* to `ready` first (the error status cannot legally re-enter `error`, so a
* failed retry could not re-mark itself otherwise) and then activates it like
* any ready plugin. A retry that fails re-records the error through markError.
*/
import { beforeEach, describe, expect, it, vi } from "vitest";
import type { Db } from "@paperclipai/db";
const mockRegistry = vi.hoisted(() => ({
getById: vi.fn(),
getByKey: vi.fn(),
list: vi.fn(),
listInstalled: vi.fn(),
listByStatus: vi.fn(),
update: vi.fn(),
updateStatus: vi.fn(),
upsertConfig: vi.fn(),
getConfig: vi.fn(),
delete: vi.fn(),
}));
vi.mock("../services/plugin-registry.js", () => ({
pluginRegistryService: () => mockRegistry,
}));
import { pluginLoader } from "../services/plugin-loader.js";
import type { PluginRuntimeServices } from "../services/plugin-loader.js";
function createPluginRecord(overrides: Record<string, unknown> = {}) {
return {
id: "plugin-err-1",
pluginKey: "example.broken-plugin",
packageName: "@example/broken-plugin",
packagePath: "/nonexistent/broken-plugin",
version: "1.0.0",
apiVersion: 1,
categories: [],
status: "error",
lastError: "Activation failed: previous boot failure",
installOrder: 1,
manifestJson: {
id: "example.broken-plugin",
apiVersion: 1,
version: "1.0.0",
displayName: "Broken Plugin",
description: "Fixture",
author: "Test",
categories: [],
capabilities: [],
entrypoints: { worker: "dist/worker.js" },
},
...overrides,
};
}
function createRuntimeServices() {
return {
lifecycleManager: {
markError: vi.fn(async () => createPluginRecord()),
},
workerManager: {},
eventBus: {},
jobScheduler: {},
jobStore: {},
toolDispatcher: {},
buildHostHandlers: vi.fn(() => ({})),
instanceInfo: { hostVersion: "0.0.0-test" },
} as unknown as PluginRuntimeServices;
}
function createLoader(runtimeServices: PluginRuntimeServices) {
return pluginLoader(
{} as unknown as Db,
{ localPluginDir: "/nonexistent/local-plugins", enableLocalFilesystem: false, enableNpmDiscovery: false },
runtimeServices,
);
}
describe("pluginLoader.loadAll error retry", () => {
beforeEach(() => {
vi.clearAllMocks();
});
it("flips an errored plugin to ready and retries its activation at boot", async () => {
const erroredPlugin = createPluginRecord();
mockRegistry.listByStatus.mockImplementation(async (status: string) => {
if (status === "error") return [erroredPlugin];
return [];
});
mockRegistry.updateStatus.mockResolvedValue({ ...erroredPlugin, status: "ready", lastError: null });
const runtimeServices = createRuntimeServices();
const loader = createLoader(runtimeServices);
const result = await loader.loadAll();
// The flip precedes activation, and clears the stale error.
expect(mockRegistry.updateStatus).toHaveBeenCalledExactlyOnceWith(erroredPlugin.id, { status: "ready" });
// The retried plugin joins the boot batch; its package is unresolvable, so
// the attempt fails and re-records a fresh error via markError.
expect(result.total).toBe(1);
expect(result.succeeded).toBe(0);
expect(result.failed).toBe(1);
expect(runtimeServices.lifecycleManager.markError).toHaveBeenCalledWith(
erroredPlugin.id,
expect.stringContaining("Activation failed"),
);
});
it("keeps loading ready plugins when the errored flip fails", async () => {
const readyPlugin = createPluginRecord({
id: "plugin-ready-1",
pluginKey: "example.ready-plugin",
status: "ready",
lastError: null,
});
const erroredPlugin = createPluginRecord();
mockRegistry.listByStatus.mockImplementation(async (status: string) => {
if (status === "ready") return [readyPlugin];
if (status === "error") return [erroredPlugin];
return [];
});
mockRegistry.updateStatus.mockRejectedValue(new Error("db write refused"));
const loader = createLoader(createRuntimeServices());
const result = await loader.loadAll();
// The failed flip skips the retry but never aborts the boot load.
expect(mockRegistry.updateStatus).toHaveBeenCalledExactlyOnceWith(erroredPlugin.id, { status: "ready" });
expect(result.total).toBe(1);
expect(result.results[0]?.plugin.id).toBe(readyPlugin.id);
});
it("returns the empty result when no plugin is ready or errored", async () => {
mockRegistry.listByStatus.mockResolvedValue([]);
const loader = createLoader(createRuntimeServices());
const result = await loader.loadAll();
expect(result).toEqual({ total: 0, succeeded: 0, failed: 0, results: [] });
});
});

View File

@ -1945,25 +1945,61 @@ export function pluginLoader(
// Fetch all plugins in ready status, ordered by installOrder
const readyPlugins = (await registry.listByStatus("ready")) as PluginRecord[];
if (readyPlugins.length === 0) {
// Retry plugins stranded in error status. An activation failure is often
// environmental — missing package dependencies, a stale build output, a
// module that moved under a pull — and the fix lands on disk without any
// write to the plugin row, so the row would otherwise stay dead until an
// operator flips it back by hand. One attempt per boot cannot crash-loop
// within a running process, and a failed attempt re-records the error
// through the normal markError path. The flip to ready must run BEFORE
// activation: the error status only legally transitions to ready or
// uninstalled, so a retry that failed while still in error status could
// not re-mark itself as errored.
const erroredPlugins = (await registry.listByStatus("error")) as PluginRecord[];
const retriedPlugins: PluginRecord[] = [];
for (const plugin of erroredPlugins) {
try {
const flipped = (await registry.updateStatus(plugin.id, { status: "ready" })) as PluginRecord | null;
if (flipped) retriedPlugins.push(flipped);
} catch (err) {
log.warn(
{
pluginId: plugin.id,
pluginKey: plugin.pluginKey,
err: err instanceof Error ? err.message : String(err),
},
"plugin-loader: could not queue errored plugin for a boot retry",
);
}
}
if (retriedPlugins.length > 0) {
log.info(
{ count: retriedPlugins.length, pluginKeys: retriedPlugins.map((plugin) => plugin.pluginKey) },
"plugin-loader: retrying plugins that failed activation on a previous boot",
);
}
const pluginsToLoad = [...readyPlugins, ...retriedPlugins];
if (pluginsToLoad.length === 0) {
log.info("plugin-loader: no ready plugins to load");
return { total: 0, succeeded: 0, failed: 0, results: [] };
}
log.info(
{ count: readyPlugins.length },
{ count: pluginsToLoad.length },
"plugin-loader: found ready plugins to load",
);
// Load plugins in parallel
const results = await Promise.allSettled(
readyPlugins.map((plugin) => activatePlugin(plugin))
pluginsToLoad.map((plugin) => activatePlugin(plugin))
);
const loadResults = results.map((r, i) => {
if (r.status === "fulfilled") return r.value;
return {
plugin: readyPlugins[i]!,
plugin: pluginsToLoad[i]!,
success: false,
error: String(r.reason),
registered: { worker: false, eventSubscriptions: 0, jobs: 0, webhooks: 0, tools: 0 },
@ -1975,7 +2011,7 @@ export function pluginLoader(
log.info(
{
total: readyPlugins.length,
total: pluginsToLoad.length,
succeeded,
failed,
},
@ -1983,7 +2019,7 @@ export function pluginLoader(
);
return {
total: readyPlugins.length,
total: pluginsToLoad.length,
succeeded,
failed,
results: loadResults,
@ -2447,8 +2483,9 @@ export function pluginLoader(
);
if (options.markErrorOnFailure) {
// Mark the plugin as errored in the database so it is not retried
// automatically on next startup without operator intervention.
// Mark the plugin as errored in the database. The running process
// leaves it inactive; the next boot's loadAll retries it once, and the
// lifecycle enable() path can revive it sooner by hand.
// markError also deactivates the plugin runtime in this process.
try {
await lifecycleManager.markError(pluginId, `Activation failed: ${errorMessage}`);