fix(sessions): freeze providers after global bootstrap

This commit is contained in:
Federico Jaramillo Martinez
2026-07-22 20:31:08 +02:00
parent c2bf595999
commit 04d8134ba4
4 changed files with 281 additions and 258 deletions
+108 -93
View File
@@ -1,110 +1,125 @@
import { mkdtemp, rm } from "node:fs/promises";
import { tmpdir } from "node:os";
import { join } from "node:path";
import { createAgentSessionServices, type ModelRuntime } from "@earendil-works/pi-coding-agent";
import {
createAgentSessionServices,
type AgentSessionRuntimeDiagnostic,
type AgentSessionServices,
type ModelRuntime,
} from "@earendil-works/pi-coding-agent";
/**
* PI WEB providers come from global sources only: Pi built-ins, agent-dir
* `models.json`, environment credentials, and providers registered by global
* (agent-dir) extensions. Provider registrations from project extensions are
* rejected — the user is told, and everything else the extension does keeps
* working.
*
* Why: all sessions share one daemon-wide {@link ModelRuntime}. Global
* extensions are identical for every session, so their providers are
* daemon-wide consistent; project extensions differ per workspace, and letting
* them mutate the shared runtime corrupts the provider set of every other
* concurrent session (issue #76).
*
* Mechanism: this deliberately shadows the `registerProvider` /
* `registerNativeProvider` / `unregisterProvider` instance methods because Pi
* 0.81.1 offers no registration hook with extension attribution (the internal
* drain, the bind-time flush, and the late `pi.registerProvider` path all
* reach the runtime through call-time property lookup with the extension path
* already dropped, so instance shadowing intercepts them identically). The
* acceptance suite that exercises the load-time and late paths is the
* tripwire: if a Pi upgrade changes these internals, those tests fail loudly
* and this shim must be revisited.
*
* Attribution: Pi drops the registering extension's path before calls reach
* the runtime, so the shim cannot tell global from project extensions per
* call. Instead the daemon learns once, at startup, which provider ids global
* extensions register ({@link learnGlobalExtensionProviderIds}) and allows
* exactly those. This is not a security boundary — extensions run in-process
* with full trust, and a project extension re-registering an allowed id would
* pass. It is a guard against accidental cross-workspace leakage.
*/
export function installGlobalProviderPolicy(
runtime: ModelRuntime,
allowedExtensionProviderIds: ReadonlySet<string>,
onRejection: (providerId: string) => void,
): void {
const registerProvider = runtime.registerProvider.bind(runtime);
const registerNativeProvider = runtime.registerNativeProvider.bind(runtime);
const unregisterProvider = runtime.unregisterProvider.bind(runtime);
runtime.registerProvider = (providerId, config) => {
if (allowedExtensionProviderIds.has(providerId)) {
registerProvider(providerId, config);
return;
}
// Swallow: the shared runtime is never mutated, the rejection is surfaced.
onRejection(providerId);
};
runtime.registerNativeProvider = (provider) => {
if (allowedExtensionProviderIds.has(provider.id)) {
registerNativeProvider(provider);
return;
}
onRejection(provider.id);
};
runtime.unregisterProvider = (providerId) => {
if (allowedExtensionProviderIds.has(providerId)) {
unregisterProvider(providerId);
return;
}
// No-op: rejected registrations never reached the runtime, so there is
// never anything of theirs to unregister.
};
/** Structured logging boundary supplied by the session daemon. */
export interface GlobalProviderBootstrapLogger {
error(details: Record<string, unknown>, message: string): void;
info(details: Record<string, unknown>, message: string): void;
warn(details: Record<string, unknown>, message: string): void;
}
/**
* Learn which provider ids the agent directory's global extensions register.
*
* Loads extensions once for a guaranteed-empty temporary cwd — so only global
* (agent-dir) extensions load — against the daemon's shared runtime, and
* returns the ids that appeared. Those registrations stay on the runtime: they
* are the daemon baseline, re-registered identically on every session load.
* Must run before {@link installGlobalProviderPolicy} is installed.
*
* Caveat: global extension code runs one extra time at daemon startup. Any
* commands/tools it registers land on the discarded scratch loader.
*/
export async function learnGlobalExtensionProviderIds(
runtime: ModelRuntime,
agentDir: string,
): Promise<ReadonlySet<string>> {
const before = new Set(runtime.getRegisteredProviderIds());
type ProviderMutationOperation = "registerNativeProvider" | "registerProvider" | "unregisterProvider";
type ProviderMutationMethods = Pick<ModelRuntime, ProviderMutationOperation>;
const LOG_CONTEXT = "global-provider-bootstrap";
async function loadGlobalExtensionServices(runtime: ModelRuntime, agentDir: string): Promise<AgentSessionServices> {
const scratchCwd = await mkdtemp(join(tmpdir(), "pi-web-global-ext-"));
try {
await createAgentSessionServices({ cwd: scratchCwd, agentDir, modelRuntime: runtime });
return await createAgentSessionServices({ cwd: scratchCwd, agentDir, modelRuntime: runtime });
} finally {
await rm(scratchCwd, { recursive: true, force: true });
}
const learned = new Set<string>();
for (const providerId of runtime.getRegisteredProviderIds()) {
if (!before.has(providerId)) learned.add(providerId);
}
function logBootstrapDiagnostic(
logger: GlobalProviderBootstrapLogger,
diagnostic: AgentSessionRuntimeDiagnostic,
): void {
const details = {
context: LOG_CONTEXT,
diagnosticType: diagnostic.type,
diagnostic: diagnostic.message,
};
if (diagnostic.type === "error") {
logger.error(details, "global extension provider bootstrap diagnostic");
} else if (diagnostic.type === "warning") {
logger.warn(details, "global extension provider bootstrap diagnostic");
} else {
logger.info(details, "global extension provider bootstrap diagnostic");
}
}
function freezeProviderMutations(runtime: ModelRuntime, logger: GlobalProviderBootstrapLogger): void {
const originalMethods: ProviderMutationMethods = {
registerNativeProvider: runtime.registerNativeProvider.bind(runtime),
registerProvider: runtime.registerProvider.bind(runtime),
unregisterProvider: runtime.unregisterProvider.bind(runtime),
};
const loggedProviderIds: Record<ProviderMutationOperation, Set<string>> = {
registerNativeProvider: new Set(),
registerProvider: new Set(),
unregisterProvider: new Set(),
};
const logIgnoredMutation = (operation: ProviderMutationOperation, providerId: string): void => {
const loggedIds = loggedProviderIds[operation];
if (loggedIds.has(providerId)) return;
loggedIds.add(providerId);
logger.info(
{ context: LOG_CONTEXT, operation, providerId },
"ignored provider mutation after global bootstrap",
);
};
const frozenMethods: ProviderMutationMethods = {
registerProvider(providerId) {
logIgnoredMutation("registerProvider", providerId);
},
registerNativeProvider(provider) {
logIgnoredMutation("registerNativeProvider", provider.id);
},
unregisterProvider(providerId) {
logIgnoredMutation("unregisterProvider", providerId);
},
};
try {
Object.assign(runtime, frozenMethods);
} catch (error: unknown) {
Object.assign(runtime, originalMethods);
throw error;
}
return learned;
}
/**
* User-facing wording for a rejected registration. `cwd` is known only for
* rejections raised while loading a workspace's services; late registrations
* from session event handlers carry no attribution beyond the provider id.
* Load global extensions once against the shared model runtime, then make its
* extension-provider baseline immutable for the rest of the daemon lifetime.
* All sessions share this runtime, so accepting project-dependent mutations
* would leak provider configuration across workspaces. This is an accidental
* contamination guard, not a sandbox for otherwise trusted extensions.
*
* The temporary cwd is guaranteed to be empty, so Pi discovers agent-dir
* extensions without loading project resources. Documented initialization-time
* config and native registrations therefore reach the runtime through Pi's
* public service factory. Pi exposes no provider-freeze hook, so the daemon
* deliberately shadows the three public instance mutation methods afterward;
* every registration replay or later call is then a logged no-op.
*/
export function providerRejectionMessage(providerId: string, cwd?: string): string {
const origin = cwd === undefined ? "registered by an extension" : `registered by an extension in ${cwd}`;
return `Provider "${providerId}" ${origin} was ignored — PI WEB providers must come from global configuration `
+ "(agent-dir models.json) or a globally installed extension. All other extension features are unaffected.";
export async function bootstrapAndFreezeGlobalExtensionProviders(
runtime: ModelRuntime,
agentDir: string,
logger: GlobalProviderBootstrapLogger,
): Promise<void> {
const services = await loadGlobalExtensionServices(runtime, agentDir);
const providerIds = Object.freeze([...runtime.getRegisteredProviderIds()].sort());
freezeProviderMutations(runtime, logger);
for (const diagnostic of services.diagnostics) logBootstrapDiagnostic(logger, diagnostic);
for (const extensionError of services.resourceLoader.getExtensions().errors) {
logger.error(
{ context: LOG_CONTEXT, error: extensionError.error },
"global extension failed during provider bootstrap",
);
}
logger.info(
{ context: LOG_CONTEXT, providerIds },
"global extension provider baseline bootstrapped and frozen",
);
}