TypeScript
// @effect-diagnostics nodeBuiltinImport:off -- unit names and RAM size come from the host.
import * as NodeCrypto from "node:crypto";
import * as NodeOS from "node:os";
import {
AgentScope,
agentScopeCommand,
type AgentScopeShape,
type AgentSystemdScope,
} from "@t3tools/shared/AgentScope";
import * as HostProcess from "@t3tools/shared/HostProcess";
import { resolveCommandPath } from "@t3tools/shared/shell";
import * as Effect from "effect/Effect";
import * as FileSystem from "effect/FileSystem";
import * as Layer from "effect/Layer";
import * as Path from "effect/Path";
import * as ServerConfig from "../config.ts";
import * as ProcessRunner from "../processRunner.ts";
const AGENT_SLICE = "app-t3code-agents.slice";
// Owns every agent slice and competes with the server's unit or scope under app.slice.
const AGENTS_PARENT_SLICE = "app-t3code.slice";
// Low weights yield CPU and disk to the server under contention, where the user manager
// has those controllers: current systemd delegates cpu, rarely io. TasksMax keeps agents
// below logind's 33% per-user task limit, leaving forks for the server.
const AGENTS_PARENT_PROPERTIES = ["CPUWeight=10", "IOWeight=10", "TasksMax=25%"];
export function ownedAgentSlice(stateDir: string): string {
const hash = NodeCrypto.createHash("sha256").update(stateDir).digest("hex").slice(0, 12);
return `app-t3code-agents-${hash}.slice`;
}
const GiB = 1024 ** 3;
/**
* Agents together leave about 6 GB free before they are throttled and 4 GB
* before the kernel kills one, so the server, sshd, and tailscaled keep
* working. Small machines still give agents at least half of RAM before
* throttling and three quarters before a kill.
*/
function agentSliceMemoryLimits(totalBytes: number) {
return {
high: Math.max(totalBytes - 6 * GiB, Math.floor(totalBytes / 2)),
max: Math.max(totalBytes - 4 * GiB, Math.floor(totalBytes * 0.75)),
};
}
export interface ScopeState {
readonly loadState: string;
readonly activeState: string;
readonly result: string;
readonly oomKills: number | undefined;
readonly populated: boolean | undefined;
}
/** Pipes can close before systemd records oom-kill; stopping, empty or OOM-counted scopes remain pending. */
export function classifyScope(state: ScopeState): "oom-killed" | "gone" | "running" | "stopping" {
if (state.result === "oom-kill") return "oom-killed";
if (
state.loadState !== "loaded" ||
state.activeState === "inactive" ||
state.activeState === "failed"
) {
return "gone";
}
if (
state.activeState === "deactivating" ||
(state.oomKills ?? 0) > 0 ||
state.populated === false
) {
return "stopping";
}
return "running";
}
function parseKeyValues(text: string): Map<string, string> {
const values = new Map<string, string>();
for (const line of text.split("\n")) {
const index = line.indexOf("=");
if (index > 0) values.set(line.slice(0, index), line.slice(index + 1).trim());
}
return values;
}
function parseCgroupCounter(text: string, key: string): number | undefined {
const match = new RegExp(`^${key} (\\d+)$`, "m").exec(text);
return match === null ? undefined : Number(match[1]);
}
// systemd's default 90 s stop timeout would outlast our leftover-agent wait.
// Delegation lets rootless containers (Podman, kind) manage their own cgroups.
const SCOPE_OPTIONS = ["--property=TimeoutStopSec=10s", "--property=Delegate=yes"];
const STOPPING_POLL_INTERVAL = "100 millis";
const STOPPING_POLL_ATTEMPTS = 20;
const CLEAR_FAILED_INTERVAL = "1 second";
const CLEAR_FAILED_ATTEMPTS = 120;
const MAX_TRACKED_THREADS = 500;
export const make = Effect.gen(function* () {
const platform = yield* HostProcess.Platform;
if (platform !== "linux") return AgentScope.defaultValue();
const runner = yield* ProcessRunner.ProcessRunner;
const fileSystem = yield* FileSystem.FileSystem;
const path = yield* Path.Path;
const hostEnvironment = yield* HostProcess.Environment;
const { stateDir } = yield* ServerConfig.ServerConfig;
const ownedSlice = ownedAgentSlice(stateDir);
const layerScope = yield* Effect.scope;
const resolve = (command: string, env: NodeJS.ProcessEnv) =>
resolveCommandPath(command, { env }).pipe(
Effect.provideService(FileSystem.FileSystem, fileSystem),
Effect.provideService(Path.Path, path),
Effect.provideService(HostProcess.Platform, platform),
);
const run = (command: string, args: ReadonlyArray<string>) =>
runner.run({ command, args, timeout: "5 seconds" }).pipe(Effect.option);
const readCgroupFile = (controlGroup: string, file: string) =>
fileSystem
.readFileString(path.join("/sys/fs/cgroup", controlGroup, file))
.pipe(Effect.orElseSucceed(() => ""));
// Old agents keep their scopes when the server is killed, and the restarted
// server resumes their sessions. Stop them first, so two copies of one
// session never work in the same worktree.
const stopLeftoverAgents = Effect.fn("agentScope.stopLeftoverAgents")(function* (
systemctl: string,
) {
const shown = yield* run(systemctl, [
"--user",
"show",
ownedSlice,
"--property=ActiveState,ControlGroup",
]);
if (shown._tag === "None") return;
const values = parseKeyValues(shown.value.stdout);
if (values.get("ActiveState") !== "active") return;
const controlGroup = values.get("ControlGroup") ?? "";
const events = controlGroup ? yield* readCgroupFile(controlGroup, "cgroup.events") : "";
if (parseCgroupCounter(events, "populated") === 0) return;
yield* Effect.logWarning("Stopping agents left by a previous server", { slice: ownedSlice });
const stopped = yield* runner
.run({ command: systemctl, args: ["--user", "stop", ownedSlice], timeout: "30 seconds" })
.pipe(Effect.option);
if (stopped._tag === "None" || stopped.value.code !== 0) {
yield* Effect.logWarning("Could not stop agents left by a previous server", {
slice: ownedSlice,
stderr: stopped._tag === "Some" ? stopped.value.stderr.trim() : undefined,
});
}
});
const systemd = yield* Effect.cached(
Effect.gen(function* () {
const runtimeDir = hostEnvironment.XDG_RUNTIME_DIR;
if (runtimeDir === undefined || runtimeDir.length === 0) return undefined;
const systemdRun = yield* resolve("systemd-run", hostEnvironment);
const systemctl = yield* resolve("systemctl", hostEnvironment);
if (systemdRun === null || systemctl === null) return undefined;
yield* stopLeftoverAgents(systemctl);
const probe = (options: ReadonlyArray<string>) =>
run(systemdRun, [
"--user",
"--scope",
"--quiet",
"--collect",
`--slice=${AGENT_SLICE}`,
...options,
"--",
"true",
]).pipe(Effect.map((result) => result._tag === "Some" && result.value.code === 0));
// Scopes take OOMPolicy from systemd 253; older managers still get scopes.
const supportsOomPolicy = yield* probe([...SCOPE_OPTIONS, "--property=OOMPolicy=stop"]);
if (!supportsOomPolicy && !(yield* probe(SCOPE_OPTIONS))) {
yield* Effect.logInfo("Agent scopes are off: no systemd user manager");
return undefined;
}
const setProperties = Effect.fn("agentScope.setProperties")(function* (
slice: string,
properties: ReadonlyArray<string>,
) {
const result = yield* run(systemctl, [
"--user",
"set-property",
"--runtime",
slice,
...properties,
]);
if (result._tag === "None" || result.value.code !== 0) {
yield* Effect.logWarning("Could not set resource limits on an agents slice", {
slice,
properties,
stderr: result._tag === "Some" ? result.value.stderr.trim() : undefined,
});
}
});
const limits = agentSliceMemoryLimits(NodeOS.totalmem());
yield* setProperties(AGENT_SLICE, [`MemoryHigh=${limits.high}`, `MemoryMax=${limits.max}`]);
yield* setProperties(AGENTS_PARENT_SLICE, AGENTS_PARENT_PROPERTIES);
yield* Effect.logInfo("Agent scopes are on", {
slice: AGENT_SLICE,
...limits,
oomPolicy: supportsOomPolicy,
});
return {
systemdRun,
systemctl,
runtimeDir,
supportsOomPolicy,
};
}),
);
// `shared` scopes serve other threads too: tracked only for OOM diagnosis on old systemd.
const scopeByThread = new Map<string, { unit: string; oomKilled: boolean; shared: boolean }>();
const readScope = Effect.fn("agentScope.readScope")(function* (systemctl: string, unit: string) {
const shown = yield* run(systemctl, [
"--user",
"show",
unit,
"--property=LoadState,ActiveState,Result,ControlGroup",
]);
if (shown._tag === "None" || shown.value.code !== 0) return undefined;
const values = parseKeyValues(shown.value.stdout);
const controlGroup = values.get("ControlGroup") ?? "";
const memoryEvents = controlGroup ? yield* readCgroupFile(controlGroup, "memory.events") : "";
const cgroupEvents = controlGroup ? yield* readCgroupFile(controlGroup, "cgroup.events") : "";
const populated = parseCgroupCounter(cgroupEvents, "populated");
return {
loadState: values.get("LoadState") ?? "",
activeState: values.get("ActiveState") ?? "",
result: values.get("Result") ?? "",
oomKills: parseCgroupCounter(memoryEvents, "oom_kill"),
populated: populated === undefined ? undefined : populated === 1,
} satisfies ScopeState;
});
const clearFailed = Effect.fn("agentScope.clearFailed")(function* (
systemctl: string,
unit: string,
) {
for (let attempt = 0; attempt < CLEAR_FAILED_ATTEMPTS; attempt++) {
const state = yield* readScope(systemctl, unit);
if (state === undefined) return;
if (state.activeState === "failed") {
yield* run(systemctl, ["--user", "reset-failed", unit]);
return;
}
if (state.activeState !== "deactivating") return;
yield* Effect.sleep(CLEAR_FAILED_INTERVAL);
}
});
const forget = (systemctl: string, unit: string) =>
clearFailed(systemctl, unit).pipe(Effect.forkIn(layerScope), Effect.asVoid);
const evictOldest = Effect.fn("agentScope.evictOldest")(function* (systemctl: string) {
const oldest = scopeByThread.entries().next();
if (oldest.done === true) return;
const [threadId, tracked] = oldest.value;
const state = yield* readScope(systemctl, tracked.unit);
if (scopeByThread.get(threadId) !== tracked) return;
scopeByThread.delete(threadId);
const status = state === undefined ? "running" : classifyScope(state);
if (status === "running" || status === "stopping") {
scopeByThread.set(threadId, tracked);
return;
}
yield* forget(systemctl, tracked.unit);
});
const track = Effect.fn("agentScope.track")(function* (
systemctl: string,
threadId: string,
unit: string | undefined,
shared = false,
) {
const previous = scopeByThread.get(threadId);
scopeByThread.delete(threadId);
if (unit !== undefined) scopeByThread.set(threadId, { unit, oomKilled: false, shared });
if (previous !== undefined) yield* forget(systemctl, previous.unit);
if (scopeByThread.size > MAX_TRACKED_THREADS) yield* evictOldest(systemctl);
});
const oomKilled: AgentScopeShape["oomKilled"] = Effect.fn("agentScope.oomKilled")(
function* (threadId) {
const tracked = scopeByThread.get(threadId);
if (tracked === undefined) return false;
if (tracked.oomKilled) return true;
const scope = yield* systemd;
if (scope === undefined) return false;
for (let attempt = 0; attempt < STOPPING_POLL_ATTEMPTS; attempt++) {
const state = yield* readScope(scope.systemctl, tracked.unit);
if (state === undefined) return false;
const status = classifyScope(state);
if (status === "running") return false;
if (status === "stopping") {
yield* Effect.sleep(STOPPING_POLL_INTERVAL);
continue;
}
yield* forget(scope.systemctl, tracked.unit);
if (status === "gone") return false;
yield* Effect.logWarning("Agent scope was killed: out of memory", {
threadId,
unit: tracked.unit,
});
tracked.oomKilled = true;
return true;
}
return false;
},
);
const stop = Effect.fn("AgentScope.stop")(function* (unit: string) {
if (!unit.startsWith("t3code-") || !unit.endsWith(".scope")) return;
const scope = yield* systemd;
if (scope === undefined) return;
const state = yield* readScope(scope.systemctl, unit);
if (
state === undefined ||
state.loadState !== "loaded" ||
state.activeState !== "active" ||
state.result === "oom-kill" ||
state.populated === false
)
return;
// The kernel counts OOM kills before signaling the victim. Let OOMPolicy
// record the result before a manual stop can overwrite it; an unreadable
// counter cannot rule out that race. Older managers need us to stop it.
if (scope.supportsOomPolicy && (state.oomKills === undefined || state.oomKills > 0)) return;
const stopped = yield* run(scope.systemctl, ["--user", "stop", "--no-block", unit]);
if (stopped._tag === "None" || stopped.value.code !== 0) {
yield* Effect.logWarning("Could not stop an exited agent's scope", {
unit,
exitCode: stopped._tag === "Some" ? stopped.value.code : undefined,
});
}
});
const kill: AgentScopeShape["kill"] = Effect.fn("AgentScope.kill")(function* (threadId) {
const tracked = scopeByThread.get(threadId);
if (tracked === undefined || tracked.shared) return false;
const scope = yield* systemd;
if (scope === undefined) return false;
const shown = yield* run(scope.systemctl, [
"--user",
"show",
tracked.unit,
"--property=ActiveState,ControlGroup",
]);
if (shown._tag === "None" || shown.value.code !== 0) return false;
const values = parseKeyValues(shown.value.stdout);
if (values.get("ActiveState") !== "active") return false;
const controlGroup = values.get("ControlGroup") ?? "";
// cgroup.kill (Linux 5.14+) kills the whole tree at once; systemd's kill is the fallback.
const killedByCgroup =
controlGroup.length > 0 &&
(yield* fileSystem
.writeFileString(path.join("/sys/fs/cgroup", controlGroup, "cgroup.kill"), "1")
.pipe(
Effect.as(true),
Effect.orElseSucceed(() => false),
));
if (killedByCgroup) return true;
const killed = yield* run(scope.systemctl, [
"--user",
"kill",
"--signal=SIGKILL",
tracked.unit,
]);
return killed._tag === "Some" && killed.value.code === 0;
});
const clear: AgentScopeShape["clear"] = Effect.fn("agentScope.clear")(function* (threadId) {
if (!scopeByThread.has(threadId)) return;
const scope = yield* systemd;
if (scope !== undefined) yield* track(scope.systemctl, threadId, undefined);
});
return {
wrap: Effect.fn("AgentScope.wrap")(function* ({
command,
args,
name,
threadId,
env,
survivesRestart,
shared,
}) {
const resolved = yield* resolve(command, env ?? hostEnvironment);
if (resolved === null) return { command, args };
const scope = yield* systemd;
if (scope === undefined) return agentScopeCommand({ command: resolved, args });
const unit = `t3code-${name}-${NodeCrypto.randomUUID().slice(0, 8)}.scope`;
// A continuing shared scope cannot attribute a tool's OOM to one thread.
// Clear the launch thread's previous scope instead.
const continuing = shared === true && scope.supportsOomPolicy;
const owned = threadId !== undefined && !continuing;
if (threadId !== undefined)
yield* track(scope.systemctl, threadId, owned ? unit : undefined, shared === true);
const systemdScope: AgentSystemdScope = {
systemdRun: scope.systemdRun,
runtimeDir: scope.runtimeDir,
slice: survivesRestart === true ? AGENT_SLICE : ownedSlice,
unit,
// Without an owning thread nothing reads the result, so let systemd
// unload the unit even when it fails.
options: [
...SCOPE_OPTIONS,
...(scope.supportsOomPolicy
? [`--property=OOMPolicy=${continuing ? "continue" : "stop"}`]
: []),
...(owned ? [] : ["--collect"]),
],
};
return { ...agentScopeCommand({ command: resolved, args, scope: systemdScope }), unit };
}),
oomKilled,
stop,
kill,
clear,
} satisfies AgentScopeShape;
});
export const layer = Layer.effect(AgentScope, make).pipe(Layer.provide(ProcessRunner.layer));