@phnx-labs/agents-cli 1.22.99 → 1.22.101

package/CHANGELOG.md

@@ -1,5 +1,17 @@
 # Changelog
 
+## 1.22.101
+
+- **`agents sync status` names what drifted, and a merged system-layer skill or subagent reaches the next `agents run` on every box (PHNX-4056).** The status readout said `11 drifted` per install and nothing more, although `AgentVersionStatus.resources` already carried the rows; both human renderers (`agents sync status`, the drift summary `agents sync` prints) now list each drifted or missing resource as `drifted skills/artifacts` under its install line, with the differ's detail when it has one. And when `~/.agents/.system/skills` or `subagents` changed since the last launch of an (agent, version), `agents run` rewrites those resources into the version home through the registered writers before spawning, filtered by the active resource profile like the full sync, with the same stat-tier sentinel shape as the rules run-sync: on 2026-09-10 a merged skill change left 31 of 33 installed copies stale until the 4-hour umbrella autosync, because the shim's `--launch` path mirrors project resources only. Source: `cli/src/lib/system-run-sync.ts`, `cli/src/lib/sync-status.ts`, `cli/src/commands/status.ts`, `cli/src/lib/drift-sync.ts`, `cli/src/lib/exec.ts`, `cli/src/commands/exec.ts`.
+
+## 1.22.100
+
+- **The daemon's 20-minute auth-probe reuse works again on a box that has uninstalled a version (PHNX-4051).** `isCachedFleetAuthProbeFresh` required EVERY cached auth row to be under 20 minutes old, including rows left behind for versions that are no longer installed. Nothing re-probes those — the probe only walks homes that exist — so their timestamp froze and the check was permanently false: the 3-minute tick live-probed the rate-limited `/api/oauth/usage` for every account on every tick, re-arming the per-account 429 backoff and keeping the usage refresher parked (the RUSH-2998 failure the reuse window exists to prevent; observed on a worker holding rows dated Sep 2 / Sep 6 for uninstalled Claude versions). Freshness now considers only rows for currently installed `(agent, version)` homes and account slots, and the tick prunes the orphan rows when it writes the cache, so `agents view` and fleet status stop reporting a verdict for a version that is gone. `agents devices ping [--strict]` still forces a genuinely live probe. Source: `cli/src/lib/daemon-ticks.ts` (`installedAuthRows`, `isCachedFleetAuthProbeFresh`, `refreshLocalFleetAuthState`), `cli/src/lib/auth-health.ts` (`authTargetKey`, `enumerateLocalAuthInstalls`, `localAuthTargetKeys`, `writeFleetAuthRows`).
+
+- **Fleet usage-sync transport no longer starves `--device auto` (PHNX-4051).** The daemon's shared-state git exchange failed on every tick two ways, so workers never received a fresh Claude usage snapshot and the 40-min placement gate turned every worker into "no ready device". Fixed both at the source: (1) the usage-sync tick is now the single committer — it also publishes the reserved-auth verdict, and auth-sync no longer runs its own `syncFleetSharedStateRepo`, so the two ticks can't contend for the one shared-repo lock ("Lock file is already being held"); (2) the untracked-collision scan is scoped to the tracked roots origin carries (`git ls-tree origin/<branch>`) — every path origin tracks stays in scope, so no path the rebase can overwrite is missed, while an untracked-only tree origin never tracks (like artifacts-cli's `~/.agents/artifact-history/`) can never push `git ls-files --others` past the 1 MiB output cap. auth-sync's credential pushes are also gated on the exchange's freshness marker, so they never push off peer state older than one tick interval. Source: `cli/src/lib/fleet-shared-repo-sync.ts`, `cli/src/lib/daemon/usage-sync-service.ts`, `cli/src/lib/daemon/auth-sync-service.ts`.
+
+- **An untitled Arc Space no longer hides every Arc Space from `agents browser`.** `parseSidebarSpaces` treated the display-only Space title as part of a Space's identity and rejected the whole sidebar when one Space had none, so `agents browser start` warned that a configured `arc-gmail` "can't launch on this machine" while `agents browser profiles doctor arc-gmail` saw Arc and the Space. Identity is the stable Space id and its profile binding; a Space Arc stored without a title (key absent or `null`) is now listed as `arc-space` (two or more get their Space-id prefix appended, as same-titled Spaces already do). A mismatched id, a duplicate id, an unknown or malformed profile mapping, and a title that is present but not text still fail discovery closed. Source: `cli/src/lib/browser/arc-discovery.ts`, `cli/src/lib/browser/arc-discovery.test.ts`, `cli/src/lib/browser/testdata/arc/`.
+
 ## 1.22.99
 
 - **`agents setup secrets` installs the published secrets CLI instead of only printing a hint (PHNX-3989).** A missing `secrets` binary tries `agents clis install secrets` then a pinned `npm i -g @phnx-labs/secrets-cli@0.1.2`, then hands off to `secrets migrate`. Docs point at the standalone engine; the broker is `secrets _agent-run` on macOS, not the agents daemon. Source: `cli/src/commands/setup-secrets.ts`, `cli/docs/secrets.md`.

package/dist/commands/exec.js

@@ -26,6 +26,7 @@ import { spawnSync } from 'child_process';
 import { randomUUID } from 'crypto';
 import { isSessionTrackedAgent } from '../lib/session/types.js';
 import { applyActiveRulesPresetAtRun } from '../lib/rules/run-sync.js';
+import { applySystemResourcesAtRun } from '../lib/system-run-sync.js';
 import { handleBroadcast } from './run-broadcast.js';
 import { bootMark } from '../lib/boot-profile.js';
 /** Validate a caller-supplied session id before it reaches tmux, paths, indexes, or remote dispatch. */
@@ -2714,6 +2715,7 @@ agents run auto --device yosemite-s0 "fix the flaky test"   # pin the device
         // path itself; there's no version home to sync into.
         if (defaultVersion) {
             applyActiveRulesPresetAtRun(agent, defaultVersion, getVersionHomePath(agent, defaultVersion));
+            applySystemResourcesAtRun(agent, defaultVersion, getVersionHomePath(agent, defaultVersion));
         }
         bootMark('rules-sync:done');
         // Login preflight (advisory, warn + continue). On a local INTERACTIVE

package/dist/commands/status.js

@@ -13,7 +13,7 @@ import chalk from 'chalk';
 import { AGENTS } from '../lib/agents.js';
 import { setHelpSections } from '../lib/help.js';
 import { addHostOption } from '../lib/hosts/option.js';
-import { computeSyncStatus } from '../lib/sync-status.js';
+import { computeSyncStatus, formatDriftRows } from '../lib/sync-status.js';
 import { promptDriftSync } from '../lib/drift-sync.js';
 import { resolveConfiguredModel } from '../lib/models.js';
 import { resolveSurface } from './utils.js';
@@ -93,6 +93,8 @@ export function registerStatusCommand(syncCmd) {
                 + (v.isDefault ? chalk.gray(' (default)') : '');
             const pad = ' '.repeat(Math.max(1, 34 - plain.length));
             console.log(`  ${shown}${pad}${versionSummary(v)}`);
+            for (const line of formatDriftRows(v))
+                console.log(chalk.gray(`      ${line}`));
         }
         if (status.totals.orphan > 0) {
             console.log(chalk.gray(`  (${status.totals.orphan} orphan${status.totals.orphan === 1 ? '' : 's'} — run \`agents prune cleanup\`)`));

package/dist/lib/auth-health.d.ts

@@ -134,6 +134,12 @@ export declare function formatCheckedAge(checkedAt: number, now?: number): strin
 export declare function authAccountLabel(info: Pick<AccountInfo, 'email' | 'accountId' | 'userId'> | null | undefined): string | undefined;
 /** Cache key: one entry per install — (host, agent, version). Unique per token. */
 export declare function authCacheKey(host: string, agent: AgentId | string, version: string): string;
+/**
+ * Host-independent identity of one probe target — the (agent, version) pair
+ * {@link authCacheKey} is keyed by. The separator cannot appear in either half,
+ * so an agent id can never run into a version the way a `:` join allows.
+ */
+export declare function authTargetKey(agent: AgentId | string, version: string): string;
 /**
  * The `version` slot an account SLOT occupies in the auth cache and the probe
  * rows: `slot:<accountId>`. A slot is a HOME-shaped dir, not an installed
@@ -156,10 +162,38 @@ export interface SlotAuthInstall extends FleetAuthInstall {
  * gone is skipped: there is nothing to probe and the row would only say so.
  */
 export declare function enumerateSlotInstalls(meta: Pick<Meta, 'accounts' | 'deviceAccounts'>, agentIds: readonly AgentId[]): SlotAuthInstall[];
+/** One probe target on this device: an installed version home, or an account slot. */
+export interface LocalAuthInstall extends FleetAuthInstall {
+    home: string;
+    accountId?: string;
+}
+/**
+ * Every (agent, version) home the local auth probe covers on this device —
+ * installed version homes plus account slots. {@link probeLocalFleetAuth} probes
+ * exactly this set, so it is also the answer to "which cached rows are still
+ * backed by something on disk" (PHNX-4051).
+ */
+export declare function enumerateLocalAuthInstalls(meta: Pick<Meta, 'accounts' | 'deviceAccounts'>, agentIds: readonly AgentId[]): LocalAuthInstall[];
+/**
+ * {@link authTargetKey} for every local probe target — what a cached row must
+ * match to still be about this device. A row outside it is an ORPHAN: its home
+ * was uninstalled, so nothing will ever re-probe it and its `checkedAt` is
+ * frozen at whatever the last probe left (PHNX-4051).
+ */
+export declare function localAuthTargetKeys(agentIds?: readonly AgentId[]): Set<string>;
 /** Read the whole cache (best-effort; a corrupt/missing file yields an empty map). */
 export declare function readAuthHealthCache(): Record<string, AuthHealth>;
 /** Read one entry, or null. */
 export declare function readAuthHealth(host: string, agent: AgentId | string, version: string): AuthHealth | null;
+/**
+ * Split a cache key back into the install it names, for one host. Returns null
+ * for a key that belongs to another host or does not name a known agent — the
+ * one place the `host:agent:version` join is undone.
+ */
+export declare function parseAuthCacheKey(key: string, host: string): {
+    agent: AgentId;
+    version: string;
+} | null;
 /** Reconstruct one host's published probe rows for a lease waiter/CLI reader. */
 export declare function readFleetAuthRows(host: string): AuthProbeRow[];
 /**
@@ -170,8 +204,15 @@ export declare function readFleetAuthRows(host: string): AuthProbeRow[];
  * behaviour promised by the `error` doc on AuthVerdict ("keep the last known
  * one"). Pure, so it's unit-tested directly. */
 export declare function mergeAuthHealthEntries(current: Record<string, AuthHealth>, incoming: Record<string, AuthHealth>): Record<string, AuthHealth>;
-/** Merge one or more entries into the cache (best-effort write). */
-export declare function writeAuthHealthEntries(entries: Record<string, AuthHealth>): void;
+/**
+ * Merge one or more entries into the cache (best-effort write).
+ *
+ * `drop` removes entries the writer knows are gone — it runs on the PRE-merge
+ * cache, so an incoming row always wins over a drop of the same key and the two
+ * can never fight. Only a writer that knows the full truth for the keys it drops
+ * may pass one (see {@link writeFleetAuthRows}).
+ */
+export declare function writeAuthHealthEntries(entries: Record<string, AuthHealth>, drop?: (key: string) => boolean): void;
 /**
  * Complete a live auth probe for one (agent, home). For claude/kimi/droid this
  * hits the provider — unless the account's usage cache already holds a fresh
@@ -261,5 +302,16 @@ export declare function probeLocalFleetAuth(opts?: {
      */
     signal?: AbortSignal;
 }): Promise<AuthProbeRow[]>;
-/** Persist a host's probed rows into the cache (keyed by host+agent+version). */
-export declare function writeFleetAuthRows(host: string, rows: AuthProbeRow[]): void;
+/**
+ * Persist a host's probed rows into the cache (keyed by host+agent+version).
+ *
+ * `installed` — the {@link authTargetKey} set of homes that still exist on
+ * `host` ({@link localAuthTargetKeys}) — prunes this host's ORPHAN entries: rows
+ * for a version that has since been uninstalled. Nothing re-probes those, so
+ * they never age out on their own; they froze the reuse window permanently false
+ * and kept surfacing in `agents view` / fleet status (PHNX-4051). Only a caller
+ * that enumerated `host`'s real installs may pass it — the `fleet ping` fan-out
+ * writes a PEER's rows and cannot enumerate that box's homes, so it omits it and
+ * merges as before.
+ */
+export declare function writeFleetAuthRows(host: string, rows: AuthProbeRow[], installed?: ReadonlySet<string>): void;

package/dist/lib/auth-health.js

@@ -21,7 +21,7 @@ import * as path from 'path';
 import { ALL_AGENT_IDS, getAccountInfo } from './agents.js';
 import { listNativeAccounts } from './account-registry.js';
 import { readSlots } from './accounts/slots.js';
-import { getCacheDir } from './state.js';
+import { getCacheDir, readMeta } from './state.js';
 import { probeClaudeStatus, probeDroidStatus, probeKimiStatus, USAGE_HEADLESS_SCOPE_MARKER, readClaudeUsageCache, } from './accounting/usage.js';
 import { getVersionHomePath, listInstalledVersions } from './installations/versions.js';
 import { atomicWriteFileSync, ensureLockTarget, withFileLock } from './fs-atomic.js';
@@ -220,6 +220,14 @@ export function authAccountLabel(info) {
 export function authCacheKey(host, agent, version) {
     return `${host}:${agent}:${version}`;
 }
+/**
+ * Host-independent identity of one probe target — the (agent, version) pair
+ * {@link authCacheKey} is keyed by. The separator cannot appear in either half,
+ * so an agent id can never run into a version the way a `:` join allows.
+ */
+export function authTargetKey(agent, version) {
+    return `${agent}@${version}`;
+}
 /**
  * The `version` slot an account SLOT occupies in the auth cache and the probe
  * rows: `slot:<accountId>`. A slot is a HOME-shaped dir, not an installed
@@ -251,6 +259,32 @@ export function enumerateSlotInstalls(meta, agentIds) {
     }
     return out;
 }
+/**
+ * Every (agent, version) home the local auth probe covers on this device —
+ * installed version homes plus account slots. {@link probeLocalFleetAuth} probes
+ * exactly this set, so it is also the answer to "which cached rows are still
+ * backed by something on disk" (PHNX-4051).
+ */
+export function enumerateLocalAuthInstalls(meta, agentIds) {
+    const installs = [];
+    for (const agent of agentIds) {
+        for (const version of listInstalledVersions(agent)) {
+            installs.push({ agent, version, home: getVersionHomePath(agent, version), account: undefined });
+        }
+    }
+    for (const slot of enumerateSlotInstalls(meta, agentIds))
+        installs.push(slot);
+    return installs;
+}
+/**
+ * {@link authTargetKey} for every local probe target — what a cached row must
+ * match to still be about this device. A row outside it is an ORPHAN: its home
+ * was uninstalled, so nothing will ever re-probe it and its `checkedAt` is
+ * frozen at whatever the last probe left (PHNX-4051).
+ */
+export function localAuthTargetKeys(agentIds = ALL_AGENT_IDS) {
+    return new Set(enumerateLocalAuthInstalls(readMeta(), agentIds).map((i) => authTargetKey(i.agent, i.version)));
+}
 function cacheFilePath() {
     return path.join(getCacheDir(), '.auth-health.json');
 }
@@ -270,23 +304,34 @@ export function readAuthHealthCache() {
 export function readAuthHealth(host, agent, version) {
     return readAuthHealthCache()[authCacheKey(host, agent, version)] ?? null;
 }
+/**
+ * Split a cache key back into the install it names, for one host. Returns null
+ * for a key that belongs to another host or does not name a known agent — the
+ * one place the `host:agent:version` join is undone.
+ */
+export function parseAuthCacheKey(key, host) {
+    const prefix = `${host}:`;
+    if (!key.startsWith(prefix))
+        return null;
+    const identity = key.slice(prefix.length);
+    const separator = identity.indexOf(':');
+    if (separator <= 0)
+        return null;
+    const agent = identity.slice(0, separator);
+    if (!ALL_AGENT_IDS.includes(agent))
+        return null;
+    return { agent: agent, version: identity.slice(separator + 1) };
+}
 /** Reconstruct one host's published probe rows for a lease waiter/CLI reader. */
 export function readFleetAuthRows(host) {
-    const prefix = `${host}:`;
     const rows = [];
     for (const [key, health] of Object.entries(readAuthHealthCache())) {
-        if (!key.startsWith(prefix))
-            continue;
-        const identity = key.slice(prefix.length);
-        const separator = identity.indexOf(':');
-        if (separator <= 0)
-            continue;
-        const agent = identity.slice(0, separator);
-        if (!ALL_AGENT_IDS.includes(agent))
+        const install = parseAuthCacheKey(key, host);
+        if (!install)
             continue;
         rows.push({
-            agent: agent,
-            version: identity.slice(separator + 1),
+            agent: install.agent,
+            version: install.version,
             account: health.account,
             accountId: health.accountId,
             health,
@@ -310,15 +355,26 @@ export function mergeAuthHealthEntries(current, incoming) {
     }
     return merged;
 }
-/** Merge one or more entries into the cache (best-effort write). */
-export function writeAuthHealthEntries(entries) {
+/**
+ * Merge one or more entries into the cache (best-effort write).
+ *
+ * `drop` removes entries the writer knows are gone — it runs on the PRE-merge
+ * cache, so an incoming row always wins over a drop of the same key and the two
+ * can never fight. Only a writer that knows the full truth for the keys it drops
+ * may pass one (see {@link writeFleetAuthRows}).
+ */
+export function writeAuthHealthEntries(entries, drop) {
     try {
         const target = cacheFilePath();
         ensureLockTarget(target, JSON.stringify({ version: 1, entries: {} }));
         withFileLock(target, () => {
+            let current = readAuthHealthCache();
+            if (drop) {
+                current = Object.fromEntries(Object.entries(current).filter(([key]) => !drop(key)));
+            }
             const merged = {
                 version: 1,
-                entries: mergeAuthHealthEntries(readAuthHealthCache(), entries),
+                entries: mergeAuthHealthEntries(current, entries),
             };
             atomicWriteFileSync(target, JSON.stringify(merged, null, 2));
         });
@@ -447,20 +503,9 @@ export async function probeLocalFleetAuth(opts) {
     // each one's account label. getAccountInfo is a local credential-file read
     // (no network), so this fan-out is cheap and cannot contribute to the rate
     // limit the probe grouping below exists to avoid.
-    const [{ readMeta }, { findNativeAccountByIdentity }] = await Promise.all([
-        import('./state.js'),
-        import('./account-registry.js'),
-    ]);
+    const { findNativeAccountByIdentity } = await import('./account-registry.js');
     const meta = readMeta();
-    const installs = [];
-    for (const agent of agentIds) {
-        for (const version of listInstalledVersions(agent)) {
-            installs.push({ agent, version, home: getVersionHomePath(agent, version), info: null, account: undefined });
-        }
-    }
-    for (const slot of enumerateSlotInstalls(meta, agentIds)) {
-        installs.push({ ...slot, info: null });
-    }
+    const installs = enumerateLocalAuthInstalls(meta, agentIds).map((inst) => ({ ...inst, info: null }));
     await Promise.all(installs.map(async (inst) => {
         inst.info = await getAccountInfo(inst.agent, inst.home).catch(() => null);
         inst.account = authAccountLabel(inst.info);
@@ -491,11 +536,28 @@ export async function probeLocalFleetAuth(opts) {
     }));
     return perGroup.flat();
 }
-/** Persist a host's probed rows into the cache (keyed by host+agent+version). */
-export function writeFleetAuthRows(host, rows) {
+/**
+ * Persist a host's probed rows into the cache (keyed by host+agent+version).
+ *
+ * `installed` — the {@link authTargetKey} set of homes that still exist on
+ * `host` ({@link localAuthTargetKeys}) — prunes this host's ORPHAN entries: rows
+ * for a version that has since been uninstalled. Nothing re-probes those, so
+ * they never age out on their own; they froze the reuse window permanently false
+ * and kept surfacing in `agents view` / fleet status (PHNX-4051). Only a caller
+ * that enumerated `host`'s real installs may pass it — the `fleet ping` fan-out
+ * writes a PEER's rows and cannot enumerate that box's homes, so it omits it and
+ * merges as before.
+ */
+export function writeFleetAuthRows(host, rows, installed) {
     const entries = {};
     for (const row of rows) {
         entries[authCacheKey(host, row.agent, row.version)] = row.health;
     }
-    writeAuthHealthEntries(entries);
+    const drop = installed
+        ? (key) => {
+            const install = parseAuthCacheKey(key, host);
+            return install != null && !installed.has(authTargetKey(install.agent, install.version));
+        }
+        : undefined;
+    writeAuthHealthEntries(entries, drop);
 }

package/dist/lib/browser/arc-discovery.d.ts

@@ -1,7 +1,7 @@
 export interface ArcSpace {
     /** Stable UUID used for every native operation. */
     id: string;
-    /** Display-only title. */
+    /** Display-only title; empty when the Space is untitled. Never part of identity. */
     title: string;
 }
 export interface ArcProfile {

package/dist/lib/browser/arc-discovery.js

@@ -65,6 +65,13 @@ function parseProfileId(value, spaceId) {
     }
     return profileId;
 }
+function parseTitle(value, spaceId) {
+    if (value === undefined || value === null)
+        return '';
+    if (typeof value === 'string')
+        return value;
+    return { error: `Arc Space ${JSON.stringify(spaceId)} has a non-text title` };
+}
 export function parseSidebarSpaces(sidebarPath) {
     let parsed;
     try {
@@ -101,9 +108,12 @@ export function parseSidebarSpaces(sidebarPath) {
             if (typeof value.id !== 'string' || value.id !== encodedId) {
                 return { error: `Arc Space key ${JSON.stringify(encodedId)} does not match its stable id` };
             }
-            if (typeof value.title !== 'string') {
-                return { error: `Arc Space ${JSON.stringify(encodedId)} has no title` };
-            }
+            // Identity is the stable id and the profile binding. The title is
+            // display-only: Arc omits it for an untitled Space, which must still be
+            // listed (as `arc-space`) instead of taking every other Space down.
+            const title = parseTitle(value.title, encodedId);
+            if (typeof title !== 'string')
+                return title;
             if (ids.has(encodedId)) {
                 return { error: `Arc Space ${JSON.stringify(encodedId)} appears more than once` };
             }
@@ -111,7 +121,7 @@ export function parseSidebarSpaces(sidebarPath) {
             if (typeof profileId !== 'string')
                 return profileId;
             ids.add(encodedId);
-            spaces.push({ id: encodedId, title: value.title, profileId });
+            spaces.push({ id: encodedId, title, profileId });
         }
     }
     return spaces;

package/dist/lib/daemon/auth-sync-service.d.ts

@@ -1,10 +1,27 @@
 /**
  * Reserved `auth` bundle fleet sync as a `PeriodicService` (PHNX-2371).
  *
- * Each daemon publishes a safe readiness verdict to the fleet-shared user repo,
- * then runs the same serialized, timeout-bounded git exchange as usage sync.
- * One deterministic ready device asynchronously provisions peers whose delivered
- * verdict says `missing`; the secret never enters Git.
+ * This tick owns only the NON-git duties of auth sync: it materializes worker
+ * slots from durable keys already on the box, then (on the elected headed
+ * publisher) provisions peers whose LAST-DELIVERED verdict says `missing` by
+ * pushing the credential over SSH. The secret never enters Git.
+ *
+ * It no longer runs its own `syncFleetSharedStateRepo` (PHNX-4051). The verdict
+ * this tick's decisions read is published and delivered by the single git
+ * committer, the usage-sync tick — folding both publishes into one caller is
+ * what stops the two ticks (30 s apart) from contending for the one shared-repo
+ * lock and starving the usage snapshot workers depend on. This tick keeps its
+ * own deadline and circuit breaker, so a hung peer SSH push parks only auth-sync
+ * and never the usage delivery. The pushes read the peer verdicts the last
+ * usage-sync exchange wrote into the local checkout; they are idempotent
+ * (push only when a peer is missing a key), so acting on at-most-one-tick-old
+ * data converges exactly as the in-tick exchange did. To keep "at-most-one-tick-
+ * old" true, the pushes are gated on the exchange's freshness marker
+ * (`readLastSuccessfulExchangeMs`): when the last usage-sync exchange is missing
+ * or older than one usage-sync tick interval (`USAGE_SYNC_TICK_MS`, the
+ * producer's cadence — not this service's own `AUTH_SYNC_TICK_MS`), this tick
+ * skips the pushes and WARNs instead of acting on peer state that may no longer
+ * hold.
  */
 import { BasePeriodicService, type DaemonContext } from './service.js';
 import type { DaemonServiceId } from '../daemon-services.js';

package/dist/lib/daemon/auth-sync-service.js

@@ -1,12 +1,30 @@
 /**
  * Reserved `auth` bundle fleet sync as a `PeriodicService` (PHNX-2371).
  *
- * Each daemon publishes a safe readiness verdict to the fleet-shared user repo,
- * then runs the same serialized, timeout-bounded git exchange as usage sync.
- * One deterministic ready device asynchronously provisions peers whose delivered
- * verdict says `missing`; the secret never enters Git.
+ * This tick owns only the NON-git duties of auth sync: it materializes worker
+ * slots from durable keys already on the box, then (on the elected headed
+ * publisher) provisions peers whose LAST-DELIVERED verdict says `missing` by
+ * pushing the credential over SSH. The secret never enters Git.
+ *
+ * It no longer runs its own `syncFleetSharedStateRepo` (PHNX-4051). The verdict
+ * this tick's decisions read is published and delivered by the single git
+ * committer, the usage-sync tick — folding both publishes into one caller is
+ * what stops the two ticks (30 s apart) from contending for the one shared-repo
+ * lock and starving the usage snapshot workers depend on. This tick keeps its
+ * own deadline and circuit breaker, so a hung peer SSH push parks only auth-sync
+ * and never the usage delivery. The pushes read the peer verdicts the last
+ * usage-sync exchange wrote into the local checkout; they are idempotent
+ * (push only when a peer is missing a key), so acting on at-most-one-tick-old
+ * data converges exactly as the in-tick exchange did. To keep "at-most-one-tick-
+ * old" true, the pushes are gated on the exchange's freshness marker
+ * (`readLastSuccessfulExchangeMs`): when the last usage-sync exchange is missing
+ * or older than one usage-sync tick interval (`USAGE_SYNC_TICK_MS`, the
+ * producer's cadence — not this service's own `AUTH_SYNC_TICK_MS`), this tick
+ * skips the pushes and WARNs instead of acting on peer state that may no longer
+ * hold.
  */
 import { BasePeriodicService } from './service.js';
+import { USAGE_SYNC_TICK_MS } from './usage-sync-service.js';
 const AUTH_SYNC_TICK_MS = 15 * 60_000;
 const AUTH_SYNC_DEADLINE_MS = 2 * 60_000;
 const AUTH_SYNC_KICKOFF_MS = 60_000;
@@ -22,13 +40,14 @@ export class AuthSyncService extends BasePeriodicService {
         // Nothing to release — the supervisor's timer teardown is the only cleanup.
     }
     async onTick(ctx) {
-        const { publishReservedAuthVerdict, reconcileLocalWorkerSlots, syncReservedAuthBundle, syncReservedStores, } = await import('../secrets-policy.js');
+        const { reconcileLocalWorkerSlots, syncReservedAuthBundle, syncReservedStores, } = await import('../secrets-policy.js');
         // Worker-side slot materialization FIRST (PHNX-3940 T6): for each registered
         // account whose durable key is already on this box, create the HOME-shaped
         // slot the picker and spawn read. It touches only local state — the registry
-        // copy and the file-backed store — so it never waits on the git exchange
-        // below. It used to sit after `if (!transport.success) return`, so a single
-        // `git rebase timed out` postponed every slot on the box by another tick.
+        // copy and the file-backed store — so it never waits on any git exchange. It
+        // used to sit after this tick's own `if (!transport.success) return`, so a
+        // single `git rebase timed out` postponed every slot on the box by another
+        // tick; that exchange has since moved to the usage-sync tick (PHNX-4051).
         // Self-gated on device role: a headed box returns immediately.
         try {
             const slots = reconcileLocalWorkerSlots();
@@ -48,20 +67,29 @@ export class AuthSyncService extends BasePeriodicService {
         catch (err) {
             ctx.log('WARN', `auth-sync: worker slot reconcile: ${err.message}`);
         }
-        const published = await publishReservedAuthVerdict();
-        if (published.error)
-            ctx.log('WARN', `auth-sync: verdict: ${published.error}`);
-        const { syncFleetSharedStateRepo } = await import('../fleet-shared-repo-sync.js');
-        const transport = await syncFleetSharedStateRepo();
-        if (transport.skipped)
-            ctx.log('WARN', `auth-sync: ${transport.skipped}`);
-        if (transport.error)
-            ctx.log('WARN', `auth-sync: shared-store transport: ${transport.error}`);
-        if (transport.untrackedBackedUp?.length) {
-            ctx.log('WARN', `auth-sync: backed up ${transport.untrackedBackedUp.length} untracked shared-store collision(s) to ${transport.untrackedBackupDir}: ${transport.untrackedBackedUp.join(', ')}`);
-        }
-        if (!transport.success)
+        // The verdict this tick's pushes read is published by the usage-sync tick,
+        // the single git committer (PHNX-4051), and delivered into the local checkout
+        // by its exchange. The pushes below act on that last-delivered peer state, so
+        // they are only sound while that state is fresh. Gate them on the exchange's
+        // freshness: if the last successful usage-sync exchange is missing or older
+        // than one tick interval, the delivered peer verdicts may be stale — a peer
+        // that already received the key still reads `missing`, or a cleared `missing`
+        // still reads stale — so skip the pushes and WARN rather than pushing off it.
+        // The worker-slot reconcile above is NOT gated: it reads only local durable
+        // keys, never delivered peer verdicts.
+        const { readLastSuccessfulExchangeMs } = await import('../fleet-shared-repo-sync.js');
+        const lastExchangeMs = readLastSuccessfulExchangeMs();
+        const ageMs = lastExchangeMs === null ? null : Date.now() - lastExchangeMs;
+        // The threshold is the PRODUCER's cadence (USAGE_SYNC_TICK_MS), not this
+        // service's own AUTH_SYNC_TICK_MS: the delivered peer verdicts are refreshed
+        // once per usage-sync exchange, so "stale" is measured against that tick, not
+        // this one. The two constants are equal today, but sourcing it here keeps the
+        // gate tracking the usage-sync cadence if either is ever retuned alone.
+        if (ageMs === null || ageMs > USAGE_SYNC_TICK_MS) {
+            const age = ageMs === null ? 'never completed' : `last completed ${Math.round(ageMs / 1000)}s ago`;
+            ctx.log('WARN', `auth-sync: skipping credential push — usage-sync exchange ${age} (need one within ${Math.round(USAGE_SYNC_TICK_MS / 1000)}s); peer verdicts may be stale`);
             return;
+        }
         const result = await syncReservedAuthBundle();
         if (result.pushed.length > 0) {
             ctx.log('INFO', `auth-sync: pushed auth to ${result.pushed.join(', ')}`);

package/dist/lib/daemon/usage-sync-service.d.ts

@@ -2,19 +2,38 @@
  * Fleet shared-state sync as a `PeriodicService` (PHNX-3392 usage-sync,
  * PHNX-3792 session mirror).
  *
- * This is the one tick that owns the bounded Git exchange over the fleet-synced
+ * This is the ONE tick that owns the bounded Git exchange over the fleet-synced
  * user repo, so every non-secret daemon-state field rides it rather than opening
  * a second committer. Each tick: (1) publishes this box's own fields into its
  * conflict-free `devices/<device>/daemon-state.json` — a headed box's Claude
- * usage snapshot, and EVERY box's lightweight session digests (PHNX-3792);
- * (2) runs one serialized, timeout-bounded commit/rebase/push; (3) consumes the
- * peer fields the exchange delivered — a worker merges usage newest-wins, and
- * every non-worker box folds peers' session digests into its local index so the
- * picker renders remote-host previews inline. No tick opens a device-to-device
- * SSH mesh.
+ * usage snapshot, EVERY box's lightweight session digests (PHNX-3792), and the
+ * reserved-auth readiness verdict (PHNX-4051, folded in from the auth-sync tick
+ * so a single caller holds the shared-repo lock per tick); (2) runs one
+ * serialized, timeout-bounded commit/rebase/push; (3) consumes the peer fields
+ * the exchange delivered — a worker merges usage newest-wins, and every
+ * non-worker box folds peers' session digests into its local index so the picker
+ * renders remote-host previews inline. No tick opens a device-to-device SSH mesh.
+ *
+ * Why the auth verdict publishes here (PHNX-4051): auth-sync used to run its OWN
+ * `syncFleetSharedStateRepo`, so on every box two ticks 30 s apart contended for
+ * the one `proper-lockfile` lock (20×100 ms ≈ 2 s of retries) while a real
+ * fetch/rebase/push on a drifted repo runs far longer — the usage tick then
+ * failed with "Lock file is already being held" (zion logged it 95× in 24 h) and
+ * workers never received a fresh usage snapshot, which the 40-min placement gate
+ * turned into "no ready device". Folding the auth verdict into this single
+ * committer removes the second caller entirely. Auth-sync keeps its non-git
+ * duties (worker-slot reconcile + the credential SSH pushes) under its own
+ * deadline and circuit breaker.
  */
 import { BasePeriodicService, type DaemonContext } from './service.js';
 import type { DaemonServiceId } from '../daemon-services.js';
+/**
+ * The usage-sync tick cadence. Exported because auth-sync's credential-push
+ * freshness gate is the CONSUMER of this producer's cadence: it skips the
+ * pushes when the last exchange is older than one usage-sync interval, so it
+ * must track this constant rather than its own equal-by-coincidence literal.
+ */
+export declare const USAGE_SYNC_TICK_MS: number;
 export declare class UsageSyncService extends BasePeriodicService {
     readonly id: DaemonServiceId;
     readonly intervalMs: number;

package/dist/lib/daemon/usage-sync-service.js

@@ -2,19 +2,37 @@
  * Fleet shared-state sync as a `PeriodicService` (PHNX-3392 usage-sync,
  * PHNX-3792 session mirror).
  *
- * This is the one tick that owns the bounded Git exchange over the fleet-synced
+ * This is the ONE tick that owns the bounded Git exchange over the fleet-synced
  * user repo, so every non-secret daemon-state field rides it rather than opening
  * a second committer. Each tick: (1) publishes this box's own fields into its
  * conflict-free `devices/<device>/daemon-state.json` — a headed box's Claude
- * usage snapshot, and EVERY box's lightweight session digests (PHNX-3792);
- * (2) runs one serialized, timeout-bounded commit/rebase/push; (3) consumes the
- * peer fields the exchange delivered — a worker merges usage newest-wins, and
- * every non-worker box folds peers' session digests into its local index so the
- * picker renders remote-host previews inline. No tick opens a device-to-device
- * SSH mesh.
+ * usage snapshot, EVERY box's lightweight session digests (PHNX-3792), and the
+ * reserved-auth readiness verdict (PHNX-4051, folded in from the auth-sync tick
+ * so a single caller holds the shared-repo lock per tick); (2) runs one
+ * serialized, timeout-bounded commit/rebase/push; (3) consumes the peer fields
+ * the exchange delivered — a worker merges usage newest-wins, and every
+ * non-worker box folds peers' session digests into its local index so the picker
+ * renders remote-host previews inline. No tick opens a device-to-device SSH mesh.
+ *
+ * Why the auth verdict publishes here (PHNX-4051): auth-sync used to run its OWN
+ * `syncFleetSharedStateRepo`, so on every box two ticks 30 s apart contended for
+ * the one `proper-lockfile` lock (20×100 ms ≈ 2 s of retries) while a real
+ * fetch/rebase/push on a drifted repo runs far longer — the usage tick then
+ * failed with "Lock file is already being held" (zion logged it 95× in 24 h) and
+ * workers never received a fresh usage snapshot, which the 40-min placement gate
+ * turned into "no ready device". Folding the auth verdict into this single
+ * committer removes the second caller entirely. Auth-sync keeps its non-git
+ * duties (worker-slot reconcile + the credential SSH pushes) under its own
+ * deadline and circuit breaker.
  */
 import { BasePeriodicService } from './service.js';
-const USAGE_SYNC_TICK_MS = 15 * 60_000;
+/**
+ * The usage-sync tick cadence. Exported because auth-sync's credential-push
+ * freshness gate is the CONSUMER of this producer's cadence: it skips the
+ * pushes when the last exchange is older than one usage-sync interval, so it
+ * must track this constant rather than its own equal-by-coincidence literal.
+ */
+export const USAGE_SYNC_TICK_MS = 15 * 60_000;
 const USAGE_SYNC_DEADLINE_MS = 2 * 60_000;
 const USAGE_SYNC_KICKOFF_MS = 90_000;
 export class UsageSyncService extends BasePeriodicService {
@@ -54,6 +72,15 @@ export class UsageSyncService extends BasePeriodicService {
             ctx.log('INFO', `session-mirror: published ${mirrored.count} session digest(s)`);
         if (mirrored.error)
             ctx.log('WARN', `session-mirror: publish: ${mirrored.error}`);
+        // The reserved-auth readiness verdict rides this single git exchange too
+        // (PHNX-4051): it is a conflict-free field in the same owned daemon-state
+        // file, so publishing it here — instead of from a second committer in
+        // auth-sync — is what keeps exactly one caller of syncFleetSharedStateRepo on
+        // the periodic path.
+        const { publishReservedAuthVerdict } = await import('../secrets-policy.js');
+        const authVerdict = await publishReservedAuthVerdict();
+        if (authVerdict.error)
+            ctx.log('WARN', `usage-sync: auth verdict: ${authVerdict.error}`);
         const { syncFleetSharedStateRepo } = await import('../fleet-shared-repo-sync.js');
         const transport = await syncFleetSharedStateRepo();
         if (transport.skipped)

package/dist/lib/daemon-ticks.d.ts

@@ -36,11 +36,28 @@ export declare function isFreshFleetAuthSnapshot(value: {
  * publishes every tick — it does not ride that endpoint.
  */
 /**
- * True when every cached auth row was probed within {@link AUTH_PROBE_MAX_AGE_MS}
- * — i.e. reusing them would not let a verdict get staler than one probe window.
- * Empty cache is never fresh (nothing to reuse). Pure — unit-tested.
+ * The cached rows still backed by a home on this device, by `authTargetKey`.
+ *
+ * A row for an uninstalled version is an ORPHAN: the probe enumerates installed
+ * homes + account slots, so nothing re-probes it and its `checkedAt` is frozen
+ * at whatever the last probe left. Pure — unit-tested.
+ */
+export declare function installedAuthRows(authRows: readonly AuthProbeRow[], installedTargets: ReadonlySet<string>): AuthProbeRow[];
+/**
+ * True when every cached auth row FOR AN INSTALLED HOME was probed within
+ * {@link AUTH_PROBE_MAX_AGE_MS} — i.e. reusing them would not let a verdict get
+ * staler than one probe window. No installed row is never fresh (nothing to
+ * reuse). Pure — unit-tested.
+ *
+ * Orphan rows are excluded rather than counted stale (PHNX-4051). Counting them
+ * made this permanently false on any box that had ever uninstalled a version:
+ * the orphan can never be re-probed, so the tick live-probed the rate-limited
+ * `/api/oauth/usage` for every account every 3 minutes, re-arming the per-account
+ * 429 backoff and parking the usage refresher — the RUSH-2998 failure the reuse
+ * window exists to prevent. Observed on yosemite-m0 (rows dated Sep 2 / Sep 6 for
+ * uninstalled Claude versions).
  */
-export declare function isCachedFleetAuthProbeFresh(authRows: AuthProbeRow[], now: number, maxAgeMs?: number): boolean;
+export declare function isCachedFleetAuthProbeFresh(authRows: readonly AuthProbeRow[], now: number, installedTargets: ReadonlySet<string>, maxAgeMs?: number): boolean;
 /**
  * Whether the tick may reuse the cached auth verdict instead of re-probing the
  * rate-limited endpoint. `force` (an on-demand `agents devices ping`) ALWAYS
@@ -49,7 +66,7 @@ export declare function isCachedFleetAuthProbeFresh(authRows: AuthProbeRow[], no
  * revoked account (the second `runFleetPing` call site, RUSH-2998). Pure —
  * unit-tested so that inversion cannot regress unnoticed.
  */
-export declare function shouldReuseCachedAuthProbe(force: boolean, cached: AuthProbeRow[], now: number, maxAgeMs?: number): boolean;
+export declare function shouldReuseCachedAuthProbe(force: boolean, cached: readonly AuthProbeRow[], now: number, installedTargets: ReadonlySet<string>, maxAgeMs?: number): boolean;
 /**
  * Fleet cache warm: publish THIS host's row for the caches `agents fleet
  * status` / `agents devices list` read (PUBLISH-OWN / READ-UNION, RUSH-2061).

package/dist/lib/daemon-ticks.js

@@ -14,7 +14,7 @@
  * by the cross-process refresh lease so an explicit CLI refresh and the daemon
  * timer converge on the same published result.
  */
-import { AUTH_PROBE_MAX_AGE_MS } from './auth-health.js';
+import { AUTH_PROBE_MAX_AGE_MS, authTargetKey } from './auth-health.js';
 export { AUTH_PROBE_MAX_AGE_MS } from './auth-health.js';
 export function isFreshFleetAuthSnapshot(value, minimumCapturedAt) {
     return value.row.capturedAt >= minimumCapturedAt
@@ -36,12 +36,32 @@ export function isFreshFleetAuthSnapshot(value, minimumCapturedAt) {
  * publishes every tick — it does not ride that endpoint.
  */
 /**
- * True when every cached auth row was probed within {@link AUTH_PROBE_MAX_AGE_MS}
- * — i.e. reusing them would not let a verdict get staler than one probe window.
- * Empty cache is never fresh (nothing to reuse). Pure — unit-tested.
+ * The cached rows still backed by a home on this device, by `authTargetKey`.
+ *
+ * A row for an uninstalled version is an ORPHAN: the probe enumerates installed
+ * homes + account slots, so nothing re-probes it and its `checkedAt` is frozen
+ * at whatever the last probe left. Pure — unit-tested.
+ */
+export function installedAuthRows(authRows, installedTargets) {
+    return authRows.filter((r) => installedTargets.has(authTargetKey(r.agent, r.version)));
+}
+/**
+ * True when every cached auth row FOR AN INSTALLED HOME was probed within
+ * {@link AUTH_PROBE_MAX_AGE_MS} — i.e. reusing them would not let a verdict get
+ * staler than one probe window. No installed row is never fresh (nothing to
+ * reuse). Pure — unit-tested.
+ *
+ * Orphan rows are excluded rather than counted stale (PHNX-4051). Counting them
+ * made this permanently false on any box that had ever uninstalled a version:
+ * the orphan can never be re-probed, so the tick live-probed the rate-limited
+ * `/api/oauth/usage` for every account every 3 minutes, re-arming the per-account
+ * 429 backoff and parking the usage refresher — the RUSH-2998 failure the reuse
+ * window exists to prevent. Observed on yosemite-m0 (rows dated Sep 2 / Sep 6 for
+ * uninstalled Claude versions).
  */
-export function isCachedFleetAuthProbeFresh(authRows, now, maxAgeMs = AUTH_PROBE_MAX_AGE_MS) {
-    return authRows.length > 0 && authRows.every((r) => now - r.health.checkedAt < maxAgeMs);
+export function isCachedFleetAuthProbeFresh(authRows, now, installedTargets, maxAgeMs = AUTH_PROBE_MAX_AGE_MS) {
+    const installed = installedAuthRows(authRows, installedTargets);
+    return installed.length > 0 && installed.every((r) => now - r.health.checkedAt < maxAgeMs);
 }
 /**
  * Whether the tick may reuse the cached auth verdict instead of re-probing the
@@ -51,8 +71,8 @@ export function isCachedFleetAuthProbeFresh(authRows, now, maxAgeMs = AUTH_PROBE
  * revoked account (the second `runFleetPing` call site, RUSH-2998). Pure —
  * unit-tested so that inversion cannot regress unnoticed.
  */
-export function shouldReuseCachedAuthProbe(force, cached, now, maxAgeMs = AUTH_PROBE_MAX_AGE_MS) {
-    return !force && isCachedFleetAuthProbeFresh(cached, now, maxAgeMs);
+export function shouldReuseCachedAuthProbe(force, cached, now, installedTargets, maxAgeMs = AUTH_PROBE_MAX_AGE_MS) {
+    return !force && isCachedFleetAuthProbeFresh(cached, now, installedTargets, maxAgeMs);
 }
 /**
  * Fleet cache warm: publish THIS host's row for the caches `agents fleet
@@ -67,7 +87,7 @@ export async function refreshLocalFleetAuthState(opts) {
     const force = opts?.force === true;
     const signal = opts?.signal;
     const { machineId } = await import('./machine-id.js');
-    const { probeLocalFleetAuth, readFleetAuthRows, writeFleetAuthRows } = await import('./auth-health.js');
+    const { probeLocalFleetAuth, readFleetAuthRows, writeFleetAuthRows, localAuthTargetKeys } = await import('./auth-health.js');
     const { getCliVersion } = await import('./version.js');
     const self = machineId();
     const requestedAt = Date.now();
@@ -91,11 +111,16 @@ export async function refreshLocalFleetAuthState(opts) {
             // Re-probe the rate-limited /oauth/usage endpoint at most every
             // AUTH_PROBE_MAX_AGE_MS; reuse the last real verdict in between (RUSH-2998).
             // Fleet status publishes every tick regardless — it does not ride that endpoint.
+            const installed = localAuthTargetKeys();
             const cached = readFleetAuthRows(self);
-            const reuse = shouldReuseCachedAuthProbe(force, cached, requestedAt);
-            const authRows = reuse ? cached : await probeLocalFleetAuth({ cliVersion: getCliVersion(), forceLive: force, signal });
-            if (!reuse)
-                writeFleetAuthRows(self, authRows);
+            const live = installedAuthRows(cached, installed);
+            const reuse = shouldReuseCachedAuthProbe(force, cached, requestedAt, installed);
+            const authRows = reuse ? live : await probeLocalFleetAuth({ cliVersion: getCliVersion(), forceLive: force, signal });
+            // Write when we probed, and ALSO when the cache holds orphan rows for homes
+            // that are gone: a reusing tick is the common case, so leaving the prune on
+            // the probe branch would keep them in `agents view` indefinitely (PHNX-4051).
+            if (!reuse || live.length !== cached.length)
+                writeFleetAuthRows(self, authRows, installed);
             const row = await publishLocalFleetStatus(self);
             return { row, authRows };
         },

package/dist/lib/drift-sync.js

@@ -24,7 +24,7 @@ import { pullRepo } from './git.js';
 import { repairAfterSync, renderRepairAfterSync } from './reconcile-and-repair.js';
 import { promptAgentVersionSelection } from './installations/versions.js';
 import { isInteractiveTerminal, isPromptCancelled } from './format.js';
-import { computeSyncStatus, } from './sync-status.js';
+import { computeSyncStatus, formatDriftRows } from './sync-status.js';
 const agentName = (id) => AGENTS[id]?.name ?? id;
 /** "claude@2.1.170  2 drifted · 1 missing" for one version. */
 function versionLine(v) {
@@ -42,8 +42,11 @@ function renderSummary(status, needing) {
     if (status.system.behind > 0) {
         console.log(`  ${'.system repo'.padEnd(28)} ${chalk.yellow(`${status.system.behind} commit${status.system.behind === 1 ? '' : 's'} behind`)} ${chalk.gray('— pull recommended')}`);
     }
-    for (const v of needing)
+    for (const v of needing) {
         console.log(versionLine(v));
+        for (const line of formatDriftRows(v))
+            console.log(chalk.gray(`      ${line}`));
+    }
     if (status.totals.orphan > 0) {
         console.log(chalk.gray(`  (${status.totals.orphan} orphan${status.totals.orphan === 1 ? '' : 's'} — run \`agents prune cleanup\`)`));
     }

package/dist/lib/exec.js

@@ -39,6 +39,7 @@ import { probeUnprivilegedUserns } from './linux-userns.js';
 import { resolveClaudeSetupToken } from './claude-account-token.js';
 import { applyAddDirs } from './add-dir.js';
 import { applyActiveRulesPresetAtRun } from './rules/run-sync.js';
+import { applySystemResourcesAtRun } from './system-run-sync.js';
 import { resolveHarnessAdapter, stripForeignConfigDir } from './harness/index.js';
 import { resolveConfigVersion } from './harness/exec-config-version.js';
 import { getAccountInfo } from './agents.js';
@@ -2406,6 +2407,7 @@ export async function runWithFallback(options) {
         const rulesVersion = version ?? resolveVersion(agent);
         if (rulesVersion) {
             applyActiveRulesPresetAtRun(agent, rulesVersion, getVersionHomePath(agent, rulesVersion));
+            applySystemResourcesAtRun(agent, rulesVersion, getVersionHomePath(agent, rulesVersion));
         }
         // Record the entry we're about to attempt so the caller (audit log) sees the
         // agent+version that actually ran, even after a rate-limit handoff.

package/dist/lib/fleet-shared-repo-sync.d.ts

@@ -1,5 +1,12 @@
 export declare const FLEET_SHARED_REPO_SYNC_DEADLINE_MS = 45000;
 export declare const FLEET_SHARED_REPO_KILL_GRACE_MS = 250;
+export declare const FLEET_SHARED_REPO_OUTPUT_MAX_BYTES: number;
+/**
+ * When the shared-state exchange last completed successfully, in epoch ms, or
+ * `null` if it never has on this box (no marker yet) or the marker is unreadable.
+ * Reads the default daemon-dir marker the periodic exchange writes.
+ */
+export declare function readLastSuccessfulExchangeMs(): number | null;
 export interface BoundedProcessResult {
     code: number | null;
     stdout: string;

package/dist/lib/fleet-shared-repo-sync.js

@@ -19,9 +19,46 @@ import { machineId } from './session/sync/config.js';
 import { getDaemonDir, getUserAgentsDir } from './state.js';
 export const FLEET_SHARED_REPO_SYNC_DEADLINE_MS = 45_000;
 export const FLEET_SHARED_REPO_KILL_GRACE_MS = 250;
-const FLEET_SHARED_REPO_OUTPUT_MAX_BYTES = 1024 * 1024;
+export const FLEET_SHARED_REPO_OUTPUT_MAX_BYTES = 1024 * 1024;
 const FLEET_SHARED_REPO_PUSH_ATTEMPTS = 3;
 const FLEET_SHARED_REPO_REBASE_CLEANUP_RESERVE_MS = 5_000;
+/**
+ * The exchange records its last SUCCESSFUL completion beside its lock so a
+ * sibling service (auth-sync) can gate work on the freshness of the peer state
+ * this exchange delivers into the local checkout (PHNX-4051). auth-sync's
+ * credential pushes read the peer verdicts the LAST usage-sync exchange wrote; if
+ * that exchange has not landed within a tick interval, those verdicts are stale
+ * and pushing off them would act on a peer that may already hold the key.
+ */
+function exchangeMarkerPath(lockPath) {
+    return `${lockPath}.last-success`;
+}
+function writeExchangeMarker(lockPath, nowMs = Date.now()) {
+    try {
+        const p = exchangeMarkerPath(lockPath);
+        fs.mkdirSync(path.dirname(p), { recursive: true });
+        fs.writeFileSync(p, JSON.stringify({ at: nowMs }), 'utf-8');
+    }
+    catch {
+        // Best-effort telemetry — never fail the exchange over its own marker.
+    }
+}
+/**
+ * When the shared-state exchange last completed successfully, in epoch ms, or
+ * `null` if it never has on this box (no marker yet) or the marker is unreadable.
+ * Reads the default daemon-dir marker the periodic exchange writes.
+ */
+export function readLastSuccessfulExchangeMs() {
+    try {
+        const lockPath = path.join(getDaemonDir(), 'fleet-shared-repo-sync');
+        const raw = fs.readFileSync(exchangeMarkerPath(lockPath), 'utf-8');
+        const at = JSON.parse(raw).at;
+        return typeof at === 'number' && Number.isFinite(at) ? at : null;
+    }
+    catch {
+        return null;
+    }
+}
 function signalProcessTree(child, signal) {
     const pid = child.pid;
     if (!pid)
@@ -177,9 +214,30 @@ function retainedAutostash(stashList, objectId) {
  * cannot create a fresh collision — so nothing is silently destroyed. Then let
  * the rebase check out origin's version. A file we cannot clear is left in
  * place; the rebase may still abort, no worse than today and never losing data.
+ *
+ * The scan is scoped to the TRACKED ROOTS origin carries — the top-level entries
+ * of `git ls-tree --name-only origin/<branch>` — NOT the whole working tree
+ * (PHNX-4051). Only a file origin tracks can be overwritten by the rebase
+ * checkout, so a tracked root is the tightest superset of the collidable paths:
+ * every path origin tracks stays in scope (a stale untracked `projects/rush.yaml`
+ * under a tracked `projects/` root is still cleared), while an untracked-ONLY tree
+ * origin never tracks — artifacts-cli's revision store
+ * (`~/.agents/artifact-history/`, ~11k files / >1 MiB of paths) is the live case —
+ * is never a tracked root, so it never enters the `git ls-files --others` listing.
+ * An unbounded listing there blows past {@link FLEET_SHARED_REPO_OUTPUT_MAX_BYTES}
+ * and kills the exchange on every tick; scoping to the tracked roots keeps the
+ * PHNX-3923 coverage whole while an unrelated untracked tree cannot push the
+ * listing over the cap. `ls-tree`'s own output is bounded — it lists only the root
+ * entries, never the recursive tree.
  */
 async function clearCollidingUntracked(git, root, branch) {
-    const others = await git(['ls-files', '--others', '--exclude-standard', '-z']);
+    const tree = await git(['ls-tree', '--name-only', '-z', `origin/${branch}`]);
+    if (tree.code !== 0)
+        return { error: tree };
+    const trackedRoots = tree.stdout.split('\0').filter(Boolean);
+    if (trackedRoots.length === 0)
+        return { cleared: 0, backedUp: [], backupDir: null };
+    const others = await git(['ls-files', '--others', '--exclude-standard', '-z', '--', ...trackedRoots]);
     if (others.code !== 0)
         return { error: others };
     const relPaths = others.stdout.split('\0').filter(Boolean);
@@ -403,7 +461,12 @@ export async function syncFleetSharedStateRepo(options = {}) {
             retries: { retries: 20, factor: 1, minTimeout: 100, maxTimeout: 100 },
             onCompromised: logAndContinueOnLockCompromised('fleet shared repo sync'),
         });
-        return await performFleetSharedRepoSync(root, device, timeoutMs);
+        const result = await performFleetSharedRepoSync(root, device, timeoutMs);
+        // Record freshness only on a real success, beside the lock so a test-redirected
+        // lockPath keeps its marker isolated too (PHNX-4051).
+        if (result.success)
+            writeExchangeMarker(lockPath);
+        return result;
     }
     catch (err) {
         return {

package/dist/lib/helper-versions.js

@@ -36,7 +36,7 @@
  * helper release now, off this table entirely.
  */
 export const HELPER_RELEASES = {
-    menubar: { tagPrefix: 'menubar', floor: '1.1.2' },
+    menubar: { tagPrefix: 'menubar', floor: '1.1.4' },
     'computer-mac': { tagPrefix: 'computer-mac', floor: '1.0.0' },
     // The Windows helper is a bare .exe, not an .app bundle, so it does not share
     // helper-download.ts's zip/codesign/notarize machinery -- but it has the same

package/dist/lib/sync-status.d.ts

@@ -115,6 +115,12 @@ export interface ResidualDrift {
 export declare function verifyVersionConverged(agent: AgentId, version: string, cwd?: string): ResidualDrift | null;
 /** One-line-per-resource description of residual drift, for the sync exit line. */
 export declare function formatResidualDrift(residual: ResidualDrift[]): string[];
+/**
+ * The rows behind a "N drifted" count, one line each, for the human renderers.
+ * A count alone cannot be acted on; the name says which skill or hook to look
+ * at, and `detail` says what differs when the differ knows.
+ */
+export declare function formatDriftRows(v: AgentVersionStatus): string[];
 export interface SyncStatusOptions {
     cwd?: string;
     /** Restrict to specific agent ids; undefined = every supported agent. */

package/dist/lib/sync-status.js

@@ -54,6 +54,17 @@ export function formatResidualDrift(residual) {
     }
     return lines;
 }
+/**
+ * The rows behind a "N drifted" count, one line each, for the human renderers.
+ * A count alone cannot be acted on; the name says which skill or hook to look
+ * at, and `detail` says what differs when the differ knows.
+ */
+export function formatDriftRows(v) {
+    return v.resources
+        .filter((r) => r.status === 'drifted' || r.status === 'missing')
+        .sort((a, b) => a.kind.localeCompare(b.kind) || a.name.localeCompare(b.name))
+        .map((r) => `${r.status.padEnd(7)} ${r.kind}/${r.name}${r.detail ? ` (${r.detail})` : ''}`);
+}
 const STATUS_MAP = {
     ok: 'synced',
     diff: 'drifted',

package/dist/lib/system-run-sync.d.ts

@@ -0,0 +1,11 @@
+import type { AgentId } from './types.js';
+export interface SystemRunSyncResult {
+    /** Names rewritten per kind; empty when the sentinel matched (skip-fast). */
+    skills: string[];
+    subagents: string[];
+}
+/**
+ * Refresh the system layer's skills and subagents into the version home when
+ * their source directories changed since the last run for (agent, version).
+ */
+export declare function applySystemResourcesAtRun(agent: AgentId, version: string, versionHome: string): SystemRunSyncResult;

package/dist/lib/system-run-sync.js

@@ -0,0 +1,114 @@
+/**
+ * System-layer skills and subagents refresh at `agents run` launch time.
+ *
+ * `agents sync <harness>@all system` rewrites every version home, but nothing
+ * runs it when the `.system` mirror moves: the shim's `--launch` path compiles
+ * project rules and mirrors project resources only, and the SessionStart
+ * autosync fires the umbrella at most once per 4 h. Rules already close that
+ * gap per run (`rules/run-sync.ts`); skills and subagents did not, so a merged
+ * system-layer skill change left 31 of 33 installed copies on the old text
+ * (PHNX-4056). This is the same shape as the rules run-sync: a small per-
+ * (agent, version) sentinel of source fingerprints, stat-tier comparison, and
+ * a write through the registered writers only when a source file changed.
+ *
+ * Scope: the system layer's `skills/` and `subagents/` directories. The
+ * writers resolve each name through the normal precedence (a user-layer
+ * shadow still wins), and the active resource profile filters both kinds the
+ * same way `syncResourcesToVersion` does. Never blocks a launch: any failure
+ * is swallowed and leaves the sentinel stale, so the next run retries.
+ */
+import * as fs from 'fs';
+import * as path from 'path';
+import { getCacheDir, getSystemSkillsDir, getSystemSubagentsDir } from './state.js';
+import { getWriter } from './staleness/registry.js';
+import { fingerprintDir, isDirStale } from './staleness/fingerprint.js';
+import { filterNamesForActiveResourceProfile } from './resource-profiles.js';
+/** ~/.agents/.cache/system-run-sync/ — regenerable; a lost sentinel costs one extra write. */
+function sentinelPath(agent, version) {
+    const key = `${agent}@${version}`.replace(/[^a-zA-Z0-9@._-]/g, '_');
+    return path.join(getCacheDir(), 'system-run-sync', `${key}.json`);
+}
+function loadSentinel(agent, version) {
+    try {
+        return JSON.parse(fs.readFileSync(sentinelPath(agent, version), 'utf-8'));
+    }
+    catch {
+        return {};
+    }
+}
+function saveSentinel(agent, version, sentinel) {
+    const p = sentinelPath(agent, version);
+    try {
+        fs.mkdirSync(path.dirname(p), { recursive: true });
+        fs.writeFileSync(p, JSON.stringify(sentinel));
+    }
+    catch {
+        // Best-effort — a failed write just means the next run redoes the copy.
+    }
+}
+/** Subdirectories of `dir` that carry `marker` (SKILL.md / AGENT.md). */
+function namesWithMarker(dir, marker) {
+    const out = new Map();
+    let entries;
+    try {
+        entries = fs.readdirSync(dir, { withFileTypes: true });
+    }
+    catch {
+        return out;
+    }
+    for (const e of entries) {
+        if (!e.isDirectory() || e.name.startsWith('.'))
+            continue;
+        const src = path.join(dir, e.name, marker);
+        if (fs.existsSync(src))
+            out.set(e.name, src);
+    }
+    return out;
+}
+/**
+ * Refresh the system layer's skills and subagents into the version home when
+ * their source directories changed since the last run for (agent, version).
+ */
+export function applySystemResourcesAtRun(agent, version, versionHome) {
+    const result = { skills: [], subagents: [] };
+    const stored = loadSentinel(agent, version);
+    const next = { ...stored };
+    const kinds = [
+        { kind: 'skills', dir: getSystemSkillsDir(), marker: 'SKILL.md' },
+        { kind: 'subagents', dir: getSystemSubagentsDir(), marker: 'AGENT.md' },
+    ];
+    for (const { kind, dir, marker } of kinds) {
+        if (!fs.existsSync(dir))
+            continue;
+        const prev = stored[kind];
+        if (prev && prev.dir === dir && !isDirStale(prev.dir, prev.files, dir))
+            continue;
+        const writer = getWriter(kind, agent);
+        if (!writer)
+            continue;
+        const sources = namesWithMarker(dir, marker);
+        // The active resource profile is the per-client scoping seam (a brand pins
+        // one); it filters every kind here exactly as `syncResourcesToVersion` does.
+        const names = filterNamesForActiveResourceProfile(kind, Array.from(sources.keys()), sources);
+        if (names.length === 0) {
+            next[kind] = { dir, files: fingerprintDir(dir) };
+            continue;
+        }
+        let written;
+        try {
+            written = writer.write({ version, versionHome, selection: names, cwd: '' });
+        }
+        catch {
+            continue; // leave the sentinel stale so the next run retries
+        }
+        result[kind] = written.synced;
+        // A per-name failure (an AGENT.md that does not parse) also leaves the
+        // sentinel stale: the next run retries instead of never seeing it again.
+        if (written.errors?.length)
+            continue;
+        next[kind] = { dir, files: fingerprintDir(dir) };
+    }
+    if (JSON.stringify(next) !== JSON.stringify(stored))
+        saveSentinel(agent, version, next);
+    return result;
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@phnx-labs/agents-cli",
-  "version": "1.22.99",
+  "version": "1.22.101",
   "description": "One CLI for all your AI coding agents - versions, config, cloud dispatch, sessions, and teams (now with first-class Grok Build CLI support)",
   "type": "module",
   "main": "dist/index.js",