chore(release): bump version to 1.6.4

feat(ai-router): test an endpoint before saving it, and re-point a live lane
Two gaps left by the AI Router feature: a profile could be saved without ever being reachable, and a lane's router could only be chosen at creation. **POST /api/ai-routers/test** answers the question saving cannot. It sends one Anthropic-style message capped at a single output token to `<baseUrl>/v1/messages` with a 15s timeout — the wire shape a claude lane's eight env keys point at, so what is tested is what will actually run. A router that answers is a 200 here whatever it replies: `{ ok, status, message }` carries the router's own words, and `status` is null when nothing answered at all. Only bad input is a 4xx. The stored key is only ever sent to the endpoint it was stored against. A blank `apiKey` falls back to the key of the profile named by `name`, but only while the submitted `baseUrl` still equals the saved one — so testing a retyped endpoint needs a retyped key (`EBADKEY`). Without that rule this route would be a way to make the server POST a saved API key to any URL a browser names. Network failures report `error.cause`, not `error.message`: node's fetch reports every transport failure as the same useless "fetch failed" and hides the part the user can act on (ECONNREFUSED, ENOTFOUND, a TLS error) one level down. **The lane card carries an AI Router select.** The Add-lane modal was the only place a router was chosen, and a lane outlives its creation. Switching goes through `POST /api/lanes/:id/router/apply` — never a lane PATCH — because the DB column and the lane's own config on disk may only move together. A failed apply puts the select back rather than leaving the card claiming a router the lane did not get. Hidden when `native` is the only profile: a select with one option is a control that cannot do anything. Adds `Lane.router_profile` to the client type (the payload already carried it).
2026-09-16 11:28:18 +07:00 · 2026-09-16 11:26:00 +07:00 · 2026-09-15 15:17:55 +07:00 · 2026-09-15 14:55:23 +07:00 · 2026-09-15 14:53:00 +07:00 · 2026-09-15 11:01:07 +07:00
31 changed files with 1089 additions and 72 deletions
@@ -1,6 +1,6 @@
 # Claude Code Working Guide

-_Last synced: v1.6.0 (2026-09-14) + AI Router profiles (`server/lib/ai-routers.js`, `server/lib/router-apply.js`, the Settings tab `/settings?tool=ai-router`, and `lanes.router_profile`) + v1.5.6 (2026-09-11) + the systemd handover rule below (`scripts/systemd-service.js`'s `writeUserService`/`restartUserService` split and `plugin-bootstrap.js`'s deferred `pendingSystemdRestart`) + eb63569 (2026-09-10) + release-based update detection (`server/lib/update-check.js` reads the forge releases API — a checkout via its `origin` remote, a plugin-cache install via the marketplace manifest on disk, since an archive-sourced cache dir has no `.git`; a `plugin_managed` install self-updates via the detached `plugin-update` chain) + the install-wide **Update all** batch and session reload (`server/lib/cc-update-all.js`, `server/lib/cc-reload.js`)._ A change that makes any statement in this file, `AGENTS.md`, or a README wrong — a moved path, a new or changed extension point, a retired rule — updates that file in the same change-set and bumps this marker.
+_Last synced: v1.6.0 (2026-09-14) + AI Router profiles (`server/lib/ai-routers.js`, `server/lib/router-apply.js`, the Settings tab `/settings?tool=ai-router`, its **Test** button / `POST /api/ai-routers/test`, the per-lane router select on `LaneCard`, and `lanes.router_profile`) + the bootstrap's superseded-build guard and plugin-cache pruner (`supersededBy`/`pruneRetiredPluginRoots` in `scripts/plugin-bootstrap.js`, `unitRoot` in `scripts/systemd-service.js`) + v1.5.6 (2026-09-11) + the systemd handover rule below (`scripts/systemd-service.js`'s `writeUserService`/`restartUserService` split and `plugin-bootstrap.js`'s deferred `pendingSystemdRestart`) + eb63569 (2026-09-10) + release-based update detection (`server/lib/update-check.js` reads the forge releases API — a checkout via its `origin` remote, a plugin-cache install via the marketplace manifest on disk, since an archive-sourced cache dir has no `.git`; a `plugin_managed` install self-updates via the detached `plugin-update` chain) + the install-wide **Update all** batch and session reload (`server/lib/cc-update-all.js`, `server/lib/cc-reload.js`)._ A change that makes any statement in this file, `AGENTS.md`, or a README wrong — a moved path, a new or changed extension point, a retired rule — updates that file in the same change-set and bumps this marker.

 ## Project mission
 - Maintain a reliable local-first dashboard for Claude Code session monitoring.
@@ -36,7 +36,7 @@ A **lane** is a durable unit of parallel agent work — one working directory, m
 - Liveness: a silent **watcher** (stage matching `/watch|poll/`) is dead after `LANE_DEAD_SEC` seconds (default 300); a silent **idle** lane is at rest, not dead.
 - **Never `rm -rf` a lane.** Destructive lane operations (`reset`, `remove`, `purge`) go through `server/lib/worktree.js`/`server/lib/lanes.js`, never a raw filesystem delete — that's what keeps the three-check destroy guard and the DB bookkeeping in lockstep.
 - **Never build a git command as a shell string.** Lane git operations use `execFile("git", [...args])` with an explicit argument array (see `server/lib/worktree.js`'s `git()` helper), never a concatenated/interpolated string passed to a shell.
- **A lane's AI Router profile is a lane fact too, and it never travels alone.** `lanes.router_profile` (nullable; null means the same as `native`) is written ONLY by `setRouterProfile`, called beside the disk write in `POST /api/lanes/:id/router/apply` — never through `PATCHABLE`. The column and the lane's own config must move together, or the row claims a router the lane's `settings.local.json` has never heard of. What is written comes from the lane row, never the request: a **claude** lane gets eight `env` keys merged into `<cwd>/.claude/settings.local.json` (`.local` because the value carries an API key; every other key in that file is preserved and a non-JSON file is refused, not rewritten), a **codex** lane gets `$CODEX_HOME/<name>.config.toml` — the layer `codex -p <name>` loads — because Codex has NO project-level config, and its base `config.toml` is never touched. Key material never leaves the server: profiles live at `<CLAUDE_HOME>/agenthub-routers/` (0700/0600), `GET /api/ai-routers` reports only `hasKey`, and there is deliberately no route that returns one profile in full. **Codex's key reaches the pane through tmux's `-e` argv and is never typed into it** — `tmux.sendCommand` takes over an idle pane with `respawn-pane` for exactly this reason; a `KEY='sk-…' codex …` line typed with `send-keys` would be rendered to every attached browser, kept in the scrollback and appended to the user's shell history. Do not reintroduce a typed command line there. **And attaching COPIES a profile**, so the store stops being the single source of truth the moment a lane exists: `PATCH`/`DELETE /api/ai-routers/:name` must push through to the lanes naming it (`reapplyToLanes`/`detachFromLanes`), or a rotated key silently never reaches the lane still using the revoked one. See `server/lib/ai-routers.js`, `server/lib/router-apply.js`.
+- **A lane's AI Router profile is a lane fact too, and it never travels alone.** `lanes.router_profile` (nullable; null means the same as `native`) is written ONLY by `setRouterProfile`, called beside the disk write in `POST /api/lanes/:id/router/apply` — never through `PATCHABLE`. The column and the lane's own config must move together, or the row claims a router the lane's `settings.local.json` has never heard of. What is written comes from the lane row, never the request: a **claude** lane gets eight `env` keys merged into `<cwd>/.claude/settings.local.json` (`.local` because the value carries an API key; every other key in that file is preserved and a non-JSON file is refused, not rewritten), a **codex** lane gets `$CODEX_HOME/<name>.config.toml` — the layer `codex -p <name>` loads — because Codex has NO project-level config, and its base `config.toml` is never touched. Key material never leaves the server: profiles live at `<CLAUDE_HOME>/agenthub-routers/` (0700/0600), `GET /api/ai-routers` reports only `hasKey`, and there is deliberately no route that returns one profile in full. **`POST /api/ai-routers/test` sends a stored key ONLY to the endpoint it was stored against** — a blank `apiKey` falls back to the profile named by `name`, but only while the submitted `baseUrl` still equals the saved one, else `EBADKEY`. Relax that and the route becomes a way to make the server POST a saved key to any URL a browser names. **Re-pointing a lane is allowed and is not a PATCH:** the lane card's AI Router select calls `POST /api/lanes/:id/router/apply` (column + disk together, same rule as above), never `PATCH /api/lanes/:id`. **Codex's key reaches the pane through tmux's `-e` argv and is never typed into it** — `tmux.sendCommand` takes over an idle pane with `respawn-pane` for exactly this reason; a `KEY='sk-…' codex …` line typed with `send-keys` would be rendered to every attached browser, kept in the scrollback and appended to the user's shell history. Do not reintroduce a typed command line there. **And attaching COPIES a profile**, so the store stops being the single source of truth the moment a lane exists: `PATCH`/`DELETE /api/ai-routers/:name` must push through to the lanes naming it (`reapplyToLanes`/`detachFromLanes`), or a rotated key silently never reaches the lane still using the revoked one. See `server/lib/ai-routers.js`, `server/lib/router-apply.js`.
 - **A lane's `tool` is a lane fact, immutable, and never read from a request.** `lanes.tool` (`claude` | `codex`, default `claude`) is set by `createLane` only — it is deliberately absent from `PATCHABLE`, so no PATCH can flip the CLI under a live pane. Every spawn path (`POST /api/lanes/:id/start`, `POST /api/run`) takes the tool from the lane row and ignores whatever the body says; the run panel's Codex-only fields (`sandbox`, `approval`, `resumeLast`) are validated against the fixed allowlists in `server/lib/pty-run.js` before they reach an argv, and the bypass flags (`--dangerously-bypass-approvals-and-sandbox`, `--dangerously-bypass-hook-trust`) are never emitted. `GET /api/run/binary?tool=` resolves the probed binary name from that same allowlist — the query string reaches `which`.
 - **Adopted lanes are not destroyable.** A lane with `kind === "adopted"` may never have its worktree reset or removed; "removing" one only drops the dashboard's own record of it. This is enforced in code (`assertDestroyable`'s first check) — do not add a path that bypasses it.
 - **A `cwd` can back at most one lane per tool, not one lane overall.** Uniqueness is the composite `idx_lanes_cwd_tool` (`cwd`, `tool`), not `cwd` alone — Repo/Adopt mode (`POST /api/lanes/ensure`) lets the same directory carry a `claude` lane and a `codex` lane side by side, shown as two separate cards. `resolveLaneByCwd(cwd, tool)` takes an optional `tool` filter for exactly this; hook ingestion (`touchLaneFromHook`, `resolveLaneForWork`) always pins it to `"claude"` because Codex has no hook mechanism in this codebase, so a hook event must never touch a `codex` lane sharing the path. Worktree mode is unaffected — it already allocates a unique directory per lane.
@@ -45,6 +45,8 @@ See `docs/LANES.md` for full guide: stage reporting, custom templates, lane acti

 ## Non-negotiable engineering rules
 - **When a `agenthub.service` unit of ours owns the server, nothing else may stop it.** The bootstrap that performs a self-update is spawned by the running server, so it lives in that server's systemd cgroup: SIGTERM-ing the server from there exits 0 (which `Restart=on-failure` read as "meant to stop") AND trips `KillMode=control-group`, reaping the bootstrap before it can repoint the unit or write `state.json` — a dashboard dead after every update until some later session bootstrapped again. So `scripts/systemd-service.js` splits the API: `writeUserService` repoints the unit and starts NOTHING, `restartUserService` uses `--no-block`, and `plugin-bootstrap.js` defers that restart past `writeState` and `releaseLock` (`pendingSystemdRestart`). The unit is `Restart=always`. Do not add a path that stops the dashboard from inside its own cgroup, and do not move the restart earlier. `agenthub service install` may keep using `installUserService` — it runs from a user's shell, outside the cgroup. `systemdOwned` additionally requires the unit to be ACTIVE (an installed-but-stopped unit owns nothing; treating it as owner skips the stop and then binds an occupied port). A unit naming a RETIRED plugin root is invisible until a reboot, so `fastPathOk` rejects it via `unitIsStale` (pure fs — that path may never spawn a process) and the bootstrap rewrites it; repairing the unit is NOT restarting the server (`pendingSystemdRestart = systemdOwned && needsServer`), and the detached-spawn fallback is guarded by `needsServer` so a stale unit alone never starts a second server onto a bound port.
+- **A bootstrap that is no longer the installed build does NOTHING, and `--force` does not override that.** `supersededBy` (`scripts/plugin-bootstrap.js`) reads Claude Code's `installed_plugins.json`: if a SIBLING cache directory is the recorded install, this build logs one line and returns `superseded` — no deps, no unit, no server. A Claude Code process that was not restarted after an update keeps loading the OLD cache dir, and without this its every session repointed `agenthub.service` at that old dir and restarted the dashboard onto the version the user just upgraded away from, once per session, forever. It may not hand over to the new build either — the process it would replace is the one loading it. Do not add a path that lets a retired root write the unit, stop the server, or "just" refresh deps.
+- **Deleting retired cache copies is the ONE thing here that touches the plugin cache, and four protections gate it.** `pruneRetiredPluginRoots` never deletes: this build; any `installPath` in `installed_plugins.json`; the dir `agenthub.service` names (`unitRoot`, foreign unit included); any dir named in the user's `settings.json` hooks (a un-restarted Claude Code is still executing from there — deleting it breaks that session's hooks). Plus the 2 newest survivors as a rollback cushion, and nothing at all outside `~/.claude/plugins/cache/`. It runs AFTER `writeState`, so a bad prune cannot also cost the record of a good bootstrap. Adding a fifth thing that pins a cache dir means adding a fifth protection here.
 - **A session reload is typed into a live pane, never spawned.** `/reload-plugins` and `/reload-skills` are scoped BY CLAUDE CODE to the session that runs them, so nothing can reload a session from the outside. `server/lib/cc-reload.js` therefore types them into the dashboard's existing tmux lane panes and spawns NO `claude` process — a one-shot `claude -p "/reload-plugins"` reloads only its own short-lived session, exits, and leaves a junk session in the dashboard's own history. Do not add a path that starts a session in order to reload one.
 - Preserve existing behavior unless explicitly asked to change it.
 - Prefer minimal, reversible diffs.
@@ -475,22 +475,31 @@ export function AddLaneModal({
              ] as const
            ).map((option) => {
              const active = option.value === "native" ? router === "native" : router !== "native";
+              // Clicking "AI Router" before the list lands would resolve back to
+              // Native and read as a dead control.
+              const disabled = option.value === "router" && !routersLoaded;
              return (
                <label
                  key={option.value}
-                  className={`flex flex-1 cursor-pointer items-center gap-2 rounded-lg border p-2.5 text-xs shadow-sm transition-colors ${
-                    active
-                      ? "border-accent bg-accent/10 text-accent"
-                      : "border-border bg-surface-2 text-fg-secondary hover:border-border-light hover:bg-surface-3"
+                  // No radio dot: the card itself lights up when selected. That
+                  // leaves the DISABLED state with nothing to show it, since the
+                  // greyed-out dot was carrying it — hence the explicit branch.
+                  // The input only goes visually hidden, so keyboard and screen
+                  // readers still get a real radio group, and the card carries
+                  // its focus ring.
+                  className={`block flex-1 rounded-lg border p-2.5 text-center text-xs shadow-sm transition-colors focus-within:ring-2 focus-within:ring-accent/50 ${
+                    disabled
+                      ? "cursor-not-allowed border-border bg-surface-2 text-fg-muted opacity-60"
+                      : active
+                        ? "cursor-pointer border-accent bg-accent/10 text-accent"
+                        : "cursor-pointer border-border bg-surface-2 text-fg-secondary hover:border-border-light hover:bg-surface-3"
                  }`}
                >
                  <input
                    type="radio"
                    name="add-lane-router"
-                    className="h-3.5 w-3.5 accent-accent"
-                    // Clicking "AI Router" before the list lands would resolve
-                    // back to Native and read as a dead control.
-                    disabled={option.value === "router" && !routersLoaded}
+                    className="sr-only"
+                    disabled={disabled}
                    checked={active}
                    onChange={() =>
                      // Switching to "AI Router" lands on the first saved
@@ -536,9 +545,12 @@ export function AddLaneModal({

        {tool === "claude" && (
          <label
-            // Card that lights up when ticked, matching LaneStripCard's
-            // selected/unselected treatment.
-            className={`flex cursor-pointer items-start gap-2.5 rounded-lg border p-3 shadow-sm transition-colors ${
+            // No tick box: the card itself lights up when selected, the same
+            // treatment LaneStripCard uses, which makes a second indicator of
+            // the same state redundant. The input stays and only goes visually
+            // hidden — it is still what a keyboard toggles and what a screen
+            // reader announces — so the card carries the focus ring for it.
+            className={`block cursor-pointer rounded-lg border p-3 shadow-sm transition-colors focus-within:ring-2 focus-within:ring-accent/50 ${
              writeDocs
                ? "border-accent bg-accent/10"
                : "border-border bg-surface-2 hover:border-border-light hover:bg-surface-3"
@@ -551,17 +563,15 @@ export function AddLaneModal({
              aria-label={t("addLaneWriteDocsLabel")}
              checked={writeDocs}
              onChange={(e) => setWriteDocs(e.target.checked)}
-              className="mt-0.5 h-3.5 w-3.5 shrink-0 accent-accent"
+              className="sr-only"
            />
-            <span>
-              <span
-                className={`block text-xs font-medium ${writeDocs ? "text-accent" : "text-fg-secondary"}`}
-              >
-                {t("addLaneWriteDocsLabel")}
-              </span>
-              <span className="mt-1 block text-[10px] text-fg-muted">
-                {t("addLaneWriteDocsHint")}
-              </span>
+            <span
+              className={`block text-xs font-medium ${writeDocs ? "text-accent" : "text-fg-secondary"}`}
+            >
+              {t("addLaneWriteDocsLabel")}
+            </span>
+            <span className="mt-1 block text-[10px] text-fg-muted">
+              {t("addLaneWriteDocsHint")}
            </span>
          </label>
        )}
@@ -7,6 +7,8 @@
 * this view exists to catch. Stage, kind, and the "auto: <stage>" chip are
 * NOT repeated here — this card only ever renders inside the Workspace
 * lane-detail section, whose header above it already shows them. The repo
+ * card also carries the lane's AI Router picker — the Add-lane modal is the
+ * only other place one is chosen, and a lane outlives its creation. The repo
 * tooling dropdown (`LaneToolsMenu`, exported from this file) and the
 * remove-lane button live in Workspace's page header / pipeline-picker row
 * instead of here, so they stay reachable and un-clipped regardless of this
@@ -20,6 +22,7 @@ import { DestructiveLaneModal } from "./DestructiveLaneModal";
 import { api } from "../../lib/api";
 import { eventBus } from "../../lib/eventBus";
 import { keepIfSame } from "../../lib/keepIfSame";
+import type { AiRouter } from "../../lib/api";
 import type { Lane, LaneGitFacts, LaneRuntime, NamedLock } from "../../lib/types";

 /** How often a mounted card re-reads its working-copy facts. Slow on purpose:
@@ -168,6 +171,92 @@ function useLaneIntegrations(
  return state as Record<(typeof INTEGRATION_NAMES)[number], boolean> | null;
 }

+/**
+ * Which AI Router this lane runs against, and the one place it can be changed
+ * after the lane exists — the Add-lane modal is the only other one, and a lane
+ * outlives the moment it was created.
+ *
+ * Applying writes the lane's own config on disk as well as the DB column, so it
+ * goes through `POST /lanes/:id/router/apply` and never a PATCH. A failure puts
+ * the select back: the row must not claim a router the lane did not get.
+ *
+ * Rendered only when there is something to choose. A user who has never opened
+ * the AI Router tab has exactly the built-in `native` profile, and a select
+ * with one option is a control that cannot do anything.
+ */
+function LaneRouterRow({ lane }: { lane: Lane }) {
+  const { t } = useTranslation(["lanes"]);
+  const [routers, setRouters] = useState<AiRouter[] | null>(null);
+  const [current, setCurrent] = useState(lane.router_profile || "native");
+  const [busy, setBusy] = useState(false);
+  const [note, setNote] = useState<{ ok: boolean; text: string } | null>(null);
+
+  // The polled lane list is the source of truth; a switch elsewhere (the lane
+  // was reset to native by a profile delete, say) has to land here too.
+  useEffect(() => {
+    setCurrent(lane.router_profile || "native");
+    setNote(null);
+  }, [lane.id, lane.router_profile]);
+
+  useEffect(() => {
+    let alive = true;
+    api.aiRouters
+      .list()
+      .then((r) => {
+        if (alive) setRouters(r.items);
+      })
+      .catch(() => {
+        // Silent, same contract as the git/runtime pollers above: a card that
+        // cannot list routers simply shows no router row.
+        if (alive) setRouters([]);
+      });
+    return () => {
+      alive = false;
+    };
+  }, []);
+
+  if (!routers || routers.length <= 1) return null;
+
+  const change = async (name: string) => {
+    const previous = current;
+    setCurrent(name);
+    setBusy(true);
+    setNote(null);
+    try {
+      await api.lanes.routerApply(lane.id, name);
+      setNote({ ok: true, text: t("router.applied") });
+    } catch (err) {
+      setCurrent(previous);
+      setNote({ ok: false, text: err instanceof Error ? err.message : String(err) });
+    } finally {
+      setBusy(false);
+    }
+  };
+
+  return (
+    <div className="mb-2 text-[11px]" data-testid={`lane-router-${lane.id}`}>
+      <label className="flex items-center gap-1.5 text-fg-muted">
+        {t("router.label")}
+        <select
+          value={current}
+          disabled={busy}
+          onChange={(e) => void change(e.target.value)}
+          className="flex-1 truncate rounded border border-border-light bg-surface-2 px-1.5 py-0.5 text-fg-secondary disabled:opacity-50"
+        >
+          {routers.map((r) => (
+            <option key={r.name} value={r.name}>
+              {r.label}
+            </option>
+          ))}
+        </select>
+      </label>
+      {note && (
+        <p className={`mt-1 ${note.ok ? "text-fg-muted" : "text-status-danger"}`}>{note.text}</p>
+      )}
+    </div>
+  );
+}
+
 export const LIVENESS_DOT: Record<Lane["liveness"], string> = {
  active: "bg-status-success",
  idle: "bg-surface-4",
@@ -502,6 +591,8 @@ export default function LaneCard({
          </div>
        </dl>

+        <LaneRouterRow lane={lane} />
+
        {childWorktrees && childWorktrees.length > 0 && (
          <div
            data-testid="lane-child-worktrees"
@@ -27,10 +27,14 @@ vi.mock("../../../lib/api", () => ({
      agentsInstall: vi.fn(),
      mcpSync: vi.fn(),
      syncBaseCheck: vi.fn(),
+      routerApply: vi.fn(),
    },
    locks: {
      list: vi.fn(),
    },
+    aiRouters: {
+      list: vi.fn(),
+    },
  },
 }));

@@ -43,6 +47,10 @@ beforeEach(() => {
  vi.mocked(api.lanes.runtime).mockResolvedValue({ available: false });
  vi.mocked(api.lanes.integration).mockReset();
  vi.mocked(api.lanes.integration).mockResolvedValue({ enabled: false });
+  // Only the built-in profile by default: a user who never opened the AI Router
+  // tab has nothing to switch between, so the row stays out of every other test.
+  vi.mocked(api.aiRouters.list).mockReset();
+  vi.mocked(api.aiRouters.list).mockResolvedValue({ items: [] });
 });
 vi.mocked(api.locks.list).mockReset();
 vi.mocked(api.locks.list).mockResolvedValue({ locks: [] });
@@ -440,3 +448,63 @@ describe("LaneToolsMenu", () => {
    expect(container).toBeEmptyDOMElement();
  });
 });
+
+describe("LaneCard — the AI Router picker", () => {
+  const native = {
+    name: "native",
+    label: "Native (Claude / Codex)",
+    builtin: true,
+    baseUrl: "",
+    hasKey: false,
+    model: "claude-sonnet-5",
+    models: {
+      opus: "claude-opus-5",
+      sonnet: "claude-sonnet-5",
+      haiku: "claude-haiku-4-5-20251001",
+      smallFast: "claude-haiku-4-5-20251001",
+      subagent: "claude-haiku-4-5-20251001",
+    },
+  };
+  const nineRouter = { ...native, name: "9router", label: "9Router", builtin: false };
+
+  it("stays out of the card when native is the only profile", async () => {
+    vi.mocked(api.aiRouters.list).mockResolvedValue({ items: [native] });
+    render(<LaneCard lane={makeLane()} onAction={vi.fn()} />);
+    await waitFor(() => expect(api.aiRouters.list).toHaveBeenCalled());
+    expect(screen.queryByTestId("lane-router-1")).toBeNull();
+  });
+
+  it("shows the lane's current profile selected", async () => {
+    vi.mocked(api.aiRouters.list).mockResolvedValue({ items: [native, nineRouter] });
+    render(<LaneCard lane={makeLane({ router_profile: "9router" })} onAction={vi.fn()} />);
+    expect(await screen.findByDisplayValue("9Router")).toBeInTheDocument();
+  });
+
+  it("applies a switch through the router route, not a lane PATCH", async () => {
+    vi.mocked(api.aiRouters.list).mockResolvedValue({ items: [native, nineRouter] });
+    vi.mocked(api.lanes.routerApply).mockResolvedValue({
+      tool: "claude",
+      file: "/work/demo/.claude/settings.local.json",
+      profile: "9router",
+    });
+    render(<LaneCard lane={makeLane()} onAction={vi.fn()} />);
+
+    await userEvent.selectOptions(await screen.findByDisplayValue("Native (Claude / Codex)"), [
+      "9router",
+    ]);
+    await waitFor(() => expect(api.lanes.routerApply).toHaveBeenCalledWith(1, "9router"));
+    expect(await screen.findByDisplayValue("9Router")).toBeInTheDocument();
+  });
+
+  it("puts the selection back when the apply fails", async () => {
+    vi.mocked(api.aiRouters.list).mockResolvedValue({ items: [native, nineRouter] });
+    vi.mocked(api.lanes.routerApply).mockRejectedValue(new Error("EBADJSON"));
+    render(<LaneCard lane={makeLane()} onAction={vi.fn()} />);
+
+    await userEvent.selectOptions(await screen.findByDisplayValue("Native (Claude / Codex)"), [
+      "9router",
+    ]);
+    await waitFor(() => expect(screen.getByText("EBADJSON")).toBeInTheDocument());
+    expect(screen.getByDisplayValue("Native (Claude / Codex)")).toBeInTheDocument();
+  });
+});
@@ -37,6 +37,7 @@ vi.mock("../../../lib/api", () => ({

 const LANE: Lane = {
  id: 1,
+  router_profile: null,
  title: "demo",
  cwd: "/workspace/a",
  branch: null,
@@ -142,5 +142,7 @@
  "addLaneRouterHint": "Native talks to the vendor's own API. An AI Router profile is written into this lane's own config when it is created.",
  "addLaneRouterEmptyHint": "No router profile saved yet — add one under Settings › AI Router.",
  "addLaneSetupRouter": "Model endpoint",
-  "addLaneRouterLoading": "Loading router profiles…"
+  "addLaneRouterLoading": "Loading router profiles…",
+  "router.label": "AI Router",
+  "router.applied": "Applied. A session already running in this lane keeps the old one until it restarts."
 }
@@ -484,6 +484,10 @@
    "cancel": "Cancel",
    "removeConfirm": "Remove the profile \"{{label}}\"? Every lane using it is reset to the vendor's own API right away — on disk and in the dashboard.",
    "loading": "Loading profiles…",
-    "lanesFailed": "Saved, but {{count}} lane could not be updated and still carries the old endpoint: {{lanes}}"
+    "lanesFailed": "Saved, but {{count}} lane could not be updated and still carries the old endpoint: {{lanes}}",
+    "test": "Test",
+    "testing": "Testing…",
+    "testOk": "The endpoint answered — {{model}} is reachable with this key.",
+    "testFailed": "No answer ({{status}}): {{message}}"
  }
 }
@@ -142,5 +142,7 @@
  "addLaneRouterHint": "Native là gọi thẳng API của hãng. Chọn profile AI Router thì cấu hình được ghi vào config riêng của lane này lúc tạo.",
  "addLaneRouterEmptyHint": "Chưa lưu profile router nào — thêm ở Settings › AI Router.",
  "addLaneSetupRouter": "Endpoint model",
-  "addLaneRouterLoading": "Đang tải danh sách router…"
+  "addLaneRouterLoading": "Đang tải danh sách router…",
+  "router.label": "AI Router",
+  "router.applied": "Đã áp dụng. Session đang chạy trong lane này vẫn dùng router cũ cho tới khi khởi động lại."
 }
@@ -484,6 +484,10 @@
    "cancel": "Huỷ",
    "removeConfirm": "Xoá profile \"{{label}}\"? Mọi lane đang dùng nó sẽ được đưa về API gốc của hãng ngay lập tức — cả trên đĩa lẫn trong dashboard.",
    "loading": "Đang tải danh sách profile…",
-    "lanesFailed": "Đã lưu, nhưng {{count}} lane không cập nhật được và vẫn đang giữ endpoint cũ: {{lanes}}"
+    "lanesFailed": "Đã lưu, nhưng {{count}} lane không cập nhật được và vẫn đang giữ endpoint cũ: {{lanes}}",
+    "test": "Kiểm tra",
+    "testing": "Đang kiểm tra…",
+    "testOk": "Endpoint đã trả lời — {{model}} dùng được với key này.",
+    "testFailed": "Không trả lời ({{status}}): {{message}}"
  }
 }
@@ -1891,6 +1891,15 @@ export const api = {
        method: "PATCH",
        body: JSON.stringify(body),
      }),
+    /**
+     * POST /api/ai-routers/test - ask the endpoint itself whether it works,
+     * before a profile that cannot reach anything is saved onto a lane.
+     */
+    test: (body: AiRouterTestInput) =>
+      request<AiRouterTestResult>("/ai-routers/test", {
+        method: "POST",
+        body: JSON.stringify(body),
+      }),
    /**
     * DELETE /api/ai-routers/:name - also resets every lane using it back to
     * native, on disk as well as in the DB.
@@ -2680,6 +2689,26 @@ export interface AiRouter {
  savedAt?: string;
 }

+/** Outcome of POST /api/ai-routers/test. A router that answered with an error
+ *  is still a successful HTTP call: `ok` is the ROUTER's verdict, `status` is
+ *  its HTTP code (null when nothing answered at all), and `message` is its own
+ *  words — truncated server-side, never the key. */
+export interface AiRouterTestResult {
+  ok: boolean;
+  status: number | null;
+  message: string;
+}
+
+/** Body for POST /api/ai-routers/test. `name` lets the server reuse the stored
+ *  key for a profile being edited — but only while `baseUrl` still matches the
+ *  saved one, so a typed-in endpoint always needs a typed-in key. */
+export interface AiRouterTestInput {
+  baseUrl: string;
+  model: string;
+  apiKey?: string;
+  name?: string;
+}
+
 /** How an edit or a delete landed on the lanes already using that profile.
 *  Attaching copies a profile, so the store alone is not the whole truth: a
 *  lane listed in `failed` still carries the OLD key and endpoint. */
@@ -2420,6 +2420,11 @@ export interface Lane {
  kind: "managed" | "adopted";
  /** The CLI this lane's runs spawn. Immutable once the lane exists. */
  tool: LaneTool;
+  /** Name of the AI Router profile this lane runs against, or null for the
+   *  vendor's own API. Unlike `tool` this IS changeable after creation, but
+   *  only through `POST /api/lanes/:id/router/apply` — the column and the
+   *  lane's own config on disk must move together. */
+  router_profile: string | null;
  /** For a managed worktree lane, the repo it was provisioned from. Null for
   *  an adopted lane (it IS a source repo, not a worktree of one). */
  source_repo: string | null;
@@ -18,6 +18,11 @@
 * An API key is write-only here by design — the server never returns one, so
 * the form shows whether a key is saved, never the key.
 *
+ * **Test** asks the endpoint itself before the profile is saved onto a lane.
+ * "Saved" and "works" are different claims, and the gap between them is only
+ * ever discovered inside a lane's session otherwise, where the error surfaces
+ * as a broken CLI rather than a bad URL.
+ *
 * @author Nguyễn Ngọc Trí Vĩ <vinnt@smartgift.vn>
 */
 import { useCallback, useEffect, useState } from "react";
@@ -51,6 +56,11 @@ export function AiRouterConfig() {
  const [form, setForm] = useState(emptyForm());
  const [open, setOpen] = useState(false);
  const [busy, setBusy] = useState(false);
+  // The router's own verdict on the form as it currently stands. Cleared on
+  // every field change: a green "works" left sitting above an edited base URL
+  // is a claim about a configuration that no longer exists.
+  const [testResult, setTestResult] = useState<{ ok: boolean; text: string } | null>(null);
+  const [testing, setTesting] = useState(false);
  // Distinct from "no profiles": an empty table rendered before the first fetch
  // resolves tells the user their profiles are gone. CodexConfig draws the same
  // distinction for the same reason.
@@ -77,6 +87,9 @@ export function AiRouterConfig() {
    void load();
  }, [load]);

+  // Any edit to the form invalidates the last verdict — see `testResult`.
+  useEffect(() => setTestResult(null), [form]);
+
  const startAdd = () => {
    setEditing(null);
    setForm(emptyForm());
@@ -105,6 +118,37 @@ export function AiRouterConfig() {
    );
  };

+  /**
+   * Ask the endpoint itself. `name` is sent when editing so a profile whose key
+   * the form cannot show is still testable — the server only honours it while
+   * the base URL is unchanged, so a retyped endpoint needs a retyped key.
+   */
+  const test = async () => {
+    setTesting(true);
+    setTestResult(null);
+    try {
+      const result = await api.aiRouters.test({
+        baseUrl: form.baseUrl.trim(),
+        model: form.model.trim(),
+        ...(form.apiKey.trim() ? { apiKey: form.apiKey.trim() } : {}),
+        ...(editing ? { name: editing.name } : {}),
+      });
+      setTestResult({
+        ok: result.ok,
+        text: result.ok
+          ? t("aiRouter.testOk", { model: result.message })
+          : t("aiRouter.testFailed", {
+              status: result.status ?? "—",
+              message: result.message,
+            }),
+      });
+    } catch (e) {
+      setTestResult({ ok: false, text: e instanceof Error ? e.message : String(e) });
+    } finally {
+      setTesting(false);
+    }
+  };
+
  const submit = async () => {
    setBusy(true);
    setError(null);
@@ -329,7 +373,29 @@ export function AiRouterConfig() {
              </details>
            </div>

-            <div className="mt-4 flex justify-end gap-2">
+            {testResult && (
+              <p
+                data-testid="router-test-result"
+                className={`mt-3 rounded-md border px-3 py-2 text-xs ${
+                  testResult.ok
+                    ? "border-status-success/40 bg-status-success/10 text-status-success"
+                    : "border-status-danger/40 bg-status-danger/10 text-status-danger"
+                }`}
+              >
+                {testResult.text}
+              </p>
+            )}
+
+            <div className="mt-4 flex items-center justify-end gap-2">
+              <button
+                type="button"
+                data-testid="router-test"
+                onClick={test}
+                disabled={testing || busy || !form.baseUrl.trim() || !form.model.trim()}
+                className="mr-auto rounded-md border border-border-light px-3 py-1.5 text-xs text-fg-secondary transition-colors hover:text-fg-primary disabled:opacity-50"
+              >
+                {testing ? t("aiRouter.testing") : t("aiRouter.test")}
+              </button>
              <button
                type="button"
                onClick={() => setOpen(false)}
@@ -14,7 +14,13 @@ import { api } from "../../lib/api";

 vi.mock("../../lib/api", () => ({
  api: {
-    aiRouters: { list: vi.fn(), create: vi.fn(), update: vi.fn(), remove: vi.fn() },
+    aiRouters: {
+      list: vi.fn(),
+      create: vi.fn(),
+      update: vi.fn(),
+      remove: vi.fn(),
+      test: vi.fn(),
+    },
  },
 }));

@@ -69,6 +75,9 @@ beforeEach(() => {
  vi.mocked(api.aiRouters.remove)
    .mockReset()
    .mockResolvedValue({ removed: "9router", lanes: { updated: 0, failed: [] } });
+  vi.mocked(api.aiRouters.test)
+    .mockReset()
+    .mockResolvedValue({ ok: true, status: 200, message: "Highest" });
 });

 describe("AiRouterConfig", () => {
@@ -176,3 +185,75 @@ describe("AiRouterConfig", () => {
    expect(screen.getByRole("dialog")).toBeInTheDocument();
  });
 });
+
+describe("AiRouterConfig — testing an endpoint before saving it", () => {
+  it("sends the typed endpoint, key and model, and reports success", async () => {
+    render(<AiRouterConfig />);
+    const user = userEvent.setup();
+
+    await user.click(await screen.findByRole("button", { name: /aiRouter\.add/ }));
+    await user.type(screen.getByLabelText("aiRouter.fieldBaseUrl"), "http://localhost:20128/v1");
+    await user.type(screen.getByLabelText("aiRouter.fieldApiKey"), "sk-typed");
+    await user.type(screen.getByLabelText("aiRouter.fieldModel"), "Highest");
+    await user.click(screen.getByTestId("router-test"));
+
+    await waitFor(() =>
+      expect(api.aiRouters.test).toHaveBeenCalledWith({
+        baseUrl: "http://localhost:20128/v1",
+        model: "Highest",
+        apiKey: "sk-typed",
+      })
+    );
+    expect(await screen.findByTestId("router-test-result")).toHaveTextContent("aiRouter.testOk");
+  });
+
+  it("names the profile being edited so the server can reuse its stored key", async () => {
+    vi.mocked(api.aiRouters.list).mockResolvedValue({ items: [native, nineRouter] });
+    render(<AiRouterConfig />);
+    const user = userEvent.setup();
+
+    await user.click(await screen.findByLabelText(/aiRouter\.edit 9Router/));
+    await user.click(screen.getByTestId("router-test"));
+
+    await waitFor(() =>
+      expect(api.aiRouters.test).toHaveBeenCalledWith({
+        baseUrl: "http://localhost:20128/v1",
+        model: "Highest",
+        name: "9router",
+      })
+    );
+  });
+
+  it("shows the router's own refusal rather than claiming it works", async () => {
+    vi.mocked(api.aiRouters.test).mockResolvedValue({
+      ok: false,
+      status: 401,
+      message: "invalid api key",
+    });
+    render(<AiRouterConfig />);
+    const user = userEvent.setup();
+
+    await user.click(await screen.findByRole("button", { name: /aiRouter\.add/ }));
+    await user.type(screen.getByLabelText("aiRouter.fieldBaseUrl"), "http://localhost:20128/v1");
+    await user.type(screen.getByLabelText("aiRouter.fieldModel"), "Highest");
+    await user.click(screen.getByTestId("router-test"));
+
+    expect(await screen.findByTestId("router-test-result")).toHaveTextContent(
+      "aiRouter.testFailed"
+    );
+  });
+
+  it("drops a stale verdict as soon as a field changes", async () => {
+    render(<AiRouterConfig />);
+    const user = userEvent.setup();
+
+    await user.click(await screen.findByRole("button", { name: /aiRouter\.add/ }));
+    await user.type(screen.getByLabelText("aiRouter.fieldBaseUrl"), "http://localhost:20128/v1");
+    await user.type(screen.getByLabelText("aiRouter.fieldModel"), "Highest");
+    await user.click(screen.getByTestId("router-test"));
+    expect(await screen.findByTestId("router-test-result")).toBeInTheDocument();
+
+    await user.type(screen.getByLabelText("aiRouter.fieldBaseUrl"), "2");
+    await waitFor(() => expect(screen.queryByTestId("router-test-result")).toBeNull());
+  });
+});
@@ -79,6 +79,7 @@ vi.mock("../../lib/api", async (importOriginal) => {
        features: { list: r({ features: [] }), show: r({ feature: null }) },
        proof: { list: r({ features: [] }), imageUrl: () => "" },
      },
+      aiRouters: { list: r({ items: [] }) },
      sessions: { list: r({ sessions: [], total: 0, limit: 50, offset: 0 }) },
    },
  };
@@ -140,6 +140,14 @@ vi.mock("../../lib/api", async (importOriginal) => {
          }),
        },
      },
+      // The lane card's router picker. Only the built-in profile, which is the
+      // shape that renders no picker — nothing here asserts on it.
+      aiRouters: {
+        list: vi.fn().mockImplementation(async () => {
+          recordCall("GET", "/api/ai-routers");
+          return { items: [] };
+        }),
+      },
      run: {
        list: vi.fn().mockImplementation(async () => {
          recordCall("GET", "/api/run/list");
@@ -9537,7 +9537,7 @@ exports[`screen snapshots > Settings 1`] = `
            <p
              class="text-[10px] text-status-warning/90 mt-1"
            >
-              UI build v1.6.1 (rebuild client to match)
+              UI build v1.6.4 (rebuild client to match)
            </p>
          </div>
          <div
@@ -1723,18 +1723,23 @@ Every history row carries `lane_id` (the lane whose run it belongs to) and `tool
 ```
 GET    /api/ai-routers                Every saved profile, the built-in `native` one first
 POST   /api/ai-routers                Create — Body: { label, baseUrl, apiKey?, model, models? }
+POST   /api/ai-routers/test           Probe an endpoint — Body: { baseUrl, model, apiKey?, name? }
 PATCH  /api/ai-routers/:name          Update the same fields, all optional
 DELETE /api/ai-routers/:name          Delete one saved profile
 ```

 `PATCH` and `DELETE` also answer with `lanes: { updated, failed[] }` — see "an edit reaches the lanes" below.

-A **router profile** is an endpoint that stands in front of Claude Code or Codex (9Router, OpenRouter, a self-hosted gateway) plus the model names to use against it. It powers the **AI Router** tab of Settings (`/settings?tool=ai-router`), which stores profiles and attaches nothing; attaching happens once, when a lane is created.
+A **router profile** is an endpoint that stands in front of Claude Code or Codex (9Router, OpenRouter, a self-hosted gateway) plus the model names to use against it. It powers the **AI Router** tab of Settings (`/settings?tool=ai-router`), which stores profiles and attaches nothing — the tab's **Test** button (`POST /api/ai-routers/test`, below) is its only outbound call. Attaching a profile to a lane happens in the "+ Add lane" modal, or later from the lane's card in Workspace.

 `models` carries the five per-tier names (`opus`, `sonnet`, `haiku`, `smallFast`, `subagent`). Any tier omitted is filled from `model`, so a router that answers to one symbolic name needs one field, not six.

 **No response here ever contains key material.** A saved profile reports `hasKey: true|false` and there is deliberately no `GET /:name` — a route returning one profile in full would be the single way an API key could leave the server. Profiles live one JSON file each under `<CLAUDE_HOME>/agenthub-routers/` (dir `0700`, files `0600`). An omitted `apiKey` on `PATCH` keeps the stored one; the edit form cannot display a key, so it has none to send back.

+**`POST /api/ai-routers/test`** answers the question saving cannot: does this endpoint, with this key, actually serve this model? It sends one Anthropic-style message capped at a single output token to `<baseUrl>/v1/messages` (a 15 s timeout), which is the wire shape a claude lane's eight env keys point at — a Codex-only router is not covered. A router that answers is a `200` here whatever it replies: `{ ok, status, message }` carries the router's own words (truncated to 300 characters), and `status` is `null` when nothing answered at all. Only bad input is a 4xx.
+
+The stored key is only ever sent to the endpoint it was stored against. A blank `apiKey` falls back to the key of the profile named by `name`, but **only** while the submitted `baseUrl` still equals the saved one — so testing a retyped endpoint needs a retyped key (`400 EBADKEY`). Without that rule this route would be a way to make the server POST a saved API key to any URL a browser names.
+
 `native` is a built-in profile, not the absence of one: empty `baseUrl`/`apiKey` and the vendor's own model names, so switching back is the same write path as switching to a router. It answers `409 EBUILTIN` to any update, delete, or attempt to shadow its name. Other errors: `400 EBADLABEL` / `EBADURL` (an endpoint must be `http:` or `https:`) / `EBADMODEL`, `404 ENOTFOUND`.

 **`POST /api/lanes/:id/router/apply`** — Body: `{ profile: "<name>" }` — is where a profile reaches disk. Only the name comes from the request; what is written, and where, is derived from the lane row:
@@ -430,11 +430,17 @@ CREATE TABLE model_pricing (

 Standard rates and intro rates are edited independently: the pricing update path writes intro columns only when the caller sends intro fields, so a standard-rate edit never disturbs a promo (and vice versa). Clearing `intro_until` also zeroes the intro rates.

-**Example default rule (Claude Sonnet 5, with its launch promo):**
+**Example default rules:**

-| Pattern | Input | Output | Intro Input | Intro Output | Intro Until |
-|---------|-------|--------|-------------|--------------|-------------|
-| `claude-sonnet-5%` | $3.00 | $15.00 | $2.00 | $10.00 | `2026-08-31` |
+| Pattern | Input | Output | Cache read | Fast in/out |
+|---------|-------|--------|------------|-------------|
+| `claude-opus-5%` | $5.00 | $25.00 | $0.50 | $10 / $50 |
+| `claude-sonnet-5%` | $2.00 | $10.00 | $0.20 | — |
+| `claude-fable-5-1%` | $10.00 | $50.00 | $0.25 | — |
+
+No rule ships an intro rate today: Sonnet 5's launch promo ran through `2026-08-31` and its discounted rate became the standard one, so `DEFAULT_INTRO_PRICING` is empty. `claude-fable-5-1%` exists alongside `claude-fable-5%` only because 5.1 caches cheaper — that overlap is what the longest-pattern-first match rule above resolves.
+
+**Retired models.** Claude 3.x, Opus 4/4.1 and Sonnet 4 were dropped from the defaults. Because the startup top-up only *adds* missing rows, those rows are also deleted once from existing databases, guarded by `PRAGMA user_version` (0 → 1) so a model you deliberately re-add in Settings survives later restarts. The same one-shot migration corrects the pre-existing Sonnet 5 row to its post-promo $2 / $10 standard rate. Deleting a rule never touches `token_usage` — only the rates used to price it.

 ---

@@ -43,7 +43,7 @@ The modal has a segmented toggle mirroring the CLI's two modes:
 - **Repo** — adopts the given directory as-is (maps to `agenthub lanes add --cwd`). No branch fields; the auto-setup summary above does not run (adopting is instant, nothing to wait on).
 - **Worktree** (default) — provisions a managed worktree (maps to `agenthub lanes add --repo`). Unlike the CLI, the modal requires you to type the new branch's name yourself rather than deriving one from the title — the underlying route still derives one when `branch` is omitted, so the CLI's behavior is unchanged.

-Under the *Pipeline template* select, a **Model endpoint** radio chooses between **Native** (the vendor's own API) and **AI Router** — a profile saved in Settings › AI Router (9Router, OpenRouter, a self-hosted gateway). Choosing one writes it into the new lane's own config as the lane is created: a claude lane gets the `env` block of its `.claude/settings.local.json`, a codex lane gets a `$CODEX_HOME/<name>.config.toml` layer plus a `-p <name>` on every spawn. The UI offers the choice only at creation, and with no profile saved the radio stays on Native and says so. Unlike the lane's tool, though, this one is not immutable: `POST /api/lanes/:id/router/apply` re-points an existing lane, rewriting its config and its `router_profile` column together. A session already running in the pane keeps the endpoint it started with — Claude Code reads `settings.local.json` at startup — so the change takes effect on the next run. Editing or deleting the profile itself in Settings also reaches every lane using it: an edit re-applies it (so a rotated key actually lands), a delete resets those lanes to Native. See `docs/API.md` § AI Router profiles.
+Under the *Pipeline template* select, a **Model endpoint** radio chooses between **Native** (the vendor's own API) and **AI Router** — a profile saved in Settings › AI Router (9Router, OpenRouter, a self-hosted gateway). Choosing one writes it into the new lane's own config as the lane is created: a claude lane gets the `env` block of its `.claude/settings.local.json`, a codex lane gets a `$CODEX_HOME/<name>.config.toml` layer plus a `-p <name>` on every spawn. With no profile saved the radio stays on Native and says so. Unlike the lane's tool, this one is not immutable: an existing lane's card in Workspace carries an **AI Router** select showing its current profile, and picking another calls `POST /api/lanes/:id/router/apply`, rewriting the lane's config and its `router_profile` column together. (The select is hidden when `native` is the only profile — there would be nothing to choose.) A failed apply puts the select back rather than leaving the card claiming a router the lane did not get. A session already running in the pane keeps the endpoint it started with — Claude Code reads `settings.local.json` at startup — so the change takes effect on the next run. Editing or deleting the profile itself in Settings also reaches every lane using it: an edit re-applies it (so a rotated key actually lands), a delete resets those lanes to Native. See `docs/API.md` § AI Router profiles.

 Under the *Pipeline template* select, a **Create/update readme file** checkbox (Claude lanes only — Codex has no slash commands) starts the new lane's terminal run right after creation with `/sg-all:sg-md-writing` as its first message, so the lane opens already writing/refreshing its `CLAUDE.md`/`AGENTS.md` set. It requires the `sg-all` plugin to be installed in Claude Code; without it the session simply reports an unknown command. Leave it unticked to create the lane without starting a run.

@@ -129,13 +129,14 @@ milliseconds — the real work happens in a detached worker, so a session never
 waits on an install:

 1. **Fast path** — recorded state matches this plugin build and a server is live → exit.
-2. **Node gate** — refuse below 22.5 with one line (see Prerequisites).
-3. **Lock** — atomic `mkdir` lock holding the PID; reclaimed when the owner is dead or the lock is older than 10 minutes, so two sessions cannot race the install or spawn two servers.
-4. **Install** — `npm install --omit=dev --ignore-scripts` into the runtime dir (never into the plugin cache, which is garbage-collected and replaced on every update). `--ignore-scripts` keeps the root `postinstall` from pulling the whole Vite client toolchain.
-5. **Legacy hook cleanup** — see above.
-6. **CLI** — write the `~/.local/bin/agenthub` launcher, never clobbering a `agenthub` the bootstrap did not write.
-7. **UI build** — `client/` is copied into the runtime dir and built there (`npm ci --include=dev`, falling back to `npm install --include=dev`, then `npm run build`), landing in `runtime/client-dist` — the same directory the server serves from. `--include=dev` is required, not tidy: the entire build toolchain (`vite`, `typescript`, `@vitejs/plugin-react`, `tailwindcss`, `@types/*`) is a devDependency, and npm reads an inherited `NODE_ENV=production` as `--omit=dev`. Every bootstrap descended from the running server carries that env — the self-update chain, and any `SessionStart` fired by a `claude` inside a tmux pane the server started — so without the flag the install drops `vite` and the build dies on missing types. Skipped when a bundle for the current version already exists; a failure here does not fail the bootstrap (the API still works, and `/agenthub-open` can retry) but does mean `/run` 404s until it's fixed — and when a bundle from the PREVIOUS version is still on disk the failure is quieter still: the dashboard keeps serving the old UI against the new API, visible only as the Settings page's `UI build vX (rebuild client to match)` line.
-8. **Server** — on Linux with `systemd --user` available, install/refresh a
+2. **Superseded check** — if Claude Code records a DIFFERENT build of this plugin as installed, this one logs a line and does nothing else (see [A retired build never acts](#a-retired-build-never-acts)).
+3. **Node gate** — refuse below 22.5 with one line (see Prerequisites).
+4. **Lock** — atomic `mkdir` lock holding the PID; reclaimed when the owner is dead or the lock is older than 10 minutes, so two sessions cannot race the install or spawn two servers.
+5. **Install** — `npm install --omit=dev --ignore-scripts` into the runtime dir (never into the plugin cache, which is garbage-collected and replaced on every update). `--ignore-scripts` keeps the root `postinstall` from pulling the whole Vite client toolchain.
+6. **Legacy hook cleanup** — see above.
+7. **CLI** — write the `~/.local/bin/agenthub` launcher, never clobbering a `agenthub` the bootstrap did not write.
+8. **UI build** — `client/` is copied into the runtime dir and built there (`npm ci --include=dev`, falling back to `npm install --include=dev`, then `npm run build`), landing in `runtime/client-dist` — the same directory the server serves from. `--include=dev` is required, not tidy: the entire build toolchain (`vite`, `typescript`, `@vitejs/plugin-react`, `tailwindcss`, `@types/*`) is a devDependency, and npm reads an inherited `NODE_ENV=production` as `--omit=dev`. Every bootstrap descended from the running server carries that env — the self-update chain, and any `SessionStart` fired by a `claude` inside a tmux pane the server started — so without the flag the install drops `vite` and the build dies on missing types. Skipped when a bundle for the current version already exists; a failure here does not fail the bootstrap (the API still works, and `/agenthub-open` can retry) but does mean `/run` 404s until it's fixed — and when a bundle from the PREVIOUS version is still on disk the failure is quieter still: the dashboard keeps serving the old UI against the new API, visible only as the Settings page's `UI build vX (rebuild client to match)` line.
+9. **Server** — on Linux with `systemd --user` available, install/refresh a
   `agenthub.service` unit (see [Auto-start on boot](#auto-start-on-boot)) instead
   of a bare spawn, so the server is supervised and comes back after a reboot
   with no Claude Code session involved. Everywhere else (macOS, Windows, or
@@ -143,10 +144,11 @@ waits on an install:
   runtime `node_modules` and `DASHBOARD_CLIENT_DIST` at the runtime
   `client-dist` — scoped to this session, same as before this fell back to
   systemd.
-9. **Record state** — `runtime/state.json`.
-10. **Hand over to systemd** — only when a unit of ours already owns the server
+10. **Record state** — `runtime/state.json`.
+11. **Prune retired cache copies** — delete the plugin-cache directories earlier updates left behind (see [Pruning the plugin cache](#pruning-the-plugin-cache)). After step 10, so a prune that goes wrong cannot also cost the record of a bootstrap that went right.
+12. **Hand over to systemd** — only when a unit of ours already owns the server
    (see [Who may stop the server](#who-may-stop-the-server)). Deliberately
-    after step 9.
+    after step 10.

 The first run takes a few minutes — most of it is the UI build, which is what
 makes client-only routes (`http://localhost:4820/run`, and every other page)
@@ -155,6 +157,66 @@ goes to `~/.claude/agenthub/runtime/bootstrap.log`; `npm run build`'s own
 output goes to `client-build.log` next to it; the server's own output goes to
 `server.log`.

+### A retired build never acts
+
+`claude plugin update` unpacks the new version into a **new** cache directory
+and prints "Restart to apply changes". A Claude Code process that was already
+running does not: it keeps loading the OLD directory — skills, commands and
+hooks included — until it is restarted.
+
+That used to be self-inflicting. Every session such a process opened ran the
+OLD `plugin-bootstrap.js`, which saw a `state.json` naming a different root,
+concluded the server had to be replaced, repointed `agenthub.service` at itself
+and restarted the dashboard onto the version the user had just upgraded away
+from. The dashboard then reported the old version and offered the update again
+— once per session, forever, while `installed_plugins.json` and the cache both
+said the new version was installed. A `/agenthub-update` from that session
+"succeeded" and changed nothing.
+
+So the bootstrap now asks first (`supersededBy`): if Claude Code's
+`~/.claude/plugins/installed_plugins.json` records a **sibling** directory as
+the installed build of this plugin, this build logs one line and returns
+`superseded` without touching dependencies, the unit, or the server. `--force`
+does not override it — the self-update chain runs `--force` from the NEW
+directory, which is never superseded, so the only `--force` this can refuse is
+one that would deliberately downgrade.
+
+It cannot hand over to the new build either: the process it would have to
+replace is the one loading it. Restarting Claude Code is the fix, and the log
+line says so.
+
+Matching is by sibling directory rather than by the `<plugin>@<marketplace>`
+key, which holds without parsing a cache path or trusting a name. An entry
+pointing at a directory that no longer exists is ignored — a phantom must never
+wedge a bootstrap shut.
+
+### Pruning the plugin cache
+
+Claude Code unpacks every update beside the last one and does not reliably
+remove the old. A machine that tracks releases closely accumulates a full copy
+of this repo — `node_modules` and all — per release; a dozen of them is
+gigabytes.
+
+Step 11 of the bootstrap (`pruneRetiredPluginRoots`) deletes them. This is the
+only place anything in the plugin touches the cache, and the rule it appears to
+break ("never write into the plugin cache") is about the *runtime* living
+outside it, not about keeping retired copies forever.
+
+Four independent readings of "still in use" are protected, and nothing matching
+any of them is ever deleted:
+
+| Protected | Why |
+|---|---|
+| This build | It is running. |
+| Every `installPath` in `installed_plugins.json`, at any scope | Claude Code says it is installed. |
+| The directory `agenthub.service` names (`unitRoot`, **foreign unit included**) | A service is executing from it. Not ours to rewrite is still not ours to delete. |
+| Any directory named in the user's `settings.json` hooks | A Claude Code process that has not been restarted is still executing from there — this is the same directory the section above describes, seen from the other side. Deleting it would break that session's hooks mid-flight. |
+
+Plus the newest `KEEP_RETIRED_ROOTS` (2) of whatever is left, as a rollback
+cushion. Outside `~/.claude/plugins/cache/` the pruner does nothing at all — for
+a git checkout the parent directory holds the user's own work. Best effort
+throughout: a directory that will not delete is left alone, never an error.
+
 ### Auto-start on boot

 On Linux, every plugin bootstrap (not just the first) installs/refreshes a
@@ -3,7 +3,7 @@
 openapi: 3.0.3
 info:
  title: Agent Dashboard for Claude Code API
-  version: 1.6.1
+  version: 1.6.4
  description: HTTP API for real-time Claude Code session monitoring, agent lifecycle tracking, analytics, pricing, hooks ingestion, and workflow intelligence.
  contact:
    name: Nguyễn Ngọc Trí Vĩ
@@ -1,12 +1,12 @@
 {
  "name": "agenthub",
-  "version": "1.6.1",
+  "version": "1.6.4",
  "lockfileVersion": 3,
  "requires": true,
  "packages": {
    "": {
      "name": "agenthub",
-      "version": "1.6.1",
+      "version": "1.6.4",
      "hasInstallScript": true,
      "license": "UNLICENSED",
      "dependencies": {
@@ -1943,7 +1943,7 @@
      "license": "MIT"
    },
    "node_modules/path-expression-matcher": {
-      "version": "1.6.1",
+      "version": "1.6.4",
      "resolved": "https://registry.npmjs.org/path-expression-matcher/-/path-expression-matcher-1.6.1.tgz",
      "integrity": "sha512-h7bxdzhHk8Knyc4Tj+jMaa7fEEoUJy7p1qtbVgkYg1Uhpe5Np5VuGXCRZnkZvU+Q42M1vStt0ifa3ueykRJPmQ==",
      "funding": [
@@ -1,6 +1,6 @@
 {
  "name": "agenthub",
-  "version": "1.6.1",
+  "version": "1.6.4",
  "bin": {
    "agenthub": "bin/agenthub.js"
  },
@@ -14,8 +14,14 @@
 * scoped to this session, same as before.
 *
 * Everything writable lives under `~/.claude/agenthub/runtime/` — NOT in
- * the plugin cache, which Claude Code garbage-collects and replaces wholesale
- * on every plugin update.
+ * the plugin cache, which Claude Code replaces wholesale on every plugin
+ * update. The one thing this file does reach into the cache for is deleting
+ * the copies those updates leave behind (`pruneRetiredPluginRoots`).
+ *
+ * A bootstrap that is no longer the installed build does nothing at all
+ * (`supersededBy`): an old cache directory, still loaded by a Claude Code
+ * process that has not been restarted, would otherwise repoint systemd at
+ * itself and downgrade the dashboard once per session, forever.
 * @author Nguyễn Ngọc Trí Vĩ <vinnt@smartgift.vn>
 */

@@ -40,14 +46,65 @@ function runtimeDir() {
  return path.join(getDataDir(), "runtime");
 }

-function pluginVersion() {
+function pluginVersion(root = PLUGIN_ROOT) {
  try {
-    return require(path.join(PLUGIN_ROOT, "package.json")).version;
+    return require(path.join(root, "package.json")).version;
  } catch {
    return "unknown";
  }
 }

+/* -------------------------------------------------- superseded plugin build */
+
+/** Claude Code's own record of which build of each plugin is installed. */
+function installedPluginsPath() {
+  return path.join(getClaudeHome(), "plugins", "installed_plugins.json");
+}
+
+/**
+ * The build Claude Code has installed IN PLACE OF this one, or null when this
+ * build is the installed one (or when there is no such record to read).
+ *
+ * `claude plugin update` unpacks the new version into a new cache directory and
+ * says "Restart to apply changes" — but a Claude Code process that was already
+ * running keeps loading the OLD directory, hooks and all. Every session it
+ * opens from then on runs the OLD bootstrap, which happily repoints the systemd
+ * unit at itself and restarts the dashboard onto the version the user just
+ * upgraded away from. The dashboard then reports the old version and offers the
+ * update again — forever, once per session, until Claude Code is restarted.
+ *
+ * So a bootstrap that is no longer the installed build does NOTHING. It cannot
+ * hand over either: the process it would have to replace is the one loading it.
+ *
+ * Matched by SIBLING directory rather than by `<plugin>@<marketplace>` key: the
+ * builds of one plugin are exactly the directories beside this one, and that
+ * holds without parsing a cache path or trusting a name. An entry naming a
+ * directory that no longer exists is ignored — a phantom must never wedge a
+ * bootstrap shut.
+ */
+function supersededBy(root = PLUGIN_ROOT) {
+  let data;
+  try {
+    data = JSON.parse(fs.readFileSync(installedPluginsPath(), "utf8"));
+  } catch {
+    return null; // no record, unreadable, or not a plugin install at all
+  }
+  const siblings = path.dirname(root);
+  const found = [];
+  for (const entries of Object.values((data && data.plugins) || {})) {
+    for (const entry of entries || []) {
+      const installed =
+        entry && typeof entry.installPath === "string" ? path.resolve(entry.installPath) : null;
+      if (!installed) continue;
+      // This build IS installed (possibly at one scope among several) — the
+      // whole question is settled, whatever any other entry says.
+      if (installed === root) return null;
+      if (path.dirname(installed) === siblings && fs.existsSync(installed)) found.push(installed);
+    }
+  }
+  return found[0] || null;
+}
+
 /* ------------------------------------------------------------------ state */

 function statePath(rt = runtimeDir()) {
@@ -447,16 +504,133 @@ function autoInstallSkillPlugins(rt, { force = false } = {}) {
  }
 }

+/* ------------------------------------------------------ plugin-cache pruning */
+
+/** How many retired builds to leave behind, newest first, as a rollback cushion. */
+const KEEP_RETIRED_ROOTS = 2;
+
+/**
+ * Delete the cache directories earlier updates left behind.
+ *
+ * Claude Code unpacks every plugin update into a NEW directory beside the old
+ * one and does not reliably remove the old. A machine that tracks releases
+ * closely ends up with a dozen full copies of this repo — node_modules and all
+ * — which is what this clears.
+ *
+ * This is the one place anything here touches the plugin cache, and the file
+ * header's "never write into the cache" rule is about the RUNTIME living
+ * outside it, not about leaving retired copies forever. Everything still in use
+ * is protected, by four independent readings of "in use":
+ *
+ *  - this build;
+ *  - every build Claude Code records as installed, at any scope;
+ *  - the directory our systemd unit runs from, foreign unit included;
+ *  - any directory still named in the user's `settings.json` hooks — a Claude
+ *    Code process that has not been restarted is still executing from there,
+ *    and deleting it would break that session's hooks mid-flight. (This is
+ *    exactly the directory `supersededBy` describes, seen from the other side.)
+ *
+ * …plus the newest `KEEP_RETIRED_ROOTS` of whatever is left, so a rollback has
+ * something to roll back to. Best effort throughout: a directory that will not
+ * delete is left alone, never an error.
+ *
+ * @returns {number} how many directories were removed
+ */
+function pruneRetiredPluginRoots(rt, root = PLUGIN_ROOT) {
+  // Only inside Claude Code's plugin cache. For a git checkout the parent
+  // directory holds the user's own work and is never ours to delete from.
+  const cacheMarker = `${path.sep}plugins${path.sep}cache${path.sep}`;
+  if (!root.includes(cacheMarker)) return 0;
+  const cacheDir = path.dirname(root);
+
+  const keep = new Set([root]);
+  try {
+    const data = JSON.parse(fs.readFileSync(installedPluginsPath(), "utf8"));
+    for (const entries of Object.values((data && data.plugins) || {})) {
+      for (const entry of entries || []) {
+        if (entry && typeof entry.installPath === "string")
+          keep.add(path.resolve(entry.installPath));
+      }
+    }
+  } catch {
+    /* no record to read — the other three protections still apply */
+  }
+  try {
+    const unitRoot = require("./systemd-service").unitRoot("agenthub");
+    if (unitRoot) keep.add(path.resolve(unitRoot));
+  } catch {
+    /* best effort */
+  }
+  let settings = "";
+  try {
+    settings = fs.readFileSync(getSettingsPath(), "utf8");
+  } catch {
+    /* no settings file — nothing pinned by a hook */
+  }
+
+  let dirs;
+  try {
+    dirs = fs.readdirSync(cacheDir, { withFileTypes: true });
+  } catch {
+    return 0;
+  }
+
+  const retired = dirs
+    .filter((d) => d.isDirectory())
+    .map((d) => path.join(cacheDir, d.name))
+    .filter((dir) => !keep.has(dir) && !settings.includes(dir))
+    .map((dir) => {
+      let at = 0;
+      try {
+        at = fs.statSync(dir).mtimeMs;
+      } catch {
+        /* unreadable sorts oldest, so it goes first */
+      }
+      return { dir, at };
+    })
+    .sort((a, b) => b.at - a.at)
+    .slice(KEEP_RETIRED_ROOTS);
+
+  let removed = 0;
+  for (const { dir } of retired) {
+    try {
+      fs.rmSync(dir, { recursive: true, force: true });
+      removed += 1;
+    } catch {
+      /* in use, read-only, whatever — leaving it is always safe */
+    }
+  }
+  if (removed) {
+    log(rt, `removed ${removed} retired plugin cache ${removed === 1 ? "copy" : "copies"}`);
+  }
+  return removed;
+}
+
 /**
 * The real work. Runs detached from the SessionStart hook, so its only channel
 * is the log file; `/agenthub-update` runs it in the foreground with `force`, which
 * reinstalls dependencies and restarts the server unconditionally.
 *
- * @returns {"ok"|"locked"|"node-too-old"|"install-failed"}
+ * @returns {"ok"|"locked"|"node-too-old"|"install-failed"|"superseded"}
 */
 function bootstrap(rt = runtimeDir(), { force = false } = {}) {
  fs.mkdirSync(rt, { recursive: true });

+  // First, and ahead of the lock: a retired build has nothing to do and must
+  // not take the lock to discover it. `force` does NOT override this — the
+  // plugin-update chain runs `--force` from the NEW cache directory, which is
+  // never superseded, so the only `--force` this can refuse is one that would
+  // downgrade the user on purpose.
+  const newer = supersededBy();
+  if (newer) {
+    log(
+      rt,
+      `this build (${pluginVersion()}) has been replaced by ${pluginVersion(newer)} — doing nothing. ` +
+        `Restart Claude Code so it loads the new one; until then this session keeps running the old plugin.`
+    );
+    return "superseded";
+  }
+
  if (!nodeVersionOk()) {
    log(
      rt,
@@ -621,6 +795,10 @@ function bootstrap(rt = runtimeDir(), { force = false } = {}) {
      },
      rt
    );
+
+    // After the state is durable, so a prune that goes wrong cannot also cost
+    // us the record of a bootstrap that went right.
+    pruneRetiredPluginRoots(rt);
    return "ok";
  } finally {
    releaseLock(rt);
@@ -657,7 +835,14 @@ if (require.main === module) {
    // npm output and the result.
    const result = bootstrap(runtimeDir(), { force });
    if (force) {
-      console.log(result === "ok" ? "AgentHub runtime refreshed." : `bootstrap: ${result}`);
+      console.log(
+        result === "ok"
+          ? "AgentHub runtime refreshed."
+          : result === "superseded"
+            ? "A newer AgentHub plugin build is already installed — this one did nothing on purpose.\n" +
+              "Restart Claude Code so it loads the new build, then try again."
+            : `bootstrap: ${result}`
+      );
      if (result !== "ok") process.exitCode = 1;
    }
  } else {
@@ -672,6 +857,11 @@ if (require.main === module) {
 module.exports = {
  MIN_NODE,
  LOCK_STALE_MS,
+  KEEP_RETIRED_ROOTS,
+  supersededBy,
+  pruneRetiredPluginRoots,
+  installedPluginsPath,
+  pluginVersion,
  runtimeDir,
  statePath,
  readState,
@@ -85,6 +85,24 @@ function unitIsStale(name, root) {
  return !body.includes(`\nWorkingDirectory=${root}\n`);
 }

+/**
+ * The directory our unit actually runs from, or null when there is no unit.
+ *
+ * Deliberately does NOT require the marker: a FOREIGN unit named `agenthub` is
+ * not ours to rewrite, but whatever directory it runs from is still a directory
+ * nothing here may delete. The one caller is the plugin-cache pruner.
+ */
+function unitRoot(name) {
+  let body;
+  try {
+    body = fs.readFileSync(unitPath(name), "utf8");
+  } catch {
+    return null;
+  }
+  const match = body.match(/^WorkingDirectory=(.*)$/m);
+  return match ? match[1].trim() : null;
+}
+
 /** Is the unit currently running? Costs a `systemctl` call — never call it
 *  from the fast path. */
 function unitIsActive(name) {
@@ -209,6 +227,7 @@ module.exports = {
  systemctlAvailable,
  unitIsOurs,
  unitIsStale,
+  unitRoot,
  unitIsActive,
  writeUserService,
  restartUserService,
@@ -188,3 +188,86 @@ describe("ai-routers — a corrupt file is skipped, never fatal", () => {
    assert.ok(!items.some((i) => i.name === "broken"));
  });
 });
+
+describe("ai-routers — testRouter", () => {
+  const realFetch = global.fetch;
+  after(() => {
+    global.fetch = realFetch;
+  });
+
+  /** Capture the one request testRouter makes, and answer it however asked. */
+  function stubFetch(answer) {
+    const calls = [];
+    global.fetch = async (url, init) => {
+      calls.push({ url, init });
+      if (answer instanceof Error) throw answer;
+      return answer;
+    };
+    return calls;
+  }
+
+  const ok = () => ({ ok: true, status: 200, statusText: "OK", text: async () => "{}" });
+
+  it("posts one capped message to the endpoint's /v1/messages", async () => {
+    const calls = stubFetch(ok());
+    const result = await routers.testRouter(nineRouter());
+    assert.equal(result.ok, true);
+    assert.equal(calls.length, 1);
+    assert.equal(calls[0].url, "http://localhost:20128/v1/v1/messages");
+    const body = JSON.parse(calls[0].init.body);
+    assert.equal(body.max_tokens, 1);
+    assert.equal(body.model, "Highest");
+    assert.equal(calls[0].init.headers["x-api-key"], "sk-secret-value");
+  });
+
+  it("reports a router that answered with an error rather than throwing", async () => {
+    stubFetch({ ok: false, status: 401, statusText: "Unauthorized", text: async () => "bad key" });
+    const result = await routers.testRouter(nineRouter());
+    assert.deepEqual(result, { ok: false, status: 401, message: "bad key" });
+  });
+
+  it("reports an endpoint that never answered", async () => {
+    // The shape node's own fetch throws: a useless outer message wrapping the
+    // one thing the user can act on.
+    stubFetch(
+      Object.assign(new Error("fetch failed"), {
+        cause: new Error("connect ECONNREFUSED 127.0.0.1:20199"),
+      })
+    );
+    const result = await routers.testRouter(nineRouter());
+    assert.equal(result.ok, false);
+    assert.equal(result.status, null);
+    assert.match(result.message, /ECONNREFUSED/);
+  });
+
+  it("falls back to the stored key for the endpoint it was stored against", async () => {
+    const created = routers.createRouter(nineRouter());
+    const calls = stubFetch(ok());
+    await routers.testRouter({
+      name: created.name,
+      baseUrl: "http://localhost:20128/v1",
+      model: "Highest",
+    });
+    assert.equal(calls[0].init.headers["x-api-key"], "sk-secret-value");
+  });
+
+  it("never sends the stored key to a different endpoint", async () => {
+    const created = routers.createRouter(nineRouter());
+    const calls = stubFetch(ok());
+    await assert.rejects(
+      routers.testRouter({
+        name: created.name,
+        baseUrl: "https://attacker.example/v1",
+        model: "Highest",
+      }),
+      { code: "EBADKEY" }
+    );
+    assert.equal(calls.length, 0);
+  });
+
+  it("refuses a base URL that is not http(s)", async () => {
+    await assert.rejects(routers.testRouter(nineRouter({ baseUrl: "file:///etc/passwd" })), {
+      code: "EBADURL",
+    });
+  });
+});
@@ -488,3 +488,127 @@ describe("buildClient dependency install", () => {
    );
  });
 });
+
+describe("supersededBy — an old cache directory must not act", () => {
+  const pluginsDir = path.join(TMP_HOME, "plugins");
+  const cacheDir = path.join(pluginsDir, "cache", "sg", "agenthub");
+  const oldRoot = path.join(cacheDir, "aaaaaaaaaaaa");
+  const newRoot = path.join(cacheDir, "bbbbbbbbbbbb");
+
+  function writeInstalled(installPath) {
+    fs.mkdirSync(pluginsDir, { recursive: true });
+    fs.writeFileSync(
+      path.join(pluginsDir, "installed_plugins.json"),
+      JSON.stringify({
+        version: 2,
+        plugins: { "agenthub@sg": [{ scope: "user", installPath }] },
+      })
+    );
+  }
+
+  beforeEach(() => {
+    fs.rmSync(pluginsDir, { recursive: true, force: true });
+    fs.mkdirSync(oldRoot, { recursive: true });
+    fs.mkdirSync(newRoot, { recursive: true });
+  });
+
+  it("names the installed build when this one is not it", () => {
+    writeInstalled(newRoot);
+    assert.equal(boot.supersededBy(oldRoot), newRoot);
+  });
+
+  it("clears the build that IS installed", () => {
+    writeInstalled(newRoot);
+    assert.equal(boot.supersededBy(newRoot), null);
+  });
+
+  it("ignores an installed path that no longer exists", () => {
+    // A phantom entry must never wedge a bootstrap shut.
+    writeInstalled(path.join(cacheDir, "cccccccccccc"));
+    assert.equal(boot.supersededBy(oldRoot), null);
+  });
+
+  it("ignores a plugin that is not a sibling of this one", () => {
+    fs.mkdirSync(pluginsDir, { recursive: true });
+    const other = path.join(pluginsDir, "cache", "sg", "sg-all", "0.1.0");
+    fs.mkdirSync(other, { recursive: true });
+    fs.writeFileSync(
+      path.join(pluginsDir, "installed_plugins.json"),
+      JSON.stringify({ plugins: { "sg-all@sg": [{ installPath: other }] } })
+    );
+    assert.equal(boot.supersededBy(oldRoot), null);
+  });
+
+  it("says nothing when there is no record to read (a git checkout)", () => {
+    assert.equal(boot.supersededBy(oldRoot), null);
+  });
+});
+
+describe("pruneRetiredPluginRoots", () => {
+  const pluginsDir = path.join(TMP_HOME, "plugins");
+  const cacheDir = path.join(pluginsDir, "cache", "sg", "agenthub");
+  const settingsPath = path.join(TMP_HOME, "settings.json");
+
+  /** Six builds, oldest first, so mtime order is deterministic. */
+  function seed(names) {
+    fs.rmSync(pluginsDir, { recursive: true, force: true });
+    let t = Date.now() - names.length * 60_000;
+    for (const n of names) {
+      const dir = path.join(cacheDir, n);
+      fs.mkdirSync(dir, { recursive: true });
+      fs.utimesSync(dir, new Date(t), new Date(t));
+      t += 60_000;
+    }
+  }
+
+  const NAMES = ["v1", "v2", "v3", "v4", "v5", "current"];
+  const dir = (n) => path.join(cacheDir, n);
+
+  beforeEach(() => {
+    resetRuntime();
+    seed(NAMES);
+    fs.rmSync(settingsPath, { force: true });
+    fs.writeFileSync(path.join(pluginsDir, "installed_plugins.json"), "{}");
+  });
+
+  it("keeps the current build and the two newest retired ones", () => {
+    const removed = boot.pruneRetiredPluginRoots(RT, dir("current"));
+    assert.equal(removed, 3); // v1 v2 v3 go; v4 v5 are the cushion
+    assert.ok(fs.existsSync(dir("current")));
+    assert.ok(fs.existsSync(dir("v5")));
+    assert.ok(fs.existsSync(dir("v4")));
+    assert.ok(!fs.existsSync(dir("v1")));
+  });
+
+  it("never deletes a build Claude Code records as installed", () => {
+    fs.writeFileSync(
+      path.join(pluginsDir, "installed_plugins.json"),
+      JSON.stringify({ plugins: { "agenthub@sg": [{ installPath: dir("v1") }] } })
+    );
+    boot.pruneRetiredPluginRoots(RT, dir("current"));
+    assert.ok(fs.existsSync(dir("v1")));
+  });
+
+  it("never deletes a build the user's hooks still point at", () => {
+    // A Claude Code process that has not been restarted is still executing
+    // from there — deleting it breaks that session's hooks mid-flight.
+    fs.writeFileSync(
+      settingsPath,
+      JSON.stringify({
+        hooks: { SessionStart: [{ hooks: [{ command: `node "${dir("v2")}/scripts/x.js"` }] }] },
+      })
+    );
+    boot.pruneRetiredPluginRoots(RT, dir("current"));
+    assert.ok(fs.existsSync(dir("v2")));
+    assert.ok(!fs.existsSync(dir("v1")));
+  });
+
+  it("refuses to touch anything outside the plugin cache", () => {
+    const checkout = path.join(TMP_HOME, "checkout", "agenthub");
+    const sibling = path.join(TMP_HOME, "checkout", "something-precious");
+    fs.mkdirSync(checkout, { recursive: true });
+    fs.mkdirSync(sibling, { recursive: true });
+    assert.equal(boot.pruneRetiredPluginRoots(RT, checkout), 0);
+    assert.ok(fs.existsSync(sibling));
+  });
+});
@@ -168,6 +168,25 @@ describe("systemd-service — unitIsStale", linuxOnly, () => {
  });
 });

+describe("systemd-service — unitRoot", linuxOnly, () => {
+  it("reports the directory the unit runs from", () => {
+    svc.writeUserService({ ...UNIT, workingDir: "/plugin/root/old" });
+    assert.equal(svc.unitRoot("agenthub"), "/plugin/root/old");
+  });
+
+  it("reports a FOREIGN unit's directory too", () => {
+    // Not ours to rewrite, but still a directory the cache pruner may not
+    // delete out from under a running service.
+    fs.mkdirSync(path.dirname(svc.unitPath("agenthub")), { recursive: true });
+    fs.writeFileSync(svc.unitPath("agenthub"), "[Service]\nWorkingDirectory=/opt/mine\n", "utf8");
+    assert.equal(svc.unitRoot("agenthub"), "/opt/mine");
+  });
+
+  it("is null when there is no unit", () => {
+    assert.equal(svc.unitRoot("agenthub"), null);
+  });
+});
+
 describe("systemd-service — unitIsActive", linuxOnly, () => {
  it("is true only for `active`", () => {
    svc.__setExecImpl((args) => ({ status: 0, stdout: args[0] === "is-active" ? "active\n" : "" }));
@@ -828,29 +828,39 @@ try {
 // Each model gets its own explicit row — no catch-all grouping.
 // Rate shape mirrors Anthropic's published table: 5m write = 1.25× input, 1h write = 2× input.
 const DEFAULT_PRICING = [
-  // Next-gen flagship
+  // Next-gen flagship (5.1 caches cheaper than 5 — needs its own row)
+  ["claude-fable-5-1%", "Claude Fable 5.1", 10, 50, 0.25, 12.5, 20, 0, 0],
+  ["claude-mythos-5-1%", "Claude Mythos 5.1", 10, 50, 0.25, 12.5, 20, 0, 0],
  ["claude-fable-5%", "Claude Fable 5", 10, 50, 1, 12.5, 20, 0, 0],
  ["claude-mythos-5%", "Claude Mythos 5", 10, 50, 1, 12.5, 20, 0, 0],
-  // Opus family (fast mode available on 4.6 / 4.7 / 4.8)
+  // Opus family (fast mode available on 5 / 4.6 / 4.7 / 4.8)
+  ["claude-opus-5%", "Claude Opus 5", 5, 25, 0.5, 6.25, 10, 10, 50],
  ["claude-opus-4-8%", "Claude Opus 4.8", 5, 25, 0.5, 6.25, 10, 10, 50],
  ["claude-opus-4-7%", "Claude Opus 4.7", 5, 25, 0.5, 6.25, 10, 30, 150],
  ["claude-opus-4-6%", "Claude Opus 4.6", 5, 25, 0.5, 6.25, 10, 30, 150],
  ["claude-opus-4-5%", "Claude Opus 4.5", 5, 25, 0.5, 6.25, 10, 0, 0],
-  ["claude-opus-4-1%", "Claude Opus 4.1", 15, 75, 1.5, 18.75, 30, 0, 0],
-  ["claude-opus-4-2%", "Claude Opus 4", 15, 75, 1.5, 18.75, 30, 0, 0],
  // Sonnet family
-  ["claude-sonnet-5%", "Claude Sonnet 5", 3, 15, 0.3, 3.75, 6, 0, 0],
+  ["claude-sonnet-5%", "Claude Sonnet 5", 2, 10, 0.2, 2.5, 4, 0, 0],
  ["claude-sonnet-4-6%", "Claude Sonnet 4.6", 3, 15, 0.3, 3.75, 6, 0, 0],
  ["claude-sonnet-4-5%", "Claude Sonnet 4.5", 3, 15, 0.3, 3.75, 6, 0, 0],
-  ["claude-sonnet-4-2%", "Claude Sonnet 4", 3, 15, 0.3, 3.75, 6, 0, 0],
-  ["claude-3-7-sonnet%", "Claude Sonnet 3.7", 3, 15, 0.3, 3.75, 6, 0, 0],
-  ["claude-3-5-sonnet%", "Claude Sonnet 3.5", 3, 15, 0.3, 3.75, 6, 0, 0],
  // Haiku family
  ["claude-haiku-4-5%", "Claude Haiku 4.5", 1, 5, 0.1, 1.25, 2, 0, 0],
-  ["claude-3-5-haiku%", "Claude Haiku 3.5", 0.8, 4, 0.08, 1, 1.6, 0, 0],
-  ["claude-3-haiku%", "Claude Haiku 3", 0.25, 1.25, 0.03, 0.3, 0.5, 0, 0],
-  // Legacy
-  ["claude-3-opus%", "Claude Opus 3", 15, 75, 1.5, 18.75, 30, 0, 0],
+];
+
+// Retired models (Claude 3.x, Opus 4/4.1, Sonnet 4): removed from the defaults
+// AND pruned ONCE from existing DBs, since the top-up below only ever adds rows.
+// Once-only (guarded by `user_version`) so a user who deliberately re-adds one
+// in Settings keeps it. Pruning is safe — these rows only hold rates; historical
+// token_usage is untouched.
+const RETIRED_PRICING_PATTERNS = [
+  "claude-opus-4-1%",
+  "claude-opus-4-2%",
+  "claude-sonnet-4-2%",
+  "claude-3-7-sonnet%",
+  "claude-3-5-sonnet%",
+  "claude-3-5-haiku%",
+  "claude-3-haiku%",
+  "claude-3-opus%",
 ];

 // Top-up: insert any default pattern that isn't already present. Preserves
@@ -858,6 +868,24 @@ const DEFAULT_PRICING = [
 // This runs every startup so new default models (e.g. Opus 4.8) appear in the
 // Settings UI automatically without requiring a manual "Reset Defaults".
 {
+  if (db.pragma("user_version", { simple: true }) < 1) {
+    const drop = db.prepare("DELETE FROM model_pricing WHERE model_pattern = ?");
+    db.transaction(() => {
+      for (const p of RETIRED_PRICING_PATTERNS) drop.run(p);
+      // Sonnet 5's launch discount expired 2026-08-31 and became the published
+      // standard rate. The top-up below only inserts missing rows, so correct
+      // the existing row here or every pre-existing install keeps billing 3/15.
+      db.prepare(
+        `UPDATE model_pricing SET
+           input_per_mtok = 2, output_per_mtok = 10, cache_read_per_mtok = 0.2,
+           cache_write_per_mtok = 2.5, cache_write_1h_per_mtok = 4,
+           intro_until = NULL
+         WHERE model_pattern = 'claude-sonnet-5%'`
+      ).run();
+    })();
+    db.pragma("user_version = 1");
+  }
+
  const existing = new Set(
    db
      .prepare("SELECT model_pattern FROM model_pricing")
@@ -877,8 +905,9 @@ const DEFAULT_PRICING = [

 // Known introductory promo rates: [pattern, in, out, cacheRead, cw5m, cw1h, until].
 // Standard rates live in the DEFAULT_PRICING row; these add the time-limited
-// discount on top. Claude Sonnet 5: 2/3-off launch pricing through 2026-08-31.
-const DEFAULT_INTRO_PRICING = [["claude-sonnet-5%", 2, 10, 0.2, 2.5, 4, "2026-08-31"]];
+// discount on top. Empty today — Sonnet 5's launch discount (through 2026-08-31)
+// became its standard rate, so it lives in DEFAULT_PRICING now.
+const DEFAULT_INTRO_PRICING = [];

 // Backfill the known intro rates. Only fills rows whose intro_until is still
 // NULL, so a user who edits or clears an intro rate in Settings is never
@@ -1594,8 +1623,11 @@ const stmts = {
    WHERE model_pattern = ?
  `),
  deletePricing: db.prepare("DELETE FROM model_pricing WHERE model_pattern = ?"),
+  // Longest pattern wins: `claude-fable-5-1%` and `claude-fable-5%` both match
+  // `claude-fable-5-1`, and without an explicit order the winner is whichever
+  // row happened to be inserted first. Most specific pattern is the right one.
  matchPricing: db.prepare(
-    "SELECT * FROM model_pricing WHERE ? LIKE REPLACE(model_pattern, '%', '%') LIMIT 1"
+    "SELECT * FROM model_pricing WHERE ? LIKE model_pattern ORDER BY LENGTH(model_pattern) DESC LIMIT 1"
  ),
  toolUsageCounts: db.prepare(`
    SELECT tool_name, COUNT(*) as count
@@ -21,6 +21,11 @@
 * native" is the same single write path as "switch to a router". It cannot be
 * edited, removed, or shadowed by a saved profile.
 *
+ * `testRouter` is the one thing here that talks to the network: the smallest
+ * real request a claude lane would make, so "saved" and "actually works" stop
+ * being the same claim. It is deliberately NOT a generic proxy — see the note
+ * on its stored-key fallback.
+ *
 * @author Nguyễn Ngọc Trí Vĩ <vinnt@smartgift.vn>
 */

@@ -307,12 +312,93 @@ function removeRouter(name) {
  return { removed: name };
 }

+/** How long a connection test waits before calling the endpoint dead. */
+const TEST_TIMEOUT_MS = 15_000;
+
+/** Enough of a failing router's own words to be actionable, not enough to flood. */
+const TEST_BODY_MAX = 300;
+
+/**
+ * Probe a router endpoint with the smallest real request a claude lane makes:
+ * one Anthropic-style message capped at a single output token.
+ *
+ * Deliberately the CLAUDE wire shape and not Codex's `responses` one. A profile
+ * is not tool-specific, so a test has to pick; `/v1/messages` is what a claude
+ * lane's eight env keys point at, and it is the path a router has to serve for
+ * this feature to work at all.
+ *
+ * **The stored key is only ever sent to the endpoint it was stored against.**
+ * A blank `apiKey` falls back to the saved profile named by `name` — but only
+ * when the submitted `baseUrl` still matches the saved one. Without that check
+ * this route would be a way to make the server POST a saved API key to any URL
+ * a browser names, which is exactly the leak `listRouters` exists to prevent.
+ *
+ * A router that answers is a success even when it answers with an error: the
+ * outcome is reported (`ok`, `status`, `message`), never thrown. Only bad INPUT
+ * throws.
+ *
+ * @param {{name?:string, baseUrl:string, apiKey?:string, model:string}} input
+ * @returns {Promise<{ok:boolean, status:number|null, message:string}>}
+ */
+async function testRouter(input = {}) {
+  const baseUrl = normalizeBaseUrl(input.baseUrl);
+  const model = normalizeModel(input.model);
+
+  let apiKey = typeof input.apiKey === "string" ? input.apiKey.trim() : "";
+  if (!apiKey && input.name) {
+    const stored = readRouter(input.name);
+    if (stored && normalizeBaseUrl(stored.baseUrl || "") === baseUrl) apiKey = stored.apiKey || "";
+  }
+  if (!apiKey) throw err("EBADKEY", "enter the API key to test this endpoint");
+
+  let res;
+  try {
+    res = await fetch(`${baseUrl}/v1/messages`, {
+      method: "POST",
+      headers: {
+        "content-type": "application/json",
+        // Both, because routers disagree on which one they read, and it is the
+        // same key going to the same endpoint either way.
+        "x-api-key": apiKey,
+        authorization: `Bearer ${apiKey}`,
+        "anthropic-version": "2023-06-01",
+      },
+      body: JSON.stringify({
+        model,
+        max_tokens: 1,
+        messages: [{ role: "user", content: "ping" }],
+      }),
+      signal: AbortSignal.timeout(TEST_TIMEOUT_MS),
+    });
+  } catch (e) {
+    // Unreachable, DNS failure, TLS refusal, timeout — a real answer about the
+    // endpoint, so it is a result and not a 500.
+    //
+    // `cause` first: node's fetch reports every transport failure as the same
+    // useless "fetch failed", and hides the part the user can act on
+    // (ECONNREFUSED, ENOTFOUND, a TLS error) one level down.
+    return {
+      ok: false,
+      status: null,
+      message: e.name === "TimeoutError" ? "timed out" : (e.cause && e.cause.message) || e.message,
+    };
+  }
+
+  const body = await res.text().catch(() => "");
+  return {
+    ok: res.ok,
+    status: res.status,
+    message: res.ok ? model : body.slice(0, TEST_BODY_MAX) || res.statusText,
+  };
+}
+
 module.exports = {
  listRouters,
  getRouter,
  createRouter,
  updateRouter,
  removeRouter,
+  testRouter,
  MODEL_SLOTS,
  BUILTIN_NAME,
  NATIVE,
@@ -27,6 +27,7 @@ const { sameOriginGuard } = require("../lib/same-origin");
 const router = express.Router();

 const ERR_TO_STATUS = {
+  EBADKEY: 400,
  EBADNAME: 400,
  EBADLABEL: 400,
  EBADURL: 400,
@@ -59,6 +60,22 @@ router.post("/", sameOriginGuard, (req, res) => {
  }
 });

+// POST /api/ai-routers/test — does this endpoint + key + model actually answer?
+// Body: `{ baseUrl, model, apiKey?, name? }`. Registered before `/:name` for
+// clarity; no collision today, since `/:name` takes no POST.
+//
+// A reachable router is a 200 here whatever it replies — `{ ok:false, status }`
+// carries the router's own words, which is the whole point. Only bad input is
+// a 4xx. The response never echoes the key, including on the path where the
+// stored one was used (see `testRouter` on why that path is narrow).
+router.post("/test", sameOriginGuard, async (req, res) => {
+  try {
+    res.json(await routers.testRouter(req.body || {}));
+  } catch (e) {
+    return fail(res, e);
+  }
+});
+
 // PATCH /api/ai-routers/:name — same fields, all optional. An omitted `apiKey`
 // keeps the stored one; the edit form cannot show a key, so it cannot echo one.
 //
Author	SHA1	Message	Date
nntrivi2001	5ab674f134	chore(release): bump version to 1.6.4	2026-09-16 11:28:18 +07:00
nntrivi2001	00a2aedb09	feat(ai-router): test an endpoint before saving it, and re-point a live lane Two gaps left by the AI Router feature: a profile could be saved without ever being reachable, and a lane's router could only be chosen at creation. POST /api/ai-routers/test answers the question saving cannot. It sends one Anthropic-style message capped at a single output token to `<baseUrl>/v1/messages` with a 15s timeout — the wire shape a claude lane's eight env keys point at, so what is tested is what will actually run. A router that answers is a 200 here whatever it replies: `{ ok, status, message }` carries the router's own words, and `status` is null when nothing answered at all. Only bad input is a 4xx. The stored key is only ever sent to the endpoint it was stored against. A blank `apiKey` falls back to the key of the profile named by `name`, but only while the submitted `baseUrl` still equals the saved one — so testing a retyped endpoint needs a retyped key (`EBADKEY`). Without that rule this route would be a way to make the server POST a saved API key to any URL a browser names. Network failures report `error.cause`, not `error.message`: node's fetch reports every transport failure as the same useless "fetch failed" and hides the part the user can act on (ECONNREFUSED, ENOTFOUND, a TLS error) one level down. The lane card carries an AI Router select. The Add-lane modal was the only place a router was chosen, and a lane outlives its creation. Switching goes through `POST /api/lanes/:id/router/apply` — never a lane PATCH — because the DB column and the lane's own config on disk may only move together. A failed apply puts the select back rather than leaving the card claiming a router the lane did not get. Hidden when `native` is the only profile: a select with one option is a control that cannot do anything. Adds `Lane.router_profile` to the client type (the payload already carried it).	2026-09-16 11:26:00 +07:00
nntrivi2001	9128352d1b	fix(pricing): price Claude Opus 5, correct Sonnet 5, drop retired models Claude Opus 5 had no pricing row at all, so every token it burned was costed at $0 — the pattern list stopped at the 4.x Opus family and nothing matched `claude-opus-5`. Add it (5/25, fast mode 10/50, same shape as 4.8). Sonnet 5's launch discount expired 2026-08-31 and the discounted rate became the published standard, so the row's 3/15 standard rate started overcharging the day after. Fold 2/10 into DEFAULT_PRICING and empty DEFAULT_INTRO_PRICING; no model ships a promo today. Fable/Mythos 5.1 fell through to the `claude-fable-5%` row, which caches at $1/MTok against 5.1's $0.25 — give them their own rows. That overlap is only resolvable if the lookup prefers the more specific pattern, and `matchPricing` took `LIMIT 1` with no ORDER BY, i.e. whichever row was inserted first. Order by pattern length descending, which is what docs/DATABASE.md already claimed. (The nested REPLACE(model_pattern, '%', '%') it dropped was a no-op.) Retire Claude 3.x, Opus 4/4.1 and Sonnet 4 from the defaults. The startup top-up only ever ADDS missing rows, so removing them from DEFAULT_PRICING would leave every existing install untouched — prune them once, guarded by PRAGMA user_version (0 -> 1) so a model deliberately re-added in Settings survives later restarts. The same one-shot migration corrects the existing Sonnet 5 row. Only rates are deleted; token_usage is never touched, and Reset Defaults still restores the full current set.	2026-09-15 15:17:55 +07:00
nntrivi2001	fb3796f064	chore(release): bump version to 1.6.3	2026-09-15 14:55:23 +07:00
nntrivi2001	c3f8dec104	refactor(lanes): drop the tick box and radio dots from the Add-lane modal Both controls already render as cards that light up when selected — the same selected/unselected treatment LaneStripCard uses — so the tick box and the two radio dots were a second indicator of a state the card was already showing. Neither input is removed, only hidden (`sr-only`): they are still what a keyboard toggles and what a screen reader announces, so the cards take a `focus-within` ring in exchange for the focus ring that went with them. The router radios needed one thing the checkbox did not. "AI Router" is disabled until the profile list lands, and the greyed-out dot was the only thing showing that; hidden, it would have read as a card you can click that does nothing — the dead control the existing comment warns about. So disabled is now its own branch: muted text, reduced opacity, no hover, not-allowed cursor. Layout follows: no flex row with a gap when there is one child left, and the two equal-width router cards centre their label rather than leaving it hard against an empty left edge. Visual only. No behaviour, API or schema change.	2026-09-15 14:53:00 +07:00
nntrivi2001	6e685ce1c3	chore(release): bump version to 1.6.2	2026-09-15 11:01:07 +07:00
nntrivi2001	29a5da3e4e	fix(bootstrap): stop a retired plugin build from downgrading the dashboard `claude plugin update` unpacks the new version into a new cache directory and says "Restart to apply changes". A Claude Code process that was already running does not: it keeps loading the OLD directory, hooks and all. Every session such a process opened then ran the OLD plugin-bootstrap, which saw a state.json naming a different root, concluded the server had to be replaced, repointed agenthub.service at itself and restarted the dashboard onto the version the user had just upgraded away from. The dashboard reported the old version and offered the update again — once per session, forever, while installed_plugins.json and the cache both said the new version was installed. A /agenthub-update from that session "succeeded" and changed nothing. - `supersededBy()` reads Claude Code's installed_plugins.json. When a SIBLING cache directory is the recorded install, the bootstrap logs one line and returns "superseded" — no deps, no unit, no server. `--force` does not override it: the self-update chain forces from the NEW directory, which is never superseded, so the only --force refused is a deliberate downgrade. It cannot hand over either — the process it would replace is the one loading it — so the log line says to restart Claude Code. - Matched by sibling directory, not by the <plugin>@<marketplace> key: correct without parsing a cache path or trusting a name. A phantom entry (installPath that no longer exists) is ignored rather than wedging the bootstrap shut. Also adds `pruneRetiredPluginRoots()`, since the same updates leave a full copy of the repo — node_modules and all — per release; twelve of them measured 1.7 GB on one machine. It runs after writeState, so a bad prune cannot cost the record of a good bootstrap, and it protects four independent readings of "in use": this build, every installPath Claude Code records, the directory agenthub.service names (`unitRoot`, foreign unit included), and any directory still named in the user's settings.json hooks — that last one is the un-restarted process above, seen from the other side. Plus the 2 newest survivors as a rollback cushion, and nothing at all outside ~/.claude/plugins/cache/.	2026-09-15 10:30:00 +07:00