diff --git a/.ai/contexts/session-cache.md b/.ai/contexts/session-cache.md
index f698cd7b..e2c74e72 100644
--- a/.ai/contexts/session-cache.md
+++ b/.ai/contexts/session-cache.md
@@ -142,6 +142,33 @@ or deleted from here.
   silently forgotten. `remote-index.js` catches per host, so one dead host does
   not stop its peers or the local scan.
 
+- **A host that keeps failing backs off per host, exponentially, capped at
+  30 min — it is never disabled (issue #215).** Field incident 2026-09-07/08:
+  a `transport disposed` cause (fixed separately) reran the plain fixed-cadence
+  loop every 300.0 s for ~19 h (226 identical `refresh failed` warnings) because
+  nothing slowed a permanently broken host down. `refreshNow()` now tracks
+  `{ failures, lastError, nextAttemptAt }` per alias; on failure the delay is
+  `min(intervalMs * 2^(failures-1), 30 min)` — base equals the host's own
+  configured cadence, so an isolated blip costs nothing extra, and the 30 min
+  ceiling was chosen so an operator never has to restart the app to get a
+  recovered host picked back up. A host past `nextAttemptAt` is skipped for
+  that cycle only: no ssh call, no log line, and its peers still run on
+  schedule — the loop `continue`s per host, it never returns early. One
+  success resets `failures`/`nextAttemptAt` to nominal immediately (issue
+  requirement: fast recovery, not a cool-down after the outage ends).
+  **Decision: slow down, never disable.** Disabling would need to flip the
+  same `enabled` flag the Settings UI owns, which is out of this issue's scope
+  and would turn a transient network problem into a silent, permanent loss of
+  mirroring that nothing in the sidebar currently surfaces — the capped
+  exponential delay already bounds the cost of a dead host to one attempt per
+  30 min, which is cheap enough to just keep trying. Logging is throttled to
+  the first failure and each change of tier (`onHostFailure` in
+  `remote-index.js`), not every attempt, so the same field incident would have
+  produced roughly 5 lines instead of 226. Per-host state is readable via
+  `getRemoteHostState(alias)` (mirrors `getRemoteSessions(alias)`); no GUI
+  reads it yet. Proven in `test/remote-index.test.js` with an injected
+  `now()` clock — no real timers, no `setTimeout` waits.
+
 - **The mirror is indexed off the main thread.** `workers/scan-projects.js` takes
   `folderPrefix` and a `folders` subset in `workerData`, and
   `sessionCache.scanFoldersViaWorker` writes each folder result through the same
diff --git a/remote-index.js b/remote-index.js
index d8371a6f..b93816fa 100644
--- a/remote-index.js
+++ b/remote-index.js
@@ -14,6 +14,9 @@ const { syncMirror } = require('./remote-mirror');
 
 const NOOP_LOG = { info() {}, warn() {}, error() {} };
 
+// see .ai/contexts/session-cache.md ("Remote hosts backoff")
+const MAX_BACKOFF_MS = 30 * 60 * 1000;
+
 /**
  * Periodic mirror + index of every declared SSH host.
  * see .ai/contexts/session-cache.md ("Remote SSH hosts")
@@ -30,16 +33,63 @@ const NOOP_LOG = { info() {}, warn() {}, error() {} };
  *   notify()                         -> push a sidebar refresh
  *   timers                           -> { setInterval, clearInterval } (test seam)
  *   sync                             -> syncMirror override (test seam)
+ *   now()                            -> current epoch ms (test seam, defaults to Date.now)
  */
 function createRemoteIndexer(ctx) {
   const log = ctx.log || NOOP_LOG;
   const timers = ctx.timers || { setInterval, clearInterval };
   const sync = ctx.sync || syncMirror;
+  const now = ctx.now || Date.now;
 
   let timer = null;
   let inFlight = false;
   let stopped = false;
   const remoteSessions = new Map(); // alias -> sessions array, from the same ssh cycle as the inventory
+  const hostBackoff = new Map(); // alias -> { failures, lastError, nextAttemptAt }
+
+  function backoffState(alias) {
+    let s = hostBackoff.get(alias);
+    if (!s) {
+      s = { failures: 0, lastError: null, nextAttemptAt: 0 };
+      hostBackoff.set(alias, s);
+    }
+    return s;
+  }
+
+  // see .ai/contexts/session-cache.md ("Remote hosts backoff")
+  function backoffDelayMs(failures, intervalMs) {
+    if (failures <= 0) return 0;
+    return Math.min(intervalMs * Math.pow(2, failures - 1), MAX_BACKOFF_MS);
+  }
+
+  function onHostSuccess(alias) {
+    const state = backoffState(alias);
+    if (state.failures > 0) {
+      log.info(`[remote:${alias}] refresh recovered after ${state.failures} consecutive failure(s)`);
+    }
+    state.failures = 0;
+    state.lastError = null;
+    state.nextAttemptAt = 0;
+  }
+
+  function onHostFailure(alias, err, intervalMs) {
+    const state = backoffState(alias);
+    const prevDelay = backoffDelayMs(state.failures, intervalMs);
+    state.failures += 1;
+    state.lastError = err.message;
+    const delay = backoffDelayMs(state.failures, intervalMs);
+    state.nextAttemptAt = now() + delay;
+    if (delay !== prevDelay) {
+      log.warn(`[remote:${alias}] refresh failed (${state.failures}x consecutive): ${err.message}; ` +
+        `retrying in ${Math.round(delay / 1000)}s`);
+    }
+  }
+
+  function getRemoteHostState(alias) {
+    const s = hostBackoff.get(alias);
+    if (!s) return { consecutiveFailures: 0, lastError: null, nextAttemptAt: 0 };
+    return { consecutiveFailures: s.failures, lastError: s.lastError, nextAttemptAt: s.nextAttemptAt };
+  }
 
   function hosts() {
     return enabledHosts(ctx.getHosts ? ctx.getHosts() : []);
@@ -66,6 +116,9 @@ function createRemoteIndexer(ctx) {
     for (const alias of [...remoteSessions.keys()]) {
       if (!known.has(alias)) remoteSessions.delete(alias);
     }
+    for (const alias of [...hostBackoff.keys()]) {
+      if (!known.has(alias)) hostBackoff.delete(alias);
+    }
     return dropped;
   }
 
@@ -140,15 +193,19 @@ function createRemoteIndexer(ctx) {
     inFlight = true;
     let changed = pruneUnknownAliases(list) > 0;
     const errors = [];
+    const intervalMs = normalizeRefreshMs(ctx.getRefreshMs ? ctx.getRefreshMs() : undefined);
     try {
       for (const host of list) {
         if (stopped) break;
+        const state = backoffState(host.alias);
+        if (now() < state.nextAttemptAt) continue; // still backing off: no attempt, no log, no ssh
         try {
           if (await refreshHost(host)) changed = true;
+          onHostSuccess(host.alias);
         } catch (err) {
           remoteSessions.set(host.alias, []);
           errors.push({ alias: host.alias, error: err.message });
-          log.warn(`[remote:${host.alias}] refresh failed: ${err.message}`);
+          onHostFailure(host.alias, err, intervalMs);
         }
       }
     } finally {
@@ -197,7 +254,12 @@ function createRemoteIndexer(ctx) {
     return remoteSessions.get(alias) || [];
   }
 
-  return { start, stop, dispose, restart, refreshNow, isRunning: () => timer !== null, getRemoteSessions };
+  return {
+    start, stop, dispose, restart, refreshNow,
+    isRunning: () => timer !== null,
+    getRemoteSessions,
+    getRemoteHostState,
+  };
 }
 
 module.exports = { createRemoteIndexer };
diff --git a/test/remote-index.test.js b/test/remote-index.test.js
index 135c2b76..2676f4e2 100644
--- a/test/remote-index.test.js
+++ b/test/remote-index.test.js
@@ -32,6 +32,14 @@ function fakeTimers() {
   };
 }
 
+/** A controllable clock for the backoff seam (`ctx.now`) — no wall-clock wait. */
+function fakeClock(start = 0) {
+  let t = start;
+  const fn = () => t;
+  fn.advance = (ms) => { t += ms; };
+  return fn;
+}
+
 test('no host declared: no timer, no transport call, no mirror directory', async () => {
   const dataDir = tmp('idx-none');
   try {
@@ -332,3 +340,161 @@ test('dispose() is terminal: it stops the timer and ends the transport', async (
     assert.equal(indexer.isRunning(), false, 'shutdown must clear the timer');
   } finally { fs.rmSync(dataDir, { recursive: true, force: true }); }
 });
+
+// Issue #215 acceptance: a host stuck in `ssh inventory failed ... transport
+// disposed` must not be retried at a fixed cadence forever (the field log
+// shows 226 identical attempts, one every 300.0 s, over ~19 h). Backoff is
+// driven by an injected clock (`ctx.now`), never a real timer.
+
+test('consecutive failures on one host back off exponentially and the delay is capped', async () => {
+  const dataDir = tmp('idx-backoff-space');
+  try {
+    const clock = fakeClock(0);
+    let attempts = 0;
+    const BASE = 60_000;
+    const CAP = 30 * 60 * 1000;
+    const indexer = createRemoteIndexer({
+      getHosts: () => [{ alias: 'dead' }],
+      getRefreshMs: () => BASE,
+      dataDir,
+      transport: {},
+      scanFolders: () => Promise.resolve({ ok: true }),
+      listIndexedFolderKeys: () => [],
+      timers: fakeTimers(),
+      now: clock,
+      sync: async () => { attempts++; throw new Error('ssh inventory failed (exit -1): transport disposed'); },
+    });
+
+    const expectedDelay = (failures) => Math.min(BASE * 2 ** (failures - 1), CAP);
+
+    for (let failures = 1; failures <= 8; failures++) {
+      await indexer.refreshNow();
+      const state = indexer.getRemoteHostState('dead');
+      assert.equal(state.consecutiveFailures, failures, `failure ${failures} counted`);
+      assert.equal(state.nextAttemptAt - clock(), expectedDelay(failures), `delay after failure ${failures}`);
+
+      // A tick before the backoff elapses must not spend another ssh attempt.
+      const before = attempts;
+      clock.advance(1);
+      await indexer.refreshNow();
+      assert.equal(attempts, before, `failure ${failures}: not due yet, must be skipped`);
+
+      // Land exactly on the next due instant for the following iteration.
+      clock.advance(state.nextAttemptAt - clock());
+    }
+
+    assert.equal(attempts, 8, 'every due cycle actually attempted the host once');
+    assert.equal(expectedDelay(8), CAP, 'sanity: by the 8th failure the exponential has saturated at the cap, ' +
+      'so the per-iteration assertion above already proved the ceiling holds');
+  } finally { fs.rmSync(dataDir, { recursive: true, force: true }); }
+});
+
+test('a success after consecutive failures resets the backoff to nominal immediately', async () => {
+  const dataDir = tmp('idx-backoff-reset');
+  try {
+    const clock = fakeClock(0);
+    let shouldFail = true;
+    const indexer = createRemoteIndexer({
+      getHosts: () => [{ alias: 'flaky' }],
+      getRefreshMs: () => 60_000,
+      dataDir,
+      transport: {},
+      scanFolders: () => Promise.resolve({ ok: true }),
+      listIndexedFolderKeys: () => [],
+      timers: fakeTimers(),
+      now: clock,
+      sync: async () => {
+        if (shouldFail) throw new Error('ssh: connect to host flaky port 22: timed out');
+        return { fetched: 0, unchanged: 0, removed: 0, failed: 0, total: 0, changedFolders: new Set() };
+      },
+    });
+
+    // Three consecutive failures widen the gap well past the nominal cadence.
+    let failedState;
+    let delayBeforeRecovery;
+    for (let i = 0; i < 3; i++) {
+      await indexer.refreshNow();
+      failedState = indexer.getRemoteHostState('flaky');
+      delayBeforeRecovery = failedState.nextAttemptAt - clock();
+      clock.advance(delayBeforeRecovery);
+    }
+    assert.equal(failedState.consecutiveFailures, 3);
+    assert.equal(delayBeforeRecovery, 240_000, 'the 3rd failure widened the delay past the 60 s nominal cadence');
+
+    // The host recovers on the next due attempt.
+    shouldFail = false;
+    await indexer.refreshNow();
+    const recovered = indexer.getRemoteHostState('flaky');
+    assert.equal(recovered.consecutiveFailures, 0, 'failure count drops to zero on success');
+    assert.equal(recovered.lastError, null, 'the stale error is cleared');
+    assert.equal(recovered.nextAttemptAt, 0, 'no artificial delay is left behind after recovery');
+
+    // Immediately eligible again — no leftover cool-down from the outage.
+    shouldFail = true; // if backoff state had survived, this would silently be skipped
+    const r = await indexer.refreshNow();
+    assert.equal(r.errors.length, 1, 'the very next cycle actually attempted the host again');
+  } finally { fs.rmSync(dataDir, { recursive: true, force: true }); }
+});
+
+test('a host backing off does not block its peers from refreshing on schedule', async () => {
+  const dataDir = tmp('idx-backoff-isolation');
+  try {
+    const clock = fakeClock(0);
+    const aliveCalls = [];
+    const indexer = createRemoteIndexer({
+      getHosts: () => [{ alias: 'dead' }, { alias: 'alive' }],
+      getRefreshMs: () => 60_000,
+      dataDir,
+      transport: {},
+      scanFolders: () => Promise.resolve({ ok: true }),
+      listIndexedFolderKeys: () => [],
+      timers: fakeTimers(),
+      now: clock,
+      sync: async ({ alias }) => {
+        if (alias === 'dead') throw new Error('ssh: dead unreachable');
+        aliveCalls.push(clock());
+        return { fetched: 0, unchanged: 0, removed: 0, failed: 0, total: 0, changedFolders: new Set() };
+      },
+    });
+
+    await indexer.refreshNow(); // dead fails once (60 s backoff), alive succeeds
+    assert.equal(aliveCalls.length, 1);
+    assert.equal(indexer.getRemoteHostState('dead').consecutiveFailures, 1);
+
+    // Advance far less than dead's backoff: dead must be skipped, alive must not.
+    clock.advance(1_000);
+    const r = await indexer.refreshNow();
+    assert.equal(aliveCalls.length, 2, 'alive is refreshed on every cycle regardless of dead backing off');
+    assert.equal(indexer.getRemoteHostState('dead').consecutiveFailures, 1, 'dead was skipped, not re-attempted');
+    assert.equal(r.errors.length, 0, 'a skipped host is not reported as a fresh error');
+  } finally { fs.rmSync(dataDir, { recursive: true, force: true }); }
+});
+
+test('failure logging is throttled: only the first failure and tier changes are logged', async () => {
+  const dataDir = tmp('idx-backoff-log');
+  try {
+    const clock = fakeClock(0);
+    const warnings = [];
+    const indexer = createRemoteIndexer({
+      getHosts: () => [{ alias: 'dead' }],
+      getRefreshMs: () => 60_000,
+      dataDir,
+      transport: {},
+      scanFolders: () => Promise.resolve({ ok: true }),
+      listIndexedFolderKeys: () => [],
+      log: { info() {}, warn: (m) => warnings.push(m), error() {} },
+      timers: fakeTimers(),
+      now: clock,
+      sync: async () => { throw new Error('ssh inventory failed (exit -1): transport disposed'); },
+    });
+
+    for (let i = 0; i < 12; i++) {
+      await indexer.refreshNow();
+      clock.advance(indexer.getRemoteHostState('dead').nextAttemptAt - clock());
+    }
+
+    assert.equal(indexer.getRemoteHostState('dead').consecutiveFailures, 12, '12 attempts actually happened');
+    assert.ok(warnings.length < 12, 'not every attempt is logged');
+    assert.ok(warnings.length <= 6, `only the tier changes are logged, got ${warnings.length}`);
+  } finally { fs.rmSync(dataDir, { recursive: true, force: true }); }
+});