diff --git a/server/services/ollamaManager.js b/server/services/ollamaManager.js
index 0a038b61f9..b88fe9c7f8 100644
--- a/server/services/ollamaManager.js
+++ b/server/services/ollamaManager.js
@@ -34,7 +34,7 @@ import { buildHfAuthHeaders, buildHfResolveUrl, HF_API } from '../lib/huggingfac
 import { isEmbeddingModel } from '../lib/localModelHeuristics.js'
 import { commandExists } from '../lib/commandExists.js'
 import {
-  OLLAMA_AGENT_MIN_CONTEXT, resolveOllamaContextLength, withOllamaContextEnv
+  OLLAMA_AGENT_MIN_CONTEXT, OLLAMA_CONTEXT_ENV_VAR, resolveOllamaContextLength, withOllamaContextEnv
 } from '../lib/ollamaContext.js'
 import { compareSemver } from '../lib/versionUtils.js'
 import { isSafeHfRepoRelativePath } from '../lib/hfCache.js'
@@ -640,7 +640,19 @@ async function getRuntimeContextLength(selectedModel = null) {
 async function ensureContextWindow(contextLength, selectedModel = null) {
   const target = Number(contextLength) > 0 ? Math.floor(Number(contextLength)) : null
   if (!target) return { applied: false, reason: 'not-configured', contextLength: null }
-  const env = withOllamaContextEnv({}, target)
+
+  // Compose the context-window env ON TOP OF the currently applied launch env
+  // instead of replacing it, so a reload preserves knobs (e.g.
+  // OLLAMA_FLASH_ATTENTION, OLLAMA_KV_CACHE_TYPE) that are not about the window.
+  // `appliedLaunchEnvValues` outlives `appliedLaunchEnv` specifically to answer
+  // "what has PortOS put in front of Ollama that has not been cleared yet", so
+  // an active tuning is preserved even if the daemon was temporarily down.
+  // When a tuning is active mid-sweep, preserving its knobs keeps the sweep's
+  // measurements comparable rather than demoting the daemon to untuned
+  // mid-sweep, while allowing the context window to expand for the harness.
+  const baseEnv = appliedLaunchEnvValues ? { ...appliedLaunchEnvValues } : {}
+  const env = withOllamaContextEnv(baseEnv, target)
+
   if (!(await checkOllamaAvailable(true))) {
     return { ...(await restartWithEnv(env, { tuning: false })), contextLength: target }
   }
@@ -817,10 +829,23 @@ async function restartWithEnv(env, { tuning = true } = {}) {
   // by construction), so anything written up front would be wiped mid-call.
   //
   // A restart that did not happen changed nothing, so the bookkeeping must not
-  // move either. A non-tuning restart that DID happen is the install's real
-  // configuration changing, which leaves nothing to undo.
-  if (result.applied === false) preTuningEnv = before
-  else preTuningEnv = tuning ? captured : null
+  // move either.
+  //
+  // When a non-tuning restart (`tuning: false`, e.g. `ensureContextWindow`) occurs
+  // while a tuning is active (`before !== null`), preserve the pre-tuning undo
+  // baseline rather than clearing it, but update its context window so that a
+  // later `clearLaunchEnv()` restores the untuned daemon with the new context
+  // length rather than reverting it. If no tuning was active (`before === null`),
+  // `preTuningEnv` remains `null`.
+  if (result.applied === false) {
+    preTuningEnv = before
+  } else if (tuning) {
+    preTuningEnv = captured
+  } else if (before) {
+    preTuningEnv = withOllamaContextEnv(before, env[OLLAMA_CONTEXT_ENV_VAR])
+  } else {
+    preTuningEnv = null
+  }
   return result
 }
 
diff --git a/server/services/ollamaManager.test.js b/server/services/ollamaManager.test.js
index d3d99163c8..7e43fa0b02 100644
--- a/server/services/ollamaManager.test.js
+++ b/server/services/ollamaManager.test.js
@@ -1164,6 +1164,102 @@ describe('ollamaManager.restartWithEnv', () => {
     expect(options.env.OLLAMA_CONTEXT_LENGTH).toBe('131072')
   })
 
+  // A context-window reload (e.g. Claude Code spawn needing numCtx 131072) must
+  // compose on top of an active tuning's launch knobs (OLLAMA_FLASH_ATTENTION,
+  // OLLAMA_KV_CACHE_TYPE) instead of wiping them, and the tuning's undo
+  // bookkeeping must survive so clearLaunchEnv() can restore the untuned state.
+  it('preserves an active tuning across ensureContextWindow on a spawned process', async () => {
+    const state = { up: false, probesBeforeStop: 0 }
+    stubReachable({ get reachable() { return state.up || state.probesBeforeStop-- > 0 } })
+    const spawn = await stubSpawnRestart(state)
+    const { ensureContextWindow, restartWithEnv } = await loadManager()
+
+    await restartWithEnv({ OLLAMA_FLASH_ATTENTION: '1', OLLAMA_KV_CACHE_TYPE: 'q8_0' })
+    state.up = false
+    state.probesBeforeStop = 3
+
+    const reload = await ensureContextWindow(131072)
+    expect(reload).toMatchObject({ applied: true, reason: 'restarted', contextLength: 131072 })
+
+    const lastSpawn = spawn.mock.calls[spawn.mock.calls.length - 1]
+    expect(lastSpawn[2].env.OLLAMA_FLASH_ATTENTION).toBe('1')
+    expect(lastSpawn[2].env.OLLAMA_KV_CACHE_TYPE).toBe('q8_0')
+    expect(lastSpawn[2].env.OLLAMA_CONTEXT_LENGTH).toBe('131072')
+
+    // Clearing the tuning removes tuning keys and preserves the context window
+    state.up = false
+    state.probesBeforeStop = 2
+    expect(await restartWithEnv({})).toMatchObject({ applied: null })
+    const clearedSpawn = spawn.mock.calls[spawn.mock.calls.length - 1]
+    expect(clearedSpawn[2].env.OLLAMA_FLASH_ATTENTION).toBeUndefined()
+    expect(clearedSpawn[2].env.OLLAMA_KV_CACHE_TYPE).toBeUndefined()
+    expect(clearedSpawn[2].env.OLLAMA_CONTEXT_LENGTH).toBe('131072')
+  })
+
+  it('preserves an active tuning and its undo bookkeeping across a service context-window reload', async () => {
+    restorePlatform = pinPlatform('darwin')
+    stubReachable({ reachable: true })
+    const calls = []
+    execMock.impl = (cmd, args, opts, cb) => {
+      const a = (args || []).join(' ')
+      calls.push(`${cmd} ${a}`)
+      if (cmd === 'brew' && a === '--version') return cb(null, { stdout: 'Homebrew 4.0.0', stderr: '' })
+      if (cmd === 'brew' && a === 'services list') return cb(null, { stdout: 'ollama started testuser plist\n', stderr: '' })
+      if (cmd === 'ps') return cb(null, { stdout: '101 /usr/local/bin/ollama serve\n', stderr: '' })
+      if (cmd === 'launchctl') return cb(null, { stdout: '', stderr: '' })
+      if (cmd === 'brew' && a === 'services restart ollama') return cb(null, { stdout: '', stderr: '' })
+      return cb(new Error(`unexpected exec: ${cmd} ${a}`))
+    }
+
+    const { ensureContextWindow, restartWithEnv } = await loadManager()
+    // 1. Apply tuning: flash attention + kv cache quantization
+    await restartWithEnv({ OLLAMA_FLASH_ATTENTION: '1', OLLAMA_KV_CACHE_TYPE: 'q8_0' })
+
+    // 2. Context window reload arrives for an agent spawn
+    calls.length = 0
+    const reloaded = await ensureContextWindow(131072)
+    expect(reloaded).toMatchObject({ applied: true, reason: 'service-restarted' })
+
+    // Both tuning knobs and the new context window are passed to the daemon
+    expect(calls).toContain('launchctl setenv OLLAMA_FLASH_ATTENTION 1')
+    expect(calls).toContain('launchctl setenv OLLAMA_KV_CACHE_TYPE q8_0')
+    expect(calls).toContain('launchctl setenv OLLAMA_CONTEXT_LENGTH 131072')
+    expect(calls).not.toContain('launchctl unsetenv OLLAMA_FLASH_ATTENTION')
+    expect(calls).not.toContain('launchctl unsetenv OLLAMA_KV_CACHE_TYPE')
+
+    // 3. Undo bookkeeping survived: clearing the tuning restores the baseline
+    // context window while unsetting the tuning keys
+    calls.length = 0
+    const cleared = await restartWithEnv({})
+    expect(cleared).toMatchObject({ applied: null })
+    expect(calls).toContain('launchctl unsetenv OLLAMA_FLASH_ATTENTION')
+    expect(calls).toContain('launchctl unsetenv OLLAMA_KV_CACHE_TYPE')
+    expect(calls).toContain('launchctl setenv OLLAMA_CONTEXT_LENGTH 131072')
+  })
+
+  it('preserves an active tuning when ensureContextWindow starts an offline daemon', async () => {
+    const state = { up: false, probesBeforeStop: 0 }
+    stubReachable({ get reachable() { return state.up || state.probesBeforeStop-- > 0 } })
+    const spawn = await stubSpawnRestart(state)
+    const { ensureContextWindow, restartWithEnv } = await loadManager()
+
+    await restartWithEnv({ OLLAMA_FLASH_ATTENTION: '1', OLLAMA_KV_CACHE_TYPE: 'q8_0' })
+
+    // Daemon goes down
+    state.up = false
+    state.probesBeforeStop = 0
+
+    // Calling ensureContextWindow when the daemon is unreachable should still
+    // preserve the tuning knobs and start the daemon with them
+    const result = await ensureContextWindow(131072)
+    expect(result).toMatchObject({ applied: true, contextLength: 131072 })
+
+    const lastSpawn = spawn.mock.calls[spawn.mock.calls.length - 1]
+    expect(lastSpawn[2].env.OLLAMA_FLASH_ATTENTION).toBe('1')
+    expect(lastSpawn[2].env.OLLAMA_KV_CACHE_TYPE).toBe('q8_0')
+    expect(lastSpawn[2].env.OLLAMA_CONTEXT_LENGTH).toBe('131072')
+  })
+
   // The domain outlives the daemon, so a stopped Ollama is no reason to leave
   // the variables in it — the next login-launched one would inherit them.
   it('unsets exported variables even when Ollama is already down', async () => {