diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index ec479a9c1..02745ad2e 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -10,8 +10,9 @@ jobs:
   test:
     runs-on: ${{ matrix.os }}
     strategy:
+      fail-fast: false
       matrix:
-        os: [ubuntu-latest, macos-latest]
+        os: [ubuntu-latest, macos-latest, windows-latest]
         go-version: ['1.26']
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
@@ -33,96 +34,6 @@ jobs:
           files: coverage.out
         continue-on-error: true
 
-  build-windows:
-    # CGO build smoke-test on native Windows. tree-sitter needs a C/C++
-    # compiler — the GitHub windows runner ships mingw-w64 on PATH, so no
-    # extra toolchain setup is required. `go build ./...` compiles every
-    # production package; the focused agents and sidecar tests also exercise
-    # Windows-only paths without opting into the unsupported full test suite.
-    runs-on: windows-latest
-    steps:
-      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
-
-      - uses: actions/setup-go@b7ad1dad31e06c5925ef5d2fc7ad053ef454303e # v7.0.0
-        with:
-          go-version-file: go.mod
-
-      - name: Build CLI
-        run: go build -o gortex.exe ./cmd/gortex/
-
-      - name: Build all packages
-        run: go build ./...
-
-      - name: Test Windows agent integrations
-        run: go test -timeout=5m -count=1 ./internal/agents
-
-      # Symlink confinement keeps out-of-repo files out of the index, and it
-      # reads differently on Windows: a symlink reparse point maps to
-      # ModeSymlink but a directory junction maps to ModeIrregular, and
-      # evalSymlinks normalises path case where Unix does not. Verifying it
-      # only on the linux/macos matrix would leave the platform whose
-      # semantics differ most as the untested one.
-      - name: Test symlink confinement
-        run: go test -timeout=5m -count=1 ./internal/pathguard
-
-      # Windows refuses to unlink an open file, so a leaked sidecar handle
-      # pins its directory and fails the enclosing t.TempDir() cleanup. That
-      # asymmetry is invisible on the linux/macos matrix, which is why this
-      # one test runs here rather than only there.
-      - name: Test sidecar handle release
-        run: go test -timeout=5m -count=1 -run TestCloseSidecar ./internal/persistence
-
-      # Same asymmetry, different holder: the plan-lock fixture opens an
-      # on-disk Store, so a missing Close pins plan_lock.sqlite and fails
-      # its t.TempDir() cleanup here while passing everywhere else. Every
-      # assertion in these tests is platform-neutral — only the cleanup
-      # distinguishes the runners — so nothing but this step can protect
-      # the fixture's Close from being dropped again.
-      - name: Test plan-lock fixture handle release
-        run: >
-          go test -timeout=10m -count=1
-          -run 'PlanLock|PlansLocked|PlansNeverScan'
-          ./internal/graph/store_sqlite
-
-      # os.UserHomeDir reads %USERPROFILE% here and $HOME everywhere else, so
-      # a test that isolates itself with HOME alone keeps resolving the real
-      # profile on Windows — and writes to it. That is invisible to the
-      # linux/macos matrix by construction, so the isolation helper and the
-      # path resolver it protects have to be verified on this runner or not
-      # at all.
-      - name: Test user-state isolation
-        run: go test -timeout=5m -count=1 ./internal/testenv ./internal/platform
-
-      # Indexed paths are repoPrefix + '/' + the rest in native separators, so
-      # on Windows a stored path reads `repo/dir\file.cs`. Every test for that
-      # shape is a no-op on the linux/macos matrix — POSIX treats '\' as an
-      # ordinary filename byte, so filepath.Clean and filepath.ToSlash both
-      # leave the fixture untouched and the assertions pass with or without
-      # the production fix. This runner is the only one where they can fail.
-      - name: Test store path separator handling
-        run: go test -timeout=5m -count=1 ./internal/graphpath
-
-      # The same asymmetry reaches the other direction too: a test that spells
-      # an expected path with '/' passes on POSIX whatever the code does, and
-      # only this runner can tell whether the value under test is a native
-      # filesystem path (filepath.Join / filepath.Clean) or a '/'-keyed store
-      # path. ParseDiffGitPaths, ParseDiffLinesNewSide and FeedbackDir all
-      # assert on native paths, so they belong here rather than nowhere.
-      # The path-domain contracts belong here for the same reason. A graph
-      # key is "<prefix>/" plus a native-separator remainder, so GraphKey,
-      # JoinFileNodes, the prefix-shadow end-to-end pair and RankFileRisk's
-      # per-domain normalization only diverge once '/' and the native
-      # separator differ — every one of them passes on the linux/macos
-      # matrix whatever the code does. internal/review joins the list
-      # because the file-risk normalizer lives there.
-      - name: Test native-separator store path comparisons
-        run: >
-          go test -timeout=10m -count=1
-          -run 'NativeSeparator|MixedSeparator|ImportAdjacency|ParseDiffGitPaths|ParseDiffLinesNewSide|FeedbackDir|ReviewRulepack|PrefixedGraphReportsRulepack|GraphKey|JoinFileNodes|ChangedSymbolsForFiles|PrefixShadow|IdsOnPrefixedGraph|RankFileRisk|PrefixedDeleteAndRename'
-          ./internal/analysis ./internal/graph/store_sqlite
-          ./internal/mcp ./internal/resolver ./internal/persistence
-          ./internal/review
-
   build-linux-static:
     # The linux release ships a statically linked binary so it runs on any
     # distro (see .goreleaser.yml). That only holds while the code stays