diff --git a/lua/code-preview/changes.lua b/lua/code-preview/changes.lua
index 4134daf..9cccbfd 100644
--- a/lua/code-preview/changes.lua
+++ b/lua/code-preview/changes.lua
@@ -54,16 +54,4 @@ function M.clear_by_status(status)
   end
 end
 
-function M.clear_by_statuses(statuses)
-  local set = {}
-  for _, s in ipairs(statuses) do
-    set[s] = true
-  end
-  for path, s in pairs(pending) do
-    if set[s] then
-      pending[path] = nil
-    end
-  end
-end
-
 return M
diff --git a/lua/code-preview/post_tool.lua b/lua/code-preview/post_tool.lua
index b8007c7..0060216 100644
--- a/lua/code-preview/post_tool.lua
+++ b/lua/code-preview/post_tool.lua
@@ -11,12 +11,13 @@
 
 local M = {}
 
-local normalisers = require("code-preview.pre_tool.normalisers")
-local changes     = require("code-preview.changes")
-local neo_tree    = require("code-preview.neo_tree")
-local diff        = require("code-preview.diff")
-local log         = require("code-preview.log")
-local apply_patch = require("code-preview.apply.patch")
+local normalisers  = require("code-preview.pre_tool.normalisers")
+local changes      = require("code-preview.changes")
+local neo_tree     = require("code-preview.neo_tree")
+local diff         = require("code-preview.diff")
+local log          = require("code-preview.log")
+local apply_patch  = require("code-preview.apply.patch")
+local shell_detect = require("code-preview.pre_tool.shell_detect")
 
 -- Extract paths from both unified-diff (+++ lines) and custom-patch
 -- (*** Update/Add/Delete File:) formats.
@@ -60,9 +61,17 @@ function M.handle(raw, backend)
 
   if tool_name == "Bash" then
     -- Bash pre-hook set deleted / bash_* markers without opening a preview.
-    -- Clear those statuses specifically so we don't clobber `modified` markers
-    -- from concurrent Edit/Write/ApplyPatch whose post-hook hasn't fired.
-    changes.clear_by_statuses({ "deleted", "bash_modified", "bash_created" })
+    -- Clear only THIS command's files, not every bash-owned marker: a global
+    -- status sweep wiped the still-pending markers of concurrent Bash commands
+    -- (issue #83). Detection is deterministic, so re-running it on the post
+    -- payload yields exactly the paths pre_tool marked — mirroring how the
+    -- ApplyPatch branch closes specific files via patch_paths. This also keeps
+    -- concurrent Edit/Write `modified`/`created` markers untouched, since those
+    -- paths never appear in a shell command's detection.
+    local cmd = input.tool_input and input.tool_input.command or ""
+    local detected = shell_detect.detect(cmd, input.cwd or "")
+    for _, p in ipairs(detected.rm_paths)    do changes.clear(p) end
+    for _, p in ipairs(detected.write_paths) do changes.clear(p) end
     neo_tree.refresh_deferred(200)
     return ""
   end
diff --git a/tests/backends/opencode/harness.ts b/tests/backends/opencode/harness.ts
index 7d70102..e317b96 100644
--- a/tests/backends/opencode/harness.ts
+++ b/tests/backends/opencode/harness.ts
@@ -9,7 +9,7 @@
 //   write_before  <socket> <dir> <file> <content>
 //   write_after   <socket> <dir> <file>
 //   bash_before   <socket> <dir> <command>
-//   bash_after    <socket> <dir>
+//   bash_after    <socket> <dir> <command>
 
 import { resolve, dirname } from "path"
 import { fileURLToPath } from "url"
@@ -94,8 +94,14 @@ async function main() {
     }
 
     case "bash_after": {
+      // OpenCode's real after-hook carries the tool args on `input` (see the
+      // "tool.execute.after" comment in backends/opencode/index.ts), so the
+      // command is available post-tool. Forward it here so the harness matches
+      // production — the post-tool re-detects the command to clear only that
+      // command's markers (issue #83).
+      const command = process.argv[5]
       await afterHook(
-        { tool: "bash", args: {} },
+        { tool: "bash", args: { command } },
         {},
       )
       console.log("OK")
diff --git a/tests/backends/opencode/test_edit.sh b/tests/backends/opencode/test_edit.sh
index fc819fa..0b0ff99 100644
--- a/tests/backends/opencode/test_edit.sh
+++ b/tests/backends/opencode/test_edit.sh
@@ -115,7 +115,7 @@ test_opencode_bash_rm() {
   change_status="$(nvim_eval "require('code-preview.changes').get('$test_file')")"
   assert_eq "deleted" "$change_status" "rm target should be marked as deleted" || return 1
 
-  run_opencode bash_after "$TEST_SOCKET" "$TEST_PROJECT_DIR" >/dev/null 2>&1
+  run_opencode bash_after "$TEST_SOCKET" "$TEST_PROJECT_DIR" "rm $test_file" >/dev/null 2>&1
   sleep 0.5
 
   local change_after
diff --git a/tests/plugin/post_tool_handle_spec.lua b/tests/plugin/post_tool_handle_spec.lua
index 71a3377..4a7cd79 100644
--- a/tests/plugin/post_tool_handle_spec.lua
+++ b/tests/plugin/post_tool_handle_spec.lua
@@ -1,9 +1,10 @@
 -- post_tool_handle_spec.lua — Smoke tests for the in-process post-tool.
 --
 -- post_tool.handle's contract:
---   * Bash: clears `deleted` / `bash_modified` / `bash_created` markers from
---     the changes registry; never touches `modified` / `created` markers
---     from concurrent Edit/Write/ApplyPatch proposals.
+--   * Bash: clears the markers for the files THIS command touched (re-detected
+--     from the post payload), never every bash-owned marker — so concurrent
+--     pending Bash commands keep their markers (issue #83), and `modified` /
+--     `created` markers from concurrent Edit/Write/ApplyPatch are untouched.
 --   * ApplyPatch: closes one preview per file referenced in the patch text.
 --   * Other tools: closes the single preview keyed by file_path.
 --   * Always returns "" (no backend reads post-tool stdout).
@@ -25,10 +26,10 @@ describe("post_tool.handle (Bash status clear)", function()
     assert.is_nil(changes.get("/proj/gone.txt"))
   end)
 
-  it("clears bash_modified / bash_created markers", function()
+  it("clears the bash_modified / bash_created markers for files it touched", function()
     changes.set("/proj/a.txt", "bash_modified")
     changes.set("/proj/b.txt", "bash_created")
-    post_tool.handle(payload("Bash", { command = "echo x > a.txt" }), "claudecode")
+    post_tool.handle(payload("Bash", { command = "echo x > a.txt; echo y > b.txt" }), "claudecode")
     assert.is_nil(changes.get("/proj/a.txt"))
     assert.is_nil(changes.get("/proj/b.txt"))
   end)
@@ -44,6 +45,28 @@ describe("post_tool.handle (Bash status clear)", function()
     assert.equals("created",  changes.get("/proj/new.lua"))
     assert.is_nil(changes.get("/proj/gone.txt"))
   end)
+
+  it("accepting one Bash delete keeps other pending Bash deletes marked", function()
+    -- Regression for #83: two separate Bash deletes are pending, each with its
+    -- own `deleted` marker. Accepting the first command (its PostToolUse fires)
+    -- must clear ONLY that command's file; the still-pending command's marker
+    -- survives. The old global status sweep wiped both.
+    changes.set("/proj/a.txt", "deleted")
+    changes.set("/proj/b.txt", "deleted")
+    post_tool.handle(payload("Bash", { command = "rm a.txt" }), "claudecode")
+    assert.is_nil(changes.get("/proj/a.txt"))
+    assert.equals("deleted", changes.get("/proj/b.txt"))
+  end)
+
+  it("scoped clear covers bash writes, not just deletes", function()
+    -- A pending shell write (`echo x > b.txt` → bash_created/modified) must
+    -- survive accepting a different command's write.
+    changes.set("/proj/a.txt", "bash_modified")
+    changes.set("/proj/b.txt", "bash_created")
+    post_tool.handle(payload("Bash", { command = "echo x > a.txt" }), "claudecode")
+    assert.is_nil(changes.get("/proj/a.txt"))
+    assert.equals("bash_created", changes.get("/proj/b.txt"))
+  end)
 end)
 
 describe("post_tool.handle (return value)", function()