diff --git a/Cargo.lock b/Cargo.lock
index b57fd5c9..69efd41c 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -1745,7 +1745,7 @@ dependencies = [
 [[package]]
 name = "collections"
 version = "0.1.0"
-source = "git+https://github.com/zed-industries/zed?rev=1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba#1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba"
+source = "git+https://github.com/zed-industries/zed?rev=cc053a4a6fa2fd0e8793201ed9099466af1be0b1#cc053a4a6fa2fd0e8793201ed9099466af1be0b1"
 dependencies = [
  "gpui_util",
  "indexmap 2.14.0",
@@ -2456,7 +2456,7 @@ dependencies = [
 [[package]]
 name = "derive_refineable"
 version = "0.1.0"
-source = "git+https://github.com/zed-industries/zed?rev=1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba#1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba"
+source = "git+https://github.com/zed-industries/zed?rev=cc053a4a6fa2fd0e8793201ed9099466af1be0b1#cc053a4a6fa2fd0e8793201ed9099466af1be0b1"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -2536,7 +2536,7 @@ dependencies = [
  "libc",
  "option-ext",
  "redox_users 0.5.2",
- "windows-sys 0.59.0",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -2816,7 +2816,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "39cab71617ae0d63f51a36d69f866391735b51691dbda63cf6f96d042b63efeb"
 dependencies = [
  "libc",
- "windows-sys 0.52.0",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -4444,7 +4444,7 @@ dependencies = [
 [[package]]
 name = "gpui"
 version = "0.2.2"
-source = "git+https://github.com/zed-industries/zed?rev=1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba#1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba"
+source = "git+https://github.com/zed-industries/zed?rev=cc053a4a6fa2fd0e8793201ed9099466af1be0b1#cc053a4a6fa2fd0e8793201ed9099466af1be0b1"
 dependencies = [
  "accesskit",
  "anyhow",
@@ -4513,6 +4513,7 @@ dependencies = [
  "sum_tree",
  "taffy",
  "thiserror 2.0.18",
+ "tracing",
  "ttf-parser",
  "url",
  "usvg 0.46.0",
@@ -4523,12 +4524,13 @@ dependencies = [
  "windows 0.61.3",
  "zed-font-kit",
  "zed-scap",
+ "ztracing",
 ]
 
 [[package]]
 name = "gpui-component"
 version = "0.5.2"
-source = "git+https://github.com/stippi/gpui-component?rev=1a66dc72424ae70300aae2ee02b04c4ce4e71d5a#1a66dc72424ae70300aae2ee02b04c4ce4e71d5a"
+source = "git+https://github.com/stippi/gpui-component?rev=abeab54c#abeab54cb264742aea8ab57705109fa49389e27b"
 dependencies = [
  "aho-corasick",
  "anyhow",
@@ -4576,7 +4578,7 @@ dependencies = [
 [[package]]
 name = "gpui-component-assets"
 version = "0.5.1"
-source = "git+https://github.com/stippi/gpui-component?rev=1a66dc72424ae70300aae2ee02b04c4ce4e71d5a#1a66dc72424ae70300aae2ee02b04c4ce4e71d5a"
+source = "git+https://github.com/stippi/gpui-component?rev=abeab54c#abeab54cb264742aea8ab57705109fa49389e27b"
 dependencies = [
  "anyhow",
  "gpui",
@@ -4590,7 +4592,7 @@ dependencies = [
 [[package]]
 name = "gpui-component-macros"
 version = "0.5.1"
-source = "git+https://github.com/stippi/gpui-component?rev=1a66dc72424ae70300aae2ee02b04c4ce4e71d5a#1a66dc72424ae70300aae2ee02b04c4ce4e71d5a"
+source = "git+https://github.com/stippi/gpui-component?rev=abeab54c#abeab54cb264742aea8ab57705109fa49389e27b"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -4600,7 +4602,7 @@ dependencies = [
 [[package]]
 name = "gpui_linux"
 version = "0.1.0"
-source = "git+https://github.com/zed-industries/zed?rev=1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba#1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba"
+source = "git+https://github.com/zed-industries/zed?rev=cc053a4a6fa2fd0e8793201ed9099466af1be0b1#cc053a4a6fa2fd0e8793201ed9099466af1be0b1"
 dependencies = [
  "accesskit",
  "accesskit_unix",
@@ -4622,6 +4624,7 @@ dependencies = [
  "itertools 0.14.0",
  "libc",
  "log",
+ "notify-rust",
  "oo7",
  "open",
  "parking_lot",
@@ -4651,13 +4654,14 @@ dependencies = [
 [[package]]
 name = "gpui_macos"
 version = "0.1.0"
-source = "git+https://github.com/zed-industries/zed?rev=1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba#1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba"
+source = "git+https://github.com/zed-industries/zed?rev=cc053a4a6fa2fd0e8793201ed9099466af1be0b1#cc053a4a6fa2fd0e8793201ed9099466af1be0b1"
 dependencies = [
  "accesskit",
  "accesskit_macos",
  "anyhow",
  "async-task",
  "block",
+ "block2 0.6.2",
  "cbindgen",
  "cocoa 0.26.0",
  "collections",
@@ -4685,6 +4689,7 @@ dependencies = [
  "objc2 0.6.4",
  "objc2-app-kit 0.3.2",
  "objc2-foundation 0.3.2",
+ "objc2-user-notifications",
  "parking_lot",
  "pathfinder_geometry",
  "raw-window-handle",
@@ -4698,7 +4703,7 @@ dependencies = [
 [[package]]
 name = "gpui_macros"
 version = "0.1.0"
-source = "git+https://github.com/zed-industries/zed?rev=1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba#1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba"
+source = "git+https://github.com/zed-industries/zed?rev=cc053a4a6fa2fd0e8793201ed9099466af1be0b1#cc053a4a6fa2fd0e8793201ed9099466af1be0b1"
 dependencies = [
  "heck 0.5.0",
  "proc-macro2",
@@ -4709,7 +4714,7 @@ dependencies = [
 [[package]]
 name = "gpui_platform"
 version = "0.1.0"
-source = "git+https://github.com/zed-industries/zed?rev=1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba#1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba"
+source = "git+https://github.com/zed-industries/zed?rev=cc053a4a6fa2fd0e8793201ed9099466af1be0b1#cc053a4a6fa2fd0e8793201ed9099466af1be0b1"
 dependencies = [
  "console_error_panic_hook",
  "gpui",
@@ -4722,7 +4727,7 @@ dependencies = [
 [[package]]
 name = "gpui_shared_string"
 version = "0.1.0"
-source = "git+https://github.com/zed-industries/zed?rev=1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba#1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba"
+source = "git+https://github.com/zed-industries/zed?rev=cc053a4a6fa2fd0e8793201ed9099466af1be0b1#cc053a4a6fa2fd0e8793201ed9099466af1be0b1"
 dependencies = [
  "schemars 1.2.1",
  "serde",
@@ -4732,7 +4737,7 @@ dependencies = [
 [[package]]
 name = "gpui_util"
 version = "0.1.0"
-source = "git+https://github.com/zed-industries/zed?rev=1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba#1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba"
+source = "git+https://github.com/zed-industries/zed?rev=cc053a4a6fa2fd0e8793201ed9099466af1be0b1#cc053a4a6fa2fd0e8793201ed9099466af1be0b1"
 dependencies = [
  "anyhow",
  "log",
@@ -4742,7 +4747,7 @@ dependencies = [
 [[package]]
 name = "gpui_web"
 version = "0.1.0"
-source = "git+https://github.com/zed-industries/zed?rev=1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba#1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba"
+source = "git+https://github.com/zed-industries/zed?rev=cc053a4a6fa2fd0e8793201ed9099466af1be0b1#cc053a4a6fa2fd0e8793201ed9099466af1be0b1"
 dependencies = [
  "anyhow",
  "console_error_panic_hook",
@@ -4754,7 +4759,6 @@ dependencies = [
  "log",
  "parking_lot",
  "raw-window-handle",
- "smallvec",
  "uuid",
  "wasm-bindgen",
  "wasm-bindgen-futures",
@@ -4766,7 +4770,7 @@ dependencies = [
 [[package]]
 name = "gpui_wgpu"
 version = "0.1.0"
-source = "git+https://github.com/zed-industries/zed?rev=1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba#1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba"
+source = "git+https://github.com/zed-industries/zed?rev=cc053a4a6fa2fd0e8793201ed9099466af1be0b1#cc053a4a6fa2fd0e8793201ed9099466af1be0b1"
 dependencies = [
  "anyhow",
  "bytemuck",
@@ -4784,6 +4788,7 @@ dependencies = [
  "raw-window-handle",
  "smallvec",
  "swash",
+ "unicode-bidi",
  "unicode-segmentation",
  "wasm-bindgen",
  "wasm-bindgen-futures",
@@ -4795,7 +4800,7 @@ dependencies = [
 [[package]]
 name = "gpui_windows"
 version = "0.1.0"
-source = "git+https://github.com/zed-industries/zed?rev=1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba#1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba"
+source = "git+https://github.com/zed-industries/zed?rev=cc053a4a6fa2fd0e8793201ed9099466af1be0b1#cc053a4a6fa2fd0e8793201ed9099466af1be0b1"
 dependencies = [
  "accesskit",
  "accesskit_windows",
@@ -5154,7 +5159,7 @@ dependencies = [
 [[package]]
 name = "http_client"
 version = "0.1.0"
-source = "git+https://github.com/zed-industries/zed?rev=1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba#1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba"
+source = "git+https://github.com/zed-industries/zed?rev=cc053a4a6fa2fd0e8793201ed9099466af1be0b1#cc053a4a6fa2fd0e8793201ed9099466af1be0b1"
 dependencies = [
  "anyhow",
  "async-compression",
@@ -5313,7 +5318,7 @@ dependencies = [
  "js-sys",
  "log",
  "wasm-bindgen",
- "windows-core 0.57.0",
+ "windows-core 0.62.2",
 ]
 
 [[package]]
@@ -5751,7 +5756,7 @@ dependencies = [
  "portable-atomic",
  "portable-atomic-util",
  "serde_core",
- "windows-sys 0.52.0",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -6288,6 +6293,20 @@ version = "0.1.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c41e0c4fef86961ac6d6f8a82609f55f31b05e4fce149ac5710e439df7619ba4"
 
+[[package]]
+name = "mac-notification-sys"
+version = "0.6.15"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "fd604973958ddcc11b561193c0fb96ba146506ef2f231ef2e7c35fd2cbc9beca"
+dependencies = [
+ "cc",
+ "log",
+ "objc2 0.6.4",
+ "objc2-foundation 0.3.2",
+ "time",
+ "uuid",
+]
+
 [[package]]
 name = "mach2"
 version = "0.5.0"
@@ -6477,7 +6496,7 @@ checksum = "490cc448043f947bae3cbee9c203358d62dbee0db12107a74be5c30ccfd09771"
 [[package]]
 name = "media"
 version = "0.1.0"
-source = "git+https://github.com/zed-industries/zed?rev=1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba#1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba"
+source = "git+https://github.com/zed-industries/zed?rev=cc053a4a6fa2fd0e8793201ed9099466af1be0b1#cc053a4a6fa2fd0e8793201ed9099466af1be0b1"
 dependencies = [
  "anyhow",
  "bindgen",
@@ -6794,6 +6813,20 @@ dependencies = [
  "windows-sys 0.52.0",
 ]
 
+[[package]]
+name = "notify-rust"
+version = "4.18.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c5b4c1b4f2aa9f25f63a7a49d3dd0ed567b3670da15330a66b29434be899b891"
+dependencies = [
+ "futures-lite 2.6.1",
+ "log",
+ "mac-notification-sys",
+ "serde",
+ "tauri-winrt-notification",
+ "zbus 5.15.0",
+]
+
 [[package]]
 name = "notify-types"
 version = "1.0.1"
@@ -6818,7 +6851,7 @@ version = "0.50.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7957b9740744892f114936ab4a57b3f487491bbeafaf8083688b16841a4240e5"
 dependencies = [
- "windows-sys 0.59.0",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -7124,6 +7157,16 @@ dependencies = [
  "objc2-foundation 0.3.2",
 ]
 
+[[package]]
+name = "objc2-core-location"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ca347214e24bc973fc025fd0d36ebb179ff30536ed1f80252706db19ee452009"
+dependencies = [
+ "objc2 0.6.4",
+ "objc2-foundation 0.3.2",
+]
+
 [[package]]
 name = "objc2-core-text"
 version = "0.3.2"
@@ -7255,6 +7298,19 @@ dependencies = [
  "objc2-metal 0.3.2",
 ]
 
+[[package]]
+name = "objc2-user-notifications"
+version = "0.3.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9df9128cbbfef73cda168416ccf7f837b62737d748333bfe9ab71c245d76613e"
+dependencies = [
+ "bitflags 2.11.1",
+ "block2 0.6.2",
+ "objc2 0.6.4",
+ "objc2-core-location",
+ "objc2-foundation 0.3.2",
+]
+
 [[package]]
 name = "objc_exception"
 version = "0.1.2"
@@ -7552,7 +7608,7 @@ checksum = "9b4f627cb1b25917193a259e49bdad08f671f8d9708acfd5fe0a8c1455d87220"
 [[package]]
 name = "perf"
 version = "0.1.0"
-source = "git+https://github.com/zed-industries/zed?rev=1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba#1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba"
+source = "git+https://github.com/zed-industries/zed?rev=cc053a4a6fa2fd0e8793201ed9099466af1be0b1#cc053a4a6fa2fd0e8793201ed9099466af1be0b1"
 dependencies = [
  "collections",
  "serde",
@@ -8311,7 +8367,7 @@ dependencies = [
  "once_cell",
  "socket2 0.6.3",
  "tracing",
- "windows-sys 0.52.0",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -8630,7 +8686,7 @@ dependencies = [
 [[package]]
 name = "refineable"
 version = "0.1.0"
-source = "git+https://github.com/zed-industries/zed?rev=1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba#1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba"
+source = "git+https://github.com/zed-industries/zed?rev=cc053a4a6fa2fd0e8793201ed9099466af1be0b1#cc053a4a6fa2fd0e8793201ed9099466af1be0b1"
 dependencies = [
  "derive_refineable",
 ]
@@ -9002,7 +9058,7 @@ dependencies = [
  "errno",
  "libc",
  "linux-raw-sys 0.12.1",
- "windows-sys 0.52.0",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -9184,7 +9240,7 @@ dependencies = [
 [[package]]
 name = "scheduler"
 version = "0.1.0"
-source = "git+https://github.com/zed-industries/zed?rev=1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba#1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba"
+source = "git+https://github.com/zed-industries/zed?rev=cc053a4a6fa2fd0e8793201ed9099466af1be0b1#cc053a4a6fa2fd0e8793201ed9099466af1be0b1"
 dependencies = [
  "async-task",
  "backtrace",
@@ -10122,7 +10178,7 @@ checksum = "13c2bddecc57b384dee18652358fb23172facb8a2c51ccc10d74c157bdea3292"
 [[package]]
 name = "sum_tree"
 version = "0.1.0"
-source = "git+https://github.com/zed-industries/zed?rev=1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba#1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba"
+source = "git+https://github.com/zed-industries/zed?rev=cc053a4a6fa2fd0e8793201ed9099466af1be0b1#cc053a4a6fa2fd0e8793201ed9099466af1be0b1"
 dependencies = [
  "heapless 0.9.3",
  "log",
@@ -10382,9 +10438,9 @@ dependencies = [
 
 [[package]]
 name = "taffy"
-version = "0.12.1"
+version = "0.12.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "73afc801dd6bd47529eaa7c7e90557f107527d1b7c9c7ed7d7803c7b8d0c357f"
+checksum = "340a09581f29809fc0df82a3955501dc7f2a21f887e5d1c13dbe288fe1c0bef4"
 dependencies = [
  "arrayvec",
  "grid",
@@ -10404,6 +10460,17 @@ dependencies = [
  "objc",
 ]
 
+[[package]]
+name = "tauri-winrt-notification"
+version = "0.7.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9ed071c670382e85fc2f48ae706492d8c338f4f89bf72520d32f8abfe880aade"
+dependencies = [
+ "thiserror 2.0.18",
+ "windows 0.61.3",
+ "windows-version",
+]
+
 [[package]]
 name = "temp-env"
 version = "0.3.6"
@@ -10424,7 +10491,7 @@ dependencies = [
  "getrandom 0.4.2",
  "once_cell",
  "rustix 1.1.4",
- "windows-sys 0.52.0",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -11517,7 +11584,7 @@ checksum = "06abde3611657adf66d383f00b093d7faecc7fa57071cce2578660c9f1010821"
 [[package]]
 name = "util_macros"
 version = "0.1.0"
-source = "git+https://github.com/zed-industries/zed?rev=1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba#1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba"
+source = "git+https://github.com/zed-industries/zed?rev=cc053a4a6fa2fd0e8793201ed9099466af1be0b1#cc053a4a6fa2fd0e8793201ed9099466af1be0b1"
 dependencies = [
  "perf",
  "quote",
@@ -11796,8 +11863,7 @@ dependencies = [
 [[package]]
 name = "wasm_thread"
 version = "0.3.3"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b7516db7f32decdadb1c3b8deb1b7d78b9df7606c5cc2f6241737c2ab3a0258e"
+source = "git+https://github.com/zed-industries/wasm_thread?rev=0cf96c7708dfb97ccf3da50347e25edcf75d6937#0cf96c7708dfb97ccf3da50347e25edcf75d6937"
 dependencies = [
  "futures",
  "js-sys",
@@ -12045,6 +12111,7 @@ dependencies = [
  "thiserror 2.0.18",
  "wgpu-core-deps-apple",
  "wgpu-core-deps-emscripten",
+ "wgpu-core-deps-wasm",
  "wgpu-core-deps-windows-linux-android",
  "wgpu-hal",
  "wgpu-naga-bridge",
@@ -12069,6 +12136,15 @@ dependencies = [
  "wgpu-hal",
 ]
 
+[[package]]
+name = "wgpu-core-deps-wasm"
+version = "29.0.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "af1fb1798be2a912497d4c224f72d39bb0cb34af50e8bcc29865bc339c943059"
+dependencies = [
+ "wgpu-hal",
+]
+
 [[package]]
 name = "wgpu-core-deps-windows-linux-android"
 version = "29.0.4"
@@ -12224,7 +12300,7 @@ version = "0.1.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c2a7b1c03c876122aa43f3020e6c3c3ee5c05081c9a00739faf7503aeba10d22"
 dependencies = [
- "windows-sys 0.52.0",
+ "windows-sys 0.48.0",
 ]
 
 [[package]]
@@ -12677,6 +12753,15 @@ dependencies = [
  "windows-link 0.2.1",
 ]
 
+[[package]]
+name = "windows-version"
+version = "0.1.7"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e4060a1da109b9d0326b7262c8e12c84df67cc0dbc9e33cf49e01ccc2eb63631"
+dependencies = [
+ "windows-link 0.2.1",
+]
+
 [[package]]
 name = "windows_aarch64_gnullvm"
 version = "0.48.5"
@@ -13635,7 +13720,7 @@ checksum = "3be3d40e40a133f9c916ee3f9f4fa2d9d63435b5fbe1bfc6d9dae0aa0ada1513"
 [[package]]
 name = "zlog"
 version = "0.1.0"
-source = "git+https://github.com/zed-industries/zed?rev=1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba#1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba"
+source = "git+https://github.com/zed-industries/zed?rev=cc053a4a6fa2fd0e8793201ed9099466af1be0b1#cc053a4a6fa2fd0e8793201ed9099466af1be0b1"
 dependencies = [
  "anyhow",
  "chrono",
@@ -13692,7 +13777,7 @@ dependencies = [
 [[package]]
 name = "ztracing"
 version = "0.1.0"
-source = "git+https://github.com/zed-industries/zed?rev=1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba#1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba"
+source = "git+https://github.com/zed-industries/zed?rev=cc053a4a6fa2fd0e8793201ed9099466af1be0b1#cc053a4a6fa2fd0e8793201ed9099466af1be0b1"
 dependencies = [
  "tracing",
  "tracing-subscriber",
@@ -13703,7 +13788,7 @@ dependencies = [
 [[package]]
 name = "ztracing_macro"
 version = "0.1.0"
-source = "git+https://github.com/zed-industries/zed?rev=1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba#1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba"
+source = "git+https://github.com/zed-industries/zed?rev=cc053a4a6fa2fd0e8793201ed9099466af1be0b1#cc053a4a6fa2fd0e8793201ed9099466af1be0b1"
 
 [[package]]
 name = "zune-core"
diff --git a/Cargo.toml b/Cargo.toml
index 31a432a6..3f25b03e 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -11,18 +11,25 @@ resolver = "2"
 transmutation = { git = "https://github.com/stippi/transmutation", rev = "87f48d18ce3ff61659d7182c1b733eae8ec6f2c9" }
 crossterm = { git = "https://github.com/nornagon/crossterm", branch = "nornagon/color-query" }
 ratatui = { git = "https://github.com/nornagon/ratatui", branch = "nornagon-v0.29.0-patch" }
-# gpui-component depends on gpui via an unpinned `git = ".../zed"` dependency.
-# The workspace must resolve to the *same* zed commit that gpui-component
-# resolves to, otherwise two copies of gpui end up in the graph and cause
-# E0053/E0277/E0599 type mismatches. We pin the zed crates to the exact commit
-# that the pinned gpui-component revision locks to (see gpui-component's own
-# Cargo.lock). At this commit the zed workspace exposes two crates literally
-# named `gpui` (a `0.0.0` internal placeholder and the real `0.2.2`), so the
-# `gpui` patch must carry `version = "=0.2.2"` to disambiguate.
-# When bumping gpui-component, update the `rev` to match gpui-component's
-# Cargo.lock, then run `cargo update -p gpui-component -p gpui`.
+# gpui-component depends on gpui via a `git = ".../zed"` dependency. On its
+# upstream `main` that dependency is UNPINNED, so it floats to zed's branch
+# tip. If we pinned our workspace `gpui` to a fixed `?rev=`, that would be a
+# different Cargo source than gpui-component's floating one, and two copies of
+# gpui would end up in the graph -> E0053/E0277/E0599 type mismatches.
+# To get a single copy we consume gpui-component from the `stippi` fork on a
+# thin "pin-zed" branch: it is upstream `main` plus one commit that pins the
+# zed git deps (gpui/gpui_platform/gpui_web/gpui_macros/reqwest_client) to the
+# exact rev below, so both gpui-component and this workspace resolve to the
+# identical `git+.../zed?rev=<rev>` source.
+# At this zed commit the workspace exposes two crates literally named `gpui`
+# (a `0.0.0` internal placeholder and the real `0.2.2`), so the `gpui` patch
+# must carry `version = "=0.2.2"` to disambiguate.
+# When bumping: rebase the pin-zed commit onto the new gpui-component main,
+# set the zed `rev` to whatever that main's Cargo.lock locks gpui to, push the
+# branch, then update both `rev`s here and run
+# `cargo update -p gpui-component -p gpui`.
 # CI builds with `--locked`. See .github/workflows/release.yml.
-gpui = { git = "https://github.com/zed-industries/zed", rev = "1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba", version = "=0.2.2" }
-gpui_platform = { git = "https://github.com/zed-industries/zed", rev = "1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba" }
-gpui_macros = { git = "https://github.com/zed-industries/zed", rev = "1a246efd7e1b83ab568ec5e3e6c1a43a42e1abba" }
-gpui-component = { git = "https://github.com/stippi/gpui-component", rev = "1a66dc72424ae70300aae2ee02b04c4ce4e71d5a" }
+gpui = { git = "https://github.com/zed-industries/zed", rev = "cc053a4a6fa2fd0e8793201ed9099466af1be0b1", version = "=0.2.2" }
+gpui_platform = { git = "https://github.com/zed-industries/zed", rev = "cc053a4a6fa2fd0e8793201ed9099466af1be0b1" }
+gpui_macros = { git = "https://github.com/zed-industries/zed", rev = "cc053a4a6fa2fd0e8793201ed9099466af1be0b1" }
+gpui-component = { git = "https://github.com/stippi/gpui-component", rev = "abeab54c" }
diff --git a/crates/code_assistant_core/src/tools/impls/execute_command.rs b/crates/code_assistant_core/src/tools/impls/execute_command.rs
index 5e42ce46..e110ab8e 100644
--- a/crates/code_assistant_core/src/tools/impls/execute_command.rs
+++ b/crates/code_assistant_core/src/tools/impls/execute_command.rs
@@ -852,6 +852,25 @@ mod tests {
         }
     }
 
+    /// Poll the mock UI's live terminal text until it contains `needle` or the
+    /// timeout elapses, returning the last observed text either way. Session
+    /// mode spawns a login shell, whose startup latency on a loaded CI runner
+    /// makes exact wall-clock timing of the first output unreliable.
+    async fn wait_for_terminal_text(
+        fixture: &ToolTestFixture,
+        needle: &str,
+        timeout: std::time::Duration,
+    ) -> String {
+        let deadline = tokio::time::Instant::now() + timeout;
+        loop {
+            let text = fixture.ui().unwrap().get_terminal_output_text();
+            if text.contains(needle) || tokio::time::Instant::now() >= deadline {
+                return text;
+            }
+            tokio::time::sleep(std::time::Duration::from_millis(50)).await;
+        }
+    }
+
     #[tokio::test(flavor = "multi_thread")]
     async fn session_mode_returns_session_id_while_running() -> Result<()> {
         let dir = tempfile::tempdir()?;
@@ -970,11 +989,14 @@ mod tests {
             .with_tool_id("tool-bg-1".to_string());
 
         // Short yield: the tool returns while the process is still running,
-        // before it prints the delayed "LATE" marker.
+        // before it prints the delayed "LATE" marker. The gap between the
+        // markers is generous so that login-shell startup latency (session
+        // mode spawns `$SHELL -l -c ...`) on a loaded CI runner cannot blur
+        // EARLY into LATE.
         let session_id = {
             let mut context = fixture.context();
             let mut input = session_mode_input(
-                "printf 'EARLY\\n'; sleep 1; printf 'LATE\\n'; sleep 30",
+                "printf 'EARLY\\n'; sleep 3; printf 'LATE\\n'; sleep 30",
                 500,
             );
             let result = ExecuteCommandTool.execute(&mut context, &mut input).await?;
@@ -984,20 +1006,24 @@ mod tests {
         // The tool call is over (context dropped). The agent would now be
         // doing other work — no tool is polling the session.
 
-        let streamed_at_return = fixture.ui().unwrap().get_terminal_output_text();
+        // The login shell may still be starting up when the tool returns, so
+        // don't demand EARLY at that exact instant; poll for it (comfortably
+        // within the 3s gap before LATE). This is the "output keeps streaming
+        // with no tool call polling the session" guarantee.
+        let streamed_early =
+            wait_for_terminal_text(&fixture, "EARLY", std::time::Duration::from_secs(2)).await;
         assert!(
-            streamed_at_return.contains("EARLY"),
-            "early output should have streamed: {streamed_at_return:?}"
+            streamed_early.contains("EARLY"),
+            "early output should have streamed: {streamed_early:?}"
         );
         assert!(
-            !streamed_at_return.contains("LATE"),
-            "the delayed output cannot have streamed yet: {streamed_at_return:?}"
+            !streamed_early.contains("LATE"),
+            "the delayed output cannot have streamed yet: {streamed_early:?}"
         );
 
         // Wait past the delay without any tool call touching the session.
-        tokio::time::sleep(std::time::Duration::from_millis(1500)).await;
-
-        let streamed_later = fixture.ui().unwrap().get_terminal_output_text();
+        let streamed_later =
+            wait_for_terminal_text(&fixture, "LATE", std::time::Duration::from_secs(5)).await;
         assert!(
             streamed_later.contains("LATE"),
             "output produced between turns should keep streaming to the card: {streamed_later:?}"
diff --git a/crates/llm/src/anthropic.rs b/crates/llm/src/anthropic.rs
index 8a04f60e..f1eb0154 100644
--- a/crates/llm/src/anthropic.rs
+++ b/crates/llm/src/anthropic.rs
@@ -81,39 +81,9 @@ impl DefaultMessageConverter {
         Self
     }
 
-    /// Get cache marker positions based on the stable prefix length.
-    ///
-    /// Messages at and after the first volatile message are excluded because they
-    /// may change or disappear between requests, which would invalidate the
-    /// provider-side cached prefix.
-    ///
-    /// 0-4 messages: no cache markers
-    /// 5-9 messages: marker at index 4
-    /// 10-14 messages: markers at indices 4 and 9
-    /// 15-19 messages: markers at indices 9 and 14
-    /// 20-24 messages: markers at indices 14 and 19
-    /// etc.
-    fn get_cache_marker_positions(&self, messages: &[Message]) -> Vec<usize> {
-        let stable_len = messages
-            .iter()
-            .position(|message| message.volatile)
-            .unwrap_or(messages.len());
-
-        if stable_len < 5 {
-            return vec![];
-        }
-        let remainder = stable_len % 5;
-        let last_marker = stable_len - remainder;
-        if last_marker > 5 {
-            vec![last_marker - 6, last_marker - 1]
-        } else {
-            vec![last_marker - 1]
-        }
-    }
-
     /// Convert generic messages to Anthropic-specific format with cache control
     fn convert_messages_with_cache(&self, messages: Vec<Message>) -> Vec<AnthropicMessage> {
-        let cache_positions = self.get_cache_marker_positions(&messages);
+        let cache_positions = crate::prompt_caching::per_request_marker_positions(&messages);
 
         messages
             .into_iter()
@@ -1432,99 +1402,55 @@ mod tests {
 
     use serde_json::json;
 
-    /// Test cache marker placement based on message count (stateless)
+    /// Test per-request cache marker placement (stateless): leading marker on
+    /// the last message, trailing marker directly before the last assistant
+    /// message — i.e. on the previous request's leading marker position.
     #[test]
-    fn test_message_count_based_cache_markers() {
+    fn test_per_request_cache_markers() {
         let converter = DefaultMessageConverter::new();
 
         // Helper to count cache markers in messages
         fn count_message_cache_markers(result: &[AnthropicMessage]) -> Vec<usize> {
-            let mut marker_positions = Vec::new();
-            for (msg_idx, msg) in result.iter().enumerate() {
-                if msg
-                    .content
-                    .iter()
-                    .any(|block| block.cache_control.is_some())
-                {
-                    marker_positions.push(msg_idx);
-                }
-            }
-            marker_positions
-        }
-
-        // Test 0-4 messages: No cache markers
-        for msg_count in 0..=4 {
-            let messages: Vec<Message> = (0..msg_count)
-                .map(|i| Message::new_user(format!("Message {i}")))
-                .collect();
-
-            let result = converter.convert_messages_with_cache(messages);
-            let markers = count_message_cache_markers(&result);
-            assert!(
-                markers.is_empty(),
-                "{msg_count} messages: Should have no cache markers"
-            );
-        }
-
-        // Test 5-9 messages: Cache marker at index 4
-        for msg_count in 5..=9 {
-            let messages: Vec<Message> = (0..msg_count)
-                .map(|i| Message::new_user(format!("Message {i}")))
-                .collect();
-
-            let result = converter.convert_messages_with_cache(messages);
-            let markers = count_message_cache_markers(&result);
-            assert_eq!(
-                markers,
-                vec![4],
-                "{msg_count} messages: Should have cache marker at index 4"
-            );
-        }
-
-        // Test 10-14 messages: Cache markers at indices 4 and 9
-        for msg_count in 10..=14 {
-            let messages: Vec<Message> = (0..msg_count)
-                .map(|i| Message::new_user(format!("Message {i}")))
-                .collect();
-
-            let result = converter.convert_messages_with_cache(messages);
-            let markers = count_message_cache_markers(&result);
-            assert_eq!(
-                markers,
-                vec![4, 9],
-                "{msg_count} messages: Should have cache markers at indices 4 and 9"
-            );
-        }
-
-        // Test 15-19 messages: Cache markers at indices 9 and 14
-        for msg_count in 15..=19 {
-            let messages: Vec<Message> = (0..msg_count)
-                .map(|i| Message::new_user(format!("Message {i}")))
-                .collect();
-
-            let result = converter.convert_messages_with_cache(messages);
-            let markers = count_message_cache_markers(&result);
-            assert_eq!(
-                markers,
-                vec![9, 14],
-                "{msg_count} messages: Should have cache markers at indices 9 and 14"
-            );
+            result
+                .iter()
+                .enumerate()
+                .filter_map(|(idx, msg)| {
+                    msg.content
+                        .iter()
+                        .any(|block| block.cache_control.is_some())
+                        .then_some(idx)
+                })
+                .collect()
         }
 
-        // Test 20-24 messages: Cache markers at indices 14 and 19
-        for msg_count in 20..=24 {
-            let messages: Vec<Message> = (0..msg_count)
-                .map(|i| Message::new_user(format!("Message {i}")))
-                .collect();
+        // No history → no markers
+        let result = converter.convert_messages_with_cache(vec![]);
+        assert!(count_message_cache_markers(&result).is_empty());
 
-            let result = converter.convert_messages_with_cache(messages);
-            let markers = count_message_cache_markers(&result);
-            assert_eq!(
-                markers,
-                vec![14, 19],
-                "{msg_count} messages: Should have cache markers at indices 14 and 19"
-            );
-        }
+        // User-only history → single marker on the last message
+        let messages: Vec<Message> = (0..3)
+            .map(|i| Message::new_user(format!("Message {i}")))
+            .collect();
+        let result = converter.convert_messages_with_cache(messages);
+        assert_eq!(count_message_cache_markers(&result), vec![2]);
+
+        // Growing conversation: each request's trailing marker must land on
+        // the previous request's leading marker position
+        let mut messages = vec![Message::new_user("u0")];
+        let result = converter.convert_messages_with_cache(messages.clone());
+        assert_eq!(count_message_cache_markers(&result), vec![0]);
+
+        messages.push(Message::new_assistant("a1"));
+        messages.push(Message::new_user("u2"));
+        let result = converter.convert_messages_with_cache(messages.clone());
+        assert_eq!(count_message_cache_markers(&result), vec![0, 2]);
+
+        // Pending user message appended on top of the regular turn
+        messages.push(Message::new_assistant("a3"));
+        messages.push(Message::new_user("u4"));
+        messages.push(Message::new_user("u5 (pending)"));
+        let result = converter.convert_messages_with_cache(messages);
+        assert_eq!(count_message_cache_markers(&result), vec![2, 5]);
     }
 
     /// Test that volatile messages cap the cacheable history prefix.
@@ -1545,25 +1471,33 @@ mod tests {
                 .collect()
         }
 
-        let mut messages: Vec<Message> = (0..18)
-            .map(|i| Message::new_user(format!("Message {i}")))
-            .collect();
+        fn alternating_messages(count: usize) -> Vec<Message> {
+            (0..count)
+                .map(|i| {
+                    if i % 2 == 0 {
+                        Message::new_user(format!("u{i}"))
+                    } else {
+                        Message::new_assistant(format!("a{i}"))
+                    }
+                })
+                .collect()
+        }
+
+        let mut messages = alternating_messages(18);
         messages[12].volatile = true;
         let result = converter.convert_messages_with_cache(messages);
         assert_eq!(
             count_message_cache_markers(&result),
-            vec![4, 9],
-            "18 messages with first volatile at 12 should use markers for the 12-message stable prefix"
+            vec![10, 11],
+            "First volatile at 12 caps markers to the 12-message stable prefix"
         );
 
-        let mut messages: Vec<Message> = (0..18)
-            .map(|i| Message::new_user(format!("Message {i}")))
-            .collect();
-        messages[4].volatile = true;
+        let mut messages = alternating_messages(18);
+        messages[0].volatile = true;
         let result = converter.convert_messages_with_cache(messages);
         assert!(
             count_message_cache_markers(&result).is_empty(),
-            "A volatile message before index 5 should suppress message-history cache markers"
+            "A fully volatile history should carry no cache markers"
         );
     }
 
@@ -1599,7 +1533,8 @@ mod tests {
 
         let result = converter.convert_messages_with_cache(messages);
 
-        // Should have cache markers at indices 9 and 14
+        // Leading marker on the last message (14), trailing marker directly
+        // before the last assistant message (13) → index 12
         let mut cache_markers = Vec::new();
         for (idx, msg) in result.iter().enumerate() {
             if msg
@@ -1612,8 +1547,8 @@ mod tests {
         }
         assert_eq!(
             cache_markers,
-            vec![9, 14],
-            "15 messages should have cache markers at indices 9 and 14"
+            vec![12, 14],
+            "15 messages should have cache markers at indices 12 and 14"
         );
 
         // Verify structured content is preserved
@@ -1693,7 +1628,7 @@ mod tests {
             })
             .collect();
 
-        // Both should have identical cache marker placement (indices 9, 14)
+        // Both should have identical cache marker placement (indices 12, 14)
         let result_a = converter.convert_messages_with_cache(messages_a);
         let result_b = converter.convert_messages_with_cache(messages_b);
 
@@ -1721,13 +1656,13 @@ mod tests {
 
         assert_eq!(
             markers_a,
-            vec![9, 14],
-            "Message set A should have markers at indices 9, 14"
+            vec![12, 14],
+            "Message set A should have markers at indices 12, 14"
         );
         assert_eq!(
             markers_b,
-            vec![9, 14],
-            "Message set B should have markers at indices 9, 14"
+            vec![12, 14],
+            "Message set B should have markers at indices 12, 14"
         );
         assert_eq!(
             markers_a, markers_b,
@@ -1743,8 +1678,8 @@ mod tests {
         let markers_short = get_markers(&result_short);
         assert_eq!(
             markers_short,
-            vec![4],
-            "7 messages should have marker at index 4"
+            vec![6],
+            "7 user-only messages should have a single marker on the last one"
         );
     }
 
@@ -1825,13 +1760,14 @@ mod tests {
                 .collect()
         };
 
-        // Test with the full 30-message conversation
+        // Test with the full 30-message conversation: leading marker on the
+        // last message (29), trailing before the last assistant message (28)
         let result30 = converter.convert_messages_with_cache(messages.clone());
         let markers30 = get_markers(&result30);
         assert_eq!(
             markers30,
-            vec![24, 29],
-            "30 messages: Should have cache markers at indices 24 and 29, found: {markers30:?}"
+            vec![27, 29],
+            "30 messages: Should have cache markers at indices 27 and 29, found: {markers30:?}"
         );
 
         // Test different message counts to demonstrate stateless behavior
@@ -1843,8 +1779,8 @@ mod tests {
         let markers_short = get_markers(&result_short);
         assert_eq!(
             markers_short,
-            vec![4],
-            "7 messages should have marker at index 4"
+            vec![3, 6],
+            "7 messages should have markers at indices 3 and 6"
         );
 
         // Agent 2: Medium conversation (12 messages)
@@ -1853,8 +1789,8 @@ mod tests {
         let markers_medium = get_markers(&result_medium);
         assert_eq!(
             markers_medium,
-            vec![4, 9],
-            "12 messages should have markers at indices 4, 9"
+            vec![9, 11],
+            "12 messages should have markers at indices 9, 11"
         );
 
         // Agent 3: Long conversation (18 messages)
@@ -1863,8 +1799,8 @@ mod tests {
         let markers_long = get_markers(&result_long);
         assert_eq!(
             markers_long,
-            vec![9, 14],
-            "18 messages should have markers at indices 9, 14"
+            vec![15, 17],
+            "18 messages should have markers at indices 15, 17"
         );
 
         // Agent 4: Very long conversation (25 messages)
@@ -1873,8 +1809,8 @@ mod tests {
         let markers_very_long = get_markers(&result_very_long);
         assert_eq!(
             markers_very_long,
-            vec![19, 24],
-            "25 messages should have markers at indices 19, 24"
+            vec![21, 24],
+            "25 messages should have markers at indices 21, 24"
         );
 
         // Verify that cache markers are only placed on first content block of structured messages
diff --git a/crates/llm/src/lib.rs b/crates/llm/src/lib.rs
index fdb61ea4..eb3675ca 100644
--- a/crates/llm/src/lib.rs
+++ b/crates/llm/src/lib.rs
@@ -30,6 +30,7 @@ pub mod openai;
 pub mod openai_responses;
 pub mod openai_responses_ws;
 pub mod openrouter;
+pub mod prompt_caching;
 pub mod provider_config;
 pub mod recording;
 pub mod streaming;
diff --git a/crates/llm/src/openai_responses.rs b/crates/llm/src/openai_responses.rs
index 711fd0f4..13d02b62 100644
--- a/crates/llm/src/openai_responses.rs
+++ b/crates/llm/src/openai_responses.rs
@@ -204,12 +204,38 @@ struct ModelCapabilities {
     supports_verbosity: bool,
     /// Default verbosity for the model
     default_verbosity: Option<Verbosity>,
+    /// Whether the model supports explicit prompt cache breakpoints
+    /// (`prompt_cache_breakpoint`, GPT-5.6 and later; older models reject the field)
+    supports_explicit_cache: bool,
+}
+
+/// Whether a model supports explicit prompt cache breakpoints.
+///
+/// Per OpenAI docs, explicit prompt caching is supported by "GPT-5.6 and later
+/// model families". Older models reject requests containing
+/// `prompt_cache_breakpoint`, so this must stay conservative.
+pub(crate) fn model_supports_explicit_cache(model_lower: &str) -> bool {
+    let Some(pos) = model_lower.find("gpt-") else {
+        return false;
+    };
+    let version: String = model_lower[pos + 4..]
+        .chars()
+        .take_while(|c| c.is_ascii_digit() || *c == '.')
+        .collect();
+    let mut parts = version.split('.');
+    let major: u32 = match parts.next().and_then(|p| p.parse().ok()) {
+        Some(m) => m,
+        None => return false,
+    };
+    let minor: u32 = parts.next().and_then(|p| p.parse().ok()).unwrap_or(0);
+    major > 5 || (major == 5 && minor >= 6)
 }
 
 impl ModelCapabilities {
     /// Get model capabilities based on model ID
     fn for_model(model: &str) -> Self {
         let model_lower = model.to_lowercase();
+        let supports_explicit_cache = model_supports_explicit_cache(&model_lower);
 
         // GPT-5 and variants - full reasoning and verbosity support
         if model_lower.contains("gpt-5") || model_lower.starts_with("gpt5") {
@@ -219,6 +245,7 @@ impl ModelCapabilities {
                 default_summary: Some("auto".to_string()),
                 supports_verbosity: true,
                 default_verbosity: Some(Verbosity::Medium),
+                supports_explicit_cache,
             };
         }
 
@@ -230,6 +257,7 @@ impl ModelCapabilities {
                 default_summary: Some("auto".to_string()),
                 supports_verbosity: false,
                 default_verbosity: None,
+                supports_explicit_cache,
             };
         }
 
@@ -241,6 +269,7 @@ impl ModelCapabilities {
                 default_summary: Some("auto".to_string()),
                 supports_verbosity: false,
                 default_verbosity: None,
+                supports_explicit_cache,
             };
         }
 
@@ -252,6 +281,7 @@ impl ModelCapabilities {
                 default_summary: None,
                 supports_verbosity: false,
                 default_verbosity: None,
+                supports_explicit_cache,
             };
         }
 
@@ -263,6 +293,7 @@ impl ModelCapabilities {
                 default_summary: None,
                 supports_verbosity: false,
                 default_verbosity: None,
+                supports_explicit_cache,
             };
         }
 
@@ -274,6 +305,28 @@ impl ModelCapabilities {
             default_summary: Some("auto".to_string()),
             supports_verbosity: false,
             default_verbosity: None,
+            supports_explicit_cache,
+        }
+    }
+}
+
+/// Place an explicit cache breakpoint on the last breakpoint-eligible content
+/// block within `items[..end]`.
+///
+/// A breakpoint marks the end of the cached prefix including everything
+/// rendered before it, but is only valid on input blocks (`input_text`,
+/// `input_image`). Function calls, tool outputs and assistant text cannot
+/// carry one, so the marker is shifted backwards to the nearest eligible
+/// block. The scan only depends on content before the anchor, which is stable
+/// across requests, so placement stays deterministic.
+fn apply_cache_breakpoint(items: &mut [ResponseInputItem], end: usize) {
+    for item in items[..end].iter_mut().rev() {
+        if let ResponseInputItem::Message { content, .. } = item {
+            for content_item in content.iter_mut().rev() {
+                if content_item.try_mark_cache_breakpoint() {
+                    return;
+                }
+            }
         }
     }
 }
@@ -302,15 +355,37 @@ enum ResponseInputItem {
     },
 }
 
+/// Explicit cache breakpoint marker on an input content block (GPT-5.6+).
+///
+/// Marks the end of a cacheable prompt prefix, including the marked block and
+/// all prompt content rendered before it. Older models reject this field, so
+/// it must only be sent when `ModelCapabilities::supports_explicit_cache`.
+#[derive(Debug, Clone, Serialize)]
+pub(crate) struct PromptCacheBreakpoint {
+    mode: String,
+}
+
+impl PromptCacheBreakpoint {
+    pub(crate) fn explicit() -> Self {
+        Self {
+            mode: "explicit".to_string(),
+        }
+    }
+}
+
 /// Content item within messages
 #[derive(Debug, Serialize)]
 #[serde(tag = "type", rename_all = "snake_case")]
 enum ResponseContentItem {
     InputText {
         text: String,
+        #[serde(skip_serializing_if = "Option::is_none")]
+        prompt_cache_breakpoint: Option<PromptCacheBreakpoint>,
     },
     InputImage {
         image_url: String,
+        #[serde(skip_serializing_if = "Option::is_none")]
+        prompt_cache_breakpoint: Option<PromptCacheBreakpoint>,
     },
     OutputText {
         text: String,
@@ -319,6 +394,41 @@ enum ResponseContentItem {
     },
 }
 
+impl ResponseContentItem {
+    fn input_text(text: String) -> Self {
+        Self::InputText {
+            text,
+            prompt_cache_breakpoint: None,
+        }
+    }
+
+    fn input_image(image_url: String) -> Self {
+        Self::InputImage {
+            image_url,
+            prompt_cache_breakpoint: None,
+        }
+    }
+
+    /// Set the explicit cache breakpoint if this is an input block that
+    /// supports one. Returns whether the marker could be applied.
+    fn try_mark_cache_breakpoint(&mut self) -> bool {
+        match self {
+            Self::InputText {
+                prompt_cache_breakpoint,
+                ..
+            }
+            | Self::InputImage {
+                prompt_cache_breakpoint,
+                ..
+            } => {
+                *prompt_cache_breakpoint = Some(PromptCacheBreakpoint::explicit());
+                true
+            }
+            Self::OutputText { .. } => false,
+        }
+    }
+}
+
 /// Response structure from the Responses API
 #[derive(Debug, Deserialize)]
 struct ResponsesResponse {
@@ -399,6 +509,32 @@ struct ResponsesUsage {
 #[derive(Debug, Deserialize)]
 struct InputTokensDetails {
     cached_tokens: u32,
+    /// Tokens written to the prompt cache (explicit caching, GPT-5.6+).
+    #[serde(default)]
+    cache_write_tokens: u32,
+}
+
+impl ResponsesUsage {
+    /// Convert to the provider-independent `Usage`.
+    ///
+    /// OpenAI reports `input_tokens` inclusive of cached (and cache-write)
+    /// tokens, while the internal `Usage` follows Anthropic semantics where
+    /// `input_tokens` counts only uncached input.
+    fn to_usage(&self) -> Usage {
+        let (cache_read, cache_write) = self
+            .input_tokens_details
+            .as_ref()
+            .map(|d| (d.cached_tokens, d.cache_write_tokens))
+            .unwrap_or((0, 0));
+        Usage {
+            input_tokens: self
+                .input_tokens
+                .saturating_sub(cache_read.saturating_add(cache_write)),
+            output_tokens: self.output_tokens,
+            cache_creation_input_tokens: cache_write,
+            cache_read_input_tokens: cache_read,
+        }
+    }
 }
 
 #[derive(Debug, Deserialize)]
@@ -607,15 +743,29 @@ impl OpenAIResponsesClient {
             .customize_url(&self.base_url, streaming)
     }
 
-    /// Convert internal messages to Responses API format
-    fn convert_messages(&self, messages: Vec<Message>) -> Vec<ResponseInputItem> {
+    /// Convert internal messages to Responses API format, optionally placing
+    /// explicit cache breakpoints at stable history positions.
+    fn convert_messages_with_cache(
+        &self,
+        messages: Vec<Message>,
+        explicit_cache: bool,
+    ) -> Vec<ResponseInputItem> {
+        let cache_positions = if explicit_cache {
+            crate::prompt_caching::cache_marker_positions(&messages)
+        } else {
+            Vec::new()
+        };
+
         let mut result = Vec::new();
+        // Number of converted items after each message, so marker positions
+        // (message indices) can be mapped back to converted input items.
+        let mut items_after_message = Vec::with_capacity(messages.len());
 
         for message in messages {
             match message.content {
                 MessageContent::Text(text) => {
                     let content_item = match message.role {
-                        MessageRole::User => ResponseContentItem::InputText { text },
+                        MessageRole::User => ResponseContentItem::input_text(text),
                         MessageRole::Assistant => ResponseContentItem::OutputText {
                             text,
                             phase: Some("final_answer".to_string()),
@@ -633,6 +783,11 @@ impl OpenAIResponsesClient {
                     self.convert_structured_message(message.role, blocks, &mut result);
                 }
             }
+            items_after_message.push(result.len());
+        }
+
+        for &position in &cache_positions {
+            apply_cache_breakpoint(&mut result, items_after_message[position]);
         }
 
         result
@@ -685,7 +840,7 @@ impl OpenAIResponsesClient {
             match block {
                 ContentBlock::Text { text, .. } => match role {
                     MessageRole::User => {
-                        current_message_content.push(ResponseContentItem::InputText { text });
+                        current_message_content.push(ResponseContentItem::input_text(text));
                     }
                     MessageRole::Assistant => {
                         current_message_content.push(ResponseContentItem::OutputText {
@@ -698,12 +853,11 @@ impl OpenAIResponsesClient {
                     media_type, data, ..
                 } => {
                     let image_url = format!("data:{media_type};base64,{data}");
-                    current_message_content.push(ResponseContentItem::InputImage { image_url });
+                    current_message_content.push(ResponseContentItem::input_image(image_url));
                 }
                 ContentBlock::Thinking { thinking, .. } => match role {
                     MessageRole::User => {
-                        current_message_content
-                            .push(ResponseContentItem::InputText { text: thinking });
+                        current_message_content.push(ResponseContentItem::input_text(thinking));
                     }
                     MessageRole::Assistant => {
                         current_message_content.push(ResponseContentItem::OutputText {
@@ -932,15 +1086,7 @@ impl OpenAIResponsesClient {
             let content = self.convert_output(responses_response.output);
             let usage = responses_response
                 .usage
-                .map_or_else(Usage::zero, |u| Usage {
-                    input_tokens: u.input_tokens,
-                    output_tokens: u.output_tokens,
-                    cache_creation_input_tokens: 0,
-                    cache_read_input_tokens: u
-                        .input_tokens_details
-                        .map(|d| d.cached_tokens)
-                        .unwrap_or(0),
-                });
+                .map_or_else(Usage::zero, |u| u.to_usage());
 
             Ok(LLMResponse {
                 content,
@@ -1044,15 +1190,7 @@ impl OpenAIResponsesClient {
         let content = self.convert_output(responses_response.output);
         let usage = responses_response
             .usage
-            .map_or_else(Usage::zero, |u| Usage {
-                input_tokens: u.input_tokens,
-                output_tokens: u.output_tokens,
-                cache_creation_input_tokens: 0,
-                cache_read_input_tokens: u
-                    .input_tokens_details
-                    .map(|d| d.cached_tokens)
-                    .unwrap_or(0),
-            });
+            .map_or_else(Usage::zero, |u| u.to_usage());
 
         Ok((
             LLMResponse {
@@ -1452,12 +1590,7 @@ impl<'a> StreamProcessor<'a> {
                     .clone(),
             )
         {
-            self.usage.input_tokens = usage_data.input_tokens;
-            self.usage.output_tokens = usage_data.output_tokens;
-            self.usage.cache_read_input_tokens = usage_data
-                .input_tokens_details
-                .map(|d| d.cached_tokens)
-                .unwrap_or(0);
+            *self.usage = usage_data.to_usage();
         }
 
         // Create a RedactedThinking block from collected reasoning items if any
@@ -1489,17 +1622,26 @@ impl LLMProvider for OpenAIResponsesClient {
         request: LLMRequest,
         streaming_callback: Option<&StreamingCallback>,
     ) -> Result<LLMResponse> {
-        let mut input = self.convert_messages(request.messages);
+        // Get model capabilities; explicit cache breakpoints are only sent to
+        // models that support them (older models reject the field).
+        let capabilities = ModelCapabilities::for_model(&self.model);
+
+        let mut input = self
+            .convert_messages_with_cache(request.messages, capabilities.supports_explicit_cache);
 
         // Add system prompt as developer message at the beginning
         if !request.system_prompt.is_empty() {
+            // The system prompt (plus tools rendered before it) is the largest
+            // stable prefix — always mark it, mirroring the Anthropic client.
+            let mut system_item = ResponseContentItem::input_text(request.system_prompt);
+            if capabilities.supports_explicit_cache {
+                system_item.try_mark_cache_breakpoint();
+            }
             input.insert(
                 0,
                 ResponseInputItem::Message {
                     role: "developer".to_string(),
-                    content: vec![ResponseContentItem::InputText {
-                        text: request.system_prompt,
-                    }],
+                    content: vec![system_item],
                 },
             );
         }
@@ -1521,8 +1663,7 @@ impl LLMProvider for OpenAIResponsesClient {
         // Configure for stateless mode with encrypted reasoning
         let store = false;
 
-        // Get model capabilities and build reasoning config
-        let capabilities = ModelCapabilities::for_model(&self.model);
+        // Build reasoning config from model capabilities
         let reasoning = if capabilities.supports_reasoning {
             Some(ReasoningConfig {
                 effort: capabilities.default_effort,
@@ -1594,7 +1735,7 @@ mod tests {
 
         let messages = vec![Message::new_user("Hello")];
 
-        let converted = client.convert_messages(messages);
+        let converted = client.convert_messages_with_cache(messages, false);
         assert_eq!(converted.len(), 1);
 
         match &converted[0] {
@@ -1602,7 +1743,7 @@ mod tests {
                 assert_eq!(role, "user");
                 assert_eq!(content.len(), 1);
                 match &content[0] {
-                    ResponseContentItem::InputText { text } => {
+                    ResponseContentItem::InputText { text, .. } => {
                         assert_eq!(text, "Hello");
                     }
                     _ => panic!("Expected InputText"),
@@ -1623,7 +1764,7 @@ mod tests {
         // Test that assistant messages with simple text use OutputText, not InputText
         let messages = vec![Message::new_assistant("Hello from assistant")];
 
-        let converted = client.convert_messages(messages);
+        let converted = client.convert_messages_with_cache(messages, false);
         assert_eq!(converted.len(), 1);
 
         match &converted[0] {
@@ -1653,7 +1794,7 @@ mod tests {
             ContentBlock::new_tool_result("test_id", "Tool output"),
         ])];
 
-        let converted = client.convert_messages(messages);
+        let converted = client.convert_messages_with_cache(messages, false);
         assert_eq!(converted.len(), 1);
 
         match &converted[0] {
@@ -1780,7 +1921,7 @@ mod tests {
             Message::new_user("What about 3+3?"),
         ];
 
-        let converted = client.convert_messages(conversation);
+        let converted = client.convert_messages_with_cache(conversation, false);
         assert_eq!(converted.len(), 4);
 
         // First: User question
@@ -1832,7 +1973,7 @@ mod tests {
                 assert_eq!(role, "user");
                 assert_eq!(content.len(), 1);
                 match &content[0] {
-                    ResponseContentItem::InputText { text } => {
+                    ResponseContentItem::InputText { text, .. } => {
                         assert_eq!(text, "What about 3+3?");
                     }
                     _ => panic!("Expected InputText"),
@@ -1862,7 +2003,7 @@ mod tests {
             ContentBlock::new_text("Third text"),
         ])];
 
-        let converted = client.convert_messages(messages);
+        let converted = client.convert_messages_with_cache(messages, false);
         assert_eq!(converted.len(), 5);
 
         // First: First text
@@ -1945,7 +2086,7 @@ mod tests {
             ContentBlock::new_text("Based on my reasoning, here's the answer."),
         ])];
 
-        let converted = client.convert_messages(messages);
+        let converted = client.convert_messages_with_cache(messages, false);
         assert_eq!(converted.len(), 2);
 
         // First should be the encrypted reasoning (maintains original order)
@@ -2167,7 +2308,7 @@ mod tests {
 
         // Simulate converting messages with a system prompt
         let messages = vec![Message::new_user("Hello")];
-        let mut input = client.convert_messages(messages);
+        let mut input = client.convert_messages_with_cache(messages, false);
 
         // Add system prompt as developer message (simulating what send_message does)
         let system_prompt = "You are a helpful assistant.";
@@ -2176,9 +2317,7 @@ mod tests {
                 0,
                 ResponseInputItem::Message {
                     role: "developer".to_string(),
-                    content: vec![ResponseContentItem::InputText {
-                        text: system_prompt.to_string(),
-                    }],
+                    content: vec![ResponseContentItem::input_text(system_prompt.to_string())],
                 },
             );
         }
@@ -2191,7 +2330,7 @@ mod tests {
                 assert_eq!(role, "developer");
                 assert_eq!(content.len(), 1);
                 match &content[0] {
-                    ResponseContentItem::InputText { text } => {
+                    ResponseContentItem::InputText { text, .. } => {
                         assert_eq!(text, "You are a helpful assistant.");
                     }
                     _ => panic!("Expected InputText"),
@@ -2206,7 +2345,7 @@ mod tests {
                 assert_eq!(role, "user");
                 assert_eq!(content.len(), 1);
                 match &content[0] {
-                    ResponseContentItem::InputText { text } => {
+                    ResponseContentItem::InputText { text, .. } => {
                         assert_eq!(text, "Hello");
                     }
                     _ => panic!("Expected InputText"),
@@ -2341,7 +2480,7 @@ mod tests {
         );
 
         let messages = vec![Message::new_assistant("Hello from assistant")];
-        let converted = client.convert_messages(messages);
+        let converted = client.convert_messages_with_cache(messages, false);
         assert_eq!(converted.len(), 1);
 
         match &converted[0] {
@@ -2375,7 +2514,7 @@ mod tests {
             ContentBlock::new_tool_use("call_1", "search", serde_json::json!({"query": "weather"})),
         ])];
 
-        let converted = client.convert_messages(messages);
+        let converted = client.convert_messages_with_cache(messages, false);
         // Should produce: Message(OutputText), FunctionCall
         assert_eq!(converted.len(), 2);
 
@@ -2419,7 +2558,7 @@ mod tests {
             ContentBlock::new_text("Second thought."),
         ])];
 
-        let converted = client.convert_messages(messages);
+        let converted = client.convert_messages_with_cache(messages, false);
         // Message("First thought."), FunctionCall, Message("Second thought.")
         assert_eq!(converted.len(), 3);
 
@@ -2447,7 +2586,7 @@ mod tests {
         );
 
         let messages = vec![Message::new_user("Hello")];
-        let converted = client.convert_messages(messages);
+        let converted = client.convert_messages_with_cache(messages, false);
         assert_eq!(converted.len(), 1);
 
         match &converted[0] {
@@ -2483,9 +2622,7 @@ mod tests {
         assert_eq!(json["phase"], "commentary");
 
         // InputText must not carry a phase key
-        let input_item = ResponseContentItem::InputText {
-            text: "hello".to_string(),
-        };
+        let input_item = ResponseContentItem::input_text("hello".to_string());
         let json = serde_json::to_value(&input_item).unwrap();
         assert!(json.get("phase").is_none());
     }
@@ -2576,4 +2713,123 @@ mod tests {
             .expect("error should be retryable");
         assert!(matches!(ctx.error, ApiError::RateLimit(_)));
     }
+
+    #[test]
+    fn test_model_supports_explicit_cache_version_gating() {
+        // GPT-5.6 and later model families support explicit cache breakpoints
+        assert!(model_supports_explicit_cache("gpt-5.6"));
+        assert!(model_supports_explicit_cache("gpt-5.6-codex"));
+        assert!(model_supports_explicit_cache("gpt-5.7-mini"));
+        assert!(model_supports_explicit_cache("gpt-6"));
+
+        // Older models reject the field and must not receive it
+        assert!(!model_supports_explicit_cache("gpt-5"));
+        assert!(!model_supports_explicit_cache("gpt-5-codex"));
+        assert!(!model_supports_explicit_cache("gpt-5.1"));
+        assert!(!model_supports_explicit_cache("gpt-4o"));
+        assert!(!model_supports_explicit_cache("o3-mini"));
+        assert!(!model_supports_explicit_cache("o1"));
+    }
+
+    fn content_items_with_breakpoint(items: &[ResponseInputItem]) -> Vec<usize> {
+        items
+            .iter()
+            .enumerate()
+            .filter_map(|(idx, item)| match item {
+                ResponseInputItem::Message { content, .. } => content
+                    .iter()
+                    .any(|c| {
+                        matches!(
+                            c,
+                            ResponseContentItem::InputText {
+                                prompt_cache_breakpoint: Some(_),
+                                ..
+                            } | ResponseContentItem::InputImage {
+                                prompt_cache_breakpoint: Some(_),
+                                ..
+                            }
+                        )
+                    })
+                    .then_some(idx),
+                _ => None,
+            })
+            .collect()
+    }
+
+    #[test]
+    fn test_explicit_cache_breakpoints_on_user_messages() {
+        let client = OpenAIResponsesClient::new(
+            "test_key".to_string(),
+            "gpt-5.6".to_string(),
+            "https://api.openai.com/v1".to_string(),
+        );
+
+        // 10 user messages → markers at message indices 4 and 9 (1:1 items)
+        let messages: Vec<Message> = (0..10)
+            .map(|i| Message::new_user(format!("Message {i}")))
+            .collect();
+        let items = client.convert_messages_with_cache(messages, true);
+        assert_eq!(content_items_with_breakpoint(&items), vec![4, 9]);
+
+        // Without explicit caching no markers are placed
+        let messages: Vec<Message> = (0..10)
+            .map(|i| Message::new_user(format!("Message {i}")))
+            .collect();
+        let items = client.convert_messages_with_cache(messages, false);
+        assert!(content_items_with_breakpoint(&items).is_empty());
+    }
+
+    #[test]
+    fn test_explicit_cache_breakpoint_shifts_to_eligible_block() {
+        let client = OpenAIResponsesClient::new(
+            "test_key".to_string(),
+            "gpt-5.6".to_string(),
+            "https://api.openai.com/v1".to_string(),
+        );
+
+        // Message index 4 is a tool result (function_call_output, no
+        // breakpoint support) — the marker must shift back to the nearest
+        // input block, the user message at index 3.
+        let messages = vec![
+            Message::new_user("Message 0"),
+            Message::new_assistant("Answer 0"),
+            Message::new_user("Message 1"),
+            Message::new_user("Message 2"),
+            Message::new_user_content(vec![ContentBlock::new_tool_result("call_1", "result")]),
+        ];
+
+        let items = client.convert_messages_with_cache(messages, true);
+        assert_eq!(content_items_with_breakpoint(&items), vec![3]);
+    }
+
+    #[test]
+    fn test_prompt_cache_breakpoint_serialization() {
+        let mut item = ResponseContentItem::input_text("hello".to_string());
+        let json = serde_json::to_value(&item).unwrap();
+        assert!(json.get("prompt_cache_breakpoint").is_none());
+
+        assert!(item.try_mark_cache_breakpoint());
+        let json = serde_json::to_value(&item).unwrap();
+        assert_eq!(json["prompt_cache_breakpoint"]["mode"], "explicit");
+    }
+
+    #[test]
+    fn test_usage_normalization_subtracts_cached_tokens() {
+        let usage: ResponsesUsage = serde_json::from_value(serde_json::json!({
+            "input_tokens": 1000,
+            "output_tokens": 50,
+            "total_tokens": 1050,
+            "input_tokens_details": {
+                "cached_tokens": 700,
+                "cache_write_tokens": 200
+            }
+        }))
+        .unwrap();
+
+        let converted = usage.to_usage();
+        assert_eq!(converted.input_tokens, 100);
+        assert_eq!(converted.cache_read_input_tokens, 700);
+        assert_eq!(converted.cache_creation_input_tokens, 200);
+        assert_eq!(converted.output_tokens, 50);
+    }
 }
diff --git a/crates/llm/src/openai_responses_ws.rs b/crates/llm/src/openai_responses_ws.rs
index 6d4b1a7b..9f7f4b7c 100644
--- a/crates/llm/src/openai_responses_ws.rs
+++ b/crates/llm/src/openai_responses_ws.rs
@@ -52,7 +52,7 @@ use tokio_tungstenite::{
 use tracing::{debug, info, warn};
 
 // Re-export types shared with the HTTP provider
-use crate::openai_responses::Verbosity;
+use crate::openai_responses::{PromptCacheBreakpoint, Verbosity, model_supports_explicit_cache};
 
 // ============================================================================
 // Request / Response types (WebSocket-specific envelope)
@@ -145,9 +145,13 @@ enum WsInputItem {
 enum WsContentItem {
     InputText {
         text: String,
+        #[serde(skip_serializing_if = "Option::is_none")]
+        prompt_cache_breakpoint: Option<PromptCacheBreakpoint>,
     },
     InputImage {
         image_url: String,
+        #[serde(skip_serializing_if = "Option::is_none")]
+        prompt_cache_breakpoint: Option<PromptCacheBreakpoint>,
     },
     OutputText {
         text: String,
@@ -156,6 +160,66 @@ enum WsContentItem {
     },
 }
 
+impl WsContentItem {
+    fn input_text(text: String) -> Self {
+        Self::InputText {
+            text,
+            prompt_cache_breakpoint: None,
+        }
+    }
+
+    fn input_image(image_url: String) -> Self {
+        Self::InputImage {
+            image_url,
+            prompt_cache_breakpoint: None,
+        }
+    }
+
+    /// Set the explicit cache breakpoint if this is an input block that
+    /// supports one. Returns whether the marker could be applied.
+    fn try_mark_cache_breakpoint(&mut self) -> bool {
+        match self {
+            Self::InputText {
+                prompt_cache_breakpoint,
+                ..
+            }
+            | Self::InputImage {
+                prompt_cache_breakpoint,
+                ..
+            } => {
+                *prompt_cache_breakpoint = Some(PromptCacheBreakpoint::explicit());
+                true
+            }
+            Self::OutputText { .. } => false,
+        }
+    }
+}
+
+/// Place an explicit cache breakpoint on the last breakpoint-eligible content
+/// block within `items[..end]`.
+///
+/// A breakpoint marks the end of the cached prefix including everything
+/// rendered before it, but is only valid on input blocks (`input_text`,
+/// `input_image`). Function calls, tool outputs and assistant text cannot
+/// carry one, so the marker is shifted backwards to the nearest eligible
+/// block. The scan only depends on content before the anchor, which is stable
+/// across requests, so placement stays deterministic.
+///
+/// Markers serialize as part of the input items, so `compute_delta` naturally
+/// forces a full resend whenever marker positions move — the previously
+/// cached token prefix stays valid because markers are not prompt content.
+fn apply_cache_breakpoint(items: &mut [WsInputItem], end: usize) {
+    for item in items[..end].iter_mut().rev() {
+        if let WsInputItem::Message { content, .. } = item {
+            for content_item in content.iter_mut().rev() {
+                if content_item.try_mark_cache_breakpoint() {
+                    return;
+                }
+            }
+        }
+    }
+}
+
 // ---------------------------------------------------------------------------
 // Output / event types (received from the API)
 // ---------------------------------------------------------------------------
@@ -249,6 +313,32 @@ struct WsResponsesUsage {
 #[derive(Debug, Deserialize)]
 struct WsInputTokensDetails {
     cached_tokens: u32,
+    /// Tokens written to the prompt cache (explicit caching, GPT-5.6+).
+    #[serde(default)]
+    cache_write_tokens: u32,
+}
+
+impl WsResponsesUsage {
+    /// Convert to the provider-independent `Usage`.
+    ///
+    /// OpenAI reports `input_tokens` inclusive of cached (and cache-write)
+    /// tokens, while the internal `Usage` follows Anthropic semantics where
+    /// `input_tokens` counts only uncached input.
+    fn to_usage(&self) -> Usage {
+        let (cache_read, cache_write) = self
+            .input_tokens_details
+            .as_ref()
+            .map(|d| (d.cached_tokens, d.cache_write_tokens))
+            .unwrap_or((0, 0));
+        Usage {
+            input_tokens: self
+                .input_tokens
+                .saturating_sub(cache_read.saturating_add(cache_write)),
+            output_tokens: self.output_tokens,
+            cache_creation_input_tokens: cache_write,
+            cache_read_input_tokens: cache_read,
+        }
+    }
 }
 
 // ============================================================================
@@ -262,11 +352,15 @@ struct ModelCapabilities {
     default_summary: Option<String>,
     supports_verbosity: bool,
     default_verbosity: Option<Verbosity>,
+    /// Whether the model supports explicit prompt cache breakpoints
+    /// (`prompt_cache_breakpoint`, GPT-5.6 and later; older models reject the field)
+    supports_explicit_cache: bool,
 }
 
 impl ModelCapabilities {
     fn for_model(model: &str) -> Self {
         let m = model.to_lowercase();
+        let supports_explicit_cache = model_supports_explicit_cache(&m);
 
         if m.contains("gpt-5") || m.starts_with("gpt5") {
             return Self {
@@ -275,6 +369,7 @@ impl ModelCapabilities {
                 default_summary: Some("auto".into()),
                 supports_verbosity: true,
                 default_verbosity: Some(Verbosity::Medium),
+                supports_explicit_cache,
             };
         }
         if m.starts_with("o3") || m.starts_with("o4") {
@@ -284,6 +379,7 @@ impl ModelCapabilities {
                 default_summary: Some("auto".into()),
                 supports_verbosity: false,
                 default_verbosity: None,
+                supports_explicit_cache,
             };
         }
         if m.starts_with("o1") {
@@ -293,6 +389,7 @@ impl ModelCapabilities {
                 default_summary: Some("auto".into()),
                 supports_verbosity: false,
                 default_verbosity: None,
+                supports_explicit_cache,
             };
         }
         if m.contains("gpt-4o") || m.contains("gpt4o") || m.contains("gpt-4") || m.contains("gpt4")
@@ -303,6 +400,7 @@ impl ModelCapabilities {
                 default_summary: None,
                 supports_verbosity: false,
                 default_verbosity: None,
+                supports_explicit_cache,
             };
         }
         // Default: assume reasoning support with conservative defaults
@@ -312,6 +410,7 @@ impl ModelCapabilities {
             default_summary: Some("auto".into()),
             supports_verbosity: false,
             default_verbosity: None,
+            supports_explicit_cache,
         }
     }
 }
@@ -587,8 +686,24 @@ impl OpenAIResponsesWsClient {
     // Message conversion (internal types → WS input items)
     // -----------------------------------------------------------------------
 
-    fn convert_messages(&self, messages: Vec<Message>) -> Vec<WsInputItem> {
+    /// Convert internal messages to WS input items, optionally placing
+    /// explicit cache breakpoints at stable history positions.
+    fn convert_messages_with_cache(
+        &self,
+        messages: Vec<Message>,
+        explicit_cache: bool,
+    ) -> Vec<WsInputItem> {
+        let cache_positions = if explicit_cache {
+            crate::prompt_caching::cache_marker_positions(&messages)
+        } else {
+            Vec::new()
+        };
+
         let mut items = Vec::new();
+        // Number of converted items after each message, so marker positions
+        // (message indices) can be mapped back to converted input items.
+        let mut items_after_message = Vec::with_capacity(messages.len());
+
         for message in messages {
             match &message.content {
                 MessageContent::Text(text) => {
@@ -605,7 +720,7 @@ impl OpenAIResponsesWsClient {
                             phase: Some("final_answer".to_string()),
                         }
                     } else {
-                        WsContentItem::InputText { text: text.clone() }
+                        WsContentItem::input_text(text.clone())
                     };
                     items.push(WsInputItem::Message {
                         role: role.to_string(),
@@ -616,7 +731,13 @@ impl OpenAIResponsesWsClient {
                     self.convert_structured_message(&message.role, blocks, &mut items);
                 }
             }
+            items_after_message.push(items.len());
         }
+
+        for &position in &cache_positions {
+            apply_cache_breakpoint(&mut items, items_after_message[position]);
+        }
+
         items
     }
 
@@ -652,7 +773,7 @@ impl OpenAIResponsesWsClient {
             match block {
                 ContentBlock::Text { text, .. } => {
                     let item = if *role == MessageRole::User {
-                        WsContentItem::InputText { text: text.clone() }
+                        WsContentItem::input_text(text.clone())
                     } else {
                         WsContentItem::OutputText {
                             text: text.clone(),
@@ -662,9 +783,10 @@ impl OpenAIResponsesWsClient {
                     current_content.push(item);
                 }
                 ContentBlock::Image { data, .. } => {
-                    current_content.push(WsContentItem::InputImage {
-                        image_url: format!("data:image/png;base64,{}", data),
-                    });
+                    current_content.push(WsContentItem::input_image(format!(
+                        "data:image/png;base64,{}",
+                        data
+                    )));
                 }
                 ContentBlock::Thinking {
                     thinking,
@@ -906,7 +1028,15 @@ impl OpenAIResponsesWsClient {
         request: LLMRequest,
         streaming_callback: Option<&StreamingCallback>,
     ) -> Result<LLMResponse> {
-        let input = self.convert_messages(request.messages);
+        let capabilities = ModelCapabilities::for_model(&self.model);
+
+        // Explicit cache breakpoints are only sent to models that support
+        // them (older models reject the field). The system prompt travels as
+        // `instructions` (a plain string, no breakpoint possible); a marker on
+        // a history message still caches instructions and tools, since a
+        // breakpoint covers everything rendered before it.
+        let input = self
+            .convert_messages_with_cache(request.messages, capabilities.supports_explicit_cache);
 
         // Add system prompt as the top-level `instructions` field (WebSocket style)
         // but also keep it as a developer message in `input` for compatibility.
@@ -930,8 +1060,6 @@ impl OpenAIResponsesWsClient {
                 .collect()
         });
 
-        let capabilities = ModelCapabilities::for_model(&self.model);
-
         let reasoning = if capabilities.supports_reasoning {
             Some(ReasoningConfig {
                 effort: capabilities.default_effort,
@@ -1277,12 +1405,7 @@ impl OpenAIResponsesWsClient {
                                         usage_val.clone(),
                                     )
                                 {
-                                    usage.input_tokens = u.input_tokens;
-                                    usage.output_tokens = u.output_tokens;
-                                    usage.cache_read_input_tokens = u
-                                        .input_tokens_details
-                                        .map(|d| d.cached_tokens)
-                                        .unwrap_or(0);
+                                    usage = u.to_usage();
                                 }
                             }
 
@@ -1632,9 +1755,7 @@ mod tests {
             previous_response_id: None,
             input: vec![WsInputItem::Message {
                 role: "user".to_string(),
-                content: vec![WsContentItem::InputText {
-                    text: "Hello".to_string(),
-                }],
+                content: vec![WsContentItem::input_text("Hello".to_string())],
             }],
             tools: None,
             tool_choice: Some("auto".to_string()),
@@ -1663,9 +1784,7 @@ mod tests {
             previous_response_id: Some("resp_abc123".to_string()),
             input: vec![WsInputItem::Message {
                 role: "user".to_string(),
-                content: vec![WsContentItem::InputText {
-                    text: "Follow-up".to_string(),
-                }],
+                content: vec![WsContentItem::input_text("Follow-up".to_string())],
             }],
             tools: None,
             tool_choice: None,
@@ -1709,9 +1828,7 @@ mod tests {
         // No previous state => no delta
         let input = vec![WsInputItem::Message {
             role: "user".to_string(),
-            content: vec![WsContentItem::InputText {
-                text: "Hello".to_string(),
-            }],
+            content: vec![WsContentItem::input_text("Hello".to_string())],
         }];
         assert!(client.compute_delta(&input).is_none());
 
@@ -1730,9 +1847,7 @@ mod tests {
         });
         extended.push(WsInputItem::Message {
             role: "user".to_string(),
-            content: vec![WsContentItem::InputText {
-                text: "How are you?".to_string(),
-            }],
+            content: vec![WsContentItem::input_text("How are you?".to_string())],
         });
 
         let result = client.compute_delta(&extended);
@@ -1832,7 +1947,7 @@ mod tests {
         );
 
         let messages = vec![Message::new_assistant("Hello from assistant")];
-        let converted = client.convert_messages(messages);
+        let converted = client.convert_messages_with_cache(messages, false);
         assert_eq!(converted.len(), 1);
 
         match &converted[0] {
@@ -1866,7 +1981,7 @@ mod tests {
             ContentBlock::new_tool_use("call_1", "search", serde_json::json!({"query": "weather"})),
         ])];
 
-        let converted = client.convert_messages(messages);
+        let converted = client.convert_messages_with_cache(messages, false);
         // Should produce: Message(OutputText), FunctionCall
         assert_eq!(converted.len(), 2);
 
@@ -1897,7 +2012,7 @@ mod tests {
         );
 
         let messages = vec![Message::new_user("Hello")];
-        let converted = client.convert_messages(messages);
+        let converted = client.convert_messages_with_cache(messages, false);
         assert_eq!(converted.len(), 1);
 
         match &converted[0] {
@@ -1931,10 +2046,92 @@ mod tests {
         let json = serde_json::to_value(&commentary_item).unwrap();
         assert_eq!(json["phase"], "commentary");
 
-        let input_item = WsContentItem::InputText {
-            text: "hello".to_string(),
-        };
+        let input_item = WsContentItem::input_text("hello".to_string());
         let json = serde_json::to_value(&input_item).unwrap();
         assert!(json.get("phase").is_none());
     }
+
+    fn ws_items_with_breakpoint(items: &[WsInputItem]) -> Vec<usize> {
+        items
+            .iter()
+            .enumerate()
+            .filter_map(|(idx, item)| match item {
+                WsInputItem::Message { content, .. } => content
+                    .iter()
+                    .any(|c| {
+                        matches!(
+                            c,
+                            WsContentItem::InputText {
+                                prompt_cache_breakpoint: Some(_),
+                                ..
+                            } | WsContentItem::InputImage {
+                                prompt_cache_breakpoint: Some(_),
+                                ..
+                            }
+                        )
+                    })
+                    .then_some(idx),
+                _ => None,
+            })
+            .collect()
+    }
+
+    #[test]
+    fn test_explicit_cache_breakpoints_on_user_messages() {
+        let client = OpenAIResponsesWsClient::new(
+            "test_key".to_string(),
+            "gpt-5.6".to_string(),
+            "https://api.openai.com/v1".to_string(),
+        );
+
+        // 10 user messages → markers at message indices 4 and 9 (1:1 items)
+        let messages: Vec<Message> = (0..10)
+            .map(|i| Message::new_user(format!("Message {i}")))
+            .collect();
+        let items = client.convert_messages_with_cache(messages, true);
+        assert_eq!(ws_items_with_breakpoint(&items), vec![4, 9]);
+
+        // Without explicit caching no markers are placed
+        let messages: Vec<Message> = (0..10)
+            .map(|i| Message::new_user(format!("Message {i}")))
+            .collect();
+        let items = client.convert_messages_with_cache(messages, false);
+        assert!(ws_items_with_breakpoint(&items).is_empty());
+    }
+
+    /// Moving cache markers must invalidate the incremental-input delta so
+    /// the full input (with markers at their new positions) is resent.
+    #[test]
+    fn test_moved_cache_marker_invalidates_delta() {
+        let mut client = OpenAIResponsesWsClient::new(
+            "test_key".to_string(),
+            "gpt-5.6".to_string(),
+            "https://api.openai.com/v1".to_string(),
+        );
+
+        let messages: Vec<Message> = (0..10)
+            .map(|i| Message::new_user(format!("Message {i}")))
+            .collect();
+        let prev_items = client.convert_messages_with_cache(messages.clone(), true);
+        client.last_response_id = Some("resp_1".to_string());
+        client.last_input_items = prev_items;
+
+        // Growing to 12 messages keeps markers at 4 and 9 → strict extension
+        let mut extended = messages.clone();
+        extended.push(Message::new_user("Message 10"));
+        extended.push(Message::new_user("Message 11"));
+        let items = client.convert_messages_with_cache(extended, true);
+        let delta = client.compute_delta(&items);
+        assert!(delta.is_some());
+        assert_eq!(delta.unwrap().1.len(), 2);
+
+        // Growing to 15 messages moves markers to 9 and 14 → full resend
+        let mut extended = messages;
+        for i in 10..15 {
+            extended.push(Message::new_user(format!("Message {i}")));
+        }
+        let items = client.convert_messages_with_cache(extended, true);
+        assert_eq!(ws_items_with_breakpoint(&items), vec![9, 14]);
+        assert!(client.compute_delta(&items).is_none());
+    }
 }
diff --git a/crates/llm/src/prompt_caching.rs b/crates/llm/src/prompt_caching.rs
new file mode 100644
index 00000000..c7f63d9e
--- /dev/null
+++ b/crates/llm/src/prompt_caching.rs
@@ -0,0 +1,143 @@
+//! Shared cache-marker placement logic for providers with explicit prompt caching.
+//!
+//! Both the Anthropic client and the OpenAI Responses clients place explicit
+//! cache breakpoints into the message history. The placement strategy is
+//! provider-independent: markers anchor at message indices derived from the
+//! stable (non-volatile) history prefix.
+
+use crate::types::{Message, MessageRole};
+
+/// Length of the stable history prefix.
+///
+/// Messages at and after the first volatile message are excluded because they
+/// may change or disappear between requests, which would invalidate the
+/// provider-side cached prefix.
+fn stable_prefix_len(messages: &[Message]) -> usize {
+    messages
+        .iter()
+        .position(|message| message.volatile)
+        .unwrap_or(messages.len())
+}
+
+/// Get cache marker positions that move on every request. Used by the
+/// Anthropic client, which resends the full message history each request.
+///
+/// Two markers are placed inside the stable prefix:
+///
+/// - the last stable message, so the entire prompt built in this request is
+///   cached as early as possible (a write pays for itself as soon as the
+///   prefix is reused once: 1.25x + 0.1x < 1x + 1x)
+/// - the message directly before the last stable assistant message
+///
+/// Every request ends on a user message and the response is appended right
+/// behind it as an assistant message, so the message before the last
+/// assistant message is exactly where the previous request placed its leading
+/// marker. The lookup there is guaranteed to hit and only the messages
+/// appended since then are written — without tracking any state, and
+/// regardless of how many messages arrived in between (tool results, pending
+/// user messages).
+pub fn per_request_marker_positions(messages: &[Message]) -> Vec<usize> {
+    let stable_len = stable_prefix_len(messages);
+    if stable_len == 0 {
+        return vec![];
+    }
+
+    let leading = stable_len - 1;
+    let trailing = messages[..stable_len]
+        .iter()
+        .rposition(|message| message.role == MessageRole::Assistant)
+        .and_then(|assistant_index| assistant_index.checked_sub(1));
+
+    match trailing {
+        // `trailing < leading` always holds: the assistant message itself
+        // sits between the two positions.
+        Some(trailing) => vec![trailing, leading],
+        None => vec![leading],
+    }
+}
+
+/// Get cache marker positions quantized to blocks of five messages. Used by
+/// the OpenAI Responses clients.
+///
+/// There the markers are stable *anchors* rather than the primary caching
+/// mechanism: OpenAI's implicit mode already caches up to the latest message
+/// on every request, and rarely-moving markers keep the WebSocket client's
+/// incremental input intact (moving a marker changes an already-sent item and
+/// forces a full resend).
+///
+/// 0-4 messages: no cache markers
+/// 5-9 messages: marker at index 4
+/// 10-14 messages: markers at indices 4 and 9
+/// 15-19 messages: markers at indices 9 and 14
+/// 20-24 messages: markers at indices 14 and 19
+/// etc.
+pub fn cache_marker_positions(messages: &[Message]) -> Vec<usize> {
+    let stable_len = stable_prefix_len(messages);
+
+    if stable_len < 5 {
+        return vec![];
+    }
+    let remainder = stable_len % 5;
+    let last_marker = stable_len - remainder;
+    if last_marker > 5 {
+        vec![last_marker - 6, last_marker - 1]
+    } else {
+        vec![last_marker - 1]
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    /// The trailing marker of each request must land on the leading marker
+    /// position of the previous request, so that cache lookups always hit.
+    #[test]
+    fn per_request_markers_chain_across_requests() {
+        // Request 1: a single user message → one marker
+        let mut messages = vec![Message::new_user("u0")];
+        assert_eq!(per_request_marker_positions(&messages), vec![0]);
+
+        // Response appended, tool results form the next request
+        messages.push(Message::new_assistant("a1"));
+        messages.push(Message::new_user("u2"));
+        assert_eq!(per_request_marker_positions(&messages), vec![0, 2]);
+
+        // Next turn adds a pending user message on top — the trailing marker
+        // still lands on the previous request's leading position (index 2)
+        messages.push(Message::new_assistant("a3"));
+        messages.push(Message::new_user("u4"));
+        messages.push(Message::new_user("u5 (pending)"));
+        assert_eq!(per_request_marker_positions(&messages), vec![2, 5]);
+    }
+
+    #[test]
+    fn per_request_markers_respect_volatile_prefix() {
+        let mut messages: Vec<Message> = (0..18)
+            .map(|i| {
+                if i % 2 == 0 {
+                    Message::new_user(format!("u{i}"))
+                } else {
+                    Message::new_assistant(format!("a{i}"))
+                }
+            })
+            .collect();
+        assert_eq!(per_request_marker_positions(&messages), vec![16, 17]);
+
+        // Volatile tail caps the markable prefix
+        messages[12].volatile = true;
+        assert_eq!(per_request_marker_positions(&messages), vec![10, 11]);
+
+        // Everything volatile → no markers
+        messages[0].volatile = true;
+        assert!(per_request_marker_positions(&messages).is_empty());
+    }
+
+    #[test]
+    fn per_request_markers_without_assistant_history() {
+        assert!(per_request_marker_positions(&[]).is_empty());
+
+        let messages: Vec<Message> = (0..3).map(|i| Message::new_user(format!("u{i}"))).collect();
+        assert_eq!(per_request_marker_positions(&messages), vec![2]);
+    }
+}
diff --git a/crates/ui_gpui/src/blocks/mod.rs b/crates/ui_gpui/src/blocks/mod.rs
index 8c6cf968..abfdd38b 100644
--- a/crates/ui_gpui/src/blocks/mod.rs
+++ b/crates/ui_gpui/src/blocks/mod.rs
@@ -7,7 +7,7 @@ pub use data::*;
 
 use gpui::prelude::*;
 use gpui::{Context, Entity, Pixels, Task, px};
-use gpui_component::text::{TextView, TextViewState};
+use gpui_component::text::{SelectionFormat, TextView, TextViewState};
 
 use std::cell::Cell;
 use std::rc::Rc;
@@ -237,9 +237,14 @@ impl BlockView {
         // selection (e.g. `**bold**`, list markers, code fences) rather than the
         // rendered plain text. The hover copy button already copies full-block
         // source; this keeps partial-selection copy consistent.
+        let selection_format = if selectable {
+            SelectionFormat::Source
+        } else {
+            SelectionFormat::Plain
+        };
         TextView::new(&state)
             .selectable(selectable)
-            .selectable_source(selectable)
+            .selection_format(selection_format)
     }
 
     /// Whether the copy button should currently render its "copied" checkmark.