chore(release): bump desktop to 0.1.6

- Modified the content security policy in `tauri.conf.json` to include `https:` in the `img-src` directive, enhancing security by allowing images from secure sources.

.github/scripts/generate-portable-update.mjs

@@ -0,0 +1,71 @@
+import { readFile, writeFile } from "node:fs/promises";
+import { basename, resolve } from "node:path";
+import { pathToFileURL } from "node:url";
+
+function readOptions(args) {
+  const options = new Map();
+  for (let index = 0; index < args.length; index += 2) {
+    const name = args[index];
+    const value = args[index + 1];
+    if (!name?.startsWith("--") || value === undefined) {
+      throw new Error(`invalid argument near ${name ?? "end of command"}`);
+    }
+    options.set(name.slice(2), value);
+  }
+  return options;
+}
+
+function required(options, name) {
+  const value = options.get(name)?.trim();
+  if (!value) throw new Error(`--${name} is required`);
+  return value;
+}
+
+export function generatePortableUpdate({ version, repository, assetName, signature }) {
+  const normalizedVersion = version.replace(/^v/, "");
+  if (!/^\d+\.\d+\.\d+(?:-[0-9A-Za-z.-]+)?$/.test(normalizedVersion)) {
+    throw new Error(`invalid semantic version: ${normalizedVersion}`);
+  }
+  if (!/^[^/\s]+\/[^/\s]+$/.test(repository)) {
+    throw new Error(`invalid GitHub repository: ${repository}`);
+  }
+  if (!assetName || basename(assetName) !== assetName) {
+    throw new Error("asset name must be a file name");
+  }
+  if (!signature.trim()) throw new Error("signature is required");
+
+  return {
+    version: normalizedVersion,
+    notes: `Cursor BYOK v${normalizedVersion}`,
+    pub_date: new Date().toISOString(),
+    platforms: {
+      "windows-x86_64": {
+        signature: signature.trim(),
+        url: `https://github.com/${repository}/releases/download/v${normalizedVersion}/${encodeURIComponent(assetName)}`,
+      },
+    },
+  };
+}
+
+async function main() {
+  const options = readOptions(process.argv.slice(2));
+  const version = required(options, "version");
+  const repository = required(options, "repository");
+  const asset = required(options, "asset");
+  const signaturePath = resolve(required(options, "signature"));
+  const output = resolve(required(options, "output"));
+  const manifest = generatePortableUpdate({
+    version,
+    repository,
+    assetName: basename(asset),
+    signature: await readFile(signaturePath, "utf8"),
+  });
+  await writeFile(output, `${JSON.stringify(manifest, null, 2)}\n`);
+}
+
+if (process.argv[1] && import.meta.url === pathToFileURL(resolve(process.argv[1])).href) {
+  main().catch((error) => {
+    console.error(error instanceof Error ? error.message : String(error));
+    process.exitCode = 1;
+  });
+}

.github/scripts/generate-portable-update.test.mjs

@@ -0,0 +1,41 @@
+import assert from "node:assert/strict";
+import test from "node:test";
+import { generatePortableUpdate } from "./generate-portable-update.mjs";
+
+test("generates a signed Windows portable updater manifest", () => {
+  const manifest = generatePortableUpdate({
+    version: "v1.2.3-beta.1",
+    repository: "owner/repository",
+    assetName: "cursor-byok-1.2.3-beta.1-windows-amd64.zip",
+    signature: "signed-payload\n",
+  });
+
+  assert.equal(manifest.version, "1.2.3-beta.1");
+  assert.deepEqual(Object.keys(manifest.platforms), ["windows-x86_64"]);
+  assert.equal(manifest.platforms["windows-x86_64"].signature, "signed-payload");
+  assert.equal(
+    manifest.platforms["windows-x86_64"].url,
+    "https://github.com/owner/repository/releases/download/v1.2.3-beta.1/cursor-byok-1.2.3-beta.1-windows-amd64.zip",
+  );
+});
+
+test("rejects invalid inputs", () => {
+  assert.throws(() => generatePortableUpdate({
+    version: "latest",
+    repository: "owner/repository",
+    assetName: "update.zip",
+    signature: "signature",
+  }), /semantic version/);
+  assert.throws(() => generatePortableUpdate({
+    version: "1.2.3",
+    repository: "owner/repository",
+    assetName: "../update.zip",
+    signature: "signature",
+  }), /file name/);
+  assert.throws(() => generatePortableUpdate({
+    version: "1.2.3",
+    repository: "owner/repository",
+    assetName: "update.zip",
+    signature: " ",
+  }), /signature/);
+});

.github/workflows/release.yml

@@ -158,14 +158,27 @@ jobs:
           mkdir -p legacy-update
           tar -czf "legacy-update/cursor-byok-${VERSION}-linux-amd64.tar.gz" -C target/release cursor-byok-desktop
 
-      - name: Package legacy Windows updater asset
+      - name: Package and sign legacy Windows updater asset
         if: matrix.platform == 'windows-x86_64'
         shell: pwsh
         env:
           VERSION: ${{ needs.prepare.outputs.version }}
+          TAURI_SIGNING_PRIVATE_KEY: ${{ secrets.TAURI_SIGNING_PRIVATE_KEY }}
+          TAURI_SIGNING_PRIVATE_KEY_PASSWORD: ${{ secrets.TAURI_SIGNING_PRIVATE_KEY_PASSWORD }}
         run: |
           New-Item -ItemType Directory -Force legacy-update | Out-Null
-          Compress-Archive -LiteralPath target/release/cursor-byok-desktop.exe -DestinationPath "legacy-update/cursor-byok-$env:VERSION-windows-amd64.zip"
+          $asset = "legacy-update/cursor-byok-$env:VERSION-windows-amd64.zip"
+          Compress-Archive -LiteralPath target/release/cursor-byok-desktop.exe -DestinationPath $asset
+          Push-Location apps/desktop
+          npm exec tauri signer sign -- "../../$asset"
+          Pop-Location
+          $entries = @(tar -tf $asset)
+          if ($entries.Count -ne 1 -or [System.IO.Path]::GetFileName($entries[0]) -ne 'cursor-byok-desktop.exe') {
+            throw "Windows updater archive must contain only cursor-byok-desktop.exe"
+          }
+          if (!(Test-Path "$asset.sig")) {
+            throw "Windows updater archive signature was not generated"
+          }
 
       - name: Package legacy macOS updater asset
         if: contains(matrix.platform, 'macos')
@@ -213,6 +226,20 @@ jobs:
             --output legacy-update/update.json \
             --notes "Cursor BYOK v${VERSION}"
 
+      - name: Generate signed Windows portable update manifest
+        env:
+          VERSION: ${{ needs.prepare.outputs.version }}
+        run: |
+          asset="cursor-byok-${VERSION}-windows-amd64.zip"
+          test -f "legacy-update/${asset}"
+          test -f "legacy-update/${asset}.sig"
+          node .github/scripts/generate-portable-update.mjs \
+            --version "${VERSION}" \
+            --repository "${GITHUB_REPOSITORY}" \
+            --asset "${asset}" \
+            --signature "legacy-update/${asset}.sig" \
+            --output legacy-update/portable-latest.json
+
       - name: Normalize Tauri updater download URLs
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}

Cargo.lock

@@ -1172,10 +1172,11 @@ checksum = "52560adf09603e58c9a7ee1fe1dcb95a16927b17c127f0ac02d6e768a0e25bc1"
 
 [[package]]
 name = "cursor-byok-desktop"
-version = "0.1.5"
+version = "0.1.6"
 dependencies = [
  "axum",
  "cursor-server",
+ "libc",
  "rfd",
  "serde",
  "serde_json",
@@ -1187,12 +1188,15 @@ dependencies = [
  "tauri-plugin-process",
  "tauri-plugin-single-instance",
  "tauri-plugin-updater",
+ "tempfile",
  "tokio",
  "tokio-util",
  "tracing",
  "tracing-appender",
  "tracing-subscriber",
  "url",
+ "windows-sys 0.61.2",
+ "zip",
 ]
 
 [[package]]

apps/desktop/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "cursor-byok-desktop",
-  "version": "0.1.5",
+  "version": "0.1.6",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "cursor-byok-desktop",
-      "version": "0.1.5",
+      "version": "0.1.6",
       "license": "MIT",
       "dependencies": {
         "@floating-ui/dom": "^1.8.0",

apps/desktop/package.json

@@ -1,6 +1,6 @@
 {
   "name": "cursor-byok-desktop",
-  "version": "0.1.5",
+  "version": "0.1.6",
   "description": "Cursor BYOK desktop management application",
   "type": "module",
   "scripts": {
@@ -8,7 +8,7 @@
     "dev": "vite",
     "typecheck": "tsc --noEmit",
     "typecheck:node": "tsc --noEmit -p tsconfig.node.json",
-    "i18n:scan": "STATIC_I18N_SCAN=true vite build",
+    "i18n:scan": "cross-env STATIC_I18N_SCAN=true vite build",
     "build": "vite build",
     "build:demo": "npm run typecheck && npm run typecheck:node && vite build --config vite.demo.config.ts",
     "check": "npm run typecheck && npm run typecheck:node && npm run build",

apps/desktop/src-tauri/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "cursor-byok-desktop"
-version = "0.1.5"
+version = "0.1.6"
 edition = "2021"
 publish = false
 
@@ -14,6 +14,7 @@ tauri-build = { version = "2", features = [] }
 [dependencies]
 axum = "0.8"
 cursor-server = { path = "../../../server" }
+libc = "0.2"
 rfd = "0.15"
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
@@ -24,9 +25,14 @@ tauri-plugin-opener = "2"
 tauri-plugin-autostart = "2"
 tauri-plugin-process = "2"
 tauri-plugin-updater = "2"
+tempfile = "3"
 tokio = { version = "1", features = ["time"] }
 tokio-util = "0.7"
 tracing = "0.1"
 tracing-appender = "0.2"
 tracing-subscriber = { version = "0.3", features = ["env-filter"] }
 url = "2"
+zip = { version = "4", default-features = false, features = ["deflate"] }
+
+[target.'cfg(windows)'.dependencies]
+windows-sys = { version = "0.61", features = ["Win32_Foundation", "Win32_System_Threading"] }

apps/desktop/src-tauri/build.rs

@@ -1,5 +1,9 @@
 fn main() {
-    let manifest = tauri_build::AppManifest::new().commands(&["open_terminal_with_command"]);
+    let manifest = tauri_build::AppManifest::new().commands(&[
+        "open_terminal_with_command",
+        "check_portable_update",
+        "install_portable_update",
+    ]);
     tauri_build::try_build(tauri_build::Attributes::new().app_manifest(manifest))
         .expect("failed to build Tauri application")
 }

apps/desktop/src-tauri/capabilities/default.json

@@ -18,6 +18,8 @@
     "core:window:allow-close",
     "core:app:allow-set-dock-visibility",
     "allow-open-terminal-with-command",
+    "allow-check-portable-update",
+    "allow-install-portable-update",
     "clipboard-manager:allow-write-text",
     "autostart:default",
     "process:allow-restart",

apps/desktop/src-tauri/permissions/autogenerated/check_portable_update.toml

@@ -0,0 +1,11 @@
+# Automatically generated - DO NOT EDIT!
+
+[[permission]]
+identifier = "allow-check-portable-update"
+description = "Enables the check_portable_update command without any pre-configured scope."
+commands.allow = ["check_portable_update"]
+
+[[permission]]
+identifier = "deny-check-portable-update"
+description = "Denies the check_portable_update command without any pre-configured scope."
+commands.deny = ["check_portable_update"]

apps/desktop/src-tauri/permissions/autogenerated/install_portable_update.toml

@@ -0,0 +1,11 @@
+# Automatically generated - DO NOT EDIT!
+
+[[permission]]
+identifier = "allow-install-portable-update"
+description = "Enables the install_portable_update command without any pre-configured scope."
+commands.allow = ["install_portable_update"]
+
+[[permission]]
+identifier = "deny-install-portable-update"
+description = "Denies the install_portable_update command without any pre-configured scope."
+commands.deny = ["install_portable_update"]

apps/desktop/src-tauri/src/desktop.rs

@@ -149,6 +149,23 @@ pub fn run() -> ExitCode {
             return ExitCode::FAILURE;
         }
     };
+    #[cfg(unix)]
+    {
+        let open_file_limit = match crate::resource_limits::raise_open_file_limit() {
+            Ok(limit) => limit,
+            Err(error) => {
+                diagnostics.report_fatal(&error);
+                return ExitCode::FAILURE;
+            }
+        };
+        tracing::info!(
+            requested = crate::resource_limits::REQUESTED_OPEN_FILE_LIMIT,
+            previous = open_file_limit.previous,
+            effective = open_file_limit.effective,
+            hard = open_file_limit.hard,
+            "open file limit configured"
+        );
+    }
     tracing::info!(
         version = env!("CARGO_PKG_VERSION"),
         os = std::env::consts::OS,
@@ -160,7 +177,11 @@ pub fn run() -> ExitCode {
     let started_by_autostart = std::env::args_os().any(|arg| arg == AUTOSTART_ARG);
 
     let app = tauri::Builder::default()
-        .invoke_handler(tauri::generate_handler![open_terminal_with_command])
+        .invoke_handler(tauri::generate_handler![
+            open_terminal_with_command,
+            crate::update::check_portable_update,
+            crate::update::install_portable_update,
+        ])
         .plugin(tauri_plugin_single_instance::init(|app, args, _| {
             if !args.iter().any(|arg| arg == AUTOSTART_ARG) {
                 tray::show_main_window(app);
@@ -228,6 +249,7 @@ pub fn run() -> ExitCode {
                 window.set_focus()?;
             }
             tray::create(app)?;
+            crate::update::signal_ready_if_requested()?;
             Ok(())
         })
         .build(tauri::generate_context!());

apps/desktop/src-tauri/src/lib.rs

@@ -1,7 +1,14 @@
 mod desktop;
 #[cfg(not(dev))]
 mod frontend;
+mod resource_limits;
 mod startup;
 mod tray;
+mod update;
 
-pub use desktop::run;
+pub fn run() -> std::process::ExitCode {
+    if let Some(exit_code) = update::run_replacement_if_requested() {
+        return exit_code;
+    }
+    desktop::run()
+}

apps/desktop/src-tauri/src/resource_limits.rs

@@ -0,0 +1,56 @@
+//! Configures process resource limits before the desktop runtime starts.
+
+#[cfg(unix)]
+use std::io;
+
+#[cfg(unix)]
+pub(crate) const REQUESTED_OPEN_FILE_LIMIT: u64 = 65_536;
+
+#[cfg(unix)]
+pub(crate) struct OpenFileLimit {
+    pub(crate) previous: u64,
+    pub(crate) effective: u64,
+    pub(crate) hard: u64,
+}
+
+#[cfg(unix)]
+pub(crate) fn raise_open_file_limit() -> io::Result<OpenFileLimit> {
+    let mut limits = libc::rlimit {
+        rlim_cur: 0,
+        rlim_max: 0,
+    };
+    // SAFETY: `limits` points to writable memory for one `rlimit` value.
+    if unsafe { libc::getrlimit(libc::RLIMIT_NOFILE, &mut limits) } != 0 {
+        return Err(io::Error::last_os_error());
+    }
+
+    let previous = limits.rlim_cur;
+    let target = limits
+        .rlim_max
+        .min(REQUESTED_OPEN_FILE_LIMIT as libc::rlim_t);
+    if previous < target {
+        let requested = libc::rlimit {
+            rlim_cur: target,
+            rlim_max: limits.rlim_max,
+        };
+        // SAFETY: `requested` is a valid `rlimit` value and does not raise the hard limit.
+        if unsafe { libc::setrlimit(libc::RLIMIT_NOFILE, &requested) } != 0 {
+            return Err(io::Error::last_os_error());
+        }
+    }
+
+    let mut effective = libc::rlimit {
+        rlim_cur: 0,
+        rlim_max: 0,
+    };
+    // SAFETY: `effective` points to writable memory for one `rlimit` value.
+    if unsafe { libc::getrlimit(libc::RLIMIT_NOFILE, &mut effective) } != 0 {
+        return Err(io::Error::last_os_error());
+    }
+
+    Ok(OpenFileLimit {
+        previous: previous as u64,
+        effective: effective.rlim_cur as u64,
+        hard: effective.rlim_max as u64,
+    })
+}

apps/desktop/src-tauri/src/update/mod.rs

@@ -0,0 +1,269 @@
+use std::{
+    fs::{self, OpenOptions},
+    io::{Cursor, Read, Write},
+    path::{Path, PathBuf},
+    process::{Command, ExitCode},
+};
+
+use serde::Serialize;
+use tauri::AppHandle;
+
+#[cfg(target_os = "windows")]
+use tauri_plugin_updater::UpdaterExt;
+
+#[cfg(target_os = "windows")]
+mod replacement;
+
+const PORTABLE_UPDATE_ENDPOINT: &str =
+    "https://github.com/leookun/cursor-byok/releases/latest/download/portable-latest.json";
+const WINDOWS_PAYLOAD_NAME: &str = "cursor-byok-desktop.exe";
+
+#[derive(Serialize)]
+#[serde(rename_all = "camelCase")]
+pub(crate) struct PortableUpdateInfo {
+    version: String,
+}
+
+pub fn run_replacement_if_requested() -> Option<ExitCode> {
+    #[cfg(target_os = "windows")]
+    {
+        match replacement::request_from_args() {
+            Ok(Some(request)) => return Some(replacement::run(request)),
+            Ok(None) => {}
+            Err(error) => {
+                eprintln!("invalid portable update replacement request: {error}");
+                return Some(ExitCode::FAILURE);
+            }
+        }
+    }
+    None
+}
+
+pub(crate) fn signal_ready_if_requested() -> std::io::Result<()> {
+    #[cfg(target_os = "windows")]
+    if let Some(path) = replacement::ready_marker_from_args() {
+        if let Some(parent) = path.parent() {
+            fs::create_dir_all(parent)?;
+        }
+        fs::write(path, b"ready")?;
+    }
+    Ok(())
+}
+
+#[tauri::command]
+pub(crate) async fn check_portable_update(
+    app: AppHandle,
+) -> Result<Option<PortableUpdateInfo>, String> {
+    #[cfg(target_os = "windows")]
+    {
+        let update = portable_update(&app).await?;
+        return Ok(update.map(|update| PortableUpdateInfo {
+            version: update.version,
+        }));
+    }
+    #[cfg(not(target_os = "windows"))]
+    {
+        let _ = app;
+        Err("portable updates are only supported on Windows".into())
+    }
+}
+
+#[tauri::command]
+pub(crate) async fn install_portable_update(
+    app: AppHandle,
+    expected_version: String,
+) -> Result<(), String> {
+    #[cfg(target_os = "windows")]
+    {
+        let update = portable_update(&app)
+            .await?
+            .ok_or_else(|| "the selected update is no longer available".to_string())?;
+        if update.version != expected_version {
+            return Err(format!(
+                "available update changed from {expected_version} to {}",
+                update.version
+            ));
+        }
+
+        let target = std::env::current_exe()
+            .map_err(|error| format!("failed to locate the running executable: {error}"))?;
+        ensure_target_writable(&target)
+            .map_err(|error| format!("the application directory is not writable: {error}"))?;
+
+        let bytes = update
+            .download(|_, _| {}, || {})
+            .await
+            .map_err(|error| format!("failed to download or verify the update: {error}"))?;
+        let payload = extract_windows_payload(&bytes)
+            .map_err(|error| format!("invalid Windows update archive: {error}"))?;
+        let staged = stage_payload(&target, &payload)
+            .map_err(|error| format!("failed to stage the update: {error}"))?;
+
+        let handshake = staged.with_extension("started");
+        let _ = fs::remove_file(&handshake);
+        let mut replacement = Command::new(&staged)
+            .arg("--apply-portable-update")
+            .arg("--update-target")
+            .arg(&target)
+            .arg("--update-wait-pid")
+            .arg(std::process::id().to_string())
+            .arg("--update-handshake")
+            .arg(&handshake)
+            .spawn()
+            .map_err(|error| format!("failed to start the update replacement process: {error}"))?;
+        wait_for_replacement_start(&mut replacement, &handshake).await?;
+
+        app.exit(0);
+        Ok(())
+    }
+    #[cfg(not(target_os = "windows"))]
+    {
+        let _ = (app, expected_version);
+        Err("portable updates are only supported on Windows".into())
+    }
+}
+
+#[cfg(target_os = "windows")]
+async fn portable_update(app: &AppHandle) -> Result<Option<tauri_plugin_updater::Update>, String> {
+    let endpoint = PORTABLE_UPDATE_ENDPOINT
+        .parse()
+        .map_err(|error| format!("invalid portable update endpoint: {error}"))?;
+    let updater = app
+        .updater_builder()
+        .endpoints(vec![endpoint])
+        .map_err(|error| format!("failed to configure the updater: {error}"))?
+        .build()
+        .map_err(|error| format!("failed to initialize the updater: {error}"))?;
+    updater
+        .check()
+        .await
+        .map_err(|error| format!("failed to check for updates: {error}"))
+}
+
+#[cfg(target_os = "windows")]
+async fn wait_for_replacement_start(
+    child: &mut std::process::Child,
+    handshake: &Path,
+) -> Result<(), String> {
+    let wait = async {
+        loop {
+            if handshake.is_file() {
+                return Ok(());
+            }
+            if let Some(status) = child
+                .try_wait()
+                .map_err(|error| format!("failed to inspect replacement process: {error}"))?
+            {
+                return Err(format!(
+                    "update replacement process exited before it was ready: {status}"
+                ));
+            }
+            tokio::time::sleep(std::time::Duration::from_millis(50)).await;
+        }
+    };
+    let result = match tokio::time::timeout(std::time::Duration::from_secs(5), wait).await {
+        Ok(result) => result,
+        Err(_) => {
+            let _ = child.kill();
+            let _ = child.wait();
+            return Err("update replacement process did not become ready".into());
+        }
+    };
+    if result.is_err() {
+        let _ = child.kill();
+        let _ = child.wait();
+    }
+    result
+}
+
+fn ensure_target_writable(target: &Path) -> std::io::Result<()> {
+    let parent = target
+        .parent()
+        .ok_or_else(|| std::io::Error::other("application executable has no parent directory"))?;
+    let probe = parent.join(format!(
+        ".cursor-byok-update-write-test-{}",
+        std::process::id()
+    ));
+    let mut file = OpenOptions::new()
+        .write(true)
+        .create_new(true)
+        .open(&probe)?;
+    file.write_all(b"test")?;
+    drop(file);
+    fs::remove_file(probe)
+}
+
+fn extract_windows_payload(bytes: &[u8]) -> Result<Vec<u8>, String> {
+    let mut archive = zip::ZipArchive::new(Cursor::new(bytes))
+        .map_err(|error| format!("failed to open ZIP: {error}"))?;
+    if archive.len() != 1 {
+        return Err("archive must contain exactly one file".into());
+    }
+    let mut entry = archive
+        .by_index(0)
+        .map_err(|error| format!("failed to read ZIP entry: {error}"))?;
+    let name = Path::new(entry.name())
+        .file_name()
+        .and_then(|name| name.to_str())
+        .ok_or_else(|| "archive entry has an invalid file name".to_string())?;
+    if name != WINDOWS_PAYLOAD_NAME || entry.is_dir() {
+        return Err(format!(
+            "expected {WINDOWS_PAYLOAD_NAME}, found {}",
+            entry.name()
+        ));
+    }
+    let mut payload = Vec::with_capacity(entry.size() as usize);
+    entry
+        .read_to_end(&mut payload)
+        .map_err(|error| format!("failed to extract executable: {error}"))?;
+    if payload.len() < 2 || &payload[..2] != b"MZ" {
+        return Err("payload is not a Windows executable".into());
+    }
+    Ok(payload)
+}
+
+fn stage_payload(target: &Path, payload: &[u8]) -> std::io::Result<PathBuf> {
+    let directory = tempfile::Builder::new()
+        .prefix("cursor-byok-portable-update-")
+        .tempdir()?;
+    let name = target
+        .file_name()
+        .ok_or_else(|| std::io::Error::other("application executable has no file name"))?;
+    let path = directory.path().join(name);
+    fs::write(&path, payload)?;
+    let _ = directory.keep();
+    Ok(path)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::io::Write;
+
+    fn update_zip(name: &str, payload: &[u8]) -> Vec<u8> {
+        let mut bytes = Cursor::new(Vec::new());
+        {
+            let mut archive = zip::ZipWriter::new(&mut bytes);
+            archive
+                .start_file(name, zip::write::SimpleFileOptions::default())
+                .unwrap();
+            archive.write_all(payload).unwrap();
+            archive.finish().unwrap();
+        }
+        bytes.into_inner()
+    }
+
+    #[test]
+    fn extracts_the_single_expected_windows_executable() {
+        let bytes = update_zip(WINDOWS_PAYLOAD_NAME, b"MZpayload");
+        assert_eq!(extract_windows_payload(&bytes).unwrap(), b"MZpayload");
+    }
+
+    #[test]
+    fn rejects_unexpected_or_non_executable_payloads() {
+        let wrong_name = update_zip("other.exe", b"MZpayload");
+        assert!(extract_windows_payload(&wrong_name).is_err());
+        let wrong_content = update_zip(WINDOWS_PAYLOAD_NAME, b"not an executable");
+        assert!(extract_windows_payload(&wrong_content).is_err());
+    }
+}

apps/desktop/src-tauri/src/update/replacement.rs

@@ -0,0 +1,296 @@
+use std::{
+    ffi::{OsStr, OsString},
+    fs, io,
+    path::{Path, PathBuf},
+    process::{Child, Command, ExitCode},
+    thread,
+    time::{Duration, Instant},
+};
+
+const APPLY_ARG: &str = "--apply-portable-update";
+const TARGET_ARG: &str = "--update-target";
+const PID_ARG: &str = "--update-wait-pid";
+const HANDSHAKE_ARG: &str = "--update-handshake";
+pub(super) const READY_ARG: &str = "--portable-update-ready";
+const PROCESS_WAIT_TIMEOUT: Duration = Duration::from_secs(30);
+const READY_WAIT_TIMEOUT: Duration = Duration::from_secs(30);
+
+pub(super) struct ReplacementRequest {
+    target: PathBuf,
+    pid: u32,
+    handshake: PathBuf,
+}
+
+pub(super) fn request_from_args() -> Result<Option<ReplacementRequest>, String> {
+    let args = std::env::args_os().collect::<Vec<_>>();
+    if !args.iter().any(|arg| arg == APPLY_ARG) {
+        return Ok(None);
+    }
+    let target = PathBuf::from(
+        argument_value(&args, TARGET_ARG).ok_or_else(|| format!("{TARGET_ARG} is required"))?,
+    );
+    let pid = argument_value(&args, PID_ARG)
+        .ok_or_else(|| format!("{PID_ARG} is required"))?
+        .to_string_lossy()
+        .parse::<u32>()
+        .map_err(|error| format!("invalid {PID_ARG}: {error}"))?;
+    let handshake = PathBuf::from(
+        argument_value(&args, HANDSHAKE_ARG)
+            .ok_or_else(|| format!("{HANDSHAKE_ARG} is required"))?,
+    );
+    validate_target(&target).map_err(|error| error.to_string())?;
+    Ok(Some(ReplacementRequest {
+        target,
+        pid,
+        handshake,
+    }))
+}
+
+pub(super) fn ready_marker_from_args() -> Option<PathBuf> {
+    let args = std::env::args_os().collect::<Vec<_>>();
+    argument_value(&args, READY_ARG).map(PathBuf::from)
+}
+
+fn argument_value<'a>(args: &'a [OsString], name: &str) -> Option<&'a OsStr> {
+    args.iter()
+        .position(|arg| arg == name)
+        .and_then(|index| args.get(index + 1))
+        .map(OsString::as_os_str)
+}
+
+fn validate_target(target: &Path) -> io::Result<()> {
+    if !target.is_absolute() || !target.is_file() {
+        return Err(io::Error::new(
+            io::ErrorKind::InvalidInput,
+            "update target must be an existing absolute file",
+        ));
+    }
+    let source_name = std::env::current_exe()?
+        .file_name()
+        .map(OsStr::to_os_string)
+        .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidInput, "updater has no file name"))?;
+    if target.file_name() != Some(source_name.as_os_str()) {
+        return Err(io::Error::new(
+            io::ErrorKind::InvalidInput,
+            "update target file name does not match the updater",
+        ));
+    }
+    Ok(())
+}
+
+pub(super) fn run(request: ReplacementRequest) -> ExitCode {
+    match run_inner(request) {
+        Ok(()) => ExitCode::SUCCESS,
+        Err(error) => {
+            eprintln!("portable update replacement failed: {error}");
+            ExitCode::FAILURE
+        }
+    }
+}
+
+fn run_inner(request: ReplacementRequest) -> io::Result<()> {
+    wait_for_process(request.pid, PROCESS_WAIT_TIMEOUT, &request.handshake)?;
+    remove_file_if_exists(&request.handshake)?;
+    let source = std::env::current_exe()?;
+    let backup = backup_path(&request.target);
+    let ready = source.with_extension("ready");
+    remove_file_if_exists(&ready)?;
+
+    if let Err(error) = install_staged(&source, &request.target, &backup) {
+        relaunch(&request.target);
+        return Err(error);
+    }
+    let mut child = match Command::new(&request.target)
+        .arg(READY_ARG)
+        .arg(&ready)
+        .spawn()
+    {
+        Ok(child) => child,
+        Err(error) => {
+            restore_backup(&request.target, &backup)?;
+            relaunch(&request.target);
+            return Err(error);
+        }
+    };
+
+    match wait_until_ready(&mut child, &ready, READY_WAIT_TIMEOUT) {
+        Ok(()) => {
+            let _ = fs::remove_file(&backup);
+            let _ = fs::remove_file(&ready);
+            Ok(())
+        }
+        Err(error) => {
+            let _ = child.kill();
+            let _ = child.wait();
+            restore_backup(&request.target, &backup)?;
+            relaunch(&request.target);
+            Err(error)
+        }
+    }
+}
+
+fn relaunch(target: &Path) {
+    if target.is_file() {
+        let _ = Command::new(target).spawn();
+    }
+}
+
+fn remove_file_if_exists(path: &Path) -> io::Result<()> {
+    match fs::remove_file(path) {
+        Ok(()) => Ok(()),
+        Err(error) if error.kind() == io::ErrorKind::NotFound => Ok(()),
+        Err(error) => Err(error),
+    }
+}
+
+fn backup_path(target: &Path) -> PathBuf {
+    path_with_suffix(target, ".old")
+}
+
+fn pending_path(target: &Path) -> PathBuf {
+    path_with_suffix(target, ".new")
+}
+
+fn path_with_suffix(target: &Path, suffix: &str) -> PathBuf {
+    let mut name = target.as_os_str().to_os_string();
+    name.push(suffix);
+    PathBuf::from(name)
+}
+
+fn install_staged(source: &Path, target: &Path, backup: &Path) -> io::Result<()> {
+    let pending = pending_path(target);
+    remove_file_if_exists(&pending)?;
+    fs::copy(source, &pending)?;
+
+    let result = activate_pending(&pending, target, backup);
+    if result.is_err() {
+        let _ = fs::remove_file(&pending);
+    }
+    result
+}
+
+fn activate_pending(pending: &Path, target: &Path, backup: &Path) -> io::Result<()> {
+    remove_file_if_exists(backup)?;
+    fs::rename(target, backup)?;
+    if let Err(error) = fs::rename(pending, target) {
+        if let Err(restore_error) = restore_backup(target, backup) {
+            return Err(io::Error::other(format!(
+                "failed to install update ({error}) and restore the original executable ({restore_error})"
+            )));
+        }
+        return Err(error);
+    }
+    Ok(())
+}
+
+fn restore_backup(target: &Path, backup: &Path) -> io::Result<()> {
+    remove_file_if_exists(target)?;
+    fs::rename(backup, target)
+}
+
+fn wait_until_ready(child: &mut Child, marker: &Path, timeout: Duration) -> io::Result<()> {
+    let deadline = Instant::now() + timeout;
+    loop {
+        if marker.is_file() {
+            return Ok(());
+        }
+        if let Some(status) = child.try_wait()? {
+            return Err(io::Error::other(format!(
+                "updated application exited before startup completed: {status}"
+            )));
+        }
+        if Instant::now() >= deadline {
+            return Err(io::Error::new(
+                io::ErrorKind::TimedOut,
+                "updated application did not report a successful startup",
+            ));
+        }
+        thread::sleep(Duration::from_millis(100));
+    }
+}
+
+#[cfg(windows)]
+fn wait_for_process(pid: u32, timeout: Duration, handshake: &Path) -> io::Result<()> {
+    use windows_sys::Win32::{
+        Foundation::{
+            CloseHandle, GetLastError, ERROR_INVALID_PARAMETER, WAIT_OBJECT_0, WAIT_TIMEOUT,
+        },
+        System::Threading::{OpenProcess, WaitForSingleObject},
+    };
+
+    const SYNCHRONIZE_ACCESS: u32 = 0x0010_0000;
+    let handle = unsafe { OpenProcess(SYNCHRONIZE_ACCESS, 0, pid) };
+    if handle.is_null() {
+        let error = unsafe { GetLastError() };
+        return if error == ERROR_INVALID_PARAMETER {
+            fs::write(handshake, b"started")
+        } else {
+            Err(io::Error::from_raw_os_error(error as i32))
+        };
+    }
+    if let Err(error) = fs::write(handshake, b"started") {
+        unsafe { CloseHandle(handle) };
+        return Err(error);
+    }
+    let milliseconds = timeout.as_millis().min(u32::MAX as u128) as u32;
+    let result = unsafe { WaitForSingleObject(handle, milliseconds) };
+    unsafe { CloseHandle(handle) };
+    match result {
+        WAIT_OBJECT_0 => Ok(()),
+        WAIT_TIMEOUT => Err(io::Error::new(
+            io::ErrorKind::TimedOut,
+            "running application did not exit before the update timeout",
+        )),
+        _ => Err(io::Error::last_os_error()),
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn staged_file_can_be_restored() {
+        let directory = tempfile::tempdir().unwrap();
+        let source = directory.path().join("source.exe");
+        let target = directory.path().join("target.exe");
+        let backup = backup_path(&target);
+        fs::write(&source, b"new").unwrap();
+        fs::write(&target, b"old").unwrap();
+
+        install_staged(&source, &target, &backup).unwrap();
+        assert_eq!(fs::read(&target).unwrap(), b"new");
+        assert_eq!(fs::read(&backup).unwrap(), b"old");
+
+        restore_backup(&target, &backup).unwrap();
+        assert_eq!(fs::read(&target).unwrap(), b"old");
+        assert!(!backup.exists());
+    }
+
+    #[test]
+    fn missing_pending_file_restores_original_after_backup() {
+        let directory = tempfile::tempdir().unwrap();
+        let target = directory.path().join("target.exe");
+        let pending = pending_path(&target);
+        let backup = backup_path(&target);
+        fs::write(&target, b"old").unwrap();
+
+        assert!(activate_pending(&pending, &target, &backup).is_err());
+        assert_eq!(fs::read(&target).unwrap(), b"old");
+        assert!(!backup.exists());
+    }
+
+    #[test]
+    fn missing_source_preserves_original_file() {
+        let directory = tempfile::tempdir().unwrap();
+        let source = directory.path().join("missing.exe");
+        let target = directory.path().join("target.exe");
+        let backup = backup_path(&target);
+        fs::write(&target, b"old").unwrap();
+
+        assert!(install_staged(&source, &target, &backup).is_err());
+        assert_eq!(fs::read(&target).unwrap(), b"old");
+        assert!(!backup.exists());
+        assert!(!pending_path(&target).exists());
+    }
+}

apps/desktop/src-tauri/tauri.conf.json

@@ -1,7 +1,7 @@
 {
   "$schema": "https://schema.tauri.app/config/2",
   "productName": "Cursor BYOK",
-  "version": "0.1.5",
+  "version": "0.1.6",
   "identifier": "dev.cursorbyok.desktop",
   "build": {
     "beforeDevCommand": "npm run dev",
@@ -12,7 +12,7 @@
   "app": {
     "windows": [],
     "security": {
-      "csp": "default-src 'self'; img-src 'self' asset: http://asset.localhost data:; style-src 'self' 'unsafe-inline'; connect-src 'self' http://127.0.0.1:*",
+      "csp": "default-src 'self'; img-src 'self' asset: http://asset.localhost data: https:; style-src 'self' 'unsafe-inline'; connect-src 'self' http://127.0.0.1:*",
       "dangerousDisableAssetCspModification": [
         "style-src"
       ]

apps/desktop/src/features/settings/AppLifecycleSettingsCard.tsx

@@ -95,7 +95,8 @@ export function AppLifecycleSettingsCard() {
       const nextVersion = await updateStore.check();
       message(nextVersion ? t("发现新版本 {version}", { version: nextVersion }) : t("当前已是最新版本"));
     } catch (cause) {
-      message(cause instanceof Error ? cause.message : String(cause));
+      const error = cause instanceof Error ? cause.message : String(cause);
+      message(t("检查更新失败：{error}", { error }));
     }
   };
 
@@ -103,7 +104,8 @@ export function AppLifecycleSettingsCard() {
     try {
       await updateStore.install();
     } catch (cause) {
-      message(cause instanceof Error ? cause.message : String(cause));
+      const error = cause instanceof Error ? cause.message : String(cause);
+      message(t("安装更新失败：{error}", { error }));
     }
   };
 

apps/desktop/src/i18n/generated/catalog.json

@@ -606,7 +606,7 @@
       "refs": [
         {
           "file": "features/settings/AppLifecycleSettingsCard.tsx",
-          "line": 156,
+          "line": 158,
           "column": 27
         }
       ]
@@ -916,7 +916,7 @@
       "refs": [
         {
           "file": "features/settings/AppLifecycleSettingsCard.tsx",
-          "line": 152,
+          "line": 154,
           "column": 13
         }
       ]
@@ -1673,7 +1673,7 @@
       "refs": [
         {
           "file": "features/settings/AppLifecycleSettingsCard.tsx",
-          "line": 126,
+          "line": 128,
           "column": 17
         }
       ]
@@ -1725,7 +1725,7 @@
       "refs": [
         {
           "file": "features/settings/AppLifecycleSettingsCard.tsx",
-          "line": 151,
+          "line": 153,
           "column": 13
         }
       ]
@@ -2353,7 +2353,7 @@
       "refs": [
         {
           "file": "shared/api.ts",
-          "line": 488,
+          "line": 486,
           "column": 43
         }
       ]
@@ -2365,7 +2365,7 @@
       "refs": [
         {
           "file": "features/settings/AppLifecycleSettingsCard.tsx",
-          "line": 149,
+          "line": 151,
           "column": 18
         }
       ]
@@ -2451,12 +2451,12 @@
       "refs": [
         {
           "file": "features/settings/AppLifecycleSettingsCard.tsx",
-          "line": 137,
+          "line": 139,
           "column": 18
         },
         {
           "file": "features/settings/AppLifecycleSettingsCard.tsx",
-          "line": 143,
+          "line": 145,
           "column": 16
         }
       ]
@@ -2680,7 +2680,7 @@
       "refs": [
         {
           "file": "shared/api.ts",
-          "line": 483,
+          "line": 481,
           "column": 43
         }
       ]
@@ -2795,7 +2795,7 @@
       "refs": [
         {
           "file": "features/settings/AppLifecycleSettingsCard.tsx",
-          "line": 160,
+          "line": 162,
           "column": 37
         }
       ]
@@ -2834,7 +2834,7 @@
       "refs": [
         {
           "file": "shared/api.ts",
-          "line": 420,
+          "line": 418,
           "column": 21
         }
       ]
@@ -2900,12 +2900,12 @@
       "refs": [
         {
           "file": "features/settings/AppLifecycleSettingsCard.tsx",
-          "line": 113,
+          "line": 115,
           "column": 18
         },
         {
           "file": "features/settings/AppLifecycleSettingsCard.tsx",
-          "line": 119,
+          "line": 121,
           "column": 16
         }
       ]
@@ -3194,6 +3194,20 @@
         }
       ]
     },
+    "92e26b27d5ea8f0e": {
+      "source": "检查更新失败：{error}",
+      "kind": "template",
+      "placeholders": [
+        "error"
+      ],
+      "refs": [
+        {
+          "file": "features/settings/AppLifecycleSettingsCard.tsx",
+          "line": 99,
+          "column": 15
+        }
+      ]
+    },
     "940a168911ade998": {
       "source": "每页条数",
       "kind": "text",
@@ -3520,7 +3534,7 @@
       "refs": [
         {
           "file": "features/settings/AppLifecycleSettingsCard.tsx",
-          "line": 110,
+          "line": 112,
           "column": 29
         }
       ]
@@ -3532,12 +3546,12 @@
       "refs": [
         {
           "file": "features/settings/AppLifecycleSettingsCard.tsx",
-          "line": 125,
+          "line": 127,
           "column": 18
         },
         {
           "file": "features/settings/AppLifecycleSettingsCard.tsx",
-          "line": 131,
+          "line": 133,
           "column": 16
         }
       ]
@@ -3749,6 +3763,20 @@
         }
       ]
     },
+    "a80b53f8848e6d27": {
+      "source": "安装更新失败：{error}",
+      "kind": "template",
+      "placeholders": [
+        "error"
+      ],
+      "refs": [
+        {
+          "file": "features/settings/AppLifecycleSettingsCard.tsx",
+          "line": 108,
+          "column": 15
+        }
+      ]
+    },
     "a98585871c5313ff": {
       "source": "显示名称",
       "kind": "text",
@@ -3814,7 +3842,7 @@
       "refs": [
         {
           "file": "features/settings/AppLifecycleSettingsCard.tsx",
-          "line": 156,
+          "line": 158,
           "column": 39
         }
       ]
@@ -3854,7 +3882,7 @@
       "refs": [
         {
           "file": "features/settings/AppLifecycleSettingsCard.tsx",
-          "line": 138,
+          "line": 140,
           "column": 17
         }
       ]
@@ -4577,7 +4605,7 @@
       "refs": [
         {
           "file": "features/settings/AppLifecycleSettingsCard.tsx",
-          "line": 114,
+          "line": 116,
           "column": 17
         }
       ]
@@ -5522,7 +5550,7 @@
         },
         {
           "file": "features/settings/AppLifecycleSettingsCard.tsx",
-          "line": 160,
+          "line": 162,
           "column": 25
         }
       ]

apps/desktop/src/i18n/locales/en-US.json

@@ -221,6 +221,7 @@
   "91aaf184cfc17ffd": "Overview",
   "91af6e57e7453fbe": "Add account",
   "92156a483d4ba248": "Only request, response, and trace attachments are deleted; call summaries, metrics, and configuration are kept.",
+  "92e26b27d5ea8f0e": "Failed to check for updates: {error}",
   "940a168911ade998": "Items per page",
   "945fb1c67eca8493": "Installing the plugin runtime",
   "946b3ffc02f026c0": "Delete this model?",
@@ -259,6 +260,7 @@
   "a748cc074f78de00": "View details",
   "a7617f42f898b2bf": "Use complete request URL",
   "a8036485f9227f2c": "Drag to reorder",
+  "a80b53f8848e6d27": "Failed to install update: {error}",
   "a98585871c5313ff": "Display name",
   "ab9084a640fbb864": "Deselect all",
   "abecab6701177721": "Launch at login enabled",

apps/desktop/src/i18n/locales/zh-CN.json

@@ -221,6 +221,7 @@
   "91aaf184cfc17ffd": "数据概览",
   "91af6e57e7453fbe": "添加账号",
   "92156a483d4ba248": "仅删除请求、响应和追踪附件等详细内容，保留调用汇总、统计指标和配置。",
+  "92e26b27d5ea8f0e": "检查更新失败：{error}",
   "940a168911ade998": "每页条数",
   "945fb1c67eca8493": "正在安装插件运行时",
   "946b3ffc02f026c0": "确定删除这个模型吗？",
@@ -259,6 +260,7 @@
   "a748cc074f78de00": "查看详情",
   "a7617f42f898b2bf": "使用完整请求地址",
   "a8036485f9227f2c": "拖动排序",
+  "a80b53f8848e6d27": "安装更新失败：{error}",
   "a98585871c5313ff": "显示名称",
   "ab9084a640fbb864": "全不选",
   "abecab6701177721": "已开启开机启动",

apps/desktop/src/shared/api.ts

@@ -235,9 +235,7 @@ export interface PluginModelDescriptor {
   description: string | null;
   icon: string;
   providerType: string;
-  contextWindowTokens: number | null;
   maxOutputTokens: number | null;
-  thinking: boolean;
   images: boolean;
 }
 

apps/desktop/src/shared/native/appLifecycle.ts

@@ -1,5 +1,5 @@
 import { getVersion, setDockVisibility } from "@tauri-apps/api/app";
-import { isTauri } from "@tauri-apps/api/core";
+import { invoke, isTauri } from "@tauri-apps/api/core";
 import { disable, enable, isEnabled } from "@tauri-apps/plugin-autostart";
 import { relaunch } from "@tauri-apps/plugin-process";
 import { check, type Update } from "@tauri-apps/plugin-updater";
@@ -46,11 +46,39 @@ export async function writeDockIconVisibility(visible: boolean): Promise<void> {
   }
 }
 
-export async function checkForUpdate(): Promise<Update | null> {
-  return check();
+export type AppUpdate = {
+  version: string;
+  install(): Promise<void>;
+  close(): Promise<void>;
+};
+
+type PortableUpdateInfo = {
+  version: string;
+};
+
+export async function checkForUpdate(): Promise<AppUpdate | null> {
+  if (desktopPlatform() === "windows") {
+    const update = await invoke<PortableUpdateInfo | null>("check_portable_update");
+    if (!update) return null;
+    return {
+      version: update.version,
+      install: () => invoke("install_portable_update", { expectedVersion: update.version }),
+      close: async () => {},
+    };
+  }
+
+  const update: Update | null = await check();
+  if (!update) return null;
+  return {
+    version: update.version,
+    install: async () => {
+      await update.downloadAndInstall();
+      await relaunch();
+    },
+    close: () => update.close(),
+  };
 }
 
-export async function installUpdate(update: Update): Promise<void> {
-  await update.downloadAndInstall();
-  await relaunch();
+export async function installUpdate(update: AppUpdate): Promise<void> {
+  await update.install();
 }

apps/desktop/src/shared/store/updateStore.ts

@@ -1,9 +1,9 @@
 import { useSyncExternalStore } from "react";
-import type { Update } from "@tauri-apps/plugin-updater";
 import {
   checkForUpdate,
   hasNativeAppLifecycle,
   installUpdate,
+  type AppUpdate,
 } from "../native/appLifecycle";
 
 export type UpdateSnapshot = {
@@ -17,7 +17,7 @@ let snapshot: UpdateSnapshot = {
   checking: false,
   installing: false,
 };
-let availableUpdate: Update | null = null;
+let availableUpdate: AppUpdate | null = null;
 let pendingCheck: Promise<string | null> | null = null;
 const listeners = new Set<() => void>();
 
@@ -26,7 +26,7 @@ function update(patch: Partial<UpdateSnapshot>) {
   listeners.forEach((listener) => listener());
 }
 
-async function replaceAvailableUpdate(next: Update | null) {
+async function replaceAvailableUpdate(next: AppUpdate | null) {
   const previous = availableUpdate;
   availableUpdate = next;
   update({ availableVersion: next?.version ?? null });

cursor.md

@@ -1,9 +1,4 @@
-以下是重构后完整目标版本
-实现时，先创建所有目录和文件固化，每个文件头部都写好注释再实现
-旧服务已被备份为server_backup,/Users/leokun/Documents/cursor-byok/server 目录已创建
-行数均为目标估算，使用 `≈` 标记；不包含测试、生成代码和空行。
-实现时可做略微调整，测试要求相对于目标文件旁边的独立文件，禁止码内测试
-本文档目录 /Users/leokun/Documents/cursor-byok/cursor.md
+
 ## 完整目录
 
 ```text
@@ -656,54 +651,7 @@ store     ─X→ cursor
 model     ─X→ cursor
 ```
 
-## 当前代码迁移
 
-```text
-当前                                      目标
-
-cursor/bidi_append.rs                  → api/cursor/bidi.rs
-cursor/run_sse.rs                      → api/cursor/run_sse.rs
-cursor/handlers.rs                     → api/cursor/handlers.rs
-cursor/proxy.rs                        → api/cursor/proxy.rs
-
-cursor/sessions.rs                     → cursor/transport/registry.rs
-                                       + cursor/transport/handle.rs
-                                       + cursor/transport/output.rs
-
-cursor/inbox.rs                        → cursor/transport/inbox.rs
-
-cursor/actor.rs                        → cursor/transport/
-                                       + cursor/conversation/runtime.rs
-                                       + cursor/conversation/delivery.rs
-
-cursor/session.rs                      → cursor/conversation/runtime.rs
-                                       + cursor/conversation/output.rs
-                                       + cursor/checkpoint/
-                                       + cursor/tools/
-
-cursor/request/prepare.rs              → cursor/compile/run.rs
-cursor/request/context.rs              → cursor/compile/context.rs
-cursor/request/background.rs           → cursor/compile/insert_messages.rs
-cursor/request/runtime.rs              → cursor/compile/break_messages.rs
-cursor/request/images.rs               → cursor/compile/images.rs
-cursor/request/model.rs                → cursor/compile/model.rs
-
-cursor/interaction/mod.rs              → cursor/protocol/events.rs
-cursor/interaction/query.rs            → cursor/tools/codec/query.rs
-cursor/interaction/render.rs           → cursor/tools/codec/render.rs
-
-cursor/projection/decode.rs            → cursor/checkpoint/messages/decode.rs
-cursor/projection/encode.rs            → cursor/checkpoint/messages/encode.rs
-cursor/projection/tests.rs             → cursor/checkpoint/messages/tests.rs
-
-cursor/presentation.rs                 → cursor/checkpoint/steps.rs
-
-run/runtime.rs RunRegistry             → cursor/conversation/registry.rs
-run/runtime.rs RunActor                → run/engine.rs + run/handle.rs
-run/port.rs                            → run/command.rs + run/event.rs + run/port.rs
-
-store/revisions.rs                     → store/checkpoints.rs
-```
 
 
 ## 最终核心
@@ -719,4 +667,4 @@ Bidi
   → Checkpoint
   → Transport
   → RunSSE
-```
+```

server/migrations/0009_add_tool_call_argument_error.sql

@@ -0,0 +1 @@
+ALTER TABLE tool_round_calls ADD COLUMN argument_error TEXT;

server/plugins/build-in/codex-auth/codex_test.ts

@@ -8,6 +8,7 @@ import type { LlmRequest, ModelEvent } from "cursor-byok:provider";
 import type { ResourceSnapshot } from "cursor-byok:resource";
 import { codexDeviceOAuth } from "./oauth.ts";
 import { parseOfficialModels } from "./models.ts";
+import { buildResponsesBody } from "cursor-byok:protocol/openai-responses";
 import { codexProvider, isQuotaError } from "./provider.ts";
 import {
   accountIdentity,
@@ -164,7 +165,10 @@ Deno.test("official model discovery excludes hidden models and puts the default
         display_name: "GPT First",
         supported_in_api: true,
         visibility: "list",
-        supported_reasoning_efforts: ["low", "medium"],
+        supported_reasoning_levels: [
+          { effort: "low", description: "Fast responses" },
+          { effort: "medium", description: "Balanced" },
+        ],
       },
       { slug: "gpt-second", supported_in_api: true, visibility: "list" },
       { slug: "gpt-hidden", supported_in_api: true, visibility: "hidden" },
@@ -172,7 +176,7 @@ Deno.test("official model discovery excludes hidden models and puts the default
     ],
   });
   assertEquals(models.map((model) => model.id), ["gpt-second", "gpt-first"]);
-  assertEquals(models[1].capabilities, { thinking: true, images: true });
+  assertEquals(models[1].capabilities, { images: true });
   assertEquals(models[1].privateData, { reasoningEfforts: ["low", "medium"] });
 });
 
@@ -302,6 +306,43 @@ Deno.test("invoke streams normalized events from the Codex Responses API", async
   ]);
 });
 
+Deno.test("reasoning replay projects response items to valid input items", () => {
+  const replayRequest = request();
+  replayRequest.messages = [{
+    role: "assistant",
+    text: "",
+    thinking: "",
+    replayState: {
+      providerKind: "openai_responses",
+      value: {
+        items: [{
+          type: "reasoning",
+          id: "item-1",
+          status: "completed",
+          summary: [{ type: "summary_text", text: "why" }],
+          content: [],
+          encrypted_content: "opaque",
+          output_only: true,
+        }],
+      },
+    },
+    toolCalls: [],
+  }];
+
+  const body = buildResponsesBody({
+    url: "https://example.com/responses",
+    model: "gpt-test",
+    request: replayRequest,
+  });
+  assertEquals(body.input, [{
+    type: "reasoning",
+    id: "item-1",
+    summary: [{ type: "summary_text", text: "why" }],
+    content: [],
+    encrypted_content: "opaque",
+  }]);
+});
+
 Deno.test("invoke streams incremental tool calls and replays reasoning items", async () => {
   const token = jwt({ "https://api.openai.com/auth": { chatgpt_account_id: "acct-1" } });
   const draft = await credentialDraft({

server/plugins/build-in/codex-auth/models.ts

@@ -24,7 +24,11 @@ function positiveInteger(value: unknown): number | null {
 }
 
 function parseReasoningEfforts(model: Record<string, unknown>): string[] {
-  const source = model.supported_reasoning_efforts ??
+  const source = model.supported_reasoning_levels ??
+    model.supportedReasoningLevels ??
+    model.reasoning_levels ??
+    model.reasoningLevels ??
+    model.supported_reasoning_efforts ??
     model.supportedReasoningEfforts ??
     model.reasoning_efforts ??
     model.reasoningEfforts;
@@ -61,10 +65,6 @@ export function parseOfficialModels(body: unknown): ModelDefinition[] {
     seen.add(id);
     const efforts = parseReasoningEfforts(model);
     const description = text(model.description);
-    const contextWindowTokens = positiveInteger(
-      model.context_window_tokens ?? model.contextWindowTokens ?? model.context_window ??
-        model.contextWindow,
-    );
     const maxOutputTokens = positiveInteger(
       model.max_output_tokens ?? model.maxOutputTokens ?? model.max_completion_tokens ??
         model.maxCompletionTokens,
@@ -74,9 +74,8 @@ export function parseOfficialModels(body: unknown): ModelDefinition[] {
       displayName: text(model.display_name ?? model.displayName ?? model.title ?? model.name) ??
         id,
       ...(description ? { description } : {}),
-      ...(contextWindowTokens !== null ? { contextWindowTokens } : {}),
       ...(maxOutputTokens !== null ? { maxOutputTokens } : {}),
-      capabilities: { thinking: efforts.length > 0, images: true },
+      capabilities: { images: true },
       privateData: { reasoningEfforts: efforts },
     });
   }

server/plugins/build-in/grok-auth/grok_test.ts

@@ -160,9 +160,8 @@ Deno.test("model discovery parses both language-models and standard list shapes"
   });
   assertEquals(richModels.map((model) => model.id), ["grok-4", "grok-3-mini"]);
   assertEquals(richModels[0].displayName, "Grok 4");
-  assertEquals(richModels[0].contextWindowTokens, 256_000);
-  assertEquals(richModels[0].capabilities, { thinking: false, images: true });
-  assertEquals(richModels[1].capabilities, { thinking: false, images: false });
+  assertEquals(richModels[0].capabilities, { images: true });
+  assertEquals(richModels[1].capabilities, { images: false });
 
   const plainModels = parseGrokModels({ data: [{ id: "grok-4-fast" }] });
   assertEquals(plainModels.map((model) => model.id), ["grok-4-fast"]);

server/plugins/build-in/grok-auth/models.ts

@@ -9,12 +9,12 @@ export const FALLBACK_MODELS: ModelDefinition[] = [
   {
     id: "grok-4.6",
     displayName: "Grok 4.6",
-    capabilities: { thinking: false, images: true },
+    capabilities: { images: true },
   },
   {
     id: "grok-4.5",
     displayName: "Grok 4.5",
-    capabilities: { thinking: false, images: true },
+    capabilities: { images: true },
   },
 ];
 
@@ -28,15 +28,6 @@ function text(value: unknown): string | null {
   return typeof value === "string" && value.trim() ? value.trim() : null;
 }
 
-function positiveInteger(value: unknown): number | null {
-  const parsed = typeof value === "number"
-    ? value
-    : typeof value === "string"
-    ? Number(value)
-    : NaN;
-  return Number.isFinite(parsed) && parsed > 0 ? Math.floor(parsed) : null;
-}
-
 function modalities(value: unknown): string[] {
   return Array.isArray(value)
     ? value.flatMap((item) => (typeof item === "string" ? [item.toLowerCase()] : []))
@@ -66,15 +57,10 @@ export function parseGrokModels(body: unknown): ModelDefinition[] {
     if (!id || seen.has(id)) continue;
     seen.add(id);
     const inputs = modalities(model?.input_modalities ?? model?.inputModalities);
-    const contextWindowTokens = positiveInteger(
-      model?.context_window ?? model?.contextWindow ?? model?.max_prompt_length,
-    );
     models.push({
       id,
       displayName: displayName(id),
-      ...(contextWindowTokens !== null ? { contextWindowTokens } : {}),
       capabilities: {
-        thinking: false,
         images: inputs.length === 0 || inputs.includes("image"),
       },
     });

server/src/api/cursor/bidi.rs

@@ -184,7 +184,11 @@ pub async fn append(
     request: DecodedAppend,
     parent: Option<TransportParent>,
 ) -> Result<ai::BidiAppendResponse> {
-    let handle = registry.get_or_create(&request.request_id).await?;
+    let replace_closing = request.model_id().is_some();
+    let handle = registry
+        .get_or_create_for_append(&request.request_id, replace_closing)
+        .await?;
+    let _admission = handle.admit()?;
     if let Some(conversation_id) = request.conversation_id() {
         handle.set_conversation_id(conversation_id)?;
     }

server/src/api/cursor/handlers.rs

@@ -19,16 +19,17 @@ use crate::{
             connect,
             proto::{agent::v1 as agent, aiserver::v1 as ai},
         },
-        services::{
-            account, analytics, knowledge, model_catalog, observability::CursorTraceRecorder, tab,
-        },
+        services::{account, analytics, knowledge, model_catalog, tab},
         transport::{TransportParent, TransportRegistry},
     },
     Result,
 };
 
-pub fn router(registry: TransportRegistry) -> Result<Router> {
-    let proxy = CursorProxy::cursor(registry.store().clone())?;
+pub fn router(
+    registry: TransportRegistry,
+    clients: crate::network::NetworkClients,
+) -> Result<Router> {
+    let proxy = CursorProxy::cursor(clients);
     let knowledge = knowledge::KnowledgeService::managed()?;
     Ok(router_with_proxy(registry, proxy, knowledge))
 }
@@ -120,16 +121,13 @@ async fn run_sse_handler(
     let (parts, body) = buffered(request).await?;
     let request: agent::BidiRequestId = connect::decode_unary(&body)?;
     let route = registry.wait_route(&request.request_id).await;
-    let trace = CursorTraceRecorder::resume(registry.store().clone(), &request.request_id).await;
-    if let Some(trace) = &trace {
-        trace
-            .request(
-                "run_sse_request",
-                &body,
-                serde_json::json!({"request_id": request.request_id}),
-            )
-            .await;
-    }
+    let trace = registry.trace(&request.request_id);
+    trace.resume();
+    trace.request(
+        "run_sse_request",
+        body.clone(),
+        serde_json::json!({"request_id": request.request_id}),
+    );
     match route {
         crate::cursor::transport::TransportRoute::Local => {
             run_sse::stream(&registry, &request.request_id).await
@@ -140,7 +138,14 @@ async fn run_sse_handler(
                 Request::from_parts(parts, Body::from(body)),
             )
             .await?;
-            Ok(run_sse::upstream(registry, request.request_id, generation, response, trace).await)
+            Ok(run_sse::upstream(
+                registry,
+                request.request_id,
+                generation,
+                response,
+                Some(trace),
+            )
+            .await)
         }
     }
 }
@@ -156,6 +161,7 @@ async fn bidi_handler(
     let first_model = decoded.model_id().map(str::to_owned);
     let conversation_id = decoded.conversation_id().map(str::to_owned);
     let trace_metadata = decoded.trace_metadata();
+    let trace = registry.trace(&decoded.request_id);
     let local = if let Some(model_id) = decoded.model_id() {
         // 插件模型 ID 只在本地有意义,永远不转发到 Cursor 官方上游。
         if model_id.starts_with(crate::plugin::ADAPTER_ID_PREFIX)
@@ -180,14 +186,18 @@ async fn bidi_handler(
     } else if registry.upstream(&decoded.request_id).await {
         false
     } else {
+        trace.resume();
+        trace.request(
+            "bidi_request",
+            body.clone(),
+            trace_outcome(trace_metadata, false, "missing_transport", None),
+        );
         return Err(crate::Error::Protocol(
             "first BidiAppend message must select a model".into(),
         ));
     };
-    let trace = if first_model.is_some() {
-        CursorTraceRecorder::begin(
-            registry.store().clone(),
-            &decoded.request_id,
+    if first_model.is_some() {
+        trace.begin(
             conversation_id.as_deref(),
             if local {
                 "local_byok"
@@ -195,26 +205,61 @@ async fn bidi_handler(
                 "cursor_official"
             },
             first_model.as_deref(),
-        )
-        .await
+        );
     } else {
-        CursorTraceRecorder::resume(registry.store().clone(), &decoded.request_id).await
-    };
-    if let Some(trace) = &trace {
-        trace.request("bidi_request", &body, trace_metadata).await;
+        trace.resume();
     }
     if !local {
         if first_model.is_some() {
             registry.mark_upstream(&decoded.request_id).await;
         }
+        trace.request(
+            "bidi_request",
+            body.clone(),
+            trace_outcome(trace_metadata, true, "upstream", None),
+        );
         return proxy::forward(
             Extension(proxy),
             Request::from_parts(parts, Body::from(body)),
         )
         .await;
     }
-    let parent = parent_headers(&parts.headers)?;
-    bidi::append(&registry, decoded, parent).await?;
+    let parent = match parent_headers(&parts.headers) {
+        Ok(parent) => parent,
+        Err(error) => {
+            trace.request(
+                "bidi_request",
+                body,
+                trace_outcome(
+                    trace_metadata,
+                    false,
+                    "invalid_parent",
+                    Some(error.to_string()),
+                ),
+            );
+            return Err(error);
+        }
+    };
+    match bidi::append(&registry, decoded, parent).await {
+        Ok(_) => trace.request(
+            "bidi_request",
+            body,
+            trace_outcome(trace_metadata, true, "local", None),
+        ),
+        Err(error) => {
+            trace.request(
+                "bidi_request",
+                body,
+                trace_outcome(
+                    trace_metadata,
+                    false,
+                    "command_rejected",
+                    Some(error.to_string()),
+                ),
+            );
+            return Err(error);
+        }
+    }
     let mut response = Response::new(axum::body::Body::empty());
     *response.status_mut() = StatusCode::OK;
     response.headers_mut().insert(
@@ -224,6 +269,22 @@ async fn bidi_handler(
     Ok(response)
 }
 
+fn trace_outcome(
+    mut metadata: serde_json::Value,
+    accepted: bool,
+    route_outcome: &str,
+    error: Option<String>,
+) -> serde_json::Value {
+    if let Some(metadata) = metadata.as_object_mut() {
+        metadata.insert("accepted".into(), accepted.into());
+        metadata.insert("route_outcome".into(), route_outcome.into());
+        if let Some(error) = error {
+            metadata.insert("error".into(), error.into());
+        }
+    }
+    metadata
+}
+
 async fn buffered(request: Request<Body>) -> Result<(axum::http::request::Parts, Bytes)> {
     let (parts, body) = request.into_parts();
     let body = to_bytes(body, usize::MAX)

server/src/api/cursor/proxy.rs

@@ -14,8 +14,7 @@ pub const UPSTREAM_URL_HEADER: &str = "x-server-upstream-url";
 
 #[derive(Clone)]
 pub struct CursorProxy {
-    client: Option<reqwest::Client>,
-    store: Option<crate::store::Store>,
+    clients: crate::network::NetworkClients,
     upstream: String,
 }
 
@@ -47,23 +46,15 @@ impl BufferedResponse {
 }
 
 impl CursorProxy {
-    pub fn cursor(store: crate::store::Store) -> Result<Self> {
-        Ok(Self {
-            client: None,
-            store: Some(store),
+    pub fn cursor(clients: crate::network::NetworkClients) -> Self {
+        Self {
+            clients,
             upstream: CURSOR_UPSTREAM.into(),
-        })
+        }
     }
 
     async fn client(&self) -> Result<reqwest::Client> {
-        match (&self.client, &self.store) {
-            (Some(client), _) => Ok(client.clone()),
-            (_, Some(store)) => Ok(crate::network::client_builder(store)
-                .await?
-                .redirect(reqwest::redirect::Policy::none())
-                .build()?),
-            _ => unreachable!("Cursor proxy always has a client or store"),
-        }
+        self.clients.cursor_client().await
     }
 }
 

server/src/api/cursor/run_sse.rs

@@ -22,7 +22,7 @@ pub async fn stream(registry: &TransportRegistry, request_id: &str) -> Result<Re
     let receiver = handle.subscribe();
     let trace = handle.trace().cloned();
     if let Some(trace) = &trace {
-        trace.response_started(StatusCode::OK.as_u16()).await;
+        trace.response_started(StatusCode::OK.as_u16());
     }
     let body_stream = local_body_stream(receiver, handle, trace);
     let mut response = Response::new(Body::from_stream(body_stream));
@@ -133,7 +133,7 @@ pub async fn upstream(
 ) -> Response<Body> {
     let (parts, body) = response.into_parts();
     if let Some(trace) = &trace {
-        trace.response_started(parts.status.as_u16()).await;
+        trace.response_started(parts.status.as_u16());
     }
     let stream = async_stream::stream! {
         let _guard = UpstreamRunGuard {
@@ -180,15 +180,15 @@ impl TraceStreamSink {
             while let Some(event) = receiver.recv().await {
                 match event {
                     TraceStreamEvent::Chunk(chunk) => {
-                        trace.response_chunk(source, &chunk).await;
+                        trace.response_chunk(source, chunk);
                     }
                     TraceStreamEvent::Finish(error) => {
-                        trace.finish(error.as_deref()).await;
+                        trace.finish(error.as_deref());
                         return;
                     }
                 }
             }
-            trace.finish(None).await;
+            trace.finish(None);
         });
         Self {
             sender: Some(sender),

server/src/api/router.rs

@@ -1,7 +1,7 @@
 //! Builds the top-level server router.
 
-use crate::{cursor::transport::TransportRegistry, Result};
+use crate::{cursor::transport::TransportRegistry, network::NetworkClients, Result};
 
-pub fn router(registry: TransportRegistry) -> Result<axum::Router> {
-    super::cursor::router(registry)
+pub fn router(registry: TransportRegistry, clients: NetworkClients) -> Result<axum::Router> {
+    super::cursor::router(registry, clients)
 }

server/src/app.rs

@@ -44,9 +44,11 @@ impl App {
             plugin_runtime.clone(),
             config.app_version.clone(),
         )?;
+        let clients = crate::network::NetworkClients::new(store.clone());
         let provider = std::sync::Arc::new(ProviderRouter::new(
             store.clone(),
             plugins.clone(),
+            clients.clone(),
             config.provider_request_timeout,
             config.provider_stream_idle_timeout,
         ));
@@ -58,10 +60,16 @@ impl App {
             plugins.clone(),
             crate::config::managed_data_dir()?.join("rules"),
         );
-        let control =
-            control::ControlService::new(store.clone(), provider, plugin_runtime, plugins)?;
+        let control = control::ControlService::new(
+            store.clone(),
+            provider,
+            plugin_runtime,
+            plugins,
+            clients.clone(),
+            config.app_version.clone(),
+        )?;
         let harness = control.cursor_harness().clone();
-        let mut router = api::router(registry.clone())?;
+        let mut router = api::router(registry.clone(), clients)?;
         router = match &config.console {
             Some(ConsoleSource::Directory(directory)) => {
                 router.merge(control::web_router(control.clone(), directory))

server/src/config.rs

@@ -46,7 +46,6 @@ pub struct ProviderConfig {
     pub custom_headers: reqwest::header::HeaderMap,
     pub max_output_tokens: Option<u64>,
     pub request_timeout: Duration,
-    pub retry_count: u32,
     pub allowed_body_fields: Option<std::collections::HashSet<String>>,
 }
 

server/src/control/ads.rs

@@ -17,6 +17,7 @@ use super::ControlService;
 // 开源项目广告为作者唯一收入来源，PR时请勿去除此广告
 
 pub(super) const ADS_ENDPOINT: &str = "https://tab.leokun.cn/api/v1/ads?placement=menu";
+// pub(super) const ADS_ENDPOINT: &str = "http://localhost:8080/api/v1/ads?placement=menu";
 pub(super) const DEVICE_ID_HEADER: &str = "X-Cursor-Assistant-Device-ID";
 pub(super) const OS_HEADER: &str = "X-Cursor-Assistant-OS";
 pub(super) const APP_VERSION_HEADER: &str = "X-Cursor-Assistant-Version";

server/src/control/service.rs

@@ -41,6 +41,8 @@ pub struct ControlService {
     provider: Arc<dyn Provider>,
     plugin_runtime: PluginRuntime,
     plugins: PluginRegistry,
+    clients: crate::network::NetworkClients,
+    app_version: String,
     model_tests: Arc<Mutex<BTreeMap<String, CancellationToken>>>,
 }
 
@@ -151,6 +153,8 @@ impl ControlService {
         provider: Arc<dyn Provider>,
         plugin_runtime: PluginRuntime,
         plugins: PluginRegistry,
+        clients: crate::network::NetworkClients,
+        app_version: String,
     ) -> Result<Self> {
         Ok(Self {
             cursor_harness: CursorHarness::new(store.clone())?,
@@ -158,6 +162,8 @@ impl ControlService {
             provider,
             plugin_runtime,
             plugins,
+            clients,
+            app_version,
             model_tests: Arc::new(Mutex::new(BTreeMap::new())),
         })
     }
@@ -256,13 +262,13 @@ impl ControlService {
         disabled_ad_ids: Option<&str>,
         language: &str,
     ) -> Result<AdRuntime> {
-        let client = crate::network::client(&self.store).await?;
+        let client = self.clients.default_client().await?;
         let installation_id = self.store.installation_id().await?;
         let mut request = client
             .get(ADS_ENDPOINT)
             .header(DEVICE_ID_HEADER, installation_id)
             .header(OS_HEADER, std::env::consts::OS)
-            .header(APP_VERSION_HEADER, env!("CARGO_PKG_VERSION"))
+            .header(APP_VERSION_HEADER, &self.app_version)
             .header(LANGUAGE_HEADER, language)
             .timeout(std::time::Duration::from_secs(60));
         if let Some(disabled_ad_ids) = disabled_ad_ids.filter(|value| !value.is_empty()) {
@@ -281,7 +287,7 @@ impl ControlService {
     }
 
     pub(super) async fn dismiss_ad(&self, ad_id: &str, input: &AdDismissalInput) -> Result<()> {
-        let client = crate::network::client(&self.store).await?;
+        let client = self.clients.default_client().await?;
         let installation_id = self.store.installation_id().await?;
         let mut endpoint = Url::parse(ADS_ENDPOINT).map_err(|error| {
             Error::Config(format!("advertisement endpoint is invalid: {error}"))
@@ -296,7 +302,7 @@ impl ControlService {
             .post(endpoint)
             .header(DEVICE_ID_HEADER, installation_id)
             .header(OS_HEADER, std::env::consts::OS)
-            .header(APP_VERSION_HEADER, env!("CARGO_PKG_VERSION"))
+            .header(APP_VERSION_HEADER, &self.app_version)
             .json(input)
             .timeout(std::time::Duration::from_secs(5))
             .send()
@@ -392,7 +398,6 @@ impl ControlService {
         if model_hash.starts_with(crate::plugin::ADAPTER_ID_PREFIX) {
             let descriptor = self.plugins.model_descriptor(model_hash).await?;
             model.display_name = Some(descriptor.display_name);
-            model.context_window_tokens = descriptor.context_window_tokens;
             model.max_output_tokens = Some(descriptor.max_output_tokens.unwrap_or(65_536));
         } else {
             let configured = self
@@ -511,7 +516,7 @@ impl ControlService {
     }
 
     pub async fn discover_models(&self, input: &ModelDiscoveryInput) -> Result<DiscoveredModels> {
-        let client = crate::network::client(&self.store).await?;
+        let client = self.clients.default_client().await?;
         let base_url = crate::model::normalize_request_url(&input.base_url)?;
         discover_models_from_endpoint(
             &client,
@@ -678,7 +683,9 @@ impl ControlService {
     }
 
     pub async fn set_proxy_settings(&self, settings: ProxySettingsInput) -> Result<ProxySettings> {
-        self.store.set_proxy_settings(settings).await
+        let settings = self.store.set_proxy_settings(settings).await?;
+        self.clients.invalidate().await;
+        Ok(settings)
     }
 
     pub async fn tab_settings(&self) -> Result<TabSettings> {

server/src/cursor/checkpoint/builder.rs

@@ -278,19 +278,17 @@ impl CheckpointBuilder {
             ),
         });
         if let Some(trace) = handle.trace() {
-            trace
-                .artifact(
-                    "checkpoint",
-                    "byok_server",
-                    &checkpoint.encode_to_vec(),
-                    serde_json::json!({
-                        "root_message_count": checkpoint.root_prompt_messages_json.len(),
-                        "turn_count": checkpoint.turns.len(),
-                        "pending_tool_call_count": checkpoint.pending_tool_calls.len(),
-                        "emit_status": if result.is_ok() { "sent" } else { "error" },
-                    }),
-                )
-                .await;
+            trace.artifact(
+                "checkpoint",
+                "byok_server",
+                &checkpoint.encode_to_vec(),
+                serde_json::json!({
+                    "root_message_count": checkpoint.root_prompt_messages_json.len(),
+                    "turn_count": checkpoint.turns.len(),
+                    "pending_tool_call_count": checkpoint.pending_tool_calls.len(),
+                    "emit_status": if result.is_ok() { "sent" } else { "error" },
+                }),
+            );
         }
         result
     }

server/src/cursor/checkpoint/messages/decode.rs

@@ -102,6 +102,12 @@ pub fn decode_pending(value: &str) -> Result<RecoveredToolRound> {
         .into_iter()
         .enumerate()
         .map(|(index, call)| {
+            let argument_error = wire
+                .pointer("/providerOptions/cursor/pendingToolExecutionContracts")
+                .and_then(|contracts| contracts.get(&call.call_id))
+                .and_then(|contract| contract.get("argumentError"))
+                .and_then(Value::as_str)
+                .map(str::to_string);
             Ok(ToolCall {
                 index,
                 call_id: call.call_id,
@@ -109,6 +115,7 @@ pub fn decode_pending(value: &str) -> Result<RecoveredToolRound> {
                 name: call.name,
                 arguments_text: serde_json::to_string(&call.arguments)?,
                 arguments: call.arguments,
+                argument_error,
             })
         })
         .collect::<Result<Vec<_>>>()?;

server/src/cursor/checkpoint/messages/encode.rs

@@ -71,6 +71,7 @@ pub fn staged_tool_round(
             allowed_tools,
             dynamic_tools,
             started_at_ms,
+            tool_calls: Some(calls),
         }),
     )?)?)
 }
@@ -96,6 +97,7 @@ pub fn staged_final(
             allowed_tools,
             dynamic_tools,
             started_at_ms,
+            tool_calls: None,
         }),
     )?)?)
 }
@@ -105,6 +107,7 @@ pub(super) struct PendingContext<'a> {
     allowed_tools: &'a [String],
     dynamic_tools: &'a HashSet<String>,
     started_at_ms: u64,
+    tool_calls: Option<&'a [ToolCall]>,
 }
 
 pub(super) fn wire_message(
@@ -132,16 +135,23 @@ pub(super) fn wire_message(
                     calls
                         .iter()
                         .map(|call| {
-                            (
-                                call.call_id.clone(),
-                                json!({
-                                    "toolCallId": call.call_id,
-                                    "outerToolName": call.name,
-                                    "toolIdentifier": tool_identifier(&call.name, pending.dynamic_tools),
-                                    "isDynamic": pending.dynamic_tools.contains(&call.name),
-                                    "allowedToolNames": pending.allowed_tools,
-                                }),
-                            )
+                            let mut contract = json!({
+                                "toolCallId": call.call_id,
+                                "outerToolName": call.name,
+                                "toolIdentifier": tool_identifier(&call.name, pending.dynamic_tools),
+                                "isDynamic": pending.dynamic_tools.contains(&call.name),
+                                "allowedToolNames": pending.allowed_tools,
+                            });
+                            if let Some(error) = pending
+                                .tool_calls
+                                .and_then(|calls| {
+                                    calls.iter().find(|candidate| candidate.call_id == call.call_id)
+                                })
+                                .and_then(|call| call.argument_error.as_deref())
+                            {
+                                contract["argumentError"] = Value::String(error.into());
+                            }
+                            (call.call_id.clone(), contract)
                         })
                         .collect(),
                 ),

server/src/cursor/checkpoint/messages/tests.rs

@@ -30,6 +30,7 @@ fn pending_tool_round_is_one_complete_assistant_message_and_round_trips() {
             name: "Read".into(),
             arguments_text: r#"{"path":"/a"}"#.into(),
             arguments: json!({"path":"/a"}),
+            argument_error: Some("Read arguments are not valid JSON".into()),
         },
         ToolCall {
             index: 1,
@@ -38,6 +39,7 @@ fn pending_tool_round_is_one_complete_assistant_message_and_round_trips() {
             name: "Grep".into(),
             arguments_text: r#"{"pattern":"x"}"#.into(),
             arguments: json!({"pattern":"x"}),
+            argument_error: None,
         },
     ];
     let pending = staged_tool_round(
@@ -55,6 +57,10 @@ fn pending_tool_round_is_one_complete_assistant_message_and_round_trips() {
         wire["providerOptions"]["cursor"]["pendingToolExecutionContracts"]["a"]["toolIdentifier"],
         "READ"
     );
+    assert_eq!(
+        wire["providerOptions"]["cursor"]["pendingToolExecutionContracts"]["a"]["argumentError"],
+        "Read arguments are not valid JSON"
+    );
     assert_eq!(wire["role"], "assistant");
     assert_eq!(
         wire["providerOptions"]["cursor"]["pendingToolExecutionContracts"]
@@ -77,6 +83,10 @@ fn pending_tool_round_is_one_complete_assistant_message_and_round_trips() {
     assert_eq!(recovered.assistant.replay_state, Some(replay_state));
     assert_eq!(recovered.calls.len(), 2);
     assert_eq!(recovered.calls[0].call_id, "a");
+    assert_eq!(
+        recovered.calls[0].argument_error.as_deref(),
+        Some("Read arguments are not valid JSON")
+    );
     assert_eq!(recovered.calls[1].call_id, "b");
 }
 

server/src/cursor/checkpoint/steps.rs

@@ -75,6 +75,11 @@ impl StepBuffer {
         });
     }
 
+    pub fn finish_model_attempt(&mut self) {
+        self.finish_text();
+        self.finish_thinking(Duration::ZERO);
+    }
+
     pub fn discard_model_output(&mut self) {
         self.text.clear();
         self.thinking.clear();
@@ -101,6 +106,28 @@ impl StepBuffer {
 mod tests {
     use super::*;
 
+    #[test]
+    fn failed_attempt_output_is_retained_for_the_next_checkpoint() {
+        let mut buffer = StepBuffer::default();
+        buffer.text_delta("partial answer");
+        buffer.thinking_delta("partial reasoning");
+
+        buffer.finish_model_attempt();
+
+        let steps = buffer.take().steps;
+        assert_eq!(steps.len(), 2);
+        assert!(matches!(
+            &steps[0].message,
+            Some(pb::conversation_step::Message::AssistantMessage(message))
+                if message.text == "partial answer"
+        ));
+        assert!(matches!(
+            &steps[1].message,
+            Some(pb::conversation_step::Message::ThinkingMessage(message))
+                if message.text == "partial reasoning"
+        ));
+    }
+
     #[test]
     fn interrupted_model_output_is_not_persisted_as_checkpoint_steps() {
         let mut buffer = StepBuffer::default();

server/src/cursor/checkpoint/summary.rs

@@ -3,7 +3,7 @@ use prost::Message;
 
 use crate::{
     cursor::{checkpoint::PendingSteps, protocol::proto::agent::v1 as pb},
-    model::CanonicalMessage,
+    model::{estimate_context_tokens, project_messages, CanonicalMessage, PromptSpec},
     store::{BlobEdge, BlobId},
     Error, Result,
 };
@@ -85,6 +85,12 @@ impl CheckpointBuilder {
                 .push(archive_id.as_bytes().to_vec());
         }
         self.base.self_summary_count = self.base.self_summary_count.saturating_add(1);
+        let projected = project_messages(messages)?;
+        let prompt = PromptSpec {
+            instructions: self.instructions.clone(),
+            tools: self.tool_definitions.clone(),
+        };
+        self.record_context_tokens(Some(estimate_context_tokens(&prompt, &projected)));
         if let Some(details) = self.base.token_details.as_mut() {
             details.breakdown = Some(crate::cursor::services::usage::breakdown(
                 details.used_tokens,

server/src/cursor/compile/context.rs

@@ -12,7 +12,7 @@ use crate::{
         protocol::proto::agent::v1 as pb, services::context_sync::RequestContextSynchronizer,
         tools::runtime::McpRoute,
     },
-    model::ToolDefinition,
+    model::{normalize_tool_name, ToolDefinition},
     store::BlobId,
     Error, Result,
 };
@@ -493,7 +493,7 @@ pub fn dynamic_mcp(
             })?),
         };
         let parameters = normalize_mcp_parameters(&wire.name, parameters)?;
-        let name = model_tool_name(&wire.name);
+        let name = normalize_tool_name(&wire.name);
         let definition = ToolDefinition {
             name: name.clone(),
             description: wire.description.clone(),
@@ -553,18 +553,6 @@ fn invalid_mcp_parameters(tool_name: &str) -> Error {
     ))
 }
 
-fn model_tool_name(name: &str) -> String {
-    name.chars()
-        .map(|character| {
-            if character.is_ascii_alphanumeric() || matches!(character, '_' | '-') {
-                character
-            } else {
-                '_'
-            }
-        })
-        .collect()
-}
-
 fn prost_value(value: &prost_types::Value) -> Value {
     use prost_types::value::Kind;
     match value.kind.as_ref() {

server/src/cursor/compile/run.rs

@@ -117,9 +117,7 @@ pub(crate) async fn prepare(
             "selected_source": "root_prompt_messages_json",
         });
         let encoded = serde_json::to_vec(&summary)?;
-        trace
-            .artifact("history_projection", "byok_server", &encoded, summary)
-            .await;
+        trace.artifact("history_projection", "byok_server", &encoded, summary);
     }
     let mut request_context = context::hydrate(request, context_sync).await?;
     if let Some(rules_dir) = local_rules_dir {

server/src/cursor/conversation/command.rs

@@ -1,6 +1,19 @@
 //! Defines commands accepted by a Conversation runtime.
 
-use crate::cursor::protocol::proto::agent::v1 as pb;
+use crate::{cursor::protocol::proto::agent::v1 as pb, Error};
+
+#[derive(Debug)]
+pub enum RunFinish {
+    TurnCompleted,
+    Transport(TransportFinish),
+}
+
+#[derive(Debug)]
+pub enum TransportFinish {
+    Success,
+    Failed(Error),
+    Cancelled,
+}
 
 #[derive(Debug)]
 pub enum TransportCommand {
@@ -8,6 +21,9 @@ pub enum TransportCommand {
         seqno: i64,
         message: Box<pb::AgentClientMessage>,
     },
+    RunFinished {
+        generation: u64,
+        finish: RunFinish,
+    },
     Disconnect,
-    Close,
 }

server/src/cursor/conversation/output.rs

@@ -21,7 +21,7 @@ use crate::{
         protocol::proto::agent::v1 as pb,
         services::blob_sync::BlobSynchronizer,
         tools::{
-            codec,
+            codec, compat,
             runtime::CursorToolRuntime,
             stream::ToolCallStream,
             tool_call_result::{ToolCompletion, ToolResultReceiver},
@@ -34,7 +34,7 @@ use crate::{
     Error, Result,
 };
 
-use super::{CompiledMessages, ConversationRegistry, MessageDelivery};
+use super::{CompiledMessages, ConversationRegistry, MessageDelivery, RunFinish, TransportFinish};
 use crate::cursor::transport::TransportHandle;
 
 pub struct ConversationOutput {
@@ -110,7 +110,7 @@ impl ConversationOutput {
         }
     }
 
-    pub async fn run(mut self) -> Result<()> {
+    pub async fn run(mut self) -> Result<RunFinish> {
         let result = self.run_inner().await;
         if let Err(error) = &result {
             if !self.superseded.is_cancelled() {
@@ -143,7 +143,7 @@ impl ConversationOutput {
         result
     }
 
-    async fn run_inner(&mut self) -> Result<()> {
+    async fn run_inner(&mut self) -> Result<RunFinish> {
         if self.context.compacting {
             self.handle.emit(&events::summary_started())?;
         }
@@ -175,7 +175,7 @@ impl ConversationOutput {
             if self.superseded.is_cancelled() {
                 worker.abort();
                 self.abort_execs().await;
-                return Ok(());
+                return Ok(RunFinish::Transport(TransportFinish::Cancelled));
             }
             let input = if let Ok(action) = self.runtime_actions.try_recv() {
                 Input::RuntimeAction(Some(Box::new(action)))
@@ -187,7 +187,7 @@ impl ConversationOutput {
                     _ = self.superseded.cancelled() => {
                         worker.abort();
                         self.abort_execs().await;
-                        return Ok(());
+                        return Ok(RunFinish::Transport(TransportFinish::Cancelled));
                     }
                     action = self.runtime_actions.recv() => Input::RuntimeAction(action.map(Box::new)),
                     event = self.core.events.recv() => Input::Event(event),
@@ -264,6 +264,32 @@ impl ConversationOutput {
                         streams.clear();
                         presentation.discard_model_output();
                     }
+                    RunEvent::ModelAttemptFailed { attempt, message } => {
+                        tracing::warn!(
+                            run_id = %self.run.run_id(),
+                            attempt,
+                            %message,
+                            "retrying model call from current checkpoint"
+                        );
+                        presentation.finish_model_attempt();
+                        for call in calls.values_mut() {
+                            if call.arguments.is_null() {
+                                call.arguments = serde_json::from_str(&call.arguments_text)
+                                    .unwrap_or_else(|_| serde_json::json!({}));
+                            }
+                            let completion = compat::failure_with_message(
+                                call,
+                                format!("Model attempt failed before tool completion: {message}"),
+                            );
+                            self.handle
+                                .emit(&codec::tool_completed(call, &completion))?;
+                            presentation.tool_completed(&completion);
+                        }
+                        response_text.clear();
+                        response_thinking.clear();
+                        calls.clear();
+                        streams.clear();
+                    }
                     RunEvent::TextStart => {}
                     RunEvent::TextEnd => {
                         if !self.context.compacting {
@@ -312,6 +338,7 @@ impl ConversationOutput {
                             name: name.clone(),
                             arguments_text: String::new(),
                             arguments: serde_json::Value::Null,
+                            argument_error: None,
                         };
                         self.emit_model_event(
                             crate::provider::ModelEvent::ToolCallStart {
@@ -335,15 +362,50 @@ impl ConversationOutput {
                         let stream = streams.get_mut(&index).ok_or_else(|| {
                             Error::Protocol(format!("missing Cursor tool stream: {index}"))
                         })?;
-                        for message in stream.arguments_delta(call, &delta)? {
-                            self.handle.emit(&message)?;
+                        match stream.arguments_delta(call, &delta) {
+                            Ok(messages) => {
+                                for message in messages {
+                                    self.handle.emit(&message)?;
+                                }
+                            }
+                            Err(Error::Protocol(message)) => {
+                                tracing::warn!(
+                                    call_id = %call.call_id,
+                                    %message,
+                                    "ignoring invalid streaming tool arguments until completion"
+                                );
+                            }
+                            Err(Error::Json(error)) => {
+                                tracing::warn!(
+                                    call_id = %call.call_id,
+                                    %error,
+                                    "ignoring invalid streaming tool arguments until completion"
+                                );
+                            }
+                            Err(error) => return Err(error),
                         }
                     }
                     RunEvent::ToolCallEnd { index } => {
                         let call = calls.get_mut(&index).ok_or_else(|| {
                             Error::Protocol(format!("unknown completed tool index: {index}"))
                         })?;
-                        call.arguments = serde_json::from_str(&call.arguments_text)?;
+                        // A tool call with no arguments streams no argument text.
+                        // Treat empty text as an empty object, matching the model
+                        // cycle, instead of failing the run on `from_str("")`.
+                        call.arguments = if call.arguments_text.trim().is_empty() {
+                            serde_json::json!({})
+                        } else {
+                            serde_json::from_str(&call.arguments_text)
+                                .unwrap_or_else(|_| serde_json::json!({}))
+                        };
+                    }
+                    RunEvent::UsageSnapshot(usage) => {
+                        if !self.context.compacting {
+                            if let Some(output_tokens) = usage.output_tokens {
+                                self.handle.emit(&events::token_delta(output_tokens))?;
+                            }
+                            context_tokens = usage.context_input_tokens;
+                        }
                     }
                     RunEvent::Usage(usage) => {
                         if !self.context.compacting {
@@ -352,10 +414,7 @@ impl ConversationOutput {
                             }
                         }
                         if !self.context.compacting {
-                            context_tokens = usage
-                                .input_tokens
-                                .zip(usage.output_tokens)
-                                .and_then(|(input, output)| input.checked_add(output));
+                            context_tokens = usage.context_input_tokens;
                         }
                         match &mut turn_usage {
                             Some(total) => *total += usage,
@@ -424,21 +483,28 @@ impl ConversationOutput {
                             streams.clear();
                         }
                         if let CommitCause::RuntimeEvent { event_id } = &state.cause {
-                            if let Some(injection_id) = event_id.strip_prefix("inject-context:") {
-                                if let Some(pending) = self.pending_injections.remove(injection_id)
-                                {
-                                    let delivered_at_ms = crate::cursor::tools::runtime::now_ms()
-                                        .min(i64::MAX as u64)
-                                        as i64;
-                                    self.handle.emit(&events::context_injection_delivered(
-                                        injection_id.to_owned(),
-                                        pending.delivery_batch_id.clone(),
-                                        delivered_at_ms,
-                                    ))?;
-                                    if let Some(user_message) = pending.user_message {
-                                        self.handle
-                                            .emit(&events::user_message_appended(user_message))?;
-                                    }
+                            // Injections key `pending_injections` by their raw
+                            // injection id and commit under `inject-context:{id}`,
+                            // while runtime user messages key it by (and commit
+                            // under) the full `user-message:{id}` event id. Strip
+                            // the injection prefix when present and otherwise use
+                            // the event id verbatim so both are cleared and emit
+                            // their delivered/appended events.
+                            let injection_id = event_id
+                                .strip_prefix("inject-context:")
+                                .unwrap_or(event_id.as_str());
+                            if let Some(pending) = self.pending_injections.remove(injection_id) {
+                                let delivered_at_ms = crate::cursor::tools::runtime::now_ms()
+                                    .min(i64::MAX as u64)
+                                    as i64;
+                                self.handle.emit(&events::context_injection_delivered(
+                                    injection_id.to_owned(),
+                                    pending.delivery_batch_id.clone(),
+                                    delivered_at_ms,
+                                ))?;
+                                if let Some(user_message) = pending.user_message {
+                                    self.handle
+                                        .emit(&events::user_message_appended(user_message))?;
                                 }
                             }
                         }
@@ -509,6 +575,7 @@ impl ConversationOutput {
                                 .map_err(|_| Error::Protocol("checkpoint worker stopped".into()))?
                             {
                                 Ok(checkpoint) => {
+                                    context_tokens = checkpoint_context_tokens(&checkpoint);
                                     compaction_checkpoint = Some(checkpoint);
                                     state.barrier.complete(Ok(()));
                                 }
@@ -652,7 +719,7 @@ impl ConversationOutput {
                         if self.superseded.is_cancelled() {
                             worker.abort();
                             self.abort_execs().await;
-                            return Ok(());
+                            return Ok(RunFinish::Transport(TransportFinish::Cancelled));
                         }
                         return match outcome {
                             RunOutcome::Completed => {
@@ -668,8 +735,7 @@ impl ConversationOutput {
                                     for _ in 0..3 {
                                         self.checkpoint.publish(&self.handle, &checkpoint).await?;
                                     }
-                                    finish_success(&self.handle);
-                                    return Ok(());
+                                    return Ok(RunFinish::TurnCompleted);
                                 }
                                 let checkpoints = final_checkpoint.take().ok_or_else(|| {
                                     Error::Protocol("Completed without final state".into())
@@ -685,18 +751,19 @@ impl ConversationOutput {
                                     ttft_breakdown: None,
                                     message: Some(pb::agent_server_message::Message::ConversationCheckpointUpdate(checkpoints.settled)),
                                 })?;
-                                finish_success(&self.handle);
-                                Ok(())
+                                Ok(RunFinish::TurnCompleted)
                             }
                             RunOutcome::Cancelled => {
                                 worker.abort();
                                 self.abort_execs().await;
-                                finish_cancelled(&self.handle)
+                                Ok(RunFinish::Transport(TransportFinish::Cancelled))
                             }
                             RunOutcome::Failed(failure) => {
                                 worker.abort();
                                 self.abort_execs().await;
-                                finish_failed(&self.handle, &cursor_error(failure))
+                                Ok(RunFinish::Transport(TransportFinish::Failed(cursor_error(
+                                    failure,
+                                ))))
                             }
                         };
                     }
@@ -1015,10 +1082,34 @@ pub(crate) fn finish_cancelled(handle: &TransportHandle) -> Result<()> {
     Ok(())
 }
 
+fn checkpoint_context_tokens(checkpoint: &pb::ConversationStateStructure) -> Option<u64> {
+    checkpoint
+        .token_details
+        .as_ref()
+        .map(|details| u64::from(details.used_tokens))
+}
+
 #[cfg(test)]
 mod tests {
-    use super::accept_tool_completion;
-    use crate::{run::CommandResult, Error};
+    use super::{accept_tool_completion, checkpoint_context_tokens};
+    use crate::{cursor::protocol::proto::agent::v1 as pb, run::CommandResult, Error};
+
+    #[test]
+    fn compacted_checkpoint_replaces_the_in_memory_context_usage() {
+        let compacted = pb::ConversationStateStructure {
+            token_details: Some(pb::ConversationTokenDetails {
+                used_tokens: 20_000,
+                ..Default::default()
+            }),
+            ..Default::default()
+        };
+
+        assert_eq!(checkpoint_context_tokens(&compacted), Some(20_000));
+        assert_eq!(
+            checkpoint_context_tokens(&pb::ConversationStateStructure::default()),
+            None
+        );
+    }
 
     #[test]
     fn closing_and_ended_runs_ignore_known_tool_completions() {

server/src/cursor/conversation/runtime.rs

@@ -11,25 +11,29 @@ use crate::{
         protocol::proto::agent::v1 as pb,
         services::{blob_sync::BlobSynchronizer, context_sync::RequestContextSynchronizer},
         tools::{
-            codec,
+            codec, compat,
             runtime::CursorToolRuntime,
             tool_call_result::{tool_result_channel, ToolResultReceiver, ToolResultSender},
             ClientToolEvent, ToolDispatcher,
         },
         transport::{OrderedInbox, TransportHandle},
     },
-    run::{CommandResult, RunEngine, RunHandle},
+    run::{CommandResult, RunEngine, RunHandle, RunPhase},
 };
 
 use super::{
     CompiledMessages, ConversationDependencies, ConversationOutput, ConversationOutputDependencies,
-    ConversationRegistry, MessageDelivery, TransportCommand,
+    ConversationRegistry, MessageDelivery, RunFinish, TransportCommand, TransportFinish,
 };
 
 pub struct ConversationRuntime;
 
+const CONTINUATION_IDLE_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(2);
+
 #[derive(Clone)]
 struct RunGeneration {
+    id: u64,
+    request: pb::AgentRunRequest,
     superseded: CancellationToken,
     finished: CancellationToken,
     run: Arc<parking_lot::Mutex<Option<RunHandle>>>,
@@ -41,6 +45,14 @@ struct RunGeneration {
 
 struct FinishGeneration(CancellationToken);
 
+struct TransportActorGuard(TransportHandle);
+
+impl Drop for TransportActorGuard {
+    fn drop(&mut self) {
+        self.0.close_transport();
+    }
+}
+
 impl Drop for FinishGeneration {
     fn drop(&mut self) {
         self.0.cancel();
@@ -54,6 +66,7 @@ impl ConversationRuntime {
         mut receiver: mpsc::Receiver<TransportCommand>,
     ) {
         tokio::spawn(async move {
+            let _actor_guard = TransportActorGuard(handle.clone());
             let dependencies = registry.dependencies().clone();
             let blob_sync = BlobSynchronizer::new(
                 handle.request_id().into(),
@@ -65,19 +78,70 @@ impl ConversationRuntime {
             let context_sync =
                 RequestContextSynchronizer::new(handle.clone(), dependencies.store.clone());
             let mut current = None::<RunGeneration>;
+            let mut next_generation = 1_u64;
+            let mut pending_finish = None::<(u64, TransportFinish)>;
+            let mut draining = false;
+            let mut waiting_for_action = false;
             loop {
-                let command = match receiver.recv().await {
-                    Some(command) => command,
-                    None => {
-                        handle.mark_disconnected();
-                        if let Some(generation) = current.as_ref() {
-                            generation.superseded.cancel();
-                            if let Some(run) = generation.run.lock().clone() {
-                                run.cancel();
+                let command = if draining {
+                    if !handle.admissions_drained() {
+                        tokio::select! {
+                            command = receiver.recv() => match command {
+                                Some(command) => command,
+                                None => {
+                                    finish_pending(&handle, &current, pending_finish.take());
+                                    break;
+                                }
+                            },
+                            _ = handle.wait_admissions_drained() => continue,
+                        }
+                    } else {
+                        handle.mark_draining();
+                        match receiver.try_recv() {
+                            Ok(command) => command,
+                            Err(mpsc::error::TryRecvError::Empty)
+                            | Err(mpsc::error::TryRecvError::Disconnected) => {
+                                finish_pending(&handle, &current, pending_finish.take());
+                                break;
                             }
                         }
-                        super::finish_cancelled(&handle).ok();
-                        break;
+                    }
+                } else if waiting_for_action {
+                    tokio::select! {
+                        command = receiver.recv() => match command {
+                            Some(command) => command,
+                            None => {
+                                handle.mark_disconnected();
+                                super::finish_success(&handle);
+                                break;
+                            }
+                        },
+                        _ = tokio::time::sleep(CONTINUATION_IDLE_TIMEOUT) => {
+                            let Some(generation) = current.as_ref() else {
+                                super::finish_success(&handle);
+                                break;
+                            };
+                            handle.begin_close();
+                            pending_finish = Some((generation.id, TransportFinish::Success));
+                            draining = true;
+                            waiting_for_action = false;
+                            continue;
+                        }
+                    }
+                } else {
+                    match receiver.recv().await {
+                        Some(command) => command,
+                        None => {
+                            handle.mark_disconnected();
+                            if let Some(generation) = current.as_ref() {
+                                generation.superseded.cancel();
+                                if let Some(run) = generation.run.lock().clone() {
+                                    run.cancel();
+                                }
+                            }
+                            super::finish_cancelled(&handle).ok();
+                            break;
+                        }
                     }
                 };
                 match command {
@@ -92,11 +156,41 @@ impl ConversationRuntime {
                                 let _ = handle.emit(&codec::abort(id));
                             }
                         }
-                        super::finish_cancelled(&handle).ok();
+                        let turn_completed = waiting_for_action
+                            || current.as_ref().is_some_and(|generation| {
+                                generation
+                                    .run
+                                    .lock()
+                                    .as_ref()
+                                    .is_none_or(|run| run.phase() != RunPhase::Running)
+                            });
+                        if turn_completed {
+                            super::finish_success(&handle);
+                        } else {
+                            super::finish_cancelled(&handle).ok();
+                        }
                         break;
                     }
-                    TransportCommand::Close => {
-                        break;
+                    TransportCommand::RunFinished { generation, finish } => {
+                        if !current
+                            .as_ref()
+                            .is_some_and(|current| current.id == generation)
+                        {
+                            continue;
+                        }
+                        match finish {
+                            RunFinish::TurnCompleted => {
+                                pending_finish = None;
+                                draining = false;
+                                waiting_for_action = true;
+                            }
+                            RunFinish::Transport(finish) => {
+                                waiting_for_action = false;
+                                handle.begin_close();
+                                pending_finish = Some((generation, finish));
+                                draining = true;
+                            }
+                        }
                     }
                     TransportCommand::Append { seqno, message } => {
                         for (_seqno, message) in inbox.push(seqno, *message) {
@@ -105,6 +199,12 @@ impl ConversationRuntime {
                                     Some(pb::agent_client_message::Message::RunRequest(
                                         request,
                                     )) => {
+                                        waiting_for_action = false;
+                                        if draining {
+                                            handle.reopen();
+                                            draining = false;
+                                            pending_finish = None;
+                                        }
                                         if let Some(conversation_id) =
                                             request.conversation_id.as_deref()
                                         {
@@ -117,60 +217,21 @@ impl ConversationRuntime {
                                                     "invalid Cursor conversation id"
                                                 );
                                                 let _ = super::finish_failed(&handle, &error);
-                                                let _ =
-                                                    handle.command(TransportCommand::Close).await;
                                                 return;
                                             }
                                         }
-                                        let previous_finished =
-                                            if let Some(previous) = current.take() {
-                                                previous.superseded.cancel();
-                                                if let Some(run) = previous.run.lock().clone() {
-                                                    run.cancel();
-                                                }
-                                                for id in previous
-                                                    .tool_runtime
-                                                    .interrupt_for_run_replacement()
-                                                    .await
-                                                {
-                                                    let _ = handle.emit(&codec::abort(id));
-                                                }
-                                                Some(previous.finished.clone())
-                                            } else {
-                                                None
-                                            };
-                                        let (results, result_receiver) = tool_result_channel();
-                                        let (runtime_actions, runtime_action_receiver) =
-                                            mpsc::unbounded_channel::<compile::RuntimeAction>();
-                                        let tool_runtime = tool_runtime_factory.next_run();
-                                        let tools = ToolDispatcher::with_results(
-                                            tool_runtime.clone(),
-                                            results.clone(),
-                                            dependencies.store.clone(),
-                                            dependencies.web_cache.clone(),
-                                        );
-                                        let generation = RunGeneration {
-                                            superseded: CancellationToken::new(),
-                                            finished: CancellationToken::new(),
-                                            run: Arc::new(parking_lot::Mutex::new(None)),
-                                            results,
-                                            runtime_actions,
-                                            tool_runtime,
-                                            tools,
-                                        };
-                                        current = Some(generation.clone());
-                                        spawn_run_request(
-                                            registry.clone(),
-                                            handle.clone(),
+                                        start_generation(
+                                            &registry,
+                                            &handle,
+                                            &dependencies,
+                                            &blob_sync,
+                                            &context_sync,
+                                            &tool_runtime_factory,
+                                            &mut current,
+                                            &mut next_generation,
                                             request,
-                                            dependencies.clone(),
-                                            blob_sync.clone(),
-                                            context_sync.clone(),
-                                            generation,
-                                            previous_finished,
-                                            result_receiver,
-                                            runtime_action_receiver,
-                                        );
+                                        )
+                                        .await;
                                     }
                                     Some(pb::agent_client_message::Message::ExecClientMessage(
                                         message,
@@ -262,17 +323,18 @@ impl ConversationRuntime {
                                                     .take_exec(throw.id)
                                                     .await
                                                 {
-                                                    Some(pending) => generation.results.send_error(
-                                                        crate::Error::Protocol(format!(
-                                                            "Exec {} failed: {}",
-                                                            pending.call.call_id, throw.error
-                                                        )),
+                                                    Some(pending) => generation.results.send(
+                                                        compat::failure_with_message(
+                                                            &pending.call,
+                                                            format!(
+                                                                "Exec {} failed: {}",
+                                                                pending.call.call_id, throw.error
+                                                            ),
+                                                        ),
                                                     ),
-                                                    None => generation.results.send_error(
-                                                        crate::Error::Protocol(format!(
-                                                            "unknown ExecClientThrow id: {}",
-                                                            throw.id
-                                                        )),
+                                                    None => tracing::warn!(
+                                                        id = throw.id,
+                                                        "ignoring failure for unknown tool execution"
                                                     ),
                                                 }
                                             }
@@ -330,27 +392,48 @@ impl ConversationRuntime {
                                     // return an explicit Protocol Error rather than falling through silently.
                                     Some(
                                         pb::agent_client_message::Message::ConversationAction(
-                                            action,
+                                            conversation_action,
                                         ),
-                                    ) => match action.action {
+                                    ) => match conversation_action.action.clone() {
                                         Some(
                                             pb::conversation_action::Action::UserMessageAction(
                                                 action,
                                             ),
                                         ) => {
-                                            let Some(generation) = current.as_ref() else {
+                                            let delivered_to_active_run =
+                                                current.as_ref().is_some_and(|generation| {
+                                                    generation.run.lock().as_ref().is_some_and(
+                                                        |run| run.phase() == RunPhase::Running,
+                                                    ) && generation
+                                                        .runtime_actions
+                                                        .send(compile::RuntimeAction::UserMessage(
+                                                            action.clone(),
+                                                        ))
+                                                        .is_ok()
+                                                });
+                                            if delivered_to_active_run {
+                                                continue;
+                                            }
+                                            let Some(previous) = current.as_ref() else {
                                                 continue;
                                             };
-                                            if generation
-                                                .runtime_actions
-                                                .send(compile::RuntimeAction::UserMessage(action))
-                                                .is_err()
-                                            {
-                                                generation.results.send_error(crate::Error::Protocol(
-                                                    "UserMessageAction arrived without an active Run"
-                                                        .into(),
-                                                ));
-                                            }
+                                            let mut request = previous.request.clone();
+                                            request.action = Some(conversation_action);
+                                            request.conversation_state = None;
+                                            request.pre_fetched_blobs.clear();
+                                            waiting_for_action = false;
+                                            start_generation(
+                                                &registry,
+                                                &handle,
+                                                &dependencies,
+                                                &blob_sync,
+                                                &context_sync,
+                                                &tool_runtime_factory,
+                                                &mut current,
+                                                &mut next_generation,
+                                                request,
+                                            )
+                                            .await;
                                         }
                                         Some(pb::conversation_action::Action::CancelAction(_)) => {
                                             if let Some(generation) = current.as_ref() {
@@ -427,6 +510,95 @@ impl ConversationRuntime {
     }
 }
 
+#[allow(clippy::too_many_arguments)]
+async fn start_generation(
+    registry: &ConversationRegistry,
+    handle: &TransportHandle,
+    dependencies: &ConversationDependencies,
+    blob_sync: &BlobSynchronizer,
+    context_sync: &RequestContextSynchronizer,
+    tool_runtime_factory: &CursorToolRuntime,
+    current: &mut Option<RunGeneration>,
+    next_generation: &mut u64,
+    request: pb::AgentRunRequest,
+) {
+    let previous_finished = if let Some(previous) = current.take() {
+        previous.superseded.cancel();
+        if let Some(run) = previous.run.lock().clone() {
+            run.cancel();
+        }
+        for id in previous.tool_runtime.interrupt_for_run_replacement().await {
+            let _ = handle.emit(&codec::abort(id));
+        }
+        Some(previous.finished.clone())
+    } else {
+        None
+    };
+    let (results, result_receiver) = tool_result_channel();
+    let (runtime_actions, runtime_action_receiver) =
+        mpsc::unbounded_channel::<compile::RuntimeAction>();
+    let tool_runtime = tool_runtime_factory.next_run();
+    let tools = ToolDispatcher::with_results(
+        tool_runtime.clone(),
+        results.clone(),
+        dependencies.store.clone(),
+        dependencies.web_cache.clone(),
+    );
+    let generation = RunGeneration {
+        id: *next_generation,
+        request: request.clone(),
+        superseded: CancellationToken::new(),
+        finished: CancellationToken::new(),
+        run: Arc::new(parking_lot::Mutex::new(None)),
+        results,
+        runtime_actions,
+        tool_runtime,
+        tools,
+    };
+    *next_generation = next_generation.saturating_add(1);
+    *current = Some(generation.clone());
+    spawn_run_request(
+        registry.clone(),
+        handle.clone(),
+        request,
+        dependencies.clone(),
+        blob_sync.clone(),
+        context_sync.clone(),
+        generation,
+        previous_finished,
+        result_receiver,
+        runtime_action_receiver,
+    );
+}
+
+fn finish_pending(
+    handle: &TransportHandle,
+    current: &Option<RunGeneration>,
+    pending: Option<(u64, TransportFinish)>,
+) {
+    let Some((generation, finish)) = pending else {
+        return;
+    };
+    if current
+        .as_ref()
+        .is_some_and(|current| current.id == generation)
+    {
+        finish_transport(handle, finish);
+    }
+}
+
+fn finish_transport(handle: &TransportHandle, finish: TransportFinish) {
+    match finish {
+        TransportFinish::Success => super::finish_success(handle),
+        TransportFinish::Failed(error) => {
+            let _ = super::finish_failed(handle, &error);
+        }
+        TransportFinish::Cancelled => {
+            let _ = super::finish_cancelled(handle);
+        }
+    }
+}
+
 #[allow(clippy::too_many_arguments)]
 fn spawn_run_request(
     registry: ConversationRegistry,
@@ -486,8 +658,12 @@ fn spawn_run_request(
                     %error,
                     "failed to prepare Cursor Run"
                 );
-                let _ = super::finish_failed(&handle, &error);
-                let _ = handle.command(TransportCommand::Close).await;
+                let _ = handle
+                    .command(TransportCommand::RunFinished {
+                        generation: generation.id,
+                        finish: RunFinish::Transport(TransportFinish::Failed(error)),
+                    })
+                    .await;
                 return;
             }
         };
@@ -519,8 +695,12 @@ fn spawn_run_request(
             {
                 CommandResult::Applied | CommandResult::Duplicate => {
                     if !generation.superseded.is_cancelled() {
-                        super::finish_success(&handle);
-                        let _ = handle.command(TransportCommand::Close).await;
+                        let _ = handle
+                            .command(TransportCommand::RunFinished {
+                                generation: generation.id,
+                                finish: RunFinish::Transport(TransportFinish::Success),
+                            })
+                            .await;
                     }
                     return;
                 }
@@ -544,8 +724,12 @@ fn spawn_run_request(
                 }
                 CommandResult::StaleTarget => {
                     if !generation.superseded.is_cancelled() {
-                        super::finish_success(&handle);
-                        let _ = handle.command(TransportCommand::Close).await;
+                        let _ = handle
+                            .command(TransportCommand::RunFinished {
+                                generation: generation.id,
+                                finish: RunFinish::Transport(TransportFinish::Success),
+                            })
+                            .await;
                     }
                     return;
                 }
@@ -604,16 +788,21 @@ fn spawn_run_request(
                 tool_runtime: generation.tool_runtime.clone(),
             },
         );
-        if let Err(error) = output.run().await {
-            if !generation.superseded.is_cancelled() {
-                tracing::error!(
-                    request_id = handle.request_id(),
-                    %error,
-                    "Cursor session failed"
-                );
-                let _ = super::finish_failed(&handle, &error);
+        let finish = match output.run().await {
+            Ok(finish) => finish,
+            Err(error) => {
+                if generation.superseded.is_cancelled() {
+                    RunFinish::Transport(TransportFinish::Cancelled)
+                } else {
+                    tracing::error!(
+                        request_id = handle.request_id(),
+                        %error,
+                        "Cursor session failed"
+                    );
+                    RunFinish::Transport(TransportFinish::Failed(error))
+                }
             }
-        }
+        };
         let _ = core_run.await;
         registry.release(&conversation_id, &run_id).await;
         if generation
@@ -625,7 +814,12 @@ fn spawn_run_request(
             *generation.run.lock() = None;
         }
         if !generation.superseded.is_cancelled() {
-            let _ = handle.command(TransportCommand::Close).await;
+            let _ = handle
+                .command(TransportCommand::RunFinished {
+                    generation: generation.id,
+                    finish,
+                })
+                .await;
         }
     });
 }

server/src/cursor/services/blob_sync.rs

@@ -20,6 +20,9 @@ use crate::{
 
 type BlobSetSender = oneshot::Sender<Result<()>>;
 
+const SET_TIMEOUT: Duration = Duration::from_secs(30 * 60);
+const GET_TIMEOUT: Duration = Duration::from_secs(10 * 60);
+
 #[derive(Clone)]
 pub struct BlobSynchronizer {
     inner: Arc<Inner>,
@@ -73,22 +76,20 @@ impl BlobSynchronizer {
         let id = self.inner.store.put_blob(data, edges).await?;
         let result = self.ensure_set(&id, data).await;
         if let Some(trace) = self.inner.handle.trace() {
-            trace
-                .linked_blob(
-                    "blob_set",
-                    "byok_server",
-                    &id,
-                    serde_json::json!({
-                        "byte_count": data.len(),
-                        "status": if result.is_ok() { "acknowledged" } else { "error" },
-                        "error": result.as_ref().err().map(ToString::to_string),
-                        "edges": edges.iter().map(|edge| serde_json::json!({
-                            "child_blob_id": edge.child.to_base64(),
-                            "field_name": edge.field_name,
-                        })).collect::<Vec<_>>(),
-                    }),
-                )
-                .await;
+            trace.linked_blob(
+                "blob_set",
+                "byok_server",
+                &id,
+                serde_json::json!({
+                    "byte_count": data.len(),
+                    "status": if result.is_ok() { "acknowledged" } else { "error" },
+                    "error": result.as_ref().err().map(ToString::to_string),
+                    "edges": edges.iter().map(|edge| serde_json::json!({
+                        "child_blob_id": edge.child.to_base64(),
+                        "field_name": edge.field_name,
+                    })).collect::<Vec<_>>(),
+                }),
+            );
         }
         result?;
         Ok(id)
@@ -130,7 +131,7 @@ impl BlobSynchronizer {
         let result = tokio::select! {
             result = receiver => result.map_err(|_| Error::Protocol("KV SET response channel closed".into()))?,
             _ = cancellation.cancelled() => Err(Error::Cancelled),
-            _ = tokio::time::sleep(Duration::from_secs(60)) => Err(Error::Protocol(format!("KV SET timed out: {}", blob_id.to_base64()))),
+            _ = tokio::time::sleep(SET_TIMEOUT) => Err(Error::Protocol(format!("KV SET timed out: {}", blob_id.to_base64()))),
         };
         if result.is_err() {
             self.inner.set_requests.lock().await.remove(&id);
@@ -141,18 +142,16 @@ impl BlobSynchronizer {
     pub async fn get(&self, blob_id: &BlobId) -> Result<Option<Vec<u8>>> {
         if let Some(data) = self.inner.store.get_blob(blob_id).await? {
             if let Some(trace) = self.inner.handle.trace() {
-                trace
-                    .linked_blob(
-                        "blob_get",
-                        "byok_server",
-                        blob_id,
-                        serde_json::json!({
-                            "byte_count": data.len(),
-                            "source": "local_store",
-                            "status": "found",
-                        }),
-                    )
-                    .await;
+                trace.linked_blob(
+                    "blob_get",
+                    "byok_server",
+                    blob_id,
+                    serde_json::json!({
+                        "byte_count": data.len(),
+                        "source": "local_store",
+                        "status": "found",
+                    }),
+                );
             }
             return Ok(Some(data));
         }
@@ -183,7 +182,7 @@ impl BlobSynchronizer {
         let result = tokio::select! {
             result = receiver => result.map_err(|_| Error::Protocol("KV GET response channel closed".into()))?,
             _ = cancellation.cancelled() => Err(Error::Cancelled),
-            _ = tokio::time::sleep(Duration::from_secs(60)) => Err(Error::Protocol(format!("KV GET timed out: {}", blob_id.to_base64()))),
+            _ = tokio::time::sleep(GET_TIMEOUT) => Err(Error::Protocol(format!("KV GET timed out: {}", blob_id.to_base64()))),
         };
         if result.is_err() {
             self.inner.get_requests.lock().await.remove(&id);
@@ -191,45 +190,39 @@ impl BlobSynchronizer {
         if let Some(trace) = self.inner.handle.trace() {
             match &result {
                 Ok(Some(data)) => {
-                    trace
-                        .linked_blob(
-                            "blob_get",
-                            "cursor_client",
-                            blob_id,
-                            serde_json::json!({
-                                "byte_count": data.len(),
-                                "source": "cursor_client",
-                                "status": "found",
-                            }),
-                        )
-                        .await;
+                    trace.linked_blob(
+                        "blob_get",
+                        "cursor_client",
+                        blob_id,
+                        serde_json::json!({
+                            "byte_count": data.len(),
+                            "source": "cursor_client",
+                            "status": "found",
+                        }),
+                    );
                 }
                 Ok(None) => {
-                    trace
-                        .artifact(
-                            "blob_get",
-                            "cursor_client",
-                            &[],
-                            serde_json::json!({
-                                "blob_id": blob_id.to_base64(),
-                                "status": "missing",
-                            }),
-                        )
-                        .await;
+                    trace.artifact(
+                        "blob_get",
+                        "cursor_client",
+                        &[],
+                        serde_json::json!({
+                            "blob_id": blob_id.to_base64(),
+                            "status": "missing",
+                        }),
+                    );
                 }
                 Err(error) => {
-                    trace
-                        .artifact(
-                            "blob_get",
-                            "cursor_client",
-                            &[],
-                            serde_json::json!({
-                                "blob_id": blob_id.to_base64(),
-                                "status": "error",
-                                "error": error.to_string(),
-                            }),
-                        )
-                        .await;
+                    trace.artifact(
+                        "blob_get",
+                        "cursor_client",
+                        &[],
+                        serde_json::json!({
+                            "blob_id": blob_id.to_base64(),
+                            "status": "error",
+                            "error": error.to_string(),
+                        }),
+                    );
                 }
             }
         }
@@ -318,3 +311,18 @@ impl BlobSynchronizer {
         Ok(())
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn set_timeout_allows_slow_cursor_acknowledgements() {
+        assert_eq!(SET_TIMEOUT, Duration::from_secs(30 * 60));
+    }
+
+    #[test]
+    fn get_timeout_allows_slow_cursor_responses() {
+        assert_eq!(GET_TIMEOUT, Duration::from_secs(10 * 60));
+    }
+}

server/src/cursor/services/model_catalog.rs

@@ -580,14 +580,9 @@ fn available_plugin_model(model: &PluginModelDescriptor) -> AvailableModel {
     let tooltip = TooltipData {
         markdown_content: model.description.clone(),
     };
-    let contexts = context_options(model.context_window_tokens);
-    let variants = model_variants(
-        &model.id,
-        &model.display_name,
-        &tooltip,
-        &contexts,
-        model.thinking,
-    );
+    // Effort 与上下文档位由宿主统一提供,与内置模型一致;插件不再声明这两项。
+    let contexts = context_options(None);
+    let variants = model_variants(&model.id, &model.display_name, &tooltip, &contexts, true);
     let legacy_slugs = variants
         .iter()
         .filter_map(|variant| variant.legacy_slug.clone())
@@ -598,7 +593,7 @@ fn available_plugin_model(model: &PluginModelDescriptor) -> AvailableModel {
         supports_agent: Some(true),
         degradation_status: Some(0),
         tooltip_data: Some(tooltip.clone()),
-        supports_thinking: Some(model.thinking),
+        supports_thinking: Some(true),
         supports_images: Some(model.images),
         supports_max_mode: Some(false),
         client_display_name: Some(model.display_name.clone()),
@@ -610,7 +605,7 @@ fn available_plugin_model(model: &PluginModelDescriptor) -> AvailableModel {
         inputbox_short_model_name: Some(model.display_name.clone()),
         supports_sandboxing: Some(true),
         supports_cmd_k: Some(false),
-        parameter_definitions: model_parameters(&contexts, model.thinking),
+        parameter_definitions: model_parameters(&contexts, true),
         variants,
         legacy_slugs,
         named_model_section_index: Some(1),
@@ -633,7 +628,7 @@ fn usable_plugin_model(model: &PluginModelDescriptor) -> agent::ModelDetails {
         display_model_id: model.id.clone(),
         display_name: model.display_name.clone(),
         display_name_short: model.display_name.clone(),
-        thinking_details: model.thinking.then(agent::ThinkingDetails::default),
+        thinking_details: Some(agent::ThinkingDetails::default()),
         ..Default::default()
     }
 }

server/src/cursor/services/observability.rs

@@ -1,225 +0,0 @@
-//! Records Cursor request traces and artifacts.
-use std::{
-    sync::{
-        atomic::{AtomicBool, Ordering},
-        Arc,
-    },
-    time::{Duration, Instant},
-};
-
-use tokio::sync::Mutex;
-
-use crate::store::{BlobId, BufferedCursorTraceChunk, Store};
-
-#[derive(Clone)]
-pub struct CursorTraceRecorder {
-    store: Store,
-    request_id: String,
-    chunks: Arc<Mutex<TraceChunkBuffer>>,
-    finished: Arc<AtomicBool>,
-}
-
-#[derive(Default)]
-struct TraceChunkBuffer {
-    chunks: Vec<BufferedCursorTraceChunk>,
-    bytes: usize,
-    first_chunk_at: Option<Instant>,
-    generation: u64,
-}
-
-const MAX_BUFFERED_CHUNKS: usize = 32;
-const MAX_BUFFERED_BYTES: usize = 256 * 1024;
-const MAX_BUFFER_AGE: Duration = Duration::from_millis(50);
-
-impl CursorTraceRecorder {
-    pub async fn begin(
-        store: Store,
-        request_id: &str,
-        conversation_id: Option<&str>,
-        route: &str,
-        model_id: Option<&str>,
-    ) -> Option<Self> {
-        match store
-            .start_cursor_trace_if_detailed(request_id, conversation_id, route, model_id)
-            .await
-        {
-            Ok(true) => Some(Self {
-                store,
-                request_id: request_id.into(),
-                chunks: Arc::new(Mutex::new(TraceChunkBuffer::default())),
-                finished: Arc::new(AtomicBool::new(false)),
-            }),
-            Ok(false) => None,
-            Err(error) => {
-                tracing::warn!(request_id, %error, "failed to start Cursor trace");
-                None
-            }
-        }
-    }
-
-    pub async fn resume(store: Store, request_id: &str) -> Option<Self> {
-        match store.cursor_trace_exists(request_id).await {
-            Ok(true) => Some(Self {
-                store,
-                request_id: request_id.into(),
-                chunks: Arc::new(Mutex::new(TraceChunkBuffer::default())),
-                finished: Arc::new(AtomicBool::new(false)),
-            }),
-            Ok(false) => None,
-            Err(error) => {
-                tracing::warn!(request_id, %error, "failed to resume Cursor trace");
-                None
-            }
-        }
-    }
-
-    pub fn request_id(&self) -> &str {
-        &self.request_id
-    }
-
-    pub async fn request(&self, artifact_type: &str, data: &[u8], metadata: serde_json::Value) {
-        if let Err(error) = self
-            .store
-            .append_cursor_trace_artifact(
-                &self.request_id,
-                artifact_type,
-                "cursor_client",
-                data,
-                &metadata,
-            )
-            .await
-        {
-            tracing::warn!(request_id = self.request_id, %error, "failed to record Cursor request artifact");
-            return;
-        }
-        if let Err(error) = self
-            .store
-            .add_cursor_trace_request_bytes(&self.request_id, data.len())
-            .await
-        {
-            tracing::warn!(request_id = self.request_id, %error, "failed to update Cursor request trace size");
-        }
-    }
-
-    pub async fn artifact(
-        &self,
-        artifact_type: &str,
-        source: &str,
-        data: &[u8],
-        metadata: serde_json::Value,
-    ) {
-        if let Err(error) = self
-            .store
-            .append_cursor_trace_artifact(&self.request_id, artifact_type, source, data, &metadata)
-            .await
-        {
-            tracing::warn!(request_id = self.request_id, %error, artifact_type, "failed to record Cursor trace artifact");
-        }
-    }
-
-    pub async fn linked_blob(
-        &self,
-        artifact_type: &str,
-        source: &str,
-        blob_id: &BlobId,
-        metadata: serde_json::Value,
-    ) {
-        if let Err(error) = self
-            .store
-            .link_cursor_trace_artifact(&self.request_id, artifact_type, source, blob_id, &metadata)
-            .await
-        {
-            tracing::warn!(request_id = self.request_id, %error, artifact_type, "failed to link Cursor trace Blob");
-        }
-    }
-
-    pub async fn response_started(&self, status: u16) {
-        if let Err(error) = self
-            .store
-            .start_cursor_trace_response(&self.request_id, status)
-            .await
-        {
-            tracing::warn!(request_id = self.request_id, %error, "failed to start Cursor response trace");
-        }
-    }
-
-    pub async fn response_chunk(&self, source: &str, data: &[u8]) {
-        let mut buffer = self.chunks.lock().await;
-        if self.finished.load(Ordering::Acquire) {
-            return;
-        }
-        let schedule_flush = if buffer.chunks.is_empty() {
-            buffer.generation = buffer.generation.wrapping_add(1);
-            buffer.first_chunk_at = Some(Instant::now());
-            Some(buffer.generation)
-        } else {
-            None
-        };
-        buffer.bytes += data.len();
-        buffer
-            .chunks
-            .push(BufferedCursorTraceChunk::new(source, data));
-        let expired = buffer
-            .first_chunk_at
-            .is_some_and(|started| started.elapsed() >= MAX_BUFFER_AGE);
-        if buffer.chunks.len() >= MAX_BUFFERED_CHUNKS
-            || buffer.bytes >= MAX_BUFFERED_BYTES
-            || expired
-        {
-            if let Err(error) = self.flush_locked(&mut buffer).await {
-                tracing::warn!(request_id = self.request_id, %error, "failed to record Cursor response chunk");
-            }
-        }
-        drop(buffer);
-        if let Some(generation) = schedule_flush {
-            let recorder = self.clone();
-            tokio::spawn(async move {
-                tokio::time::sleep(MAX_BUFFER_AGE).await;
-                let mut buffer = recorder.chunks.lock().await;
-                if buffer.generation == generation {
-                    if let Err(error) = recorder.flush_locked(&mut buffer).await {
-                        tracing::warn!(request_id = recorder.request_id, %error, "failed to flush Cursor response chunks");
-                    }
-                }
-            });
-        }
-    }
-
-    pub async fn finish(&self, error: Option<&str>) {
-        if self.finished.swap(true, Ordering::AcqRel) {
-            return;
-        }
-        let mut buffer = self.chunks.lock().await;
-        if let Err(store_error) = self.flush_locked(&mut buffer).await {
-            tracing::warn!(request_id = self.request_id, %store_error, "failed to flush Cursor response chunks");
-        }
-        drop(buffer);
-        if let Err(store_error) = self
-            .store
-            .finish_cursor_trace(&self.request_id, error)
-            .await
-        {
-            tracing::warn!(request_id = self.request_id, %store_error, "failed to finish Cursor trace");
-        }
-    }
-
-    async fn flush_locked(&self, buffer: &mut TraceChunkBuffer) -> crate::Result<()> {
-        if buffer.chunks.is_empty() {
-            return Ok(());
-        }
-        let chunks = std::mem::take(&mut buffer.chunks);
-        buffer.bytes = 0;
-        buffer.first_chunk_at = None;
-        if let Err(error) = self
-            .store
-            .add_cursor_trace_response_chunks(&self.request_id, &chunks)
-            .await
-        {
-            buffer.bytes = chunks.iter().map(|chunk| chunk.data.len()).sum();
-            buffer.first_chunk_at = Some(Instant::now());
-            buffer.chunks = chunks;
-            return Err(error);
-        }
-        Ok(())
-    }
-}

server/src/cursor/services/observability/event.rs

@@ -0,0 +1,71 @@
+use std::sync::{atomic::AtomicU8, Arc};
+
+use bytes::Bytes;
+
+use crate::store::BlobId;
+
+pub(super) const TRACE_UNKNOWN: u8 = 0;
+pub(super) const TRACE_ACTIVE: u8 = 1;
+pub(super) const TRACE_DISABLED: u8 = 2;
+
+pub(super) enum TraceEvent {
+    Begin {
+        request_id: String,
+        activation: Arc<AtomicU8>,
+        conversation_id: Option<String>,
+        route: String,
+        model_id: Option<String>,
+    },
+    Resume {
+        request_id: String,
+        activation: Arc<AtomicU8>,
+    },
+    Request {
+        request_id: String,
+        artifact_type: String,
+        data: Bytes,
+        metadata: serde_json::Value,
+    },
+    Artifact {
+        request_id: String,
+        artifact_type: String,
+        source: String,
+        data: Bytes,
+        metadata: serde_json::Value,
+    },
+    LinkedBlob {
+        request_id: String,
+        artifact_type: String,
+        source: String,
+        blob_id: BlobId,
+        metadata: serde_json::Value,
+    },
+    ResponseStarted {
+        request_id: String,
+        status: u16,
+    },
+    ResponseChunk {
+        request_id: String,
+        source: String,
+        data: Bytes,
+    },
+    Finish {
+        request_id: String,
+        error: Option<String>,
+    },
+}
+
+impl TraceEvent {
+    pub(super) fn request_id(&self) -> &str {
+        match self {
+            Self::Begin { request_id, .. }
+            | Self::Resume { request_id, .. }
+            | Self::Request { request_id, .. }
+            | Self::Artifact { request_id, .. }
+            | Self::LinkedBlob { request_id, .. }
+            | Self::ResponseStarted { request_id, .. }
+            | Self::ResponseChunk { request_id, .. }
+            | Self::Finish { request_id, .. } => request_id,
+        }
+    }
+}

server/src/cursor/services/observability/mod.rs

@@ -0,0 +1,157 @@
+//! Records Cursor request traces without blocking request or runtime paths.
+
+mod event;
+mod worker;
+
+use std::sync::{
+    atomic::{AtomicBool, AtomicU8, Ordering},
+    Arc,
+};
+
+use bytes::Bytes;
+use tokio::sync::mpsc;
+
+use crate::store::{BlobId, Store};
+
+use event::{TraceEvent, TRACE_DISABLED, TRACE_UNKNOWN};
+
+const TRACE_QUEUE_CAPACITY: usize = 512;
+
+#[derive(Clone)]
+pub struct CursorTraceService {
+    sender: mpsc::Sender<TraceEvent>,
+}
+
+impl CursorTraceService {
+    pub fn new(store: Store) -> Self {
+        let (sender, receiver) = mpsc::channel(TRACE_QUEUE_CAPACITY);
+        tokio::spawn(worker::run(store, receiver));
+        Self { sender }
+    }
+
+    pub fn recorder(&self, request_id: &str) -> CursorTraceRecorder {
+        CursorTraceRecorder {
+            request_id: Arc::from(request_id),
+            sender: self.sender.clone(),
+            finished: Arc::new(AtomicBool::new(false)),
+            activation: Arc::new(AtomicU8::new(TRACE_UNKNOWN)),
+        }
+    }
+}
+
+#[derive(Clone)]
+pub struct CursorTraceRecorder {
+    request_id: Arc<str>,
+    sender: mpsc::Sender<TraceEvent>,
+    finished: Arc<AtomicBool>,
+    activation: Arc<AtomicU8>,
+}
+
+impl CursorTraceRecorder {
+    pub fn request_id(&self) -> &str {
+        &self.request_id
+    }
+
+    pub fn begin(&self, conversation_id: Option<&str>, route: &str, model_id: Option<&str>) {
+        self.send_control(TraceEvent::Begin {
+            request_id: self.request_id.to_string(),
+            activation: self.activation.clone(),
+            conversation_id: conversation_id.map(str::to_owned),
+            route: route.to_owned(),
+            model_id: model_id.map(str::to_owned),
+        });
+    }
+
+    pub fn resume(&self) {
+        self.send_control(TraceEvent::Resume {
+            request_id: self.request_id.to_string(),
+            activation: self.activation.clone(),
+        });
+    }
+
+    pub fn request(&self, artifact_type: &str, data: Bytes, metadata: serde_json::Value) {
+        self.send(TraceEvent::Request {
+            request_id: self.request_id.to_string(),
+            artifact_type: artifact_type.to_owned(),
+            data,
+            metadata,
+        });
+    }
+
+    pub fn artifact(
+        &self,
+        artifact_type: &str,
+        source: &str,
+        data: &[u8],
+        metadata: serde_json::Value,
+    ) {
+        self.send(TraceEvent::Artifact {
+            request_id: self.request_id.to_string(),
+            artifact_type: artifact_type.to_owned(),
+            source: source.to_owned(),
+            data: Bytes::copy_from_slice(data),
+            metadata,
+        });
+    }
+
+    pub fn linked_blob(
+        &self,
+        artifact_type: &str,
+        source: &str,
+        blob_id: &BlobId,
+        metadata: serde_json::Value,
+    ) {
+        self.send(TraceEvent::LinkedBlob {
+            request_id: self.request_id.to_string(),
+            artifact_type: artifact_type.to_owned(),
+            source: source.to_owned(),
+            blob_id: blob_id.clone(),
+            metadata,
+        });
+    }
+
+    pub fn response_started(&self, status: u16) {
+        self.send(TraceEvent::ResponseStarted {
+            request_id: self.request_id.to_string(),
+            status,
+        });
+    }
+
+    pub fn response_chunk(&self, source: &str, data: Bytes) {
+        if self.finished.load(Ordering::Acquire) {
+            return;
+        }
+        self.send(TraceEvent::ResponseChunk {
+            request_id: self.request_id.to_string(),
+            source: source.to_owned(),
+            data,
+        });
+    }
+
+    pub fn finish(&self, error: Option<&str>) {
+        if self.finished.swap(true, Ordering::AcqRel) {
+            return;
+        }
+        self.send_control(TraceEvent::Finish {
+            request_id: self.request_id.to_string(),
+            error: error.map(str::to_owned),
+        });
+    }
+
+    fn send(&self, event: TraceEvent) {
+        if self.activation.load(Ordering::Acquire) == TRACE_DISABLED {
+            return;
+        }
+        self.send_control(event);
+    }
+
+    fn send_control(&self, event: TraceEvent) {
+        if let Err(error) = self.sender.try_send(event) {
+            tracing::warn!(
+                request_id = %self.request_id,
+                %error,
+                "dropping Cursor trace event"
+            );
+        }
+    }
+}

server/src/cursor/services/observability/worker.rs

@@ -0,0 +1,349 @@
+use std::{
+    collections::{BTreeMap, HashMap},
+    sync::atomic::Ordering,
+    time::Duration,
+};
+
+use tokio::sync::mpsc;
+
+use crate::store::{BufferedCursorTraceChunk, Store};
+
+use super::event::{TraceEvent, TRACE_ACTIVE, TRACE_DISABLED};
+
+const MAX_BUFFERED_CHUNKS: usize = 32;
+const MAX_BUFFERED_BYTES: usize = 256 * 1024;
+const FLUSH_INTERVAL: Duration = Duration::from_millis(50);
+
+#[derive(Clone, Copy)]
+enum TraceState {
+    Active,
+    Disabled,
+}
+
+#[derive(Default)]
+struct ResponseBuffer {
+    chunks: Vec<BufferedCursorTraceChunk>,
+    bytes: usize,
+}
+
+struct BufferedRequest {
+    artifact_type: String,
+    data: bytes::Bytes,
+    metadata: serde_json::Value,
+}
+
+#[derive(Default)]
+struct RequestOrder {
+    next: i64,
+    pending: BTreeMap<i64, Vec<BufferedRequest>>,
+}
+
+pub(super) async fn run(store: Store, mut receiver: mpsc::Receiver<TraceEvent>) {
+    let mut states = HashMap::<String, TraceState>::new();
+    let mut buffers = HashMap::<String, ResponseBuffer>::new();
+    let mut request_orders = HashMap::<String, RequestOrder>::new();
+    let mut interval = tokio::time::interval(FLUSH_INTERVAL);
+    interval.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Skip);
+
+    loop {
+        tokio::select! {
+            event = receiver.recv() => {
+                let Some(event) = event else {
+                    flush_all(&store, &mut buffers).await;
+                    return;
+                };
+                process(&store, &mut states, &mut buffers, &mut request_orders, event).await;
+            }
+            _ = interval.tick() => flush_all(&store, &mut buffers).await,
+        }
+    }
+}
+
+async fn process(
+    store: &Store,
+    states: &mut HashMap<String, TraceState>,
+    buffers: &mut HashMap<String, ResponseBuffer>,
+    request_orders: &mut HashMap<String, RequestOrder>,
+    event: TraceEvent,
+) {
+    let request_id = event.request_id().to_owned();
+    let finishes_trace = matches!(&event, TraceEvent::Finish { .. });
+    match event {
+        TraceEvent::Begin {
+            request_id,
+            activation,
+            conversation_id,
+            route,
+            model_id,
+        } => {
+            let state = match store
+                .start_cursor_trace_if_detailed(
+                    &request_id,
+                    conversation_id.as_deref(),
+                    &route,
+                    model_id.as_deref(),
+                )
+                .await
+            {
+                Ok(true) => TraceState::Active,
+                Ok(false) => TraceState::Disabled,
+                Err(error) => {
+                    tracing::warn!(%request_id, %error, "failed to start Cursor trace");
+                    TraceState::Disabled
+                }
+            };
+            activation.store(
+                match state {
+                    TraceState::Active => TRACE_ACTIVE,
+                    TraceState::Disabled => TRACE_DISABLED,
+                },
+                Ordering::Release,
+            );
+            states.insert(request_id, state);
+            return;
+        }
+        TraceEvent::Resume {
+            request_id,
+            activation,
+        } => {
+            let state = ensure_state(store, states, &request_id).await;
+            activation.store(
+                match state {
+                    TraceState::Active => TRACE_ACTIVE,
+                    TraceState::Disabled => TRACE_DISABLED,
+                },
+                Ordering::Release,
+            );
+            return;
+        }
+        _ => {}
+    }
+
+    if !matches!(
+        ensure_state(store, states, &request_id).await,
+        TraceState::Active
+    ) {
+        if finishes_trace {
+            states.remove(&request_id);
+            buffers.remove(&request_id);
+            request_orders.remove(&request_id);
+        }
+        return;
+    }
+
+    let result = match event {
+        TraceEvent::Request {
+            artifact_type,
+            data,
+            metadata,
+            ..
+        } => {
+            append_request(
+                store,
+                request_orders,
+                &request_id,
+                artifact_type,
+                data,
+                metadata,
+            )
+            .await
+        }
+        TraceEvent::Artifact {
+            artifact_type,
+            source,
+            data,
+            metadata,
+            ..
+        } => {
+            store
+                .append_cursor_trace_artifact(
+                    &request_id,
+                    &artifact_type,
+                    &source,
+                    &data,
+                    &metadata,
+                )
+                .await
+        }
+        TraceEvent::LinkedBlob {
+            artifact_type,
+            source,
+            blob_id,
+            metadata,
+            ..
+        } => {
+            store
+                .link_cursor_trace_artifact(
+                    &request_id,
+                    &artifact_type,
+                    &source,
+                    &blob_id,
+                    &metadata,
+                )
+                .await
+        }
+        TraceEvent::ResponseStarted { status, .. } => {
+            store.start_cursor_trace_response(&request_id, status).await
+        }
+        TraceEvent::ResponseChunk { source, data, .. } => {
+            let buffer = buffers.entry(request_id.clone()).or_default();
+            buffer.bytes += data.len();
+            buffer
+                .chunks
+                .push(BufferedCursorTraceChunk::new(&source, &data));
+            if buffer.chunks.len() >= MAX_BUFFERED_CHUNKS || buffer.bytes >= MAX_BUFFERED_BYTES {
+                flush_one(store, buffers, &request_id).await;
+            }
+            return;
+        }
+        TraceEvent::Finish { error, .. } => {
+            flush_request_order(store, request_orders, &request_id).await;
+            flush_one(store, buffers, &request_id).await;
+            store
+                .finish_cursor_trace(&request_id, error.as_deref())
+                .await
+        }
+        TraceEvent::Begin { .. } | TraceEvent::Resume { .. } => unreachable!(),
+    };
+    if let Err(error) = result {
+        tracing::warn!(%request_id, %error, "failed to record Cursor trace event");
+    }
+    if finishes_trace {
+        states.remove(&request_id);
+        buffers.remove(&request_id);
+    }
+}
+
+async fn append_request(
+    store: &Store,
+    request_orders: &mut HashMap<String, RequestOrder>,
+    request_id: &str,
+    artifact_type: String,
+    data: bytes::Bytes,
+    metadata: serde_json::Value,
+) -> crate::Result<()> {
+    let append_seqno = metadata
+        .get("append_seqno")
+        .and_then(serde_json::Value::as_i64);
+    let ordered = artifact_type == "bidi_request"
+        && metadata
+            .get("accepted")
+            .and_then(serde_json::Value::as_bool)
+            == Some(true)
+        && metadata
+            .get("route_outcome")
+            .and_then(serde_json::Value::as_str)
+            == Some("local");
+    let Some(append_seqno) = append_seqno.filter(|_| ordered) else {
+        return store
+            .append_cursor_trace_request(
+                request_id,
+                &artifact_type,
+                "cursor_client",
+                &data,
+                &metadata,
+            )
+            .await;
+    };
+
+    let request = BufferedRequest {
+        artifact_type,
+        data,
+        metadata,
+    };
+    let order = request_orders.entry(request_id.to_owned()).or_default();
+    if append_seqno < order.next {
+        return store
+            .append_cursor_trace_request(
+                request_id,
+                &request.artifact_type,
+                "cursor_client",
+                &request.data,
+                &request.metadata,
+            )
+            .await;
+    }
+    order.pending.entry(append_seqno).or_default().push(request);
+    while let Some(requests) = order.pending.remove(&order.next) {
+        for request in requests {
+            store
+                .append_cursor_trace_request(
+                    request_id,
+                    &request.artifact_type,
+                    "cursor_client",
+                    &request.data,
+                    &request.metadata,
+                )
+                .await?;
+        }
+        order.next = order.next.saturating_add(1);
+    }
+    Ok(())
+}
+
+async fn flush_request_order(
+    store: &Store,
+    request_orders: &mut HashMap<String, RequestOrder>,
+    request_id: &str,
+) {
+    let Some(order) = request_orders.remove(request_id) else {
+        return;
+    };
+    for requests in order.pending.into_values() {
+        for request in requests {
+            if let Err(error) = store
+                .append_cursor_trace_request(
+                    request_id,
+                    &request.artifact_type,
+                    "cursor_client",
+                    &request.data,
+                    &request.metadata,
+                )
+                .await
+            {
+                tracing::warn!(%request_id, %error, "failed to flush ordered Cursor request trace");
+            }
+        }
+    }
+}
+
+async fn ensure_state(
+    store: &Store,
+    states: &mut HashMap<String, TraceState>,
+    request_id: &str,
+) -> TraceState {
+    if let Some(state) = states.get(request_id).copied() {
+        return state;
+    }
+    let state = match store.cursor_trace_exists(request_id).await {
+        Ok(true) => TraceState::Active,
+        Ok(false) => TraceState::Disabled,
+        Err(error) => {
+            tracing::warn!(%request_id, %error, "failed to resume Cursor trace");
+            TraceState::Disabled
+        }
+    };
+    states.insert(request_id.to_owned(), state);
+    state
+}
+
+async fn flush_one(store: &Store, buffers: &mut HashMap<String, ResponseBuffer>, request_id: &str) {
+    let Some(mut buffer) = buffers.remove(request_id) else {
+        return;
+    };
+    if let Err(error) = store
+        .add_cursor_trace_response_chunks(request_id, &buffer.chunks)
+        .await
+    {
+        tracing::warn!(%request_id, %error, "failed to flush Cursor response chunks");
+        buffer.bytes = buffer.chunks.iter().map(|chunk| chunk.data.len()).sum();
+        buffers.insert(request_id.to_owned(), buffer);
+    }
+}
+
+async fn flush_all(store: &Store, buffers: &mut HashMap<String, ResponseBuffer>) {
+    let request_ids = buffers.keys().cloned().collect::<Vec<_>>();
+    for request_id in request_ids {
+        flush_one(store, buffers, &request_id).await;
+    }
+}

server/src/cursor/tools/codec/mod.rs

@@ -12,7 +12,7 @@ use crate::{
 };
 
 pub use query::tool_query;
-pub(crate) use render::{create_plan_partial, edit_content_delta, edit_path_partial};
+pub(crate) use render::{create_plan_partial, edit_content_delta, edit_path_partial, task_partial};
 pub use render::{dynamic_mcp_placeholder, render_dynamic_mcp, tool_completed};
 pub use request::{abort, mcp_request, mcp_state_request, request};
 pub(crate) use request::{edit_read_request, json_object_to_prost, mcp_meta_request};

server/src/cursor/tools/codec/render.rs

@@ -54,6 +54,49 @@ pub(crate) fn edit_content_delta(call: &ToolCall, content: String) -> pb::AgentS
     )))
 }
 
+pub(crate) fn task_partial(
+    call: &ToolCall,
+    description: &str,
+    prompt: &str,
+    subagent: &str,
+    model: &str,
+    resume: &str,
+    environment: &str,
+) -> pb::AgentServerMessage {
+    server_interaction(pb::interaction_update::Message::PartialToolCall(
+        pb::PartialToolCallUpdate {
+            call_id: call.call_id.clone(),
+            tool_call: Some(pb::ToolCall {
+                hook_additional_contexts: Vec::new(),
+                tool_call_id: Some(call.call_id.clone()),
+                started_at_ms: None,
+                completed_at_ms: None,
+                tool: Some(pb::tool_call::Tool::TaskToolCall(pb::TaskToolCall {
+                    args: Some(pb::TaskArgs {
+                        description: description.into(),
+                        prompt: prompt.into(),
+                        subagent_type: Some(subagent_type(subagent)),
+                        model: (!model.is_empty()).then(|| model.into()),
+                        resume: (!resume.is_empty()).then(|| resume.into()),
+                        agent_id: None,
+                        attachments: Vec::new(),
+                        mode: 0,
+                        responding_to_message_ids: Vec::new(),
+                        environment: execution_environment(
+                            (!environment.is_empty()).then_some(environment),
+                        ),
+                        machine: None,
+                    }),
+                    result: None,
+                    ..Default::default()
+                })),
+            }),
+            args_text_delta: String::new(),
+            model_call_id: call.model_call_id.clone(),
+        },
+    ))
+}
+
 pub(crate) fn create_plan_partial(
     call: &ToolCall,
     name: &str,
@@ -167,7 +210,7 @@ pub fn tool_completed(call: &ToolCall, completion: &ToolCompletion) -> pb::Agent
 pub fn tool_placeholder(name: &str, call_id: &str) -> Result<pb::ToolCall> {
     use pb::tool_call::Tool;
     let tool = match normalized(name).as_str() {
-        "shell" => Tool::ShellToolCall(pb::ShellToolCall::default()),
+        "shell" | "bash" => Tool::ShellToolCall(pb::ShellToolCall::default()),
         "delete" => Tool::DeleteToolCall(pb::DeleteToolCall::default()),
         "glob" => Tool::GlobToolCall(pb::GlobToolCall::default()),
         "grep" => Tool::GrepToolCall(pb::GrepToolCall::default()),
@@ -527,3 +570,23 @@ fn now_ms() -> u64 {
         .unwrap_or_default()
         .as_millis() as u64
 }
+
+#[cfg(test)]
+mod tests {
+    use super::tool_placeholder;
+    use crate::cursor::protocol::proto::agent::v1 as pb;
+
+    #[test]
+    fn bash_renders_as_a_shell_placeholder() {
+        // The dispatcher treats `bash`/`Bash` as a Shell alias, so the streaming
+        // placeholder must too; otherwise a `Bash` tool call aborts the turn with
+        // `unsupported tool: bash` before it ever runs.
+        for name in ["shell", "Shell", "bash", "Bash"] {
+            let tool = tool_placeholder(name, "call-1").unwrap().tool;
+            assert!(
+                matches!(tool, Some(pb::tool_call::Tool::ShellToolCall(_))),
+                "{name} should render as a Shell tool"
+            );
+        }
+    }
+}

server/src/cursor/tools/codec/request.rs

@@ -35,7 +35,7 @@ pub fn request(id: u32, call: &ToolCall, context: &ExecContext) -> Result<pb::Ag
             .map(|v| v as i32)
     };
     let message = match normalize(&call.name).as_str() {
-        "shell" => {
+        "shell" | "bash" => {
             let command = string("command")?;
             let (simple_commands, parsing_result) = shell_command_metadata(&command);
             Message::ShellStreamArgs(pb::ShellArgs {
@@ -520,3 +520,38 @@ fn prost_value(value: &Value) -> prost_types::Value {
     };
     ProstValue { kind: Some(kind) }
 }
+
+#[cfg(test)]
+mod tests {
+    use serde_json::json;
+
+    use super::request;
+    use crate::cursor::protocol::proto::agent::v1 as pb;
+    use crate::cursor::tools::runtime::ExecContext;
+    use crate::model::ToolCall;
+
+    #[test]
+    fn bash_is_encoded_as_a_shell_exec_request() {
+        // The dispatcher routes `bash`/`Bash` to the shell executor, so the
+        // request codec must encode it as a Shell stream instead of erroring
+        // with `tool bash is not executed through ExecServerMessage`.
+        let call = ToolCall {
+            index: 0,
+            call_id: "call-1".into(),
+            model_call_id: "model-1".into(),
+            name: "Bash".into(),
+            arguments_text: String::new(),
+            arguments: json!({ "command": "ls -la" }),
+            argument_error: None,
+        };
+        let message = request(1, &call, &ExecContext::default()).unwrap();
+        let Some(pb::agent_server_message::Message::ExecServerMessage(exec)) = message.message
+        else {
+            panic!("expected an ExecServerMessage");
+        };
+        let Some(pb::exec_server_message::Message::ShellStreamArgs(args)) = exec.message else {
+            panic!("expected ShellStreamArgs");
+        };
+        assert_eq!(args.command, "ls -la");
+    }
+}

server/src/cursor/tools/codec/response.rs

@@ -3,7 +3,7 @@ use crate::{
     cursor::{
         protocol::{events, proto::agent::v1 as pb},
         tools::{
-            edit,
+            compat, edit,
             runtime::{CursorToolRuntime, ExecStage, PendingExec},
             tool_call_result::{self as result, ToolCompletion},
         },
@@ -34,16 +34,15 @@ pub async fn client_event(
     let call = match pending.exec_call(message.id).await {
         Some(call) => call,
         None if pending.completed_call(message.id).await.is_some() => {
-            return Err(Error::Protocol(format!(
-                "duplicate terminal ExecClientMessage id: {}",
-                message.id
-            )))
+            tracing::warn!(id = message.id, "ignoring duplicate terminal tool response");
+            return Ok(ClientExecEvent::Pending);
         }
         None => {
-            return Err(Error::Protocol(format!(
-                "unknown ExecClientMessage id: {}",
-                message.id
-            )))
+            tracing::warn!(
+                id = message.id,
+                "ignoring response for unknown tool execution"
+            );
+            return Ok(ClientExecEvent::Pending);
         }
     };
     let Some(wire_result) = &message.message else {
@@ -144,7 +143,8 @@ pub async fn stream_closed(id: u32, pending: &CursorToolRuntime) -> Result<Optio
         return Ok(None);
     };
     let error = "Cursor Exec stream closed before returning a terminal result";
-    if entry.call.name.eq_ignore_ascii_case("Shell") {
+    if entry.call.name.eq_ignore_ascii_case("Shell") || entry.call.name.eq_ignore_ascii_case("Bash")
+    {
         let command = entry
             .call
             .arguments
@@ -173,9 +173,22 @@ pub async fn stream_closed(id: u32, pending: &CursorToolRuntime) -> Result<Optio
     }
     let rendered = match &entry.stage {
         ExecStage::DynamicMcp(definition) => {
-            super::render_dynamic_mcp(&entry.call, definition, false)
+            Ok(super::render_dynamic_mcp(&entry.call, definition, false))
         }
-        _ => super::render_tool_call(&entry.call, false)?,
+        _ => super::render_tool_call(&entry.call, false),
+    };
+    let rendered = match rendered {
+        Ok(rendered) => rendered,
+        Err(Error::Protocol(message)) => {
+            return Ok(Some(compat::failure_with_message(&entry.call, message)));
+        }
+        Err(Error::Json(error)) => {
+            return Ok(Some(compat::failure_with_message(
+                &entry.call,
+                error.to_string(),
+            )));
+        }
+        Err(error) => return Err(error),
     };
     Ok(Some(ToolCompletion::from_rendered(
         &entry.call,
@@ -211,10 +224,10 @@ async fn advance_edit(
         pb::exec_client_message::Message::ReadResult(result)
         | pb::exec_client_message::Message::RedactedReadResult(result) => result,
         _ => {
-            return Err(Error::Protocol(format!(
-                "expected ReadResult for edit tool {}",
-                entry.call.name
-            )))
+            let message = format!("expected ReadResult for edit tool {}", entry.call.name);
+            return Ok(ClientExecEvent::Completed(Box::new(
+                compat::failure_with_message(&entry.call, message),
+            )));
         }
     };
     let write = match edit::after_read(&entry.call, read) {
@@ -259,9 +272,14 @@ fn completed(
     pending: PendingExec,
     result: pb::exec_client_message::Message,
 ) -> Result<ClientExecEvent> {
-    Ok(ClientExecEvent::Completed(Box::new(result::from_exec(
-        pending, &result,
-    )?)))
+    let call = pending.call.clone();
+    let completion = match result::from_exec(pending, &result) {
+        Ok(completion) => completion,
+        Err(Error::Protocol(message)) => compat::failure_with_message(&call, message),
+        Err(Error::Json(error)) => compat::failure_with_message(&call, error.to_string()),
+        Err(error) => return Err(error),
+    };
+    Ok(ClientExecEvent::Completed(Box::new(completion)))
 }
 
 fn shell_exit_result(

server/src/cursor/tools/compat.rs

@@ -49,7 +49,10 @@ pub(crate) fn render(call: &ToolCall, completed: bool) -> pb::ToolCall {
 }
 
 pub(crate) fn failure(call: &ToolCall) -> ToolCompletion {
-    let error = failure_message(&call.name);
+    failure_with_message(call, failure_message(&call.name))
+}
+
+pub(crate) fn failure_with_message(call: &ToolCall, error: String) -> ToolCompletion {
     let arguments = call
         .arguments
         .as_object()

server/src/cursor/tools/edit.rs

@@ -183,6 +183,9 @@ fn edit_notebook(call: &ToolCall, before: &str) -> std::result::Result<String, S
             .unwrap_or_default();
         let old =
             normalize_newlines(&string(call, "old_string").map_err(|error| error.to_string())?);
+        if old.is_empty() {
+            return Err("old_string must not be empty".into());
+        }
         let occurrences = source.match_indices(&old).count();
         let edited = match occurrences {
             0 => return Err("old_string was not found in the notebook cell".into()),
@@ -241,3 +244,50 @@ fn normalized(value: &str) -> String {
         .flat_map(char::to_lowercase)
         .collect()
 }
+
+#[cfg(test)]
+mod tests {
+    use serde_json::json;
+
+    use super::edit_notebook;
+    use crate::model::ToolCall;
+
+    fn notebook_call(old_string: &str) -> ToolCall {
+        ToolCall {
+            index: 0,
+            call_id: "call".into(),
+            model_call_id: "model".into(),
+            name: "EditNotebook".into(),
+            arguments_text: String::new(),
+            arguments: json!({
+                "target_notebook": "/notebook.ipynb",
+                "cell_idx": 0,
+                "old_string": old_string,
+                "new_string": "replacement",
+            }),
+            argument_error: None,
+        }
+    }
+
+    fn single_cell_notebook() -> String {
+        json!({
+            "cells": [{"cell_type": "code", "source": ["print('hi')\n"]}],
+        })
+        .to_string()
+    }
+
+    #[test]
+    fn edit_notebook_rejects_empty_old_string() {
+        // StrReplace rejects an empty old_string; EditNotebook must do the same
+        // instead of prepending new_string (empty cell) or reporting a
+        // misleading "not unique" error (non-empty cell).
+        let error = edit_notebook(&notebook_call(""), &single_cell_notebook()).unwrap_err();
+        assert_eq!(error, "old_string must not be empty");
+    }
+
+    #[test]
+    fn edit_notebook_replaces_a_unique_old_string() {
+        let edited = edit_notebook(&notebook_call("hi"), &single_cell_notebook()).unwrap();
+        assert!(edited.contains("print('replacement')"));
+    }
+}

server/src/cursor/tools/mod.rs

@@ -103,10 +103,20 @@ impl ToolDispatcher {
             }
             let message_index = first_tool_index + position;
             let publish_started = !state.started.contains(&call.call_id);
+            if let Some(error) = &call.argument_error {
+                dispatched.push(validation_failure(call, error.clone()));
+                continue;
+            }
             let edit_path = if dynamic_mcp.contains_key(&call.name) {
                 None
             } else {
-                edit::execution_path(call)?
+                match edit::execution_path(call) {
+                    Ok(path) => path,
+                    Err(error) => {
+                        dispatched.push(recover_validation_failure(call, error)?);
+                        continue;
+                    }
+                }
             };
             if let Some(path) = edit_path {
                 let next = self.edit_schedule.lock().await.start_or_defer(
@@ -121,22 +131,28 @@ impl ToolDispatcher {
                 let Some(next) = next else {
                     continue;
                 };
-                dispatched.push(
-                    self.start(
+                let started = self
+                    .start(
                         &next.call,
                         next.message_index,
                         next.publish_started,
                         dynamic_mcp,
                         &next.context,
                     )
-                    .await?,
-                );
+                    .await;
+                dispatched.push(match started {
+                    Ok(started) => started,
+                    Err(error) => recover_validation_failure(&next.call, error)?,
+                });
                 continue;
             }
-            dispatched.push(
-                self.start(call, message_index, publish_started, dynamic_mcp, context)
-                    .await?,
-            );
+            let started = self
+                .start(call, message_index, publish_started, dynamic_mcp, context)
+                .await;
+            dispatched.push(match started {
+                Ok(started) => started,
+                Err(error) => recover_validation_failure(call, error)?,
+            });
         }
         Ok(dispatched)
     }
@@ -146,15 +162,19 @@ impl ToolDispatcher {
         let Some(next) = next else {
             return Ok(None);
         };
-        self.start(
-            &next.call,
-            next.message_index,
-            next.publish_started,
-            &BTreeMap::new(),
-            &next.context,
-        )
-        .await
-        .map(Some)
+        match self
+            .start(
+                &next.call,
+                next.message_index,
+                next.publish_started,
+                &BTreeMap::new(),
+                &next.context,
+            )
+            .await
+        {
+            Ok(started) => Ok(Some(started)),
+            Err(error) => recover_validation_failure(&next.call, error).map(Some),
+        }
     }
 
     pub async fn interrupt_for_message(&self) -> Vec<u32> {
@@ -203,34 +223,64 @@ impl ToolDispatcher {
         let pending = match self.runtime.take_interaction(response.id).await {
             Some(pending) => pending,
             None if self.runtime.completed_call(response.id).await.is_some() => {
-                return Err(Error::Protocol(format!(
-                    "duplicate terminal InteractionResponse id: {}",
-                    response.id
-                )));
+                tracing::warn!(
+                    id = response.id,
+                    "ignoring duplicate terminal interaction response"
+                );
+                return Ok(ClientToolEvent::Pending);
             }
             None => {
-                return Err(Error::Protocol(format!(
-                    "unknown InteractionResponse id: {}",
-                    response.id
-                )));
+                tracing::warn!(
+                    id = response.id,
+                    "ignoring response for unknown interaction"
+                );
+                return Ok(ClientToolEvent::Pending);
             }
         };
-        Ok(
-            match tool_call_dispatch::resume_interaction(
-                &self.results,
-                &self.search,
-                &self.fetch,
-                pending,
-                response,
-            )
-            .await?
-            {
-                tool_call_dispatch::InteractionContinuation::Completed(completion) => {
-                    ClientToolEvent::Completed(completion)
-                }
-                tool_call_dispatch::InteractionContinuation::Pending => ClientToolEvent::Pending,
-            },
+        let call = pending.call.clone();
+        let continuation = match tool_call_dispatch::resume_interaction(
+            &self.results,
+            &self.search,
+            &self.fetch,
+            pending,
+            response,
         )
+        .await
+        {
+            Ok(continuation) => continuation,
+            Err(Error::Protocol(message)) => {
+                return Ok(ClientToolEvent::Completed(Box::new(
+                    compat::failure_with_message(&call, message),
+                )));
+            }
+            Err(Error::Json(error)) => {
+                return Ok(ClientToolEvent::Completed(Box::new(
+                    compat::failure_with_message(&call, error.to_string()),
+                )));
+            }
+            Err(error) => return Err(error),
+        };
+        Ok(match continuation {
+            tool_call_dispatch::InteractionContinuation::Completed(completion) => {
+                ClientToolEvent::Completed(completion)
+            }
+            tool_call_dispatch::InteractionContinuation::Pending => ClientToolEvent::Pending,
+        })
+    }
+}
+
+fn validation_failure(call: &ToolCall, message: String) -> DispatchedTool {
+    DispatchedTool {
+        messages: Vec::new(),
+        completion: Some(compat::failure_with_message(call, message)),
+    }
+}
+
+fn recover_validation_failure(call: &ToolCall, error: Error) -> Result<DispatchedTool> {
+    match error {
+        Error::Protocol(message) => Ok(validation_failure(call, message)),
+        Error::Json(error) => Ok(validation_failure(call, error.to_string())),
+        error => Err(error),
     }
 }
 

server/src/cursor/tools/stream.rs

@@ -20,6 +20,7 @@ enum Presentation {
     DynamicMcp(pb::McpToolDefinition),
     Edit(EditProjection),
     CreatePlan(CreatePlanProjection),
+    Task(TaskProjection),
 }
 
 struct EditProjection {
@@ -38,6 +39,17 @@ struct CreatePlanProjection {
     overview: String,
 }
 
+#[derive(Default)]
+struct TaskProjection {
+    fields: JsonStringFields,
+    description: String,
+    prompt: String,
+    subagent_type: String,
+    model: String,
+    resume: String,
+    environment: String,
+}
+
 impl ToolCallStream {
     pub fn new(name: &str, dynamic_mcp: Option<&pb::McpToolDefinition>) -> Self {
         let presentation = match dynamic_mcp {
@@ -49,6 +61,7 @@ impl ToolCallStream {
                     Presentation::Edit(EditProjection::new("target_notebook", "new_string"))
                 }
                 "createplan" => Presentation::CreatePlan(CreatePlanProjection::default()),
+                "task" => Presentation::Task(TaskProjection::default()),
                 _ => Presentation::Plain,
             },
         };
@@ -73,10 +86,48 @@ impl ToolCallStream {
                 Ok(messages)
             }
             Presentation::CreatePlan(plan) => plan.project(call, raw_delta),
+            Presentation::Task(task) => task.project(call, raw_delta),
         }
     }
 }
 
+impl TaskProjection {
+    fn project(&mut self, call: &ToolCall, raw_delta: &str) -> Result<Vec<pb::AgentServerMessage>> {
+        let mut description_completed = false;
+        for event in self.fields.push(raw_delta)? {
+            match event {
+                StringFieldEvent::Delta { name, text } => match name.as_str() {
+                    "description" => self.description.push_str(&text),
+                    "prompt" => self.prompt.push_str(&text),
+                    "subagent_type" => self.subagent_type.push_str(&text),
+                    "model" => self.model.push_str(&text),
+                    "resume" => self.resume.push_str(&text),
+                    "environment" => self.environment.push_str(&text),
+                    _ => {}
+                },
+                StringFieldEvent::End { name } if name == "description" => {
+                    description_completed = true
+                }
+                _ => {}
+            }
+        }
+        Ok(description_completed
+            .then(|| {
+                interaction::task_partial(
+                    call,
+                    &self.description,
+                    &self.prompt,
+                    &self.subagent_type,
+                    &self.model,
+                    &self.resume,
+                    &self.environment,
+                )
+            })
+            .into_iter()
+            .collect())
+    }
+}
+
 impl CreatePlanProjection {
     fn project(&mut self, call: &ToolCall, raw_delta: &str) -> Result<Vec<pb::AgentServerMessage>> {
         let mut completed_field = false;
@@ -184,3 +235,54 @@ fn normalized(value: &str) -> String {
         .flat_map(char::to_lowercase)
         .collect()
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn task_call(arguments_text: &str) -> ToolCall {
+        ToolCall {
+            index: 0,
+            call_id: "task-1".into(),
+            model_call_id: "model-1".into(),
+            name: "Task".into(),
+            arguments_text: arguments_text.into(),
+            arguments: serde_json::Value::Null,
+            argument_error: None,
+        }
+    }
+
+    #[test]
+    fn task_description_projects_a_visible_partial_card_before_execution() {
+        let mut stream = ToolCallStream::new("Task", None);
+        let first = r#"{"description":"Review K10"#;
+        assert!(stream
+            .arguments_delta(&task_call(first), first)
+            .unwrap()
+            .is_empty());
+
+        let closing = "\",";
+        let messages = stream
+            .arguments_delta(&task_call(&format!("{first}{closing}")), closing)
+            .unwrap();
+        assert_eq!(messages.len(), 1);
+        let Some(pb::agent_server_message::Message::InteractionUpdate(update)) =
+            messages[0].message.as_ref()
+        else {
+            panic!("expected interaction update")
+        };
+        let Some(pb::interaction_update::Message::PartialToolCall(partial)) =
+            update.message.as_ref()
+        else {
+            panic!("expected partial tool call")
+        };
+        let tool_call = partial.tool_call.as_ref().expect("expected tool call");
+        assert_eq!(tool_call.started_at_ms, None);
+        let Some(pb::tool_call::Tool::TaskToolCall(task)) = tool_call.tool.as_ref() else {
+            panic!("expected Task tool call")
+        };
+        let args = task.args.as_ref().expect("expected partial Task args");
+        assert_eq!(args.description, "Review K10");
+        assert_eq!(args.prompt, "");
+    }
+}

server/src/cursor/tools/tool_call_result/interaction.rs

@@ -406,6 +406,7 @@ mod tests {
                 name: "WebFetch".into(),
                 arguments_text: r#"{"url":"https://example.com"}"#.into(),
                 arguments: json!({"url": "https://example.com"}),
+                argument_error: None,
             },
             started_at_ms: 1,
         }

server/src/cursor/transport/handle.rs

@@ -15,7 +15,7 @@ use crate::{
     Error, Result,
 };
 
-use super::OutputHub;
+use super::{OutputHub, TransportAdmission, TransportLifecycle};
 
 #[derive(Clone, Debug, PartialEq, Eq)]
 pub struct TransportParent {
@@ -30,7 +30,8 @@ pub struct TransportHandle {
     output: Arc<OutputHub>,
     conversation_id: Arc<OnceLock<String>>,
     parent: Arc<OnceLock<TransportParent>>,
-    trace: Option<CursorTraceRecorder>,
+    trace: CursorTraceRecorder,
+    lifecycle: TransportLifecycle,
     disconnect: CancellationToken,
 }
 
@@ -39,7 +40,7 @@ impl TransportHandle {
         request_id: String,
         commands: mpsc::Sender<TransportCommand>,
         output: Arc<OutputHub>,
-        trace: Option<CursorTraceRecorder>,
+        trace: CursorTraceRecorder,
     ) -> Self {
         Self {
             request_id,
@@ -48,6 +49,7 @@ impl TransportHandle {
             conversation_id: Arc::new(OnceLock::new()),
             parent: Arc::new(OnceLock::new()),
             trace,
+            lifecycle: TransportLifecycle::new(),
             disconnect: CancellationToken::new(),
         }
     }
@@ -127,12 +129,46 @@ impl TransportHandle {
         self.output.close()
     }
 
-    pub(crate) async fn wait_closed(&self) {
-        self.output.wait_closed().await;
+    pub(crate) fn trace(&self) -> Option<&CursorTraceRecorder> {
+        Some(&self.trace)
     }
 
-    pub(crate) fn trace(&self) -> Option<&CursorTraceRecorder> {
-        self.trace.as_ref()
+    pub(crate) fn accepting_appends(&self) -> bool {
+        self.lifecycle.is_open()
+    }
+
+    pub(crate) fn admit(&self) -> Result<TransportAdmission> {
+        self.lifecycle
+            .admit()
+            .ok_or_else(|| Error::RunNotFound(self.request_id.clone()))
+    }
+
+    pub(crate) fn begin_close(&self) {
+        self.lifecycle.begin_close();
+    }
+
+    pub(crate) fn admissions_drained(&self) -> bool {
+        self.lifecycle.admissions_drained()
+    }
+
+    pub(crate) async fn wait_admissions_drained(&self) {
+        self.lifecycle.wait_admissions_drained().await;
+    }
+
+    pub(crate) fn mark_draining(&self) {
+        self.lifecycle.mark_draining();
+    }
+
+    pub(crate) fn reopen(&self) {
+        self.lifecycle.reopen();
+    }
+
+    pub(crate) fn close_transport(&self) {
+        self.lifecycle.close();
+    }
+
+    pub(crate) async fn wait_transport_closed(&self) {
+        self.lifecycle.wait_closed().await;
     }
 
     pub(crate) fn disconnect_token(&self) -> CancellationToken {

server/src/cursor/transport/lifecycle.rs

@@ -0,0 +1,170 @@
+//! Coordinates append admission with transport shutdown.
+
+use std::sync::Arc;
+
+use tokio::sync::Notify;
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub(crate) enum TransportState {
+    Open,
+    Closing,
+    Draining,
+    Closed,
+}
+
+#[derive(Clone)]
+pub(crate) struct TransportLifecycle {
+    inner: Arc<LifecycleInner>,
+}
+
+struct LifecycleInner {
+    state: parking_lot::Mutex<LifecycleState>,
+    admissions_drained: Notify,
+    closed: Notify,
+}
+
+struct LifecycleState {
+    state: TransportState,
+    admissions: usize,
+}
+
+pub(crate) struct TransportAdmission {
+    inner: Arc<LifecycleInner>,
+}
+
+impl TransportLifecycle {
+    pub(crate) fn new() -> Self {
+        Self {
+            inner: Arc::new(LifecycleInner {
+                state: parking_lot::Mutex::new(LifecycleState {
+                    state: TransportState::Open,
+                    admissions: 0,
+                }),
+                admissions_drained: Notify::new(),
+                closed: Notify::new(),
+            }),
+        }
+    }
+
+    pub(crate) fn is_open(&self) -> bool {
+        self.inner.state.lock().state == TransportState::Open
+    }
+
+    pub(crate) fn admit(&self) -> Option<TransportAdmission> {
+        let mut lifecycle = self.inner.state.lock();
+        if lifecycle.state != TransportState::Open {
+            return None;
+        }
+        lifecycle.admissions += 1;
+        Some(TransportAdmission {
+            inner: self.inner.clone(),
+        })
+    }
+
+    pub(crate) fn begin_close(&self) {
+        let mut lifecycle = self.inner.state.lock();
+        if lifecycle.state != TransportState::Open {
+            return;
+        }
+        lifecycle.state = TransportState::Closing;
+        let drained = lifecycle.admissions == 0;
+        drop(lifecycle);
+        if drained {
+            self.inner.admissions_drained.notify_waiters();
+        }
+    }
+
+    pub(crate) fn admissions_drained(&self) -> bool {
+        self.inner.state.lock().admissions == 0
+    }
+
+    pub(crate) async fn wait_admissions_drained(&self) {
+        loop {
+            let notified = self.inner.admissions_drained.notified();
+            if self.inner.state.lock().admissions == 0 {
+                return;
+            }
+            notified.await;
+        }
+    }
+
+    pub(crate) fn mark_draining(&self) {
+        let mut lifecycle = self.inner.state.lock();
+        if lifecycle.state == TransportState::Closing && lifecycle.admissions == 0 {
+            lifecycle.state = TransportState::Draining;
+        }
+    }
+
+    pub(crate) fn reopen(&self) {
+        let mut lifecycle = self.inner.state.lock();
+        if matches!(
+            lifecycle.state,
+            TransportState::Closing | TransportState::Draining
+        ) {
+            lifecycle.state = TransportState::Open;
+        }
+    }
+
+    pub(crate) fn close(&self) {
+        let mut lifecycle = self.inner.state.lock();
+        if lifecycle.state == TransportState::Closed {
+            return;
+        }
+        lifecycle.state = TransportState::Closed;
+        drop(lifecycle);
+        self.inner.closed.notify_waiters();
+    }
+
+    pub(crate) async fn wait_closed(&self) {
+        loop {
+            let notified = self.inner.closed.notified();
+            if self.inner.state.lock().state == TransportState::Closed {
+                return;
+            }
+            notified.await;
+        }
+    }
+}
+
+impl Drop for TransportAdmission {
+    fn drop(&mut self) {
+        let mut lifecycle = self.inner.state.lock();
+        lifecycle.admissions = lifecycle.admissions.saturating_sub(1);
+        let drained = lifecycle.state == TransportState::Closing && lifecycle.admissions == 0;
+        drop(lifecycle);
+        if drained {
+            self.inner.admissions_drained.notify_waiters();
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{TransportLifecycle, TransportState};
+
+    #[tokio::test]
+    async fn closing_waits_for_existing_admissions() {
+        let lifecycle = TransportLifecycle::new();
+        let admission = lifecycle.admit().unwrap();
+        lifecycle.begin_close();
+        assert!(lifecycle.admit().is_none());
+        drop(admission);
+        lifecycle.wait_admissions_drained().await;
+        lifecycle.mark_draining();
+        assert_eq!(lifecycle.inner.state.lock().state, TransportState::Draining);
+    }
+
+    #[tokio::test]
+    async fn an_admitted_continuation_reopens_the_transport() {
+        let lifecycle = TransportLifecycle::new();
+        let admission = lifecycle.admit().unwrap();
+        lifecycle.begin_close();
+        drop(admission);
+        lifecycle.wait_admissions_drained().await;
+        lifecycle.mark_draining();
+        lifecycle.reopen();
+
+        assert_eq!(lifecycle.inner.state.lock().state, TransportState::Open);
+        assert!(lifecycle.admit().is_some());
+    }
+}

server/src/cursor/transport/mod.rs

@@ -2,10 +2,12 @@
 
 mod handle;
 mod inbox;
+mod lifecycle;
 mod output;
 mod registry;
 
 pub use handle::*;
 pub use inbox::*;
+pub(crate) use lifecycle::*;
 pub use output::*;
 pub use registry::*;

server/src/cursor/transport/output.rs

@@ -1,12 +1,11 @@
 //! Buffers, replays, broadcasts, and atomically closes downstream output.
 
 use bytes::Bytes;
-use tokio::sync::{mpsc, Notify};
+use tokio::sync::mpsc;
 
 #[derive(Default)]
 pub struct OutputHub {
     state: parking_lot::Mutex<OutputState>,
-    closed: Notify,
 }
 
 #[derive(Default)]
@@ -49,17 +48,6 @@ impl OutputHub {
         state.closed = true;
         state.subscribers.clear();
         drop(state);
-        self.closed.notify_waiters();
         true
     }
-
-    pub async fn wait_closed(&self) {
-        loop {
-            let notified = self.closed.notified();
-            if self.state.lock().closed {
-                return;
-            }
-            notified.await;
-        }
-    }
 }

server/src/cursor/transport/registry.rs

@@ -1,13 +1,19 @@
 //! Maps request IDs to active transport handles.
 
-use std::{collections::HashMap, sync::Arc};
+use std::{
+    collections::HashMap,
+    sync::{
+        atomic::{AtomicU64, Ordering},
+        Arc,
+    },
+};
 
 use tokio::sync::{mpsc, Mutex, Notify};
 
 use crate::{
     cursor::{
         conversation::ConversationRegistry, prompting::PromptCompiler,
-        services::observability::CursorTraceRecorder,
+        services::observability::CursorTraceService,
     },
     plugin::PluginRegistry,
     provider::Provider,
@@ -24,15 +30,23 @@ pub struct TransportRegistry {
 }
 
 struct RegistryInner {
-    local: Mutex<HashMap<String, TransportHandle>>,
+    local: Mutex<HashMap<String, LocalTransport>>,
+    next_local_generation: AtomicU64,
     upstream: Mutex<HashMap<String, u64>>,
     route_changed: Notify,
     store: Store,
+    traces: CursorTraceService,
     web_cache: WebCache,
     plugins: Option<PluginRegistry>,
     conversations: ConversationRegistry,
 }
 
+#[derive(Clone)]
+struct LocalTransport {
+    generation: u64,
+    handle: TransportHandle,
+}
+
 #[derive(Clone, Copy, Debug, PartialEq, Eq)]
 pub enum TransportRoute {
     Local,
@@ -99,8 +113,10 @@ impl TransportRegistry {
         Self {
             inner: Arc::new(RegistryInner {
                 local: Mutex::new(HashMap::new()),
+                next_local_generation: AtomicU64::new(1),
                 upstream: Mutex::new(HashMap::new()),
                 route_changed: Notify::new(),
+                traces: CursorTraceService::new(store.clone()),
                 conversations: ConversationRegistry::new(
                     store.clone(),
                     provider,
@@ -119,6 +135,13 @@ impl TransportRegistry {
         &self.inner.store
     }
 
+    pub fn trace(
+        &self,
+        request_id: &str,
+    ) -> crate::cursor::services::observability::CursorTraceRecorder {
+        self.inner.traces.recorder(request_id)
+    }
+
     pub fn web_cache(&self) -> &WebCache {
         &self.inner.web_cache
     }
@@ -132,18 +155,37 @@ impl TransportRegistry {
     }
 
     pub async fn get_or_create(&self, request_id: &str) -> Result<TransportHandle> {
-        if let Some(handle) = self.inner.local.lock().await.get(request_id).cloned() {
-            return Ok(handle);
+        self.get_or_create_for_append(request_id, false).await
+    }
+
+    pub(crate) async fn get_or_create_for_append(
+        &self,
+        request_id: &str,
+        replace_closing: bool,
+    ) -> Result<TransportHandle> {
+        let mut local = self.inner.local.lock().await;
+        if let Some(transport) = local.get(request_id) {
+            if transport.handle.accepting_appends() || !replace_closing {
+                return Ok(transport.handle.clone());
+            }
         }
+        local.remove(request_id);
         let (commands, receiver) = mpsc::channel(128);
         let output = Arc::new(OutputHub::default());
-        let trace = CursorTraceRecorder::resume(self.inner.store.clone(), request_id).await;
-        let handle = TransportHandle::new(request_id.into(), commands, output.clone(), trace);
-        let mut local = self.inner.local.lock().await;
-        if let Some(existing) = local.get(request_id).cloned() {
-            return Ok(existing);
-        }
-        local.insert(request_id.into(), handle.clone());
+        let trace = self.inner.traces.recorder(request_id);
+        trace.resume();
+        let handle = TransportHandle::new(request_id.into(), commands, output, trace);
+        let generation = self
+            .inner
+            .next_local_generation
+            .fetch_add(1, Ordering::Relaxed);
+        local.insert(
+            request_id.into(),
+            LocalTransport {
+                generation,
+                handle: handle.clone(),
+            },
+        );
         drop(local);
         self.inner.route_changed.notify_waiters();
         self.inner
@@ -152,17 +194,29 @@ impl TransportRegistry {
 
         let registry = Arc::downgrade(&self.inner);
         let request_id = request_id.to_string();
+        let lifecycle = handle.clone();
         tokio::spawn(async move {
-            output.wait_closed().await;
+            lifecycle.wait_transport_closed().await;
             if let Some(registry) = registry.upgrade() {
-                registry.local.lock().await.remove(&request_id);
+                let mut local = registry.local.lock().await;
+                if local
+                    .get(&request_id)
+                    .is_some_and(|transport| transport.generation == generation)
+                {
+                    local.remove(&request_id);
+                }
             }
         });
         Ok(handle)
     }
 
     pub async fn local(&self, request_id: &str) -> Option<TransportHandle> {
-        self.inner.local.lock().await.get(request_id).cloned()
+        self.inner
+            .local
+            .lock()
+            .await
+            .get(request_id)
+            .map(|transport| transport.handle.clone())
     }
 
     pub async fn mark_upstream(&self, request_id: &str) {
@@ -206,10 +260,13 @@ impl TransportRegistry {
         self.inner.conversations.shutdown().await;
         let handles = std::mem::take(&mut *self.inner.local.lock().await);
         self.inner.upstream.lock().await.clear();
-        for handle in handles.into_values() {
-            handle.disconnect().await;
-            let _ =
-                tokio::time::timeout(std::time::Duration::from_secs(2), handle.wait_closed()).await;
+        for transport in handles.into_values() {
+            transport.handle.disconnect().await;
+            let _ = tokio::time::timeout(
+                std::time::Duration::from_secs(2),
+                transport.handle.wait_transport_closed(),
+            )
+            .await;
         }
     }
 }

server/src/model/observability.rs

@@ -6,11 +6,10 @@ mod usage {
 
     use serde::{Deserialize, Serialize};
 
-    use super::ProviderType;
-
     #[derive(Clone, Copy, Debug, Default, Serialize, Deserialize, PartialEq, Eq)]
     pub struct Usage {
         pub input_tokens: Option<u64>,
+        pub context_input_tokens: Option<u64>,
         pub output_tokens: Option<u64>,
         pub total_tokens: Option<u64>,
         pub cache_read_tokens: Option<u64>,
@@ -18,24 +17,10 @@ mod usage {
         pub reasoning_tokens: Option<u64>,
     }
 
-    impl Usage {
-        /// Returns the provider-visible input context without counting cached tokens twice.
-        pub(crate) fn context_input_tokens(self, provider: ProviderType) -> Option<u64> {
-            let input = self.input_tokens?;
-            match provider {
-                ProviderType::OpenAiChat | ProviderType::OpenAiResponses | ProviderType::Plugin => {
-                    Some(input)
-                }
-                ProviderType::Anthropic => input
-                    .checked_add(self.cache_read_tokens.unwrap_or_default())?
-                    .checked_add(self.cache_write_tokens.unwrap_or_default()),
-            }
-        }
-    }
-
     impl AddAssign for Usage {
         fn add_assign(&mut self, rhs: Self) {
             self.input_tokens = sum(self.input_tokens, rhs.input_tokens);
+            self.context_input_tokens = sum(self.context_input_tokens, rhs.context_input_tokens);
             self.output_tokens = sum(self.output_tokens, rhs.output_tokens);
             self.total_tokens = sum(self.total_tokens, rhs.total_tokens);
             self.cache_read_tokens = sum(self.cache_read_tokens, rhs.cache_read_tokens);
@@ -53,7 +38,7 @@ pub use usage::*;
 mod llm_call {
     use serde::Serialize;
 
-    use super::{ProviderType, Usage};
+    use super::ProviderType;
 
     #[derive(Clone, Debug)]
     pub struct NewLlmCall {
@@ -75,14 +60,6 @@ mod llm_call {
         pub detailed: bool,
     }
 
-    #[derive(Clone, Copy, Debug, PartialEq, Eq)]
-    pub(crate) struct LlmCallUsageAnchor {
-        pub request_type: ProviderType,
-        pub usage: Usage,
-        pub message_count: usize,
-        pub tool_count: usize,
-    }
-
     #[derive(Clone, Debug, Serialize)]
     pub struct LlmCallSummary {
         pub call_id: String,

server/src/model/projection.rs

@@ -6,8 +6,8 @@ use serde::{Deserialize, Serialize};
 use crate::{Error, Result};
 
 use super::{
-    CanonicalMessage, ContentPart, MessageContent, ProviderReplayState, Role, ToolCallContent,
-    ToolResultContent,
+    normalize_tool_name, CanonicalMessage, ContentPart, MessageContent, ProviderReplayState, Role,
+    ToolCallContent, ToolResultContent,
 };
 
 #[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
@@ -91,7 +91,7 @@ fn project_tool_round(
                 "tool round repeats provider replay state".into(),
             ));
         }
-        calls.extend(part_calls.iter().cloned());
+        calls.extend(part_calls.iter().map(normalized_tool_call));
         cursor += 1;
 
         while cursor < messages.len() {
@@ -139,7 +139,7 @@ fn project_tool_round(
             .map(|(message_id, result)| ProjectedMessage {
                 message_id,
                 role: Role::Tool,
-                content: ProjectedContent::ToolResult(result),
+                content: ProjectedContent::ToolResult(normalized_tool_result(&result)),
             }),
     );
     Ok(Some((output, cursor)))
@@ -158,9 +158,11 @@ fn project_message(message: &CanonicalMessage) -> ProjectedMessage {
             text: text.clone(),
             thinking: thinking.clone(),
             replay_state: replay_state.clone(),
-            calls: tool_calls.clone(),
+            calls: tool_calls.iter().map(normalized_tool_call).collect(),
         },
-        MessageContent::ToolResult(result) => ProjectedContent::ToolResult(result.clone()),
+        MessageContent::ToolResult(result) => {
+            ProjectedContent::ToolResult(normalized_tool_result(result))
+        }
     };
     ProjectedMessage {
         message_id: message.message_id.clone(),
@@ -168,3 +170,50 @@ fn project_message(message: &CanonicalMessage) -> ProjectedMessage {
         content,
     }
 }
+
+fn normalized_tool_call(call: &ToolCallContent) -> ToolCallContent {
+    let mut call = call.clone();
+    call.name = normalize_tool_name(&call.name);
+    call
+}
+
+fn normalized_tool_result(result: &ToolResultContent) -> ToolResultContent {
+    let mut result = result.clone();
+    result.name = normalize_tool_name(&result.name);
+    result
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::model::Origin;
+    use serde_json::json;
+
+    #[test]
+    fn tool_names_are_normalized_before_provider_dispatch() {
+        let messages = [CanonicalMessage {
+            message_id: "assistant-1".into(),
+            role: Role::Assistant,
+            origin: Origin::Assistant,
+            content: MessageContent::Assistant {
+                text: String::new(),
+                thinking: String::new(),
+                tool_round_id: None,
+                replay_state: None,
+                tool_calls: vec![ToolCallContent {
+                    index: 0,
+                    call_id: "call-1".into(),
+                    name: "multi_tool_use.parallel".into(),
+                    arguments: json!({}),
+                }],
+            },
+            runtime_event_id: None,
+        }];
+
+        let projected = project_messages(&messages).unwrap();
+        let ProjectedContent::Assistant { calls, .. } = &projected[0].content else {
+            panic!("expected assistant projection");
+        };
+        assert_eq!(calls[0].name, "multi_tool_use_parallel");
+    }
+}

server/src/model/token_count.rs

@@ -1,4 +1,99 @@
-//! Estimates and records model token usage.
+//! Estimates provider-visible context size and formats configured token counts.
+
+use super::{ContentPart, ProjectedContent, ProjectedMessage, PromptSpec};
+
+const TOKENS_PER_MESSAGE_OVERHEAD: u64 = 8;
+const TOKENS_PER_TOOL_CALL_OVERHEAD: u64 = 6;
+const TOKENS_PER_IMAGE: u64 = 1_024;
+
+pub(crate) fn estimate_context_tokens(prompt: &PromptSpec, messages: &[ProjectedMessage]) -> u64 {
+    let instructions = estimate_text_tokens(&prompt.instructions);
+    let tools = prompt.tools.iter().fold(0_u64, |total, tool| {
+        total.saturating_add(estimate_json_tokens(tool))
+    });
+    instructions
+        .saturating_add(tools)
+        .saturating_add(estimate_projected_messages_tokens(messages))
+}
+
+pub(crate) fn estimate_projected_messages_tokens(messages: &[ProjectedMessage]) -> u64 {
+    messages.iter().fold(0_u64, |total, message| {
+        total.saturating_add(estimate_message_tokens(message))
+    })
+}
+
+fn estimate_message_tokens(message: &ProjectedMessage) -> u64 {
+    let content = match &message.content {
+        ProjectedContent::Parts(parts) => estimate_parts_tokens(parts),
+        ProjectedContent::Assistant {
+            text,
+            thinking,
+            replay_state: _,
+            calls,
+        } => {
+            let calls = calls.iter().fold(0_u64, |total, call| {
+                total
+                    .saturating_add(TOKENS_PER_TOOL_CALL_OVERHEAD)
+                    .saturating_add(estimate_text_tokens(&call.call_id))
+                    .saturating_add(estimate_text_tokens(&call.name))
+                    .saturating_add(estimate_json_tokens(&call.arguments))
+            });
+            estimate_text_tokens(text)
+                .saturating_add(estimate_text_tokens(thinking))
+                .saturating_add(calls)
+        }
+        ProjectedContent::ToolResult(result) => {
+            let content = if result.provider_parts.is_empty() {
+                estimate_text_tokens(&result.content).saturating_add(
+                    result
+                        .image
+                        .as_ref()
+                        .map(|image| {
+                            TOKENS_PER_IMAGE.saturating_add(estimate_text_tokens(&image.mime_type))
+                        })
+                        .unwrap_or_default(),
+                )
+            } else {
+                estimate_parts_tokens(&result.provider_parts)
+            };
+            estimate_text_tokens(&result.call_id)
+                .saturating_add(estimate_text_tokens(&result.name))
+                .saturating_add(content)
+        }
+    };
+    TOKENS_PER_MESSAGE_OVERHEAD.saturating_add(content)
+}
+
+fn estimate_parts_tokens(parts: &[ContentPart]) -> u64 {
+    parts.iter().fold(0_u64, |total, part| {
+        let tokens = match part {
+            ContentPart::Text { text } => estimate_text_tokens(text),
+            ContentPart::Image { mime_type, .. } => {
+                TOKENS_PER_IMAGE.saturating_add(estimate_text_tokens(mime_type))
+            }
+        };
+        total.saturating_add(tokens)
+    })
+}
+
+fn estimate_json_tokens(value: &impl serde::Serialize) -> u64 {
+    serde_json::to_string(value)
+        .map(|value| estimate_text_tokens(&value))
+        .unwrap_or_default()
+}
+
+fn estimate_text_tokens(text: &str) -> u64 {
+    let text = text.trim();
+    if text.is_empty() {
+        return 0;
+    }
+    let characters = text.chars().count() as u64;
+    characters
+        .div_ceil(4)
+        .saturating_add(text.bytes().filter(|byte| *byte == b'\n').count() as u64)
+        .max(1)
+}
+
 pub(crate) fn parse_token_count(value: &str) -> Option<u64> {
     let value = value.trim().to_ascii_lowercase();
     let (number, multiplier) = match value.chars().last()? {
@@ -18,3 +113,136 @@ pub(crate) fn format_token_count(tokens: u64) -> String {
         tokens.to_string()
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::model::{
+        ProviderReplayState, Role, ToolCallContent, ToolDefinition, ToolResultContent,
+    };
+
+    fn prompt() -> PromptSpec {
+        PromptSpec {
+            instructions: "system instructions".into(),
+            tools: vec![ToolDefinition {
+                name: "Read".into(),
+                description: "Read a file".into(),
+                parameters: serde_json::json!({"type": "object", "properties": {"path": {"type": "string"}}}),
+            }],
+        }
+    }
+
+    #[test]
+    fn context_estimate_grows_with_provider_visible_text_and_tools() {
+        let short = vec![ProjectedMessage {
+            message_id: "short".into(),
+            role: Role::User,
+            content: ProjectedContent::Parts(vec![ContentPart::Text {
+                text: "hello".into(),
+            }]),
+        }];
+        let long = vec![ProjectedMessage {
+            message_id: "long".into(),
+            role: Role::User,
+            content: ProjectedContent::Parts(vec![ContentPart::Text {
+                text: "x".repeat(40_000),
+            }]),
+        }];
+        let without_tools = PromptSpec {
+            instructions: prompt().instructions,
+            tools: Vec::new(),
+        };
+
+        assert!(
+            estimate_context_tokens(&prompt(), &short)
+                > estimate_context_tokens(&without_tools, &short)
+        );
+        assert!(
+            estimate_context_tokens(&prompt(), &long) > estimate_context_tokens(&prompt(), &short)
+        );
+    }
+
+    #[test]
+    fn context_estimate_counts_tool_calls_results_and_images() {
+        let assistant = ProjectedMessage {
+            message_id: "assistant".into(),
+            role: Role::Assistant,
+            content: ProjectedContent::Assistant {
+                text: String::new(),
+                thinking: "reasoning".into(),
+                replay_state: None,
+                calls: vec![ToolCallContent {
+                    index: 0,
+                    call_id: "call-1".into(),
+                    name: "Read".into(),
+                    arguments: serde_json::json!({"path": "/tmp/file"}),
+                }],
+            },
+        };
+        let text_result = ProjectedMessage {
+            message_id: "result-text".into(),
+            role: Role::Tool,
+            content: ProjectedContent::ToolResult(ToolResultContent {
+                call_id: "call-1".into(),
+                name: "Read".into(),
+                content: "file contents".into(),
+                is_error: false,
+                image: None,
+                provider_parts: Vec::new(),
+            }),
+        };
+        let image_result = ProjectedMessage {
+            message_id: "result-image".into(),
+            role: Role::Tool,
+            content: ProjectedContent::ToolResult(ToolResultContent {
+                call_id: "call-1".into(),
+                name: "Read".into(),
+                content: "file contents".into(),
+                is_error: false,
+                image: None,
+                provider_parts: vec![ContentPart::Image {
+                    mime_type: "image/png".into(),
+                    data: vec![0; 32],
+                }],
+            }),
+        };
+
+        let base = estimate_context_tokens(&prompt(), &[]);
+        let with_call = estimate_context_tokens(&prompt(), std::slice::from_ref(&assistant));
+        let with_text = estimate_context_tokens(&prompt(), &[assistant.clone(), text_result]);
+        let with_image = estimate_context_tokens(&prompt(), &[assistant, image_result]);
+        assert!(with_call > base);
+        assert!(with_text > with_call);
+        assert!(with_image > with_text);
+    }
+
+    #[test]
+    fn assistant_replay_state_does_not_duplicate_thinking_or_count_signature() {
+        let assistant = |replay_state| ProjectedMessage {
+            message_id: "assistant".into(),
+            role: Role::Assistant,
+            content: ProjectedContent::Assistant {
+                text: "answer".into(),
+                thinking: "reasoning".repeat(1_000),
+                replay_state,
+                calls: Vec::new(),
+            },
+        };
+        let without_replay = assistant(None);
+        let with_replay = assistant(Some(ProviderReplayState {
+            provider_kind: "anthropic".into(),
+            value: serde_json::json!({
+                "blocks": [{
+                    "type": "thinking",
+                    "thinking": "reasoning".repeat(1_000),
+                    "signature": "s".repeat(282_100)
+                }]
+            }),
+        }));
+
+        assert_eq!(
+            estimate_projected_messages_tokens(&[without_replay]),
+            estimate_projected_messages_tokens(&[with_replay])
+        );
+    }
+}

server/src/model/tool.rs

@@ -4,6 +4,24 @@ use serde_json::Value;
 
 use super::ProviderReplayState;
 
+pub fn normalize_tool_name(name: &str) -> String {
+    let normalized = name
+        .chars()
+        .map(|character| {
+            if character.is_ascii_alphanumeric() || matches!(character, '_' | '-') {
+                character
+            } else {
+                '_'
+            }
+        })
+        .collect::<String>();
+    if normalized.is_empty() {
+        "_".into()
+    } else {
+        normalized
+    }
+}
+
 #[derive(Clone, Debug, Serialize, Deserialize, PartialEq)]
 pub struct ToolDefinition {
     pub name: String,
@@ -19,6 +37,7 @@ pub struct ToolCall {
     pub name: String,
     pub arguments_text: String,
     pub arguments: Value,
+    pub argument_error: Option<String>,
 }
 
 #[derive(Clone, Debug, Serialize, Deserialize, PartialEq, Eq)]

server/src/network.rs

@@ -1,8 +1,93 @@
-//! Provides shared network client and transport configuration.
-//! Outbound HTTP clients configured from persisted application proxy settings.
+//! Owns reusable outbound HTTP clients configured from persisted proxy settings.
+
+use std::{sync::Arc, time::Duration};
+
+use tokio::sync::RwLock;
 
 use crate::{store::Store, Result};
 
+#[derive(Clone)]
+pub struct NetworkClients {
+    store: Store,
+    cache: Arc<RwLock<ClientCache>>,
+}
+
+#[derive(Default)]
+struct ClientCache {
+    default: Option<reqwest::Client>,
+    cursor: Option<reqwest::Client>,
+    provider: Option<(Duration, reqwest::Client)>,
+}
+
+impl NetworkClients {
+    pub fn new(store: Store) -> Self {
+        Self {
+            store,
+            cache: Arc::new(RwLock::new(ClientCache::default())),
+        }
+    }
+
+    pub async fn default_client(&self) -> Result<reqwest::Client> {
+        if let Some(client) = self.cache.read().await.default.clone() {
+            return Ok(client);
+        }
+        let mut cache = self.cache.write().await;
+        if let Some(client) = cache.default.clone() {
+            return Ok(client);
+        }
+        let client = client_builder(&self.store).await?.build()?;
+        cache.default = Some(client.clone());
+        Ok(client)
+    }
+
+    pub async fn cursor_client(&self) -> Result<reqwest::Client> {
+        if let Some(client) = self.cache.read().await.cursor.clone() {
+            return Ok(client);
+        }
+        let mut cache = self.cache.write().await;
+        if let Some(client) = cache.cursor.clone() {
+            return Ok(client);
+        }
+        let client = client_builder(&self.store)
+            .await?
+            .redirect(reqwest::redirect::Policy::none())
+            .build()?;
+        cache.cursor = Some(client.clone());
+        Ok(client)
+    }
+
+    pub async fn provider_client(&self, timeout: Duration) -> Result<reqwest::Client> {
+        if let Some((_, client)) = self
+            .cache
+            .read()
+            .await
+            .provider
+            .as_ref()
+            .filter(|(cached_timeout, _)| *cached_timeout == timeout)
+        {
+            return Ok(client.clone());
+        }
+        let mut cache = self.cache.write().await;
+        if let Some((_, client)) = cache
+            .provider
+            .as_ref()
+            .filter(|(cached_timeout, _)| *cached_timeout == timeout)
+        {
+            return Ok(client.clone());
+        }
+        let client = client_builder(&self.store)
+            .await?
+            .timeout(timeout)
+            .build()?;
+        cache.provider = Some((timeout, client.clone()));
+        Ok(client)
+    }
+
+    pub async fn invalidate(&self) {
+        *self.cache.write().await = ClientCache::default();
+    }
+}
+
 pub async fn client_builder(store: &Store) -> Result<reqwest::ClientBuilder> {
     let settings = store.proxy_settings_secret().await?;
     // Use the platform TLS stack for compatibility with provider gateways that

server/src/plugin/descriptor.rs

@@ -107,9 +107,7 @@ pub struct PluginModelDescriptor {
     pub description: Option<String>,
     pub icon: String,
     pub provider_type: String,
-    pub context_window_tokens: Option<u64>,
     pub max_output_tokens: Option<u64>,
-    pub thinking: bool,
     pub images: bool,
 }
 
@@ -209,9 +207,7 @@ impl PluginModelDescriptor {
             description: model.description.clone(),
             icon: icon.to_owned(),
             provider_type: provider.provider_type.clone(),
-            context_window_tokens: model.context_window_tokens,
             max_output_tokens: model.max_output_tokens,
-            thinking: model.thinking,
             images: model.images,
         }
     }

server/src/plugin/mod.rs

@@ -20,7 +20,6 @@ pub use descriptor::{
 };
 pub use registry::{ImportResponse, OAuthBeginResponse, OAuthPollResponse, PluginRegistry};
 pub use runtime::{PluginRuntime, PluginRuntimePhase, PluginRuntimeState, PluginRuntimeStatus};
-pub(crate) use wire::llm_request as plugin_llm_request;
 
 /// Windows 下阻止 Deno 子进程弹出控制台窗口(CREATE_NO_WINDOW)。
 #[cfg(windows)]

server/src/plugin/registry.rs

@@ -20,8 +20,11 @@ use super::{
     worker::{PluginWorker, WorkerStreamItem},
 };
 use crate::{
-    model::ModelInvocation, provider::ModelEvent, provider::ProviderStream, store::Store, Error,
-    Result,
+    model::ModelInvocation,
+    provider::ProviderStream,
+    provider::{CallRecorder, ModelEvent},
+    store::Store,
+    Error, Result,
 };
 
 const OAUTH_SLOW_DOWN_STEP_MS: i64 = 5_000;
@@ -203,6 +206,7 @@ impl PluginRegistry {
         &self,
         invocation: ModelInvocation,
         cancellation: CancellationToken,
+        recorder: CallRecorder,
     ) -> ProviderStream {
         let registry = self.clone();
         Box::pin(try_stream! {
@@ -232,7 +236,7 @@ impl PluginRegistry {
                 "request": request,
             });
             let worker = registry.worker(&entry, &executable).await;
-            let mut items = worker.invoke_streaming("provider.invoke", params, cancellation.clone()).await?;
+            let mut items = worker.invoke_streaming("provider.invoke", params, cancellation.clone(), Some(recorder)).await?;
             yield ModelEvent::Start { model_call_id: invocation.call_id.clone() };
             while let Some(item) = items.recv().await {
                 match item {

server/src/plugin/sdk/model.ts

@@ -2,7 +2,6 @@ import type { JsonValue, PluginContext } from "./plugin.ts";
 import type { ResourceSnapshot } from "./resource.ts";
 
 export type ModelCapabilities = {
-  thinking?: boolean;
   images?: boolean;
 };
 
@@ -10,7 +9,6 @@ export type ModelDefinition = {
   id: string;
   displayName: string;
   description?: string;
-  contextWindowTokens?: number;
   maxOutputTokens?: number;
   capabilities?: ModelCapabilities;
   /** 之后的调用原样传回;永远不会展示给用户。 */

server/src/plugin/sdk/protocol/openai_responses.ts

@@ -53,7 +53,17 @@ function replayItems(value: JsonValue): JsonValue[] {
   if (!Array.isArray(items)) {
     throw new Error("OpenAI Responses replay state is missing items");
   }
-  return items;
+  return items.map((item) => {
+    const source = record(item);
+    if (source?.type !== "reasoning") {
+      throw new Error("OpenAI Responses replay state contains a non-reasoning item");
+    }
+    const projected: Record<string, JsonValue> = { type: "reasoning" };
+    for (const field of ["id", "summary", "content", "encrypted_content"] as const) {
+      if (field in source) projected[field] = source[field] as JsonValue;
+    }
+    return projected;
+  });
 }
 
 export function buildResponsesBody(call: OpenAiResponsesCall): Record<string, JsonValue> {

server/src/plugin/state.rs

@@ -135,12 +135,8 @@ pub struct StoredModel {
     #[serde(default)]
     pub description: Option<String>,
     #[serde(default)]
-    pub context_window_tokens: Option<u64>,
-    #[serde(default)]
     pub max_output_tokens: Option<u64>,
     #[serde(default)]
-    pub thinking: bool,
-    #[serde(default)]
     pub images: bool,
     #[serde(default)]
     pub private_data: serde_json::Value,
@@ -179,13 +175,9 @@ impl StoredModel {
                 .get("description")
                 .and_then(serde_json::Value::as_str)
                 .map(str::to_owned),
-            context_window_tokens: object
-                .get("contextWindowTokens")
-                .and_then(serde_json::Value::as_u64),
             max_output_tokens: object
                 .get("maxOutputTokens")
                 .and_then(serde_json::Value::as_u64),
-            thinking: capability("thinking"),
             images: capability("images"),
             private_data: object
                 .get("privateData")
@@ -200,9 +192,8 @@ impl StoredModel {
             "id": self.id,
             "displayName": self.display_name,
             "description": self.description,
-            "contextWindowTokens": self.context_window_tokens,
             "maxOutputTokens": self.max_output_tokens,
-            "capabilities": { "thinking": self.thinking, "images": self.images },
+            "capabilities": { "images": self.images },
             "privateData": self.private_data,
         })
     }
@@ -450,7 +441,7 @@ mod tests {
         let model = StoredModel::from_definition(&serde_json::json!({
             "id": "gpt-test",
             "displayName": "GPT Test",
-            "capabilities": {"thinking": true},
+            "capabilities": {"images": true},
             "privateData": {"reasoningEfforts": ["low"]},
         }))
         .unwrap();
@@ -460,7 +451,7 @@ mod tests {
             .unwrap();
         let models = store.models("dev.example", "codex").await.unwrap();
         assert_eq!(models.len(), 1);
-        assert!(models[0].thinking);
+        assert!(models[0].images);
         assert_eq!(models[0].private_data["reasoningEfforts"][0], "low");
     }
 }

server/src/plugin/wire.rs

@@ -147,8 +147,10 @@ pub fn model_event(value: &serde_json::Value) -> Result<ModelEvent> {
                 .get("usage")
                 .ok_or_else(|| Error::Protocol("plugin usage event requires usage".into()))?;
             let tokens = |name: &str| usage.get(name).and_then(serde_json::Value::as_u64);
+            let input_tokens = tokens("inputTokens");
             ModelEvent::Usage(Usage {
-                input_tokens: tokens("inputTokens"),
+                input_tokens,
+                context_input_tokens: input_tokens,
                 output_tokens: tokens("outputTokens"),
                 total_tokens: tokens("totalTokens"),
                 cache_read_tokens: tokens("cacheReadTokens"),
@@ -285,6 +287,7 @@ mod tests {
             usage,
             ModelEvent::Usage(Usage {
                 input_tokens: Some(10),
+                context_input_tokens: Some(10),
                 output_tokens: Some(2),
                 total_tokens: None,
                 cache_read_tokens: Some(4),

server/src/plugin/worker.rs

@@ -3,7 +3,10 @@ use std::{
     collections::{HashMap, HashSet},
     path::PathBuf,
     process::Stdio,
-    sync::Arc,
+    sync::{
+        atomic::{AtomicBool, Ordering},
+        Arc,
+    },
     time::Duration,
 };
 
@@ -19,7 +22,7 @@ use super::{
     definition::{file_url, PluginDefinitionLoader},
     protocol::{HostMessage, WorkerMessage},
 };
-use crate::{store::Store, Error, Result};
+use crate::{provider::CallRecorder, store::Store, Error, Result};
 
 const INVOCATION_TIMEOUT: Duration = Duration::from_secs(10 * 60);
 const MAX_NETWORK_RESPONSE_BYTES: u64 = 16 * 1024 * 1024;
@@ -56,12 +59,29 @@ struct WorkerProcess {
     stdin: Arc<Mutex<ChildStdin>>,
 }
 
+struct InvocationState {
+    cancellation: CancellationToken,
+    recorder: Option<CallRecorder>,
+    recorder_claimed: AtomicBool,
+}
+
+impl InvocationState {
+    fn claim_recorder(&self) -> Option<CallRecorder> {
+        self.recorder.as_ref().and_then(|recorder| {
+            self.recorder_claimed
+                .compare_exchange(false, true, Ordering::AcqRel, Ordering::Acquire)
+                .ok()
+                .map(|_| recorder.clone())
+        })
+    }
+}
+
 #[derive(Clone)]
 struct HostContext {
     plugin_id: String,
     network_hosts: Arc<HashSet<String>>,
     store: Store,
-    cancellations: Arc<Mutex<HashMap<String, CancellationToken>>>,
+    invocations: Arc<Mutex<HashMap<String, Arc<InvocationState>>>>,
     streams: Arc<Mutex<HashMap<String, StreamLines>>>,
 }
 
@@ -87,7 +107,7 @@ impl PluginWorker {
                             .collect(),
                     ),
                     store,
-                    cancellations: Arc::new(Mutex::new(HashMap::new())),
+                    invocations: Arc::new(Mutex::new(HashMap::new())),
                     streams: Arc::new(Mutex::new(HashMap::new())),
                 },
                 plugin_id,
@@ -108,7 +128,9 @@ impl PluginWorker {
         params: serde_json::Value,
         cancellation: CancellationToken,
     ) -> Result<serde_json::Value> {
-        let mut items = self.invoke_streaming(method, params, cancellation).await?;
+        let mut items = self
+            .invoke_streaming(method, params, cancellation, None)
+            .await?;
         let result = tokio::time::timeout(INVOCATION_TIMEOUT, async {
             while let Some(item) = items.recv().await {
                 if let WorkerStreamItem::Result(result) = item {
@@ -137,15 +159,18 @@ impl PluginWorker {
         method: &str,
         params: serde_json::Value,
         cancellation: CancellationToken,
+        recorder: Option<CallRecorder>,
     ) -> Result<mpsc::UnboundedReceiver<WorkerStreamItem>> {
         let id = uuid::Uuid::new_v4().to_string();
         let request_cancellation = CancellationToken::new();
-        self.inner
-            .host
-            .cancellations
-            .lock()
-            .await
-            .insert(id.clone(), request_cancellation.clone());
+        self.inner.host.invocations.lock().await.insert(
+            id.clone(),
+            Arc::new(InvocationState {
+                cancellation: request_cancellation.clone(),
+                recorder,
+                recorder_claimed: AtomicBool::new(false),
+            }),
+        );
         let (sender, receiver) = mpsc::unbounded_channel();
         self.inner
             .pending
@@ -182,10 +207,10 @@ impl PluginWorker {
                     }
                     let _ = sender.send(WorkerStreamItem::Result(Err(Error::Cancelled)));
                     inner.pending.lock().await.remove(&request_id);
-                    inner.host.cancellations.lock().await.remove(&request_id);
+                    inner.host.invocations.lock().await.remove(&request_id);
                 }
                 _ = sender.closed() => {
-                    inner.host.cancellations.lock().await.remove(&request_id);
+                    inner.host.invocations.lock().await.remove(&request_id);
                 }
             }
         });
@@ -201,7 +226,7 @@ impl PluginWorker {
 
     async fn cleanup(&self, id: &str) {
         self.inner.pending.lock().await.remove(id);
-        self.inner.host.cancellations.lock().await.remove(id);
+        self.inner.host.invocations.lock().await.remove(id);
     }
 
     async fn stdin(&self) -> Result<Arc<Mutex<ChildStdin>>> {
@@ -393,6 +418,28 @@ async fn fail_pending(pending: &Pending, message: &str) {
     }
 }
 
+fn recorded_network_request(
+    params: &serde_json::Value,
+) -> Result<(serde_json::Value, serde_json::Value)> {
+    let mut recorded_headers = serde_json::Map::new();
+    if let Some(headers) = params.get("headers").and_then(serde_json::Value::as_object) {
+        for (name, value) in headers {
+            let value = value.as_str().ok_or_else(|| {
+                Error::Config(format!("plugin HTTP header '{name}' must be a string"))
+            })?;
+            if !crate::model::is_sensitive_header(name) {
+                recorded_headers.insert(name.clone(), value.into());
+            }
+        }
+    }
+    let body = params
+        .get("body")
+        .and_then(serde_json::Value::as_str)
+        .map(|body| serde_json::from_str(body).unwrap_or_else(|_| body.into()))
+        .unwrap_or(serde_json::Value::Null);
+    Ok((serde_json::Value::Object(recorded_headers), body))
+}
+
 impl HostContext {
     async fn call(
         &self,
@@ -421,7 +468,11 @@ impl HostContext {
         &self,
         request_id: &str,
         params: &serde_json::Value,
-    ) -> Result<(reqwest::RequestBuilder, CancellationToken)> {
+    ) -> Result<(
+        reqwest::RequestBuilder,
+        CancellationToken,
+        Option<CallRecorder>,
+    )> {
         let raw_url = required_string(params, "url")?;
         let url = url::Url::parse(raw_url)
             .map_err(|error| Error::Config(format!("invalid plugin network URL: {error}")))?;
@@ -463,14 +514,17 @@ impl HostContext {
         if let Some(body) = params.get("body").and_then(serde_json::Value::as_str) {
             request = request.body(body.to_owned());
         }
-        let cancellation = self
-            .cancellations
-            .lock()
-            .await
-            .get(request_id)
-            .cloned()
+        let invocation = self.invocations.lock().await.get(request_id).cloned();
+        let cancellation = invocation
+            .as_ref()
+            .map(|state| state.cancellation.clone())
             .unwrap_or_default();
-        Ok((request, cancellation))
+        let recorder = invocation.and_then(|state| state.claim_recorder());
+        if let Some(recorder) = &recorder {
+            let (headers, body) = recorded_network_request(params)?;
+            recorder.request(headers, &body).await?;
+        }
+        Ok((request, cancellation, recorder))
     }
 
     async fn fetch(
@@ -478,13 +532,16 @@ impl HostContext {
         request_id: &str,
         params: serde_json::Value,
     ) -> Result<serde_json::Value> {
-        let (request, cancellation) = self.request(request_id, &params).await?;
+        let (request, cancellation, recorder) = self.request(request_id, &params).await?;
         let request = request.timeout(Duration::from_secs(60));
         let response = tokio::select! {
             _ = cancellation.cancelled() => return Err(Error::Cancelled),
             response = request.send() => response?,
         };
         let status = response.status().as_u16();
+        if let Some(recorder) = &recorder {
+            recorder.response_headers(status).await?;
+        }
         if response
             .content_length()
             .is_some_and(|size| size > MAX_NETWORK_RESPONSE_BYTES)
@@ -503,6 +560,9 @@ impl HostContext {
                 "plugin network response is larger than allowed".into(),
             ));
         }
+        if let Some(recorder) = &recorder {
+            recorder.response_chunk(&body).await?;
+        }
         Ok(
             serde_json::json!({ "status": status, "headers": headers, "body": String::from_utf8_lossy(&body) }),
         )
@@ -514,12 +574,15 @@ impl HostContext {
         request_id: &str,
         params: serde_json::Value,
     ) -> Result<serde_json::Value> {
-        let (request, cancellation) = self.request(request_id, &params).await?;
+        let (request, cancellation, recorder) = self.request(request_id, &params).await?;
         let response = tokio::select! {
             _ = cancellation.cancelled() => return Err(Error::Cancelled),
             response = request.send() => response?,
         };
         let status = response.status().as_u16();
+        if let Some(recorder) = &recorder {
+            recorder.response_headers(status).await?;
+        }
         let headers = header_map(&response);
         let (sender, receiver) = mpsc::channel::<Result<String>>(256);
         tokio::spawn(async move {
@@ -552,6 +615,12 @@ impl HostContext {
                         .await;
                     return;
                 }
+                if let Some(recorder) = &recorder {
+                    if let Err(error) = recorder.response_chunk(&chunk).await {
+                        let _ = sender.send(Err(error)).await;
+                        return;
+                    }
+                }
                 buffered.extend_from_slice(&chunk);
                 while let Some(position) = buffered.iter().position(|byte| *byte == b'\n') {
                     let mut line = buffered.drain(..=position).collect::<Vec<u8>>();
@@ -645,3 +714,179 @@ fn required_string<'a>(params: &'a serde_json::Value, key: &str) -> Result<&'a s
         .and_then(serde_json::Value::as_str)
         .ok_or_else(|| Error::Protocol(format!("plugin host call requires string '{key}'")))
 }
+
+#[cfg(test)]
+mod tests {
+    use crate::{
+        model::{NewLlmCall, ProviderType},
+        provider::{CallRecorder, FinishReason},
+        store::Store,
+    };
+
+    use super::*;
+
+    async fn recorder(detailed: bool, call_id: &str) -> (tempfile::TempDir, Store, CallRecorder) {
+        let directory = tempfile::tempdir().unwrap();
+        let store = Store::connect(&format!(
+            "sqlite://{}",
+            directory.path().join("test.db").display()
+        ))
+        .await
+        .unwrap();
+        store.set_detailed_logging(detailed).await.unwrap();
+        let recorder = CallRecorder::start(
+            store.clone(),
+            NewLlmCall {
+                call_id: call_id.into(),
+                run_id: "run".into(),
+                conversation_id: "conversation".into(),
+                provider_call_index: 0,
+                model_hash: "plugin:test/provider/model".into(),
+                provider_type: ProviderType::Plugin,
+                provider_url: "plugin://test/provider".into(),
+                request_type: ProviderType::Plugin,
+                request_url: "plugin://test/provider".into(),
+                model_id: "model".into(),
+                display_name: "Model".into(),
+                reasoning_effort: None,
+                fast: false,
+                message_count: 1,
+                tool_count: 0,
+                detailed: false,
+            },
+        )
+        .await
+        .unwrap();
+        (directory, store, recorder)
+    }
+
+    fn network_params() -> serde_json::Value {
+        serde_json::json!({
+            "url": "https://example.com/v1/responses",
+            "method": "POST",
+            "headers": {
+                "Authorization": "Bearer secret",
+                "X-Api-Key": "secret-key",
+                "Cookie": "session=secret",
+                "content-type": "application/json",
+                "x-client-request-id": "request-1"
+            },
+            "body": "{\"model\":\"test\",\"stream\":true}"
+        })
+    }
+
+    async fn host_with_recorder(store: Store, recorder: CallRecorder) -> HostContext {
+        let invocations = Arc::new(Mutex::new(HashMap::new()));
+        invocations.lock().await.insert(
+            "invocation".into(),
+            Arc::new(InvocationState {
+                cancellation: CancellationToken::new(),
+                recorder: Some(recorder),
+                recorder_claimed: AtomicBool::new(false),
+            }),
+        );
+        HostContext {
+            plugin_id: "test".into(),
+            network_hosts: Arc::new(HashSet::from(["example.com".into()])),
+            store,
+            invocations,
+            streams: Arc::new(Mutex::new(HashMap::new())),
+        }
+    }
+
+    #[test]
+    fn recorded_plugin_request_omits_sensitive_headers() {
+        let (headers, body) = recorded_network_request(&network_params()).unwrap();
+
+        assert_eq!(
+            headers,
+            serde_json::json!({
+                "content-type": "application/json",
+                "x-client-request-id": "request-1"
+            })
+        );
+        assert_eq!(body, serde_json::json!({ "model": "test", "stream": true }));
+    }
+
+    #[tokio::test]
+    async fn detailed_plugin_network_recording_persists_request_and_raw_response() {
+        let (_directory, store, recorder) = recorder(true, "detailed-plugin").await;
+        let host = host_with_recorder(store.clone(), recorder.clone()).await;
+        let params = network_params();
+        let (_, _, first_recorder) = host.request("invocation", &params).await.unwrap();
+        let (_, _, second_recorder) = host.request("invocation", &params).await.unwrap();
+        let (_, body) = recorded_network_request(&params).unwrap();
+
+        assert!(first_recorder.is_some());
+        assert!(second_recorder.is_none());
+        recorder.response_headers(200).await.unwrap();
+        recorder
+            .response_chunk(b"data: {\"type\":\"response.created\"}\n\n")
+            .await
+            .unwrap();
+        recorder.response_chunk(b"data: [DONE]\n\n").await.unwrap();
+        recorder.completed(FinishReason::Stop).await.unwrap();
+
+        let request = store
+            .llm_call_request("detailed-plugin")
+            .await
+            .unwrap()
+            .unwrap();
+        assert_eq!(
+            request.headers,
+            serde_json::json!({
+                "content-type": "application/json",
+                "x-client-request-id": "request-1"
+            })
+        );
+        assert_eq!(request.body, body);
+        let chunks = store.llm_call_chunks("detailed-plugin").await.unwrap();
+        let expected_response = "data: {\"type\":\"response.created\"}\n\ndata: [DONE]\n\n";
+        assert_eq!(chunks.len(), 2);
+        assert_eq!(
+            chunks
+                .iter()
+                .map(|chunk| chunk.data.as_str())
+                .collect::<String>(),
+            expected_response
+        );
+        let summary = store.llm_call("detailed-plugin").await.unwrap().unwrap();
+        assert_eq!(summary.http_status, Some(200));
+        assert_eq!(summary.stream_event_count, 2);
+        assert_eq!(summary.response_bytes, expected_response.len() as i64);
+        assert!(summary.detailed);
+    }
+
+    #[tokio::test]
+    async fn standard_plugin_network_recording_keeps_metrics_without_payloads() {
+        let (_directory, store, recorder) = recorder(false, "standard-plugin").await;
+        let host = host_with_recorder(store.clone(), recorder.clone()).await;
+        let params = network_params();
+        let (_, body) = recorded_network_request(&params).unwrap();
+        let request_bytes = serde_json::to_string(&body).unwrap().len() as i64;
+        let response = b"data: [DONE]\n\n";
+
+        let (_, _, observed) = host.request("invocation", &params).await.unwrap();
+        assert!(observed.is_some());
+        recorder.response_headers(204).await.unwrap();
+        recorder.response_chunk(response).await.unwrap();
+        recorder.completed(FinishReason::Stop).await.unwrap();
+
+        assert!(store
+            .llm_call_request("standard-plugin")
+            .await
+            .unwrap()
+            .is_none());
+        assert!(store
+            .llm_call_chunks("standard-plugin")
+            .await
+            .unwrap()
+            .is_empty());
+        let summary = store.llm_call("standard-plugin").await.unwrap().unwrap();
+        assert_eq!(summary.http_status, Some(204));
+        assert_eq!(summary.request_bytes, Some(request_bytes));
+        assert_eq!(summary.response_bytes, response.len() as i64);
+        assert_eq!(summary.stream_event_count, 1);
+        assert!(!summary.detailed);
+    }
+}

server/src/provider/anthropic.rs

@@ -12,9 +12,9 @@ use crate::{
 };
 
 use super::{
+    attempt::{send_once, Attempt},
     map_sse_error, merge_extra_params, provider_event_error,
     recorder::recorded_headers,
-    retry::{send_with_retry, Attempt, RetryPolicy},
     CallRecorder, FinishReason, ModelEvent, Provider, ProviderStream,
 };
 
@@ -90,17 +90,14 @@ impl Provider for AnthropicProvider {
             if let Some(recorder) = &recorder {
                 recorder.request(request_headers.clone(), &body).await?;
             }
-            let attempt = send_with_retry(
+            let attempt = send_once(
                 "Anthropic",
                 || client.post(&config.request_url)
                     .header("x-api-key", &config.api_key).header("anthropic-version", "2023-06-01")
                     .headers(config.custom_headers.clone())
                     .json(&body),
-                RetryPolicy { retries: config.retry_count, ..RetryPolicy::default() },
                 &cancellation,
                 recorder.as_ref(),
-                request_headers,
-                &body,
             ).await?;
             let Attempt::Response(response) = attempt else { return };
             yield ModelEvent::Start { model_call_id: call_id };
@@ -321,10 +318,19 @@ fn apply_model(body: &mut Value, model: &crate::model::ModelSpec) -> Result<()>
 
 fn merge_usage(total: &mut Usage, update: Usage) {
     merge_usage_field(&mut total.input_tokens, update.input_tokens);
+    merge_usage_field(&mut total.context_input_tokens, update.context_input_tokens);
     merge_usage_field(&mut total.output_tokens, update.output_tokens);
+    merge_usage_field(&mut total.total_tokens, update.total_tokens);
     merge_usage_field(&mut total.cache_read_tokens, update.cache_read_tokens);
     merge_usage_field(&mut total.cache_write_tokens, update.cache_write_tokens);
     merge_usage_field(&mut total.reasoning_tokens, update.reasoning_tokens);
+    if let Some(request_tokens) = total
+        .context_input_tokens
+        .zip(total.output_tokens)
+        .and_then(|(input, output)| input.checked_add(output))
+    {
+        total.total_tokens = Some(request_tokens);
+    }
 }
 
 fn merge_usage_field(total: &mut Option<u64>, update: Option<u64>) {
@@ -475,14 +481,67 @@ fn required_u64(value: &Value, name: &str) -> Result<u64> {
 }
 
 fn anthropic_usage(value: &Value) -> Usage {
+    let input_tokens = value.get("input_tokens").and_then(Value::as_u64);
+    let cache_read_tokens = value.get("cache_read_input_tokens").and_then(Value::as_u64);
+    let cache_write_tokens = value
+        .get("cache_creation_input_tokens")
+        .and_then(Value::as_u64);
+    let context_input_tokens = input_tokens.map(|input| {
+        input
+            .saturating_add(cache_read_tokens.unwrap_or_default())
+            .saturating_add(cache_write_tokens.unwrap_or_default())
+    });
     Usage {
-        input_tokens: value.get("input_tokens").and_then(Value::as_u64),
+        input_tokens,
+        context_input_tokens,
         output_tokens: value.get("output_tokens").and_then(Value::as_u64),
         total_tokens: value.get("total_tokens").and_then(Value::as_u64),
-        cache_read_tokens: value.get("cache_read_input_tokens").and_then(Value::as_u64),
-        cache_write_tokens: value
-            .get("cache_creation_input_tokens")
-            .and_then(Value::as_u64),
+        cache_read_tokens,
+        cache_write_tokens,
         reasoning_tokens: None,
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn cached_tokens_are_included_once_in_anthropic_context_input() {
+        let usage = anthropic_usage(&serde_json::json!({
+            "input_tokens": 10,
+            "output_tokens": 5,
+            "cache_read_input_tokens": 20,
+            "cache_creation_input_tokens": 30
+        }));
+
+        assert_eq!(usage.input_tokens, Some(10));
+        assert_eq!(usage.context_input_tokens, Some(60));
+        assert_eq!(usage.output_tokens, Some(5));
+    }
+
+    #[test]
+    fn streamed_anthropic_usage_includes_cached_input_in_total() {
+        let mut usage = Usage::default();
+        merge_usage(
+            &mut usage,
+            anthropic_usage(&serde_json::json!({
+                "input_tokens": 414,
+                "output_tokens": 0,
+                "cache_read_input_tokens": 100_352,
+                "cache_creation_input_tokens": 0
+            })),
+        );
+        merge_usage(
+            &mut usage,
+            anthropic_usage(&serde_json::json!({"output_tokens": 191})),
+        );
+
+        assert_eq!(usage.input_tokens, Some(414));
+        assert_eq!(usage.cache_read_tokens, Some(100_352));
+        assert_eq!(usage.cache_write_tokens, Some(0));
+        assert_eq!(usage.context_input_tokens, Some(100_766));
+        assert_eq!(usage.output_tokens, Some(191));
+        assert_eq!(usage.total_tokens, Some(100_957));
+    }
+}

server/src/provider/attempt.rs

@@ -0,0 +1,100 @@
+//! Sends one Provider HTTP attempt without applying retry policy.
+
+use tokio_util::sync::CancellationToken;
+
+use crate::{Error, Result};
+
+use super::CallRecorder;
+
+#[derive(Debug)]
+pub(crate) enum Attempt {
+    Response(reqwest::Response),
+    Cancelled,
+}
+
+pub(crate) async fn send_once<F>(
+    label: &str,
+    build: F,
+    cancellation: &CancellationToken,
+    recorder: Option<&CallRecorder>,
+) -> Result<Attempt>
+where
+    F: FnOnce() -> reqwest::RequestBuilder,
+{
+    let response = tokio::select! {
+        _ = cancellation.cancelled() => return Ok(Attempt::Cancelled),
+        response = build().send() => response,
+    }?;
+    if let Some(recorder) = recorder {
+        recorder
+            .response_headers(response.status().as_u16())
+            .await?;
+    }
+    if response.status().is_success() {
+        return Ok(Attempt::Response(response));
+    }
+    let status = response.status();
+    let bytes = tokio::select! {
+        _ = cancellation.cancelled() => return Ok(Attempt::Cancelled),
+        bytes = response.bytes() => bytes,
+    }?;
+    Err(Error::Provider(format!(
+        "{label} {status}: {}",
+        String::from_utf8_lossy(&bytes)
+    )))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use tokio::io::{AsyncReadExt, AsyncWriteExt};
+
+    async fn server(response: &'static [u8]) -> String {
+        let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+        let address = listener.local_addr().unwrap();
+        tokio::spawn(async move {
+            let (mut socket, _) = listener.accept().await.unwrap();
+            let mut request = [0_u8; 1024];
+            let _ = socket.read(&mut request).await;
+            socket.write_all(response).await.unwrap();
+        });
+        format!("http://{address}")
+    }
+
+    #[tokio::test]
+    async fn non_success_status_is_one_failed_attempt() {
+        let url =
+            server(b"HTTP/1.1 503 Service Unavailable\r\nContent-Length: 4\r\n\r\ndown").await;
+        let client = reqwest::Client::new();
+        let error = send_once("test", || client.get(&url), &CancellationToken::new(), None)
+            .await
+            .unwrap_err();
+        assert!(
+            matches!(error, Error::Provider(message) if message.contains("503") && message.contains("down"))
+        );
+    }
+
+    #[tokio::test]
+    async fn response_body_transport_failure_is_one_failed_attempt() {
+        let url =
+            server(b"HTTP/1.1 500 Internal Server Error\r\nContent-Length: 100\r\n\r\nshort").await;
+        let client = reqwest::Client::new();
+        let error = send_once("test", || client.get(&url), &CancellationToken::new(), None)
+            .await
+            .unwrap_err();
+        assert!(matches!(error, Error::Http(_)));
+    }
+
+    #[tokio::test]
+    async fn request_transport_failure_is_one_failed_attempt() {
+        let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.unwrap();
+        let address = listener.local_addr().unwrap();
+        drop(listener);
+        let url = format!("http://{address}");
+        let client = reqwest::Client::new();
+        let error = send_once("test", || client.get(&url), &CancellationToken::new(), None)
+            .await
+            .unwrap_err();
+        assert!(matches!(error, Error::Http(_)));
+    }
+}

server/src/provider/mod.rs

@@ -1,11 +1,11 @@
 //! Defines the provider interface and exports provider implementations.
 mod anthropic;
+mod attempt;
 mod event;
 mod normalize;
 mod openai_chat;
 mod openai_responses;
 mod recorder;
-mod retry;
 mod router;
 
 use std::pin::Pin;

server/src/provider/openai_chat.rs

@@ -17,10 +17,10 @@ use crate::{
 };
 
 use super::{
-    apply_body_allowlist, apply_openai_prompt_cache_key, map_sse_error, merge_extra_params,
-    provider_event_error,
+    apply_body_allowlist, apply_openai_prompt_cache_key,
+    attempt::{send_once, Attempt},
+    map_sse_error, merge_extra_params, provider_event_error,
     recorder::recorded_headers,
-    retry::{send_with_retry, Attempt, RetryPolicy},
     CallRecorder, FinishReason, ModelEvent, Provider, ProviderStream,
 };
 
@@ -92,15 +92,12 @@ impl Provider for OpenAiChatProvider {
             if let Some(recorder) = &recorder {
                 recorder.request(request_headers.clone(), &body).await?;
             }
-            let attempt = send_with_retry(
+            let attempt = send_once(
                 "OpenAI Chat",
                 || client.post(&config.request_url)
                     .bearer_auth(&config.api_key).headers(config.custom_headers.clone()).json(&body),
-                RetryPolicy { retries: config.retry_count, ..RetryPolicy::default() },
                 &cancellation,
                 recorder.as_ref(),
-                request_headers,
-                &body,
             ).await?;
             let Attempt::Response(response) = attempt else { return };
             yield ModelEvent::Start { model_call_id: call_id };
@@ -421,8 +418,10 @@ fn merge_chat_fragment(target: &mut String, fragment: &str) {
 }
 
 pub(crate) fn openai_usage(value: &Value) -> Usage {
+    let input_tokens = value.get("prompt_tokens").and_then(Value::as_u64);
     Usage {
-        input_tokens: value.get("prompt_tokens").and_then(Value::as_u64),
+        input_tokens,
+        context_input_tokens: input_tokens,
         output_tokens: value.get("completion_tokens").and_then(Value::as_u64),
         total_tokens: value.get("total_tokens").and_then(Value::as_u64),
         cache_read_tokens: value

server/src/provider/openai_responses.rs

@@ -14,10 +14,10 @@ use crate::{
 };
 
 use super::{
-    apply_body_allowlist, apply_openai_prompt_cache_key, map_sse_error, merge_extra_params,
-    provider_event_error,
+    apply_body_allowlist, apply_openai_prompt_cache_key,
+    attempt::{send_once, Attempt},
+    map_sse_error, merge_extra_params, provider_event_error,
     recorder::recorded_headers,
-    retry::{send_with_retry, Attempt, RetryPolicy},
     CallRecorder, FinishReason, ModelEvent, Provider, ProviderStream,
 };
 
@@ -89,15 +89,12 @@ impl Provider for OpenAiResponsesProvider {
             if let Some(recorder) = &recorder {
                 recorder.request(request_headers.clone(), &body).await?;
             }
-            let attempt = send_with_retry(
+            let attempt = send_once(
                 "OpenAI Responses",
                 || client.post(&config.request_url)
                     .bearer_auth(&config.api_key).headers(config.custom_headers.clone()).json(&body),
-                RetryPolicy { retries: config.retry_count, ..RetryPolicy::default() },
                 &cancellation,
                 recorder.as_ref(),
-                request_headers,
-                &body,
             ).await?;
             let Attempt::Response(response) = attempt else { return };
             yield ModelEvent::Start { model_call_id: call_id };
@@ -423,7 +420,12 @@ fn responses_input(messages: &[ProjectedMessage]) -> Result<Vec<Value>> {
                         .ok_or_else(|| {
                             Error::Protocol("OpenAI Responses replay state is missing items".into())
                         })?;
-                    input.extend(items.iter().cloned());
+                    input.extend(
+                        items
+                            .iter()
+                            .map(response_reasoning_input)
+                            .collect::<Result<Vec<_>>>()?,
+                    );
                 }
                 push_responses_text(&mut input, &message.role, text);
                 for call in calls {
@@ -440,6 +442,23 @@ fn responses_input(messages: &[ProjectedMessage]) -> Result<Vec<Value>> {
     Ok(input)
 }
 
+fn response_reasoning_input(item: &Value) -> Result<Value> {
+    let source = item
+        .as_object()
+        .filter(|object| object.get("type").and_then(Value::as_str) == Some("reasoning"))
+        .ok_or_else(|| {
+            Error::Protocol("OpenAI Responses replay state contains a non-reasoning item".into())
+        })?;
+    let mut projected = Map::new();
+    projected.insert("type".into(), json!("reasoning"));
+    for field in ["id", "summary", "content", "encrypted_content"] {
+        if let Some(value) = source.get(field) {
+            projected.insert(field.into(), value.clone());
+        }
+    }
+    Ok(Value::Object(projected))
+}
+
 fn push_responses_parts(input: &mut Vec<Value>, role: &Role, parts: &[ContentPart]) -> Result<()> {
     let text_type = if *role == Role::Assistant {
         "output_text"
@@ -505,8 +524,10 @@ fn required_u64(value: &Value, name: &str) -> Result<u64> {
 }
 
 fn responses_usage(value: &Value) -> Usage {
+    let input_tokens = value.get("input_tokens").and_then(Value::as_u64);
     Usage {
-        input_tokens: value.get("input_tokens").and_then(Value::as_u64),
+        input_tokens,
+        context_input_tokens: input_tokens,
         output_tokens: value.get("output_tokens").and_then(Value::as_u64),
         total_tokens: value.get("total_tokens").and_then(Value::as_u64),
         cache_read_tokens: value
@@ -518,3 +539,47 @@ fn responses_usage(value: &Value) -> Usage {
             .and_then(Value::as_u64),
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::model::ProviderReplayState;
+
+    #[test]
+    fn reasoning_replay_projects_response_items_to_valid_input_items() {
+        let messages = [ProjectedMessage {
+            message_id: "assistant-1".into(),
+            role: Role::Assistant,
+            content: ProjectedContent::Assistant {
+                text: String::new(),
+                thinking: String::new(),
+                replay_state: Some(ProviderReplayState {
+                    provider_kind: "openai_responses".into(),
+                    value: json!({
+                        "items": [{
+                            "type": "reasoning",
+                            "id": "item-1",
+                            "status": "completed",
+                            "summary": [{"type": "summary_text", "text": "why"}],
+                            "content": [],
+                            "encrypted_content": "opaque",
+                            "output_only": true
+                        }]
+                    }),
+                }),
+                calls: Vec::new(),
+            },
+        }];
+
+        assert_eq!(
+            responses_input(&messages).unwrap(),
+            vec![json!({
+                "type": "reasoning",
+                "id": "item-1",
+                "summary": [{"type": "summary_text", "text": "why"}],
+                "content": [],
+                "encrypted_content": "opaque"
+            })]
+        );
+    }
+}

server/src/provider/recorder.rs

@@ -1,7 +1,7 @@
 //! Records provider requests, responses, usage, and timing.
 use std::{
     sync::{
-        atomic::{AtomicBool, AtomicI64, AtomicU32, AtomicU64, Ordering},
+        atomic::{AtomicBool, AtomicI64, AtomicU64, Ordering},
         Arc,
     },
     time::Instant,
@@ -71,7 +71,6 @@ struct Inner {
     base_call: NewLlmCall,
     detailed: bool,
     attempt: Mutex<AttemptState>,
-    next_attempt: AtomicU32,
     next_generation: AtomicU64,
     finished: AtomicBool,
 }
@@ -120,7 +119,6 @@ impl CallRecorder {
                 base_call: call.clone(),
                 detailed: call.detailed,
                 attempt: Mutex::new(AttemptState::new(call.call_id.clone())),
-                next_attempt: AtomicU32::new(0),
                 next_generation: AtomicU64::new(0),
                 finished: AtomicBool::new(false),
             }),
@@ -284,34 +282,6 @@ impl CallRecorder {
         self.finish("cancelled", None, None, None).await
     }
 
-    pub async fn retry(
-        &self,
-        error: &crate::Error,
-        headers: serde_json::Value,
-        body: &serde_json::Value,
-    ) -> Result<()> {
-        self.failed(error).await?;
-
-        let attempt_number = self.inner.next_attempt.fetch_add(1, Ordering::Relaxed) + 1;
-        let mut call = self.inner.base_call.clone();
-        call.call_id = format!("{}:retry-{attempt_number}", self.inner.base_call.call_id);
-        {
-            let mut attempt = self.inner.attempt.lock().await;
-            *attempt = AttemptState::new(call.call_id.clone());
-            self.inner.finished.store(false, Ordering::Release);
-            if let Err(error) = self.inner.store.start_llm_call(&call).await {
-                self.inner.finished.store(true, Ordering::Release);
-                return Err(error);
-            }
-        }
-
-        if let Err(error) = self.request(headers, body).await {
-            self.failed(&error).await?;
-            return Err(error);
-        }
-        Ok(())
-    }
-
     async fn finish(
         &self,
         status: &str,

server/src/provider/retry.rs

@@ -1,84 +0,0 @@
-//! Applies provider retry and backoff behavior.
-use std::time::Duration;
-
-use tokio_util::sync::CancellationToken;
-
-use crate::{Error, Result};
-
-use super::CallRecorder;
-
-#[derive(Clone, Copy, Debug)]
-pub(crate) struct RetryPolicy {
-    pub retries: u32,
-    pub delay: Duration,
-}
-
-impl Default for RetryPolicy {
-    fn default() -> Self {
-        Self {
-            retries: 5,
-            delay: Duration::from_secs(5),
-        }
-    }
-}
-
-#[derive(Debug)]
-pub(crate) enum Attempt {
-    Response(reqwest::Response),
-    Cancelled,
-}
-
-pub(crate) async fn send_with_retry<F>(
-    label: &str,
-    build: F,
-    policy: RetryPolicy,
-    cancellation: &CancellationToken,
-    recorder: Option<&CallRecorder>,
-    request_headers: serde_json::Value,
-    request_body: &serde_json::Value,
-) -> Result<Attempt>
-where
-    F: Fn() -> reqwest::RequestBuilder,
-{
-    for attempt in 0..=policy.retries {
-        let response = tokio::select! {
-            _ = cancellation.cancelled() => return Ok(Attempt::Cancelled),
-            response = build().send() => response,
-        }?;
-        if let Some(recorder) = recorder {
-            recorder
-                .response_headers(response.status().as_u16())
-                .await?;
-        }
-        if response.status().is_success() {
-            return Ok(Attempt::Response(response));
-        }
-        let status = response.status();
-        let bytes = response.bytes().await?;
-        let error = Error::Provider(format!(
-            "{label} {status}: {}",
-            String::from_utf8_lossy(&bytes)
-        ));
-        if attempt == policy.retries {
-            return Err(error);
-        }
-        tracing::warn!(
-            provider = label,
-            status = status.as_u16(),
-            attempt = attempt + 1,
-            retries = policy.retries,
-            delay_ms = policy.delay.as_millis(),
-            "provider returned a non-success status, retrying"
-        );
-        if let Some(recorder) = recorder {
-            recorder
-                .retry(&error, request_headers.clone(), request_body)
-                .await?;
-        }
-        tokio::select! {
-            _ = cancellation.cancelled() => return Ok(Attempt::Cancelled),
-            _ = tokio::time::sleep(policy.delay) => {}
-        }
-    }
-    unreachable!("the retry loop returns on the final attempt")
-}

server/src/provider/router.rs

@@ -18,11 +18,10 @@ use super::{
     OpenAiChatProvider, OpenAiResponsesProvider, Provider, ProviderStream,
 };
 
-const BUILTIN_PROVIDER_RETRIES: u32 = 5;
-
 pub struct ProviderRouter {
     store: Store,
     plugins: PluginRegistry,
+    clients: crate::network::NetworkClients,
     request_timeout: Duration,
     stream_idle_timeout: Duration,
 }
@@ -31,12 +30,14 @@ impl ProviderRouter {
     pub fn new(
         store: Store,
         plugins: PluginRegistry,
+        clients: crate::network::NetworkClients,
         request_timeout: Duration,
         stream_idle_timeout: Duration,
     ) -> Self {
         Self {
             store,
             plugins,
+            clients,
             request_timeout,
             stream_idle_timeout,
         }
@@ -51,6 +52,7 @@ impl Provider for ProviderRouter {
     ) -> ProviderStream {
         let store = self.store.clone();
         let plugins = self.plugins.clone();
+        let clients = self.clients.clone();
         let request_timeout = self.request_timeout;
         let stream_idle_timeout = self.stream_idle_timeout;
         Box::pin(try_stream! {
@@ -64,17 +66,14 @@ impl Provider for ProviderRouter {
                     let plan = plugins.plan_model(&selected).await?;
                     let recorder = start_recorder(&store, &invocation, &selected, &plan.model.display_name, ProviderType::Plugin, &plan.request_url, &plan.model.model_id).await?;
                     let guard = recorder.cancel_on_drop();
-                    recorder.request(serde_json::json!({}), &crate::plugin::plugin_llm_request(&invocation)?).await?;
                     let mut routed = invocation.clone();
                     routed.request.model.display_name = Some(plan.model.display_name.clone());
-                    if let Some(tokens) = plan.model.context_window_tokens {
-                        routed.request.model.context_window_tokens.get_or_insert(tokens);
-                    }
                     if let Some(tokens) = plan.model.max_output_tokens {
                         routed.request.model.max_output_tokens.get_or_insert(tokens);
                     }
                     let provider: Arc<dyn Provider> = Arc::new(NormalizedProvider::new(Arc::new(PluginModelProvider {
                         registry: plugins.clone(),
+                        recorder: recorder.clone(),
                     })));
                     (recorder, guard, provider.stream(routed, cancellation.clone()))
                 } else {
@@ -94,10 +93,9 @@ impl Provider for ProviderRouter {
                         custom_headers: if model.custom_headers_enabled { custom_headers(&model.custom_headers)? } else { reqwest::header::HeaderMap::new() },
                         max_output_tokens: model.max_output_tokens(),
                         request_timeout,
-                        retry_count: BUILTIN_PROVIDER_RETRIES,
                         allowed_body_fields: None,
                     };
-                    let client = crate::network::client_builder(&store).await?.timeout(request_timeout).build()?;
+                    let client = clients.provider_client(request_timeout).await?;
                     let provider = build_observed(&config, recorder.clone(), client)?;
                     (recorder, guard, provider.stream(routed, cancellation.clone()))
                 };
@@ -233,6 +231,7 @@ async fn finish_stream(recorder: &CallRecorder, cancellation: &CancellationToken
 /// 插件模型的 Provider 实现;对路由与规范化层完全等同于内置 Provider。
 struct PluginModelProvider {
     registry: PluginRegistry,
+    recorder: CallRecorder,
 }
 
 impl Provider for PluginModelProvider {
@@ -241,7 +240,8 @@ impl Provider for PluginModelProvider {
         invocation: ModelInvocation,
         cancellation: CancellationToken,
     ) -> ProviderStream {
-        self.registry.stream_model(invocation, cancellation)
+        self.registry
+            .stream_model(invocation, cancellation, self.recorder.clone())
     }
 }
 

server/src/run/compaction.rs

@@ -1,67 +1,78 @@
-//! Decides when to compact context and builds a stable fallback summary.
+//! Decides when to compact provider-visible context and builds a stable fallback summary.
 
 use std::collections::HashSet;
 
-use crate::model::{
-    CanonicalMessage, LlmCallUsageAnchor, PreparedRun, ProjectedMessage, RunAction,
+use crate::{
+    model::{
+        estimate_context_tokens, estimate_projected_messages_tokens, CanonicalMessage, PreparedRun,
+        ProjectedMessage,
+    },
+    store::ContextUsageAnchor,
 };
 
 const FALLBACK_CHARS: usize = 12_000;
 
+pub(super) const RESERVE_TOKENS: u64 = 10_000;
 pub(super) const OUTPUT_TOKENS: u64 = 4_096;
 pub(super) const INSTRUCTIONS: &str = "Summarize the conversation for the next model turn. Preserve goals, constraints, decisions, files, commands, errors, results, and unfinished work. Do not call tools. Return only the concise durable summary.";
 
-#[derive(Clone, Copy, Debug, PartialEq, Eq)]
-pub(super) struct ContextUsageAnchor {
-    input_tokens: u64,
-    message_count: usize,
-    tool_count: usize,
+pub(super) fn input_budget(prepared: &PreparedRun) -> Option<u64> {
+    prepared
+        .model
+        .context_window_tokens
+        .map(|window| window.saturating_sub(RESERVE_TOKENS))
 }
 
-impl ContextUsageAnchor {
-    pub(super) fn from_llm_call(anchor: LlmCallUsageAnchor) -> Option<Self> {
-        Some(Self {
-            input_tokens: anchor.usage.context_input_tokens(anchor.request_type)?,
-            message_count: anchor.message_count,
-            tool_count: anchor.tool_count,
+pub(super) fn estimated_tokens(
+    prepared: &PreparedRun,
+    projected_messages: &[ProjectedMessage],
+    anchor: Option<ContextUsageAnchor>,
+) -> u64 {
+    anchor
+        .filter(|anchor| anchor.message_count <= projected_messages.len())
+        .map(|anchor| {
+            anchor
+                .context_input_tokens
+                .saturating_add(estimate_projected_messages_tokens(
+                    &projected_messages[anchor.message_count..],
+                ))
         })
-    }
+        .unwrap_or_else(|| estimate_context_tokens(&prepared.prompt, projected_messages))
 }
 
+pub(super) fn compaction_estimate(
+    prepared: &PreparedRun,
+    projected_messages: &[ProjectedMessage],
+    anchor: Option<ContextUsageAnchor>,
+) -> Option<u64> {
+    let budget = input_budget(prepared)?;
+    let estimated = estimated_tokens(prepared, projected_messages, anchor);
+    (estimated > budget).then_some(estimated)
+}
+
+#[cfg(test)]
 pub(super) fn should_compact(
     prepared: &PreparedRun,
-    messages: &[CanonicalMessage],
     projected_messages: &[ProjectedMessage],
     anchor: Option<ContextUsageAnchor>,
 ) -> bool {
-    if prepared.action != RunAction::Start {
-        return false;
-    }
-    let Some(context_window) = prepared.model.context_window_tokens else {
-        return false;
+    compaction_estimate(prepared, projected_messages, anchor).is_some()
+}
+
+pub(super) fn validate_compacted(
+    prepared: &PreparedRun,
+    projected_messages: &[ProjectedMessage],
+) -> std::result::Result<u64, String> {
+    let estimated = estimate_context_tokens(&prepared.prompt, projected_messages);
+    let Some(budget) = input_budget(prepared) else {
+        return Ok(estimated);
     };
-    if context_window == 0 || messages.len() <= prepared.initial_messages.len() {
-        return false;
+    if estimated <= budget {
+        return Ok(estimated);
     }
-    let estimated_input = anchor
-        .filter(|anchor| {
-            anchor.message_count <= projected_messages.len()
-                && anchor.tool_count == prepared.prompt.tools.len()
-        })
-        .map(|anchor| {
-            anchor
-                .input_tokens
-                .saturating_add(estimate_serialized_tokens(
-                    &serde_json::to_string(&projected_messages[anchor.message_count..])
-                        .unwrap_or_default(),
-                ))
-        })
-        .unwrap_or_else(|| {
-            estimate_serialized_tokens(
-                &serde_json::to_string(&(&prepared.prompt, messages)).unwrap_or_default(),
-            )
-        });
-    estimated_input > context_window
+    Err(format!(
+        "context overflow after compaction: estimated input {estimated} tokens exceeds budget {budget} tokens"
+    ))
 }
 
 pub(super) fn partition(
@@ -100,28 +111,18 @@ pub(super) fn fallback_summary(messages: &[CanonicalMessage]) -> String {
     )
 }
 
-fn estimate_serialized_tokens(serialized: &str) -> u64 {
-    serialized
-        .chars()
-        .fold(0_u64, |units, character| {
-            units.saturating_add(if character.is_ascii() { 273 } else { 550 })
-        })
-        .div_ceil(1_000)
-}
-
 #[cfg(test)]
 mod tests {
     use super::*;
     use crate::model::{
-        project_messages, CheckpointId, ConversationId, ModelSpec, Origin, PromptSpec, Role, RunId,
-        RunKind,
+        project_messages, CheckpointId, ConversationId, ModelSpec, Origin, PromptSpec, Role,
+        RunAction, RunId, RunKind,
     };
 
-    #[test]
-    fn automatic_compaction_starts_only_after_the_context_window_is_exceeded() {
+    fn prepared(context_window_tokens: u64) -> PreparedRun {
         let mut model = ModelSpec::new("model");
-        model.context_window_tokens = Some(200_000);
-        let prepared = PreparedRun {
+        model.context_window_tokens = Some(context_window_tokens);
+        PreparedRun {
             run_id: RunId::new("run"),
             cursor_request_id: None,
             conversation_id: ConversationId::new("conversation"),
@@ -134,40 +135,133 @@ mod tests {
             initial_messages: Vec::new(),
             action: RunAction::Start,
             base_checkpoint_id: CheckpointId(1),
-        };
+        }
+    }
+
+    #[test]
+    fn automatic_compaction_uses_fixed_reserve_for_every_action() {
         let messages = vec![CanonicalMessage::text(
             "user",
             Role::User,
             Origin::Runtime,
-            "hello",
+            "x".repeat(40_000),
         )];
         let projected = project_messages(&messages).unwrap();
-        let tail_tokens = estimate_serialized_tokens(&serde_json::to_string(&projected).unwrap());
-        let anchor = |estimated_input| {
-            Some(ContextUsageAnchor {
-                input_tokens: estimated_input - tail_tokens,
-                message_count: 0,
-                tool_count: 0,
-            })
-        };
+        let estimated = estimate_context_tokens(&prepared(1).prompt, &projected);
+        let mut prepared = prepared(estimated + RESERVE_TOKENS);
 
+        assert!(!should_compact(&prepared, &projected, None));
+        prepared.model.context_window_tokens = Some(estimated + RESERVE_TOKENS - 1);
+        assert!(should_compact(&prepared, &projected, None));
+
+        prepared.action = RunAction::Resume {
+            pending_tool_round: None,
+        };
+        assert!(should_compact(&prepared, &projected, None));
+    }
+
+    #[test]
+    fn provider_usage_anchor_only_estimates_messages_added_after_last_request() {
+        let messages = vec![
+            CanonicalMessage::text("old", Role::User, Origin::Runtime, "x".repeat(400_000)),
+            CanonicalMessage::text("new", Role::User, Origin::Runtime, "short follow-up"),
+        ];
+        let projected = project_messages(&messages).unwrap();
+        let anchor = ContextUsageAnchor {
+            context_input_tokens: 103_904,
+            message_count: 1,
+        };
+        let expected = 103_904 + estimate_projected_messages_tokens(&projected[1..]);
+
+        assert_eq!(
+            estimated_tokens(&prepared(200_000), &projected, Some(anchor)),
+            expected
+        );
         assert!(!should_compact(
-            &prepared,
-            &messages,
+            &prepared(200_000),
             &projected,
-            anchor(199_999)
-        ));
-        assert!(!should_compact(
-            &prepared,
-            &messages,
-            &projected,
-            anchor(200_000)
-        ));
-        assert!(should_compact(
-            &prepared,
-            &messages,
-            &projected,
-            anchor(200_001)
+            Some(anchor)
         ));
     }
+
+    #[test]
+    fn provider_usage_anchor_triggers_after_new_messages_cross_budget() {
+        let messages = vec![
+            CanonicalMessage::text("old", Role::User, Origin::Runtime, "old"),
+            CanonicalMessage::text("new", Role::User, Origin::Runtime, "x".repeat(80_000)),
+        ];
+        let projected = project_messages(&messages).unwrap();
+
+        assert!(should_compact(
+            &prepared(200_000),
+            &projected,
+            Some(ContextUsageAnchor {
+                context_input_tokens: 180_000,
+                message_count: 1,
+            })
+        ));
+    }
+
+    #[test]
+    fn missing_anchor_uses_full_fallback() {
+        let messages = vec![CanonicalMessage::text(
+            "user",
+            Role::User,
+            Origin::Runtime,
+            "x".repeat(40_000),
+        )];
+        let projected = project_messages(&messages).unwrap();
+        let prepared = prepared(200_000);
+
+        assert_eq!(
+            estimated_tokens(&prepared, &projected, None),
+            estimate_context_tokens(&prepared.prompt, &projected)
+        );
+    }
+
+    #[test]
+    fn invalid_anchor_message_count_uses_full_fallback() {
+        let messages = vec![CanonicalMessage::text(
+            "user",
+            Role::User,
+            Origin::Runtime,
+            "x".repeat(40_000),
+        )];
+        let projected = project_messages(&messages).unwrap();
+        let expected = estimate_context_tokens(&prepared(200_000).prompt, &projected);
+
+        assert_eq!(
+            estimated_tokens(
+                &prepared(200_000),
+                &projected,
+                Some(ContextUsageAnchor {
+                    context_input_tokens: 1,
+                    message_count: 2,
+                })
+            ),
+            expected
+        );
+    }
+
+    #[test]
+    fn compacted_history_is_validated_against_the_same_budget() {
+        let messages = vec![CanonicalMessage::text(
+            "user",
+            Role::User,
+            Origin::Runtime,
+            "x".repeat(40_000),
+        )];
+        let projected = project_messages(&messages).unwrap();
+        let estimated = estimate_context_tokens(&prepared(1).prompt, &projected);
+
+        assert_eq!(
+            validate_compacted(&prepared(estimated + RESERVE_TOKENS), &projected),
+            Ok(estimated)
+        );
+        assert!(
+            validate_compacted(&prepared(estimated + RESERVE_TOKENS - 1), &projected)
+                .unwrap_err()
+                .contains("context overflow after compaction")
+        );
+    }
 }

server/src/run/engine.rs

@@ -10,12 +10,14 @@ use crate::{
         ToolRoundId, Usage,
     },
     provider::Provider,
-    store::{RunStatus, Store},
+    store::{ContextUsageAnchor, RunStatus, Store},
 };
 
 use super::{
-    consume_model_cycle, CommitBarrier, CommitCause, MessagesCommitted, ModelCycleFailure,
-    RunCommand, RunEvent, RunFailure, RunOutcome, RunPort,
+    consume_model_cycle,
+    model_retry::{should_retry, MODEL_RETRY_DELAY},
+    CommitBarrier, CommitCause, MessagesCommitted, RunCommand, RunEvent, RunFailure, RunOutcome,
+    RunPort,
 };
 
 pub struct RunEngine {
@@ -90,6 +92,14 @@ impl RunEngine {
         cancellation: &CancellationToken,
     ) -> (RunOutcome, Option<Usage>) {
         let mut usage = None;
+        let mut context_usage_anchor = match self
+            .store
+            .latest_context_usage(prepared.conversation_id.as_str())
+            .await
+        {
+            Ok(anchor) => anchor,
+            Err(error) => return (RunOutcome::Failed(error.into()), usage),
+        };
         tracing::info!(
             checkpoint_id = checkpoint.0,
             "Run claimed conversation ownership"
@@ -161,7 +171,6 @@ impl RunEngine {
             };
         }
 
-        let mut auto_compacted = prepared.action == RunAction::Compact;
         'model: loop {
             if cancellation.is_cancelled() {
                 return (RunOutcome::Cancelled, usage);
@@ -170,37 +179,32 @@ impl RunEngine {
                 Ok(messages) => messages,
                 Err(error) => return (RunOutcome::Failed(error.into()), usage),
             };
-            let context_anchor = if !auto_compacted && prepared.action == RunAction::Start {
-                match self
-                    .store
-                    .latest_llm_call_usage_anchor(
-                        &prepared.conversation_id,
-                        &prepared.model.model_id,
-                    )
-                    .await
-                {
-                    Ok(anchor) => {
-                        anchor.and_then(super::compaction::ContextUsageAnchor::from_llm_call)
-                    }
-                    Err(error) => return (RunOutcome::Failed(error.into()), usage),
-                }
-            } else {
-                None
-            };
             let history = match crate::model::project_messages(&messages) {
                 Ok(history) => history,
                 Err(error) => return (RunOutcome::Failed(error.into()), usage),
             };
-            if !auto_compacted
-                && super::compaction::should_compact(prepared, &messages, &history, context_anchor)
-            {
-                auto_compacted = true;
+            let compaction_estimate = (prepared.action != RunAction::Compact)
+                .then(|| {
+                    super::compaction::compaction_estimate(prepared, &history, context_usage_anchor)
+                })
+                .flatten();
+            if let Some(estimated_tokens) = compaction_estimate {
+                if emit(
+                    client,
+                    RunEvent::UsageSnapshot(context_usage_snapshot(estimated_tokens)),
+                )
+                .await
+                .is_err()
+                {
+                    return (client_failure(), usage);
+                }
                 match self
                     .auto_compact(prepared, checkpoint, &messages, client, cancellation)
                     .await
                 {
                     Ok((next_checkpoint, compaction_usage)) => {
                         checkpoint = next_checkpoint;
+                        context_usage_anchor = None;
                         if let Some(compaction_usage) = compaction_usage {
                             accumulate_usage(&mut usage, compaction_usage);
                         }
@@ -227,122 +231,236 @@ impl RunEngine {
                 model: prepared.model.clone(),
                 history,
             };
-            let invocation = crate::model::ModelInvocation {
-                call_id: format!("{}:{provider_call_index}", prepared.run_id),
-                run_id: prepared.run_id.to_string(),
-                conversation_id: prepared.conversation_id.to_string(),
-                provider_call_index,
-                request,
-            };
-            let cycle_cancellation = cancellation.child_token();
-            let cycle_events = client.events.clone();
-            let cycle = consume_model_cycle(
-                self.provider.stream(invocation, cycle_cancellation.clone()),
-                &cycle_events,
-                &cycle_cancellation,
-            );
-            tokio::pin!(cycle);
+            let mut retries = 0_u32;
             let mut pending_insertions = Vec::new();
-            let cycle = loop {
-                tokio::select! {
-                    biased;
-                    command = client.commands.recv() => {
-                        let interruption = match command {
-                            Some(RunCommand::InsertMessages(insertion)) => {
-                                pending_insertions.push(insertion);
-                                continue;
+            let cycle = 'attempt: loop {
+                let call_id = if retries == 0 {
+                    format!("{}:{provider_call_index}", prepared.run_id)
+                } else {
+                    format!("{}:{provider_call_index}:retry-{retries}", prepared.run_id)
+                };
+                let invocation = crate::model::ModelInvocation {
+                    call_id,
+                    run_id: prepared.run_id.to_string(),
+                    conversation_id: prepared.conversation_id.to_string(),
+                    provider_call_index,
+                    request: request.clone(),
+                };
+                let cycle_cancellation = cancellation.child_token();
+                let cycle_events = client.events.clone();
+                let cycle = consume_model_cycle(
+                    self.provider.stream(invocation, cycle_cancellation.clone()),
+                    &cycle_events,
+                    &cycle_cancellation,
+                );
+                tokio::pin!(cycle);
+                let cycle = loop {
+                    tokio::select! {
+                        biased;
+                        command = client.commands.recv() => {
+                            let interruption = match command {
+                                Some(RunCommand::InsertMessages(insertion)) => {
+                                    pending_insertions.push(insertion);
+                                    continue;
+                                }
+                                Some(RunCommand::BreakMessages(messages)) => messages,
+                                Some(RunCommand::Cancel) => {
+                                    cycle_cancellation.cancel();
+                                    let _ = cycle.await;
+                                    let _ = emit(client, RunEvent::CycleInterrupted).await;
+                                    return (RunOutcome::Cancelled, usage);
+                                }
+                                Some(RunCommand::ToolResult(_)) => {
+                                    cycle_cancellation.cancel();
+                                    let _ = cycle.await;
+                                    let _ = emit(client, RunEvent::CycleInterrupted).await;
+                                    return (
+                                        RunOutcome::Failed(RunFailure::Protocol(
+                                            "received a tool result while the model was running".into(),
+                                        )),
+                                        usage,
+                                    );
+                                }
+                                None => {
+                                    cycle_cancellation.cancel();
+                                    let _ = cycle.await;
+                                    let _ = emit(client, RunEvent::CycleInterrupted).await;
+                                    return (client_failure(), usage);
+                                }
+                            };
+                            cycle_cancellation.cancel();
+                            let interrupted = cycle.await;
+                            match interrupted {
+                                Ok(cycle) => {
+                                    if let Some(cycle_usage) = cycle.usage {
+                                        update_context_usage_anchor(
+                                            &mut context_usage_anchor,
+                                            cycle_usage,
+                                            request.history.len(),
+                                        );
+                                        accumulate_usage(&mut usage, cycle_usage);
+                                    }
+                                }
+                                Err(failure) => {
+                                    if let Some(cycle_usage) = failure.usage {
+                                        update_context_usage_anchor(
+                                            &mut context_usage_anchor,
+                                            cycle_usage,
+                                            request.history.len(),
+                                        );
+                                        accumulate_usage(&mut usage, cycle_usage);
+                                    }
+                                }
                             }
-                            Some(RunCommand::BreakMessages(messages)) => messages,
-                            Some(RunCommand::Cancel) => {
-                                cycle_cancellation.cancel();
-                                let _ = cycle.await;
-                                let _ = emit(client, RunEvent::CycleInterrupted).await;
-                                return (RunOutcome::Cancelled, usage);
-                            }
-                            Some(RunCommand::ToolResult(_)) => {
-                                cycle_cancellation.cancel();
-                                let _ = cycle.await;
-                                let _ = emit(client, RunEvent::CycleInterrupted).await;
-                                return (
-                                    RunOutcome::Failed(RunFailure::Protocol(
-                                        "received a tool result while the model was running".into(),
-                                    )),
-                                    usage,
-                                );
-                            }
-                            None => {
-                                cycle_cancellation.cancel();
-                                let _ = cycle.await;
-                                let _ = emit(client, RunEvent::CycleInterrupted).await;
+                            if emit(client, RunEvent::CycleInterrupted).await.is_err() {
                                 return (client_failure(), usage);
                             }
-                        };
-                        cycle_cancellation.cancel();
-                        let interrupted = cycle.await;
-                        match interrupted {
-                            Ok(cycle) => {
-                                if let Some(cycle_usage) = cycle.usage {
-                                    accumulate_usage(&mut usage, cycle_usage);
-                                }
-                            }
-                            Err(failure) => {
-                                if let Some(cycle_usage) = failure.usage {
-                                    accumulate_usage(&mut usage, cycle_usage);
-                                }
-                            }
+                            checkpoint = match super::messages::append_batches(
+                                &self.store,
+                                prepared,
+                                client,
+                                cancellation,
+                                checkpoint,
+                                std::mem::take(&mut pending_insertions),
+                            )
+                            .await
+                            {
+                                Ok((checkpoint, _)) => checkpoint,
+                                Err(outcome) => return (outcome, usage),
+                            };
+                            checkpoint = match super::messages::append_batches(
+                                &self.store,
+                                prepared,
+                                client,
+                                cancellation,
+                                checkpoint,
+                                vec![interruption],
+                            )
+                            .await
+                            {
+                                Ok((checkpoint, _)) => checkpoint,
+                                Err(outcome) => return (outcome, usage),
+                            };
+                            continue 'model;
+                        },
+                        result = &mut cycle => break result,
+                    }
+                };
+                match cycle {
+                    Ok(cycle) => break 'attempt cycle,
+                    Err(cycle_failure) => {
+                        if let Some(cycle_usage) = cycle_failure.usage {
+                            update_context_usage_anchor(
+                                &mut context_usage_anchor,
+                                cycle_usage,
+                                request.history.len(),
+                            );
+                            accumulate_usage(&mut usage, cycle_usage);
                         }
-                        if emit(client, RunEvent::CycleInterrupted).await.is_err() {
+                        if cancellation.is_cancelled() {
+                            let _ = emit(client, RunEvent::CycleInterrupted).await;
+                            return (RunOutcome::Cancelled, usage);
+                        }
+                        if !should_retry(&cycle_failure, retries) {
+                            return (RunOutcome::Failed(cycle_failure.failure), usage);
+                        }
+                        retries += 1;
+                        let message = failure_message(&cycle_failure.failure);
+                        tracing::warn!(
+                            provider_call_index,
+                            retries,
+                            max_retries = super::model_retry::MAX_MODEL_RETRIES,
+                            delay_ms = MODEL_RETRY_DELAY.as_millis() as u64,
+                            %message,
+                            checkpoint_id = checkpoint.0,
+                            "model attempt failed; retrying from current checkpoint"
+                        );
+                        if emit(
+                            client,
+                            RunEvent::ModelAttemptFailed {
+                                attempt: retries,
+                                message,
+                            },
+                        )
+                        .await
+                        .is_err()
+                        {
                             return (client_failure(), usage);
                         }
-                        checkpoint = match super::messages::append_batches(
-                            &self.store,
-                            prepared,
-                            client,
-                            cancellation,
-                            checkpoint,
-                            std::mem::take(&mut pending_insertions),
-                        )
-                        .await
-                        {
-                            Ok((checkpoint, _)) => checkpoint,
-                            Err(outcome) => return (outcome, usage),
-                        };
-                        checkpoint = match super::messages::append_batches(
-                            &self.store,
-                            prepared,
-                            client,
-                            cancellation,
-                            checkpoint,
-                            vec![interruption],
-                        )
-                        .await
-                        {
-                            Ok((checkpoint, _)) => checkpoint,
-                            Err(outcome) => return (outcome, usage),
-                        };
-                        continue 'model;
-                    },
-                    result = &mut cycle => break result,
-                }
-            };
-            let cycle = match cycle {
-                Ok(cycle) => cycle,
-                Err(ModelCycleFailure {
-                    failure,
-                    usage: cycle_usage,
-                    ..
-                }) => {
-                    if let Some(cycle_usage) = cycle_usage {
-                        accumulate_usage(&mut usage, cycle_usage);
+
+                        let delay = tokio::time::sleep(MODEL_RETRY_DELAY);
+                        tokio::pin!(delay);
+                        loop {
+                            tokio::select! {
+                                biased;
+                                command = client.commands.recv() => {
+                                    let interruption = match command {
+                                        Some(RunCommand::InsertMessages(insertion)) => {
+                                            pending_insertions.push(insertion);
+                                            continue;
+                                        }
+                                        Some(RunCommand::BreakMessages(messages)) => messages,
+                                        Some(RunCommand::Cancel) => {
+                                            let _ = emit(client, RunEvent::CycleInterrupted).await;
+                                            return (RunOutcome::Cancelled, usage);
+                                        }
+                                        Some(RunCommand::ToolResult(_)) => {
+                                            return (
+                                                RunOutcome::Failed(RunFailure::Protocol(
+                                                    "received a tool result while waiting to retry the model".into(),
+                                                )),
+                                                usage,
+                                            );
+                                        }
+                                        None => return (client_failure(), usage),
+                                    };
+                                    if emit(client, RunEvent::CycleInterrupted).await.is_err() {
+                                        return (client_failure(), usage);
+                                    }
+                                    checkpoint = match super::messages::append_batches(
+                                        &self.store,
+                                        prepared,
+                                        client,
+                                        cancellation,
+                                        checkpoint,
+                                        std::mem::take(&mut pending_insertions),
+                                    )
+                                    .await
+                                    {
+                                        Ok((checkpoint, _)) => checkpoint,
+                                        Err(outcome) => return (outcome, usage),
+                                    };
+                                    checkpoint = match super::messages::append_batches(
+                                        &self.store,
+                                        prepared,
+                                        client,
+                                        cancellation,
+                                        checkpoint,
+                                        vec![interruption],
+                                    )
+                                    .await
+                                    {
+                                        Ok((checkpoint, _)) => checkpoint,
+                                        Err(outcome) => return (outcome, usage),
+                                    };
+                                    continue 'model;
+                                }
+                                _ = cancellation.cancelled() => {
+                                    let _ = emit(client, RunEvent::CycleInterrupted).await;
+                                    return (RunOutcome::Cancelled, usage);
+                                }
+                                _ = &mut delay => break,
+                            }
+                        }
                     }
-                    if cancellation.is_cancelled() {
-                        let _ = emit(client, RunEvent::CycleInterrupted).await;
-                        return (RunOutcome::Cancelled, usage);
-                    }
-                    return (RunOutcome::Failed(failure), usage);
                 }
             };
             if let Some(cycle_usage) = cycle.usage {
+                update_context_usage_anchor(
+                    &mut context_usage_anchor,
+                    cycle_usage,
+                    request.history.len(),
+                );
                 accumulate_usage(&mut usage, cycle_usage);
             }
 
@@ -583,7 +701,15 @@ impl RunEngine {
         let (compactable, retained_request_context) =
             super::compaction::partition(messages, &current_ids);
         if compactable.is_empty() {
-            return Ok((checkpoint, None));
+            let projected = crate::model::project_messages(messages)
+                .map_err(|error| RunOutcome::Failed(error.into()))?;
+            let message = super::compaction::validate_compacted(prepared, &projected)
+                .err()
+                .unwrap_or_else(|| {
+                    "context overflow after compaction: no conversation history can be compacted"
+                        .into()
+                });
+            return Err(RunOutcome::Failed(RunFailure::Protocol(message)));
         }
 
         emit(client, RunEvent::AutoCompactionStarted)
@@ -689,7 +815,7 @@ impl RunEngine {
                 )
             }
         };
-        let event_id = format!("summary:auto:{}", prepared.run_id);
+        let event_id = format!("summary:auto:{}:{provider_call_index}", prepared.run_id);
         let summary_message = CanonicalMessage {
             message_id: format!("runtime:{event_id}"),
             role: Role::User,
@@ -704,6 +830,10 @@ impl RunEngine {
         let mut replacement = retained_request_context.into_iter().collect::<Vec<_>>();
         replacement.push(summary_message);
         replacement.extend(prepared.initial_messages.iter().cloned());
+        let projected_replacement = crate::model::project_messages(&replacement)
+            .map_err(|error| RunOutcome::Failed(error.into()))?;
+        super::compaction::validate_compacted(prepared, &projected_replacement)
+            .map_err(|message| RunOutcome::Failed(RunFailure::Protocol(message)))?;
         let mut checkpoint = self
             .store
             .replace_checkpoint(
@@ -730,6 +860,9 @@ impl RunEngine {
         emit(client, RunEvent::AutoCompactionCompleted)
             .await
             .map_err(|_| client_failure())?;
+        emit(client, RunEvent::UsageSnapshot(context_usage_snapshot(0)))
+            .await
+            .map_err(|_| client_failure())?;
         checkpoint = super::messages::append_batches(
             &self.store,
             prepared,
@@ -790,6 +923,31 @@ async fn hydrate_tool_images(
     Ok(())
 }
 
+fn context_usage_snapshot(tokens: u64) -> Usage {
+    Usage {
+        input_tokens: Some(tokens),
+        context_input_tokens: Some(tokens),
+        output_tokens: Some(0),
+        total_tokens: Some(tokens),
+        cache_read_tokens: Some(0),
+        cache_write_tokens: Some(0),
+        reasoning_tokens: Some(0),
+    }
+}
+
+fn update_context_usage_anchor(
+    anchor: &mut Option<ContextUsageAnchor>,
+    usage: Usage,
+    message_count: usize,
+) {
+    if let Some(context_input_tokens) = usage.context_input_tokens {
+        *anchor = Some(ContextUsageAnchor {
+            context_input_tokens,
+            message_count,
+        });
+    }
+}
+
 fn accumulate_usage(total: &mut Option<Usage>, usage: Usage) {
     match total {
         Some(total) => *total += usage,

server/src/run/event.rs

@@ -104,6 +104,10 @@ pub enum RunEvent {
     AutoCompactionStarted,
     AutoCompactionCompleted,
     CycleInterrupted,
+    ModelAttemptFailed {
+        attempt: u32,
+        message: String,
+    },
     TextStart,
     TextDelta(String),
     TextEnd,
@@ -125,6 +129,7 @@ pub enum RunEvent {
     ToolCallEnd {
         index: usize,
     },
+    UsageSnapshot(Usage),
     Usage(Usage),
     ExecuteToolRound {
         round_id: ToolRoundId,

server/src/run/mod.rs

@@ -7,6 +7,7 @@ mod event;
 mod handle;
 mod messages;
 mod model_cycle;
+mod model_retry;
 mod port;
 mod tool_round;
 

server/src/run/model_cycle.rs

@@ -6,7 +6,7 @@ use tokio::sync::mpsc;
 use tokio_util::sync::CancellationToken;
 
 use crate::{
-    model::{ProviderReplayState, ToolCall, Usage},
+    model::{normalize_tool_name, ProviderReplayState, ToolCall, Usage},
     provider::{FinishReason, ModelEvent, ProviderStream},
 };
 
@@ -29,6 +29,7 @@ pub struct ModelCycleFailure {
     pub partial_text: String,
     pub partial_reasoning: String,
     pub usage: Option<Usage>,
+    pub retryable: bool,
 }
 
 struct OpenTool {
@@ -164,6 +165,7 @@ pub async fn consume_model_cycle(
                 call_id,
                 name,
             } => {
+                let name = normalize_tool_name(&name);
                 let Some(model_call_id) = model_call_id.as_ref() else {
                     return Err(failure(
                         RunFailure::Protocol("provider emitted content before Start".into()),
@@ -186,6 +188,7 @@ pub async fn consume_model_cycle(
                                 name: name.clone(),
                                 arguments_text: String::new(),
                                 arguments: serde_json::Value::Null,
+                                argument_error: None,
                             },
                             ended: false,
                         });
@@ -218,13 +221,26 @@ pub async fn consume_model_cycle(
                         serde_json::from_str(&tool.call.arguments_text)
                     };
                     match arguments {
-                        Ok(arguments) => {
+                        Ok(arguments) if arguments.is_object() => {
                             tool.call.arguments = arguments;
-                            tool.ended = true;
-                            send(client, RunEvent::ToolCallEnd { index }).await
                         }
-                        Err(_) => Err("provider ended a tool call with invalid JSON arguments"),
+                        Ok(_) => {
+                            tool.call.arguments = serde_json::json!({});
+                            tool.call.argument_error = Some(format!(
+                                "{} arguments must be a JSON object",
+                                tool.call.name
+                            ));
+                        }
+                        Err(error) => {
+                            tool.call.arguments = serde_json::json!({});
+                            tool.call.argument_error = Some(format!(
+                                "{} arguments are not valid JSON: {error}",
+                                tool.call.name
+                            ));
+                        }
                     }
+                    tool.ended = true;
+                    send(client, RunEvent::ToolCallEnd { index }).await
                 }
                 Some(_) => Err("provider emitted duplicate ToolCallEnd"),
                 None => Err("provider ended an unknown tool index"),
@@ -239,6 +255,13 @@ pub async fn consume_model_cycle(
             ModelEvent::Usage(value) => {
                 if usage.replace(value).is_some() {
                     Err("provider emitted duplicate Usage")
+                } else if send(client, RunEvent::Usage(value)).await.is_err() {
+                    return Err(failure(
+                        RunFailure::Client("client event channel closed".into()),
+                        text,
+                        reasoning,
+                        usage,
+                    ));
                 } else {
                     Ok(())
                 }
@@ -280,7 +303,7 @@ pub async fn consume_model_cycle(
         .map(|tool| tool.call)
         .collect::<Vec<_>>();
     if finish_reason == FinishReason::Length {
-        return Err(failure(
+        return Err(terminal_failure(
             RunFailure::Provider("model stopped before completing the response".into()),
             text,
             reasoning,
@@ -296,16 +319,6 @@ pub async fn consume_model_cycle(
             usage,
         ));
     }
-    if let Some(usage) = usage {
-        if send(client, RunEvent::Usage(usage)).await.is_err() {
-            return Err(failure(
-                RunFailure::Client("client event channel closed".into()),
-                text,
-                reasoning,
-                Some(usage),
-            ));
-        }
-    }
     let model_call_id = model_call_id.ok_or_else(|| {
         failure(
             RunFailure::Protocol("provider completed without Start".into()),
@@ -360,10 +373,111 @@ fn failure(
     partial_reasoning: String,
     usage: Option<Usage>,
 ) -> ModelCycleFailure {
+    let retryable = matches!(failure, RunFailure::Protocol(_) | RunFailure::Provider(_));
     ModelCycleFailure {
         failure,
         partial_text,
         partial_reasoning,
         usage,
+        retryable,
+    }
+}
+
+fn terminal_failure(
+    failure: RunFailure,
+    partial_text: String,
+    partial_reasoning: String,
+    usage: Option<Usage>,
+) -> ModelCycleFailure {
+    ModelCycleFailure {
+        failure,
+        partial_text,
+        partial_reasoning,
+        usage,
+        retryable: false,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::{
+        model::Usage,
+        provider::{FinishReason, ModelEvent},
+    };
+    use tokio_stream::wrappers::ReceiverStream;
+
+    #[tokio::test]
+    async fn provider_tool_names_are_normalized_when_received() {
+        let events = vec![
+            Ok(ModelEvent::Start {
+                model_call_id: "call".into(),
+            }),
+            Ok(ModelEvent::ToolCallStart {
+                index: 0,
+                call_id: "tool-call".into(),
+                name: "multi_tool_use.parallel".into(),
+            }),
+            Ok(ModelEvent::ToolCallEnd { index: 0 }),
+            Ok(ModelEvent::Done(FinishReason::ToolUse)),
+        ];
+        let stream = Box::pin(tokio_stream::iter(events));
+        let (event_tx, mut event_rx) = tokio::sync::mpsc::channel(4);
+
+        let result = consume_model_cycle(stream, &event_tx, &CancellationToken::new())
+            .await
+            .unwrap();
+
+        assert_eq!(result.calls[0].name, "multi_tool_use_parallel");
+        assert!(matches!(
+            event_rx.recv().await,
+            Some(RunEvent::ToolCallStart { name, .. }) if name == "multi_tool_use_parallel"
+        ));
+    }
+
+    #[tokio::test]
+    async fn usage_is_forwarded_before_the_provider_call_finishes() {
+        let (provider_tx, provider_rx) = tokio::sync::mpsc::channel(4);
+        let stream = Box::pin(ReceiverStream::new(provider_rx));
+        let (event_tx, mut event_rx) = tokio::sync::mpsc::channel(4);
+        let cancellation = CancellationToken::new();
+        let cycle_cancellation = cancellation.clone();
+        let cycle = tokio::spawn(async move {
+            consume_model_cycle(stream, &event_tx, &cycle_cancellation).await
+        });
+        let usage = Usage {
+            input_tokens: Some(100),
+            context_input_tokens: Some(100),
+            output_tokens: Some(20),
+            total_tokens: Some(120),
+            ..Default::default()
+        };
+
+        provider_tx
+            .send(Ok(ModelEvent::Start {
+                model_call_id: "call".into(),
+            }))
+            .await
+            .unwrap();
+        provider_tx
+            .send(Ok(ModelEvent::Usage(usage)))
+            .await
+            .unwrap();
+
+        let event = tokio::time::timeout(std::time::Duration::from_secs(1), event_rx.recv())
+            .await
+            .unwrap()
+            .unwrap();
+        assert!(matches!(event, RunEvent::Usage(value) if value == usage));
+        assert!(!cycle.is_finished(), "usage must arrive before Done");
+
+        provider_tx
+            .send(Ok(ModelEvent::Done(FinishReason::Stop)))
+            .await
+            .unwrap();
+        drop(provider_tx);
+        let result = cycle.await.unwrap().unwrap();
+        assert_eq!(result.usage, Some(usage));
+        assert!(event_rx.try_recv().is_err(), "usage must be forwarded once");
     }
 }

server/src/run/model_retry.rs

@@ -0,0 +1,42 @@
+//! Defines retry policy for one logical model call.
+
+use std::time::Duration;
+
+use super::ModelCycleFailure;
+
+pub(super) const MAX_MODEL_RETRIES: u32 = 8;
+pub(super) const MODEL_RETRY_DELAY: Duration = Duration::from_secs(5);
+
+pub(super) fn should_retry(failure: &ModelCycleFailure, retries: u32) -> bool {
+    failure.retryable && retries < MAX_MODEL_RETRIES
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::run::{ModelCycleFailure, RunFailure};
+
+    fn failure(retryable: bool) -> ModelCycleFailure {
+        ModelCycleFailure {
+            failure: RunFailure::Provider("failed".into()),
+            partial_text: String::new(),
+            partial_reasoning: String::new(),
+            usage: None,
+            retryable,
+        }
+    }
+
+    #[test]
+    fn permits_eight_retries_after_the_initial_attempt() {
+        let retryable = failure(true);
+        for retries in 0..MAX_MODEL_RETRIES {
+            assert!(should_retry(&retryable, retries));
+        }
+        assert!(!should_retry(&retryable, MAX_MODEL_RETRIES));
+    }
+
+    #[test]
+    fn terminal_failures_never_retry() {
+        assert!(!should_retry(&failure(false), 0));
+    }
+}

server/src/store/cursor_traces.rs

@@ -62,6 +62,40 @@ impl Store {
         .await?)
     }
 
+    pub async fn append_cursor_trace_request(
+        &self,
+        request_id: &str,
+        artifact_type: &str,
+        source: &str,
+        data: &[u8],
+        metadata: &serde_json::Value,
+    ) -> Result<()> {
+        let metadata_json = serde_json::to_string(metadata)?;
+        let blob_id = BlobId::digest(data);
+        let _write = self.writes.lock().await;
+        let mut tx = self.pool.begin_with("BEGIN IMMEDIATE").await?;
+        Self::put_blob_tx(&mut tx, &blob_id, data, &[]).await?;
+        Self::link_cursor_trace_artifact_tx(
+            &mut tx,
+            request_id,
+            artifact_type,
+            source,
+            &blob_id,
+            &metadata_json,
+        )
+        .await?;
+        sqlx::query(
+            "UPDATE cursor_run_traces
+             SET request_bytes = request_bytes + ? WHERE request_id = ?",
+        )
+        .bind(as_i64(data.len()))
+        .bind(request_id)
+        .execute(&mut *tx)
+        .await?;
+        tx.commit().await?;
+        Ok(())
+    }
+
     pub async fn append_cursor_trace_artifact(
         &self,
         request_id: &str,
@@ -144,23 +178,6 @@ impl Store {
         Ok(())
     }
 
-    pub async fn add_cursor_trace_request_bytes(
-        &self,
-        request_id: &str,
-        bytes: usize,
-    ) -> Result<()> {
-        let _write = self.writes.lock().await;
-        sqlx::query(
-            "UPDATE cursor_run_traces
-             SET request_bytes = request_bytes + ? WHERE request_id = ?",
-        )
-        .bind(as_i64(bytes))
-        .bind(request_id)
-        .execute(&self.pool)
-        .await?;
-        Ok(())
-    }
-
     pub async fn start_cursor_trace_response(&self, request_id: &str, status: u16) -> Result<()> {
         let now = now_ms();
         let _write = self.writes.lock().await;

server/src/store/llm_calls.rs

@@ -1,18 +1,19 @@
 //! Persists provider call payloads, timing, and usage.
-use std::str::FromStr;
-
 use sqlx::Row;
 
 use crate::{
-    model::{
-        ConversationId, LlmCallRequest, LlmCallResponseChunk, LlmCallSummary, LlmCallUsageAnchor,
-        NewLlmCall, ProviderType, Usage,
-    },
+    model::{LlmCallRequest, LlmCallResponseChunk, LlmCallSummary, NewLlmCall, Usage},
     Result,
 };
 
 use super::{now_ms, Store};
 
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub(crate) struct ContextUsageAnchor {
+    pub(crate) context_input_tokens: u64,
+    pub(crate) message_count: usize,
+}
+
 #[derive(Clone, Debug)]
 pub(crate) struct BufferedLlmChunk {
     pub(crate) seq: i64,
@@ -271,6 +272,37 @@ impl Store {
         Ok(())
     }
 
+    pub(crate) async fn latest_context_usage(
+        &self,
+        conversation_id: &str,
+    ) -> Result<Option<ContextUsageAnchor>> {
+        let row = sqlx::query(
+            "SELECT usage_json, message_count FROM llm_calls
+             WHERE conversation_id = ?
+               AND json_extract(usage_json, '$.context_input_tokens') IS NOT NULL
+             ORDER BY created_at_ms DESC, rowid DESC
+             LIMIT 1",
+        )
+        .bind(conversation_id)
+        .fetch_optional(&self.pool)
+        .await?;
+        let Some(row) = row else {
+            return Ok(None);
+        };
+        let usage: Usage = serde_json::from_str(row.try_get("usage_json")?)?;
+        let Some(context_input_tokens) = usage.context_input_tokens else {
+            return Ok(None);
+        };
+        let message_count = row.try_get::<i64, _>("message_count")?;
+        let Ok(message_count) = usize::try_from(message_count) else {
+            return Ok(None);
+        };
+        Ok(Some(ContextUsageAnchor {
+            context_input_tokens,
+            message_count,
+        }))
+    }
+
     pub async fn llm_calls(&self, limit: i64) -> Result<Vec<LlmCallSummary>> {
         let rows = sqlx::query("SELECT * FROM llm_calls ORDER BY created_at_ms DESC LIMIT ?")
             .bind(limit.clamp(1, 500))
@@ -288,41 +320,6 @@ impl Store {
             .transpose()
     }
 
-    pub(crate) async fn latest_llm_call_usage_anchor(
-        &self,
-        conversation_id: &ConversationId,
-        model_hash: &str,
-    ) -> Result<Option<LlmCallUsageAnchor>> {
-        let row = sqlx::query(
-            r#"SELECT request_type, usage_json, message_count, tool_count
-               FROM llm_calls
-               WHERE conversation_id = ?
-                 AND model_hash = ?
-                 AND status = 'completed'
-                 AND input_tokens IS NOT NULL
-                 AND usage_json IS NOT NULL
-               ORDER BY rowid DESC
-               LIMIT 1"#,
-        )
-        .bind(conversation_id.as_str())
-        .bind(model_hash)
-        .fetch_optional(&self.pool)
-        .await?;
-        row.map(|row| {
-            let message_count =
-                usize::try_from(row.try_get::<i64, _>("message_count")?).unwrap_or(usize::MAX);
-            let tool_count =
-                usize::try_from(row.try_get::<i64, _>("tool_count")?).unwrap_or(usize::MAX);
-            Ok(LlmCallUsageAnchor {
-                request_type: ProviderType::from_str(row.try_get("request_type")?)?,
-                usage: serde_json::from_str(row.try_get("usage_json")?)?,
-                message_count,
-                tool_count,
-            })
-        })
-        .transpose()
-    }
-
     pub async fn llm_call_request(&self, call_id: &str) -> Result<Option<LlmCallRequest>> {
         let row = sqlx::query(
             "SELECT headers_json, body_json, byte_count FROM llm_call_requests WHERE call_id = ?",
@@ -416,6 +413,7 @@ fn summary_from_row(row: sqlx::sqlite::SqliteRow) -> Result<LlmCallSummary> {
 #[cfg(test)]
 mod tests {
     use super::*;
+    use crate::model::ProviderType;
 
     /// 插件模型不在 model_configs 中,调用记录必须照常落库并可按其稳定 ID 筛选。
     #[tokio::test]
@@ -460,4 +458,65 @@ mod tests {
         assert_eq!(overview.metrics.llm_calls, 1);
         assert_eq!(overview.metrics.successful_calls, 1);
     }
+
+    #[tokio::test]
+    async fn latest_context_usage_follows_conversation_chronology() {
+        let directory = tempfile::tempdir().unwrap();
+        let store = Store::connect(&format!(
+            "sqlite://{}",
+            directory.path().join("test.db").display()
+        ))
+        .await
+        .unwrap();
+
+        for (call_id, model_id, context_input_tokens, message_count) in [
+            ("call-a-1", "model-a", 100_u64, 3_usize),
+            ("call-b", "model-b", 200_u64, 5_usize),
+            ("call-a-2", "model-a", 300_u64, 7_usize),
+        ] {
+            store
+                .start_llm_call(&NewLlmCall {
+                    call_id: call_id.into(),
+                    run_id: format!("run-{call_id}"),
+                    conversation_id: "conversation".into(),
+                    provider_call_index: 0,
+                    model_hash: model_id.into(),
+                    provider_type: ProviderType::Plugin,
+                    provider_url: "plugin://test".into(),
+                    request_type: ProviderType::Plugin,
+                    request_url: "plugin://test".into(),
+                    model_id: model_id.into(),
+                    display_name: model_id.into(),
+                    reasoning_effort: None,
+                    fast: false,
+                    message_count,
+                    tool_count: 0,
+                    detailed: false,
+                })
+                .await
+                .unwrap();
+            store
+                .record_llm_usage(
+                    call_id,
+                    Usage {
+                        input_tokens: Some(context_input_tokens),
+                        context_input_tokens: Some(context_input_tokens),
+                        output_tokens: Some(10),
+                        total_tokens: Some(context_input_tokens + 10),
+                        ..Default::default()
+                    },
+                )
+                .await
+                .unwrap();
+        }
+
+        assert_eq!(
+            store.latest_context_usage("conversation").await.unwrap(),
+            Some(ContextUsageAnchor {
+                context_input_tokens: 300,
+                message_count: 7,
+            })
+        );
+        assert_eq!(store.latest_context_usage("other").await.unwrap(), None);
+    }
 }

server/src/store/migrations.rs

@@ -445,9 +445,20 @@ mod tests {
             .await
             .unwrap();
 
+            let argument_error_column_exists: i64 = sqlx::query_scalar(
+                "SELECT EXISTS(
+                    SELECT 1 FROM pragma_table_info('tool_round_calls')
+                    WHERE name = 'argument_error'
+                 )",
+            )
+            .fetch_one(&pool)
+            .await
+            .unwrap();
+
             assert_eq!(checksum_after, checksum_before);
-            assert_eq!(versions, vec![1, 2, 3, 4, 5, 6, 7, 8]);
+            assert_eq!(versions, vec![1, 2, 3, 4, 5, 6, 7, 8, 9]);
             assert_eq!(checkpoint_table_exists, 1);
+            assert_eq!(argument_error_column_exists, 1);
         }
     }