v0.8.6: general mode, clipboard text, PDF text extraction

Working mode: --mode coding|general (CST_MODE, settings.json, /mode, WebUI run settings). general swaps the default framing for a general assistant one -- workflow, safety notes, tool wording (run_shell, the subagents) -- and /status shows it; minimal (the paper profile) stays byte-frozen. The prompt state carries the mode so a switch leaves a system note in history. Clipboard: clipboard_read / clipboard_write tools (16 MiB cap read, default 64 KiB window; writing is a WRITE permission with a preview) on pbcopy / wl-copy / xclip / pbpaste / wl-paste / xclip and the Win32 clipboard. In the TUI, /attach paste falls back to clipboard text when there is no image on it. PDF: a dependency-free text extractor (src/pdf_text.c): inflate, Flate/ASCIIHex/ASCII85, PNG/TIFF predictors, layout-independent object scan, object streams, page tree walk with inherited resources, content lexer (Tj/TJ/'/" and the Td/TD/T*/Tm spacing heuristics), per-font ToUnicode CMaps and inline images skipped. Encrypted files and pages with no text layer are refused with the reason. Attachments: /attach (a path or files dropped on the terminal) now take images, PDFs and text. A PDF is stored as its text layer with the mime still saying application/pdf and its page count; anything else has to be UTF-8 text. Documents travel inline as text parts, so CST_NO_IMAGES and text-only models no longer drop them, while images keep their old path. Tests: test_pdf_text (plain, bare and zlib-framed Flate, inline image, CID + ToUnicode, object stream, error paths), documents in test_attach, inlined documents in test_llm_body, mode in test_config.
build: finish cross-platform release tooling
2026-10-11 11:09:39 +08:00 · 2026-10-10 16:21:00 +08:00 · 2026-10-10 14:54:36 +08:00 · 2026-10-10 13:57:17 +08:00 · 2026-10-09 15:54:53 +08:00 · 2026-10-09 13:51:34 +08:00
76 changed files with 11868 additions and 1030 deletions
@@ -11,13 +11,11 @@ concurrency:
  cancel-in-progress: true

 jobs:
-  linux:
+  # Verification: TinyCC builds and tests everything in seconds. tcc warns on
+  # `#pragma GCC`, so no -Werror here; the gcc release builds below carry it.
+  tcc-check:
    runs-on: ubuntu-latest
    timeout-minutes: 30
-    strategy:
-      fail-fast: false
-      matrix:
-        cc: [gcc, clang]
    steps:
      - name: Checkout
        env:
@@ -34,21 +32,14 @@ jobs:
          $SUDO bash scripts/ci_pkg_mirror.sh
          $SUDO apt-get update
          $SUDO apt-get install -y --no-install-recommends \
-            build-essential clang cmake pkg-config python3 \
+            build-essential pkg-config python3 curl ca-certificates \
            libcurl4-openssl-dev libncurses-dev
-      - name: Build (warnings are errors)
-        run: make CC=${{ matrix.cc }} CFLAGS="-std=c11 -Wall -Wextra -O2 -g -Werror" -j"$(nproc)"
+      - name: Build TinyCC
+        run: bash scripts/ci_tcc.sh
+      - name: Build
+        run: make CC=tcc AR=ar CFLAGS="-std=c11 -Wall -O2 -g" -j"$(nproc)"
      - name: Unit tests
-        run: |
-          # Same -Werror as the build step so undeclared libc helpers cannot
-          # silently return int and truncate pointers (see test_subagent).
-          make CC=${{ matrix.cc }} \
-            CFLAGS="-std=c11 -Wall -Wextra -O2 -g -Werror" check
-      - name: CMake build + ctest
-        run: |
-          cmake -B build -DCMAKE_C_COMPILER=${{ matrix.cc }} -DCST_BUILD_TUI=ON
-          cmake --build build -j"$(nproc)"
-          ctest --test-dir build --output-on-failure
+        run: make CC=tcc AR=ar CFLAGS="-std=c11 -Wall -O2 -g" check
      - name: Smoke test
        run: |
          ./cst-agent --help
@@ -57,39 +48,22 @@ jobs:
      - name: WebUI end-to-end (recorded stream, no API key)
        run: sh scripts/e2e_webui.sh

-  musl:
+  # Release: static musl binaries, gcc with -Werror.
+  # dist/cst-agent-VERSION-linux-ARCH.tar.gz
+  linux:
+    needs: tcc-check
    runs-on: ubuntu-latest
-    timeout-minutes: 30
-    container: alpine:3.20
-    steps:
-      - name: Install tools + mirror
-        run: |
-          sed -i 's#https\?://dl-cdn.alpinelinux.org/alpine#https://mirrors.tuna.tsinghua.edu.cn/alpine#g' \
-            /etc/apk/repositories
-          apk add --no-cache curl git bash
-      - name: Checkout
-        env:
-          GITHUB_SERVER_URL: ${{ github.server_url }}
-          GITHUB_REPOSITORY: ${{ github.repository }}
-          GITHUB_SHA: ${{ github.sha }}
-        run: |
-          curl -fsSL "$GITHUB_SERVER_URL/$GITHUB_REPOSITORY/raw/commit/$GITHUB_SHA/scripts/ci_checkout.sh" \
-            -o /tmp/ci_checkout.sh
-          bash /tmp/ci_checkout.sh
-      - name: Install build dependencies
-        run: |
-          # linux-headers provides linux/audit.h for the seccomp probe path
-          apk add --no-cache build-base pkgconf curl-dev ncurses-dev \
-            linux-headers git nodejs bash python3
-      - name: Build + unit tests
-        run: |
-          make CFLAGS="-std=c11 -Wall -Wextra -O2 -Werror" -j"$(nproc)"
-          make check
-
-  # MinGW-w64 UCRT cross build → dist/cst-agent-*-windows-x86_64.zip
-  windows-cross:
-    runs-on: ubuntu-latest
-    timeout-minutes: 120
+    timeout-minutes: 60
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - arch: x86_64
+            check: 1
+          - arch: x86
+            check: 1
+          - arch: aarch64
+            check: 0
    steps:
      - name: Checkout
        env:
@@ -105,20 +79,29 @@ jobs:
          SUDO=$(command -v sudo || true)
          $SUDO bash scripts/ci_pkg_mirror.sh
          $SUDO apt-get update
+          # ncurses-bin/-term: tic, infocmp and the fallback terminal entries
          $SUDO apt-get install -y --no-install-recommends \
-            ca-certificates curl git cmake ninja-build make pkg-config \
-            python3 zip unzip perl xz-utils
-      - name: Cross-build Windows zip
-        run: bash scripts/ci_windows_cross.sh
-      - name: Verify zip
+            build-essential cmake ninja-build pkg-config python3 perl \
+            curl ca-certificates xz-utils bzip2 file ncurses-bin ncurses-term
+          if [ "${{ matrix.arch }}" = aarch64 ]; then
+            $SUDO apt-get install -y --no-install-recommends qemu-user-static
+          fi
+      - name: Build (static, -Werror)
+        env:
+          CST_LINUX_CHECK: ${{ matrix.check }}
+        run: bash scripts/package_linux.sh ${{ matrix.arch }}
+      - name: Smoke test
        run: |
-          ls -lah dist/windows-x86_64/
-          ls -lah dist/cst-agent-*-windows-x86_64.zip
-          cat dist/cst-agent-*-windows-x86_64.zip.sha256
-          unzip -l dist/cst-agent-*-windows-x86_64.zip | head -20
-      # Auto-publish to Gitea Releases (v$CST_VERSION from include/cst_version.h).
-      # Prefer repo secret RELEASE_TOKEN (write:repository); fall back to the
-      # job token. Skipped on pull_request and when no token is available.
+          d=$(ls -d dist/cst-agent-*-linux-${{ matrix.arch }})
+          file "$d/cst-agent"
+          run=""
+          [ "${{ matrix.arch }}" = aarch64 ] && run=qemu-aarch64-static
+          $run "$d/cst-agent" --version
+          $run "$d/cst-agent" --help >/dev/null
+          $run "$d/cst-agent" --workspace "$PWD" --list-sessions
+          cat dist/cst-agent-*-linux-${{ matrix.arch }}.tar.gz.sha256
+      # Release vVERSION (from include/cst_version.h) on Gitea. Prefer repo
+      # secret RELEASE_TOKEN (write:repository); fall back to the job token.
      - name: Publish Gitea Release
        if: github.event_name == 'push'
        env:
@@ -134,4 +117,59 @@ jobs:
            echo "::warning::no RELEASE_TOKEN / GITHUB_TOKEN; skip auto-release"
            exit 0
          fi
-          bash scripts/release_windows.sh
+          bash scripts/release.sh dist/cst-agent-*-linux-${{ matrix.arch }}.tar.gz
+
+  # Release: mingw-w64 UCRT gcc (Debian trixie) with -Werror.
+  # dist/cst-agent-VERSION-windows-x86_64.zip and -setup.exe (NSIS)
+  windows:
+    needs: tcc-check
+    runs-on: ubuntu-latest
+    timeout-minutes: 120
+    container: debian:trixie
+    steps:
+      - name: Install tools
+        run: |
+          # No CA certificates in the image yet, so the mirror is plain http.
+          sed -i 's#https\?://deb.debian.org#http://mirrors.tuna.tsinghua.edu.cn#g' \
+            /etc/apt/sources.list.d/debian.sources
+          apt-get update
+          apt-get install -y --no-install-recommends \
+            ca-certificates curl git build-essential cmake ninja-build make \
+            pkg-config python3 perl zip unzip xz-utils patch file \
+            gcc-mingw-w64-ucrt64 binutils-mingw-w64-ucrt64 nsis
+      - name: Checkout
+        env:
+          GITHUB_SERVER_URL: ${{ github.server_url }}
+          GITHUB_REPOSITORY: ${{ github.repository }}
+          GITHUB_SHA: ${{ github.sha }}
+        run: |
+          curl -fsSL "$GITHUB_SERVER_URL/$GITHUB_REPOSITORY/raw/commit/$GITHUB_SHA/scripts/ci_checkout.sh" \
+            -o /tmp/ci_checkout.sh
+          bash /tmp/ci_checkout.sh
+      - name: Cross-build zip and installer
+        env:
+          CST_WIN_CFLAGS: -Wall -Wextra -Werror
+        run: bash scripts/ci_windows_cross.sh
+      - name: Verify
+        run: |
+          ls -lah dist/windows-x86_64/
+          file dist/windows-x86_64/cst-agent.exe dist/cst-agent-*-windows-x86_64-setup.exe
+          cat dist/cst-agent-*-windows-x86_64.zip.sha256 dist/cst-agent-*-windows-x86_64-setup.exe.sha256
+          unzip -l dist/cst-agent-*-windows-x86_64.zip
+      - name: Publish Gitea Release
+        if: github.event_name == 'push'
+        env:
+          RELEASE_TOKEN: ${{ secrets.RELEASE_TOKEN }}
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          CST_GITEA: ${{ github.server_url }}
+          CST_REPO: ${{ github.repository }}
+          GITHUB_REF: ${{ github.ref }}
+          CI: "true"
+        run: |
+          export GITEA_TOKEN="${RELEASE_TOKEN:-${GITHUB_TOKEN:-}}"
+          if [[ -z "${GITEA_TOKEN}" ]]; then
+            echo "::warning::no RELEASE_TOKEN / GITHUB_TOKEN; skip auto-release"
+            exit 0
+          fi
+          bash scripts/release.sh dist/cst-agent-*-windows-x86_64.zip \
+            dist/cst-agent-*-windows-x86_64-setup.exe
@@ -26,6 +26,9 @@ third_party/ncurses-*.tar.gz
 third_party/busybox-w32/
 third_party/busybox-src/
 third_party/prefix-mingw/
+third_party/prefix-linux-*/
+third_party/build-linux-*/
+third_party/src/
 vendor/sashidown/*.o
 vendor/sashidown/src/*.o
 vendor/sashidown/libsashidown.a
@@ -95,6 +95,7 @@ set(CST_CORE_SOURCES
  src/subagent.c
  src/attach.c
  src/clipboard.c
+  src/pdf_text.c
  src/session.c
  src/tools/tools.c
  src/tools/patch.c
@@ -111,6 +112,7 @@ set(CST_CORE_SOURCES
  src/agent_store.c
  src/json_baseline.c
  src/creds.c
+  src/provider_switch.c
  src/mm_tool.c
  src/models.c
  src/model_registry.c
@@ -171,9 +173,12 @@ endif()
 if(WIN32)
  # MinGW has no <regex.h>: musl's TRE engine, UTF-8 aware (vendor/tre-regex)
  include_directories(vendor/tre-regex)
-  list(APPEND CST_CORE_SOURCES
+  set(_CST_TRE_SOURCES
    vendor/tre-regex/regcomp.c vendor/tre-regex/regexec.c
    vendor/tre-regex/regerror.c vendor/tre-regex/tre-mem.c)
+  list(APPEND CST_CORE_SOURCES ${_CST_TRE_SOURCES})
+  # Upstream code, kept as is: CST_WIN_CFLAGS=-Werror must not apply to it.
+  set_source_files_properties(${_CST_TRE_SOURCES} PROPERTIES COMPILE_OPTIONS "-w")
 endif()

 set(CST_SOURCES ${CST_CORE_SOURCES} src/main.c)
@@ -275,7 +280,7 @@ endif()
 enable_testing()
 set(_CST_SMALL_TESTS test_ctl_frontier test_ctl_heredoc test_json_baseline
    test_replay_starvation test_replay_real test_mm_tool test_textsel
-    test_markdown)
+    test_pdf_text test_markdown)
 foreach(t ${_CST_SMALL_TESTS})
  if(t STREQUAL test_textsel)
    add_executable(${t} tests/${t}.c src/textsel.c src/os_compat.c src/os_win.c
@@ -283,6 +288,12 @@ foreach(t ${_CST_SMALL_TESTS})
    if(WIN32)
      target_link_libraries(${t} PRIVATE advapi32 user32)
    endif()
+  elseif(t STREQUAL test_pdf_text)
+    add_executable(${t} tests/${t}.c src/pdf_text.c src/os_compat.c src/os_win.c
+      src/util.c)
+    if(WIN32)
+      target_link_libraries(${t} PRIVATE advapi32 user32)
+    endif()
  elseif(t STREQUAL test_markdown)
    add_executable(${t} tests/${t}.c)
    target_include_directories(${t} PRIVATE "${SD_DIR}/include")
@@ -306,7 +317,7 @@ foreach(t ${_CST_SMALL_TESTS})
  target_include_directories(${t} PRIVATE include "${CJSON_DIR}")
  target_compile_definitions(${t} PRIVATE _GNU_SOURCE)
  if(NOT WIN32)
-    target_link_libraries(${t} PRIVATE m)
+    target_link_libraries(${t} PRIVATE m Threads::Threads)
  endif()
  cst_win_exe(${t})
  add_test(NAME ${t} COMMAND ${t} WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}")
@@ -7,6 +7,7 @@ PREFIX ?= /usr/local
 DESTDIR ?=
 UNAME_S := $(shell uname -s 2>/dev/null)
 UNAME_O := $(shell uname -o 2>/dev/null)
+comma := ,

 # Android (Termux): pthreads live in libc.
 PTHREAD_LIBS := -lpthread
@@ -26,6 +27,15 @@ CPPFLAGS += -D_DARWIN_C_SOURCE
 endif
 PKGC = $(PKG_ENV) $(PKG_CONFIG)

+# STATIC=1: fully static link (the musl release builds).
+STATIC ?= 0
+LINK_STATIC :=
+ifeq ($(STATIC),1)
+PKGC += --static
+LINK_STATIC := -static
+endif
+LDFLAGS += $(LINK_STATIC)
+
 # Wide-character curses: ncursesw, else ncurses (macOS/BSD build it wide).
 CURSES_PC := $(shell for p in ncursesw ncurses; do $(PKGC) --exists $$p 2>/dev/null && { echo $$p; exit 0; }; done)
 ifneq ($(CURSES_PC),)
@@ -50,9 +60,19 @@ CST_WEBUI ?= 1
 PKG_CFLAGS := $(CURL_CFLAGS) $(CURSES_CFLAGS)
 PKG_LIBS := $(CURL_LIBS) $(CURSES_LIBS)

-CPPFLAGS += -Iinclude -I.gen -I$(SD_DIR)/include -I$(CJSON_DIR) -D_XOPEN_SOURCE_EXTENDED $(PKG_CFLAGS)
+# tcc writes no rpath of its own, so shared libs found via pkg-config -L
+# (Nix store, /opt prefixes) would not load at run time.
+RPATH_LIBS :=
+ifneq ($(findstring tcc,$(notdir $(CC))),)
+ifneq ($(STATIC),1)
+RPATH_LIBS := $(patsubst -L%,-Wl$(comma)-rpath$(comma)%,$(filter -L%,$(PKG_LIBS)))
+endif
+endif
+
+# Keep POSIX declarations visible for shared test helpers (notably symlink()).
+CPPFLAGS += -Iinclude -I.gen -I$(SD_DIR)/include -I$(CJSON_DIR) -D_POSIX_C_SOURCE=200809L -D_XOPEN_SOURCE_EXTENDED $(PKG_CFLAGS)
 LDFLAGS += -L$(SD_DIR)
-LDLIBS += -lsashidown -lm $(PKG_LIBS) $(PTHREAD_LIBS)
+LDLIBS += -lsashidown -lm $(PKG_LIBS) $(PTHREAD_LIBS) $(RPATH_LIBS)

 WEB_DIR := web
 WEB_ASSETS := $(wildcard $(WEB_DIR)/*.html) $(wildcard $(WEB_DIR)/*.js) \
@@ -76,6 +96,7 @@ CORE_SRC = \
 	src/keymap.c \
 	src/skills.c \
 	src/extensions.c src/plugins.c src/mcp.c src/subagent.c src/attach.c src/clipboard.c \
+	src/pdf_text.c \
 	src/session.c \
 	src/tools/tools.c \
 	src/tools/patch.c \
@@ -91,7 +112,7 @@ CORE_SRC = \
 	src/agent_loop.c \
 	src/agent_store.c \
 	src/json_baseline.c \
-	src/creds.c \
+	src/creds.c src/provider_switch.c \
 	src/mm_tool.c \
 	src/models.c src/model_registry.c \
 	src/metrics.c \
@@ -115,6 +136,7 @@ TESTS = \
 	tests/test_apply_patch \
 	tests/test_session_roundtrip \
 	tests/test_attach \
+	tests/test_pdf_text \
 	tests/test_shell \
 	tests/test_diff \
 	tests/test_tools_fs \
@@ -141,7 +163,11 @@ ifneq ($(CST_WEBUI),0)
 TESTS += tests/test_webui
 endif

-TEST_LIBS = -L$(SD_DIR) -lsashidown -lm $(CURL_LIBS) $(PTHREAD_LIBS)
+TEST_SYS_LIBS = -lm $(PTHREAD_LIBS) $(LINK_STATIC)
+TEST_LIBS = -L$(SD_DIR) -lsashidown $(CURL_LIBS) $(RPATH_LIBS) $(TEST_SYS_LIBS)
+# cst_test_util.h has static os_* helpers; gcc drops them when unused, tcc
+# keeps them, so the small tests link the OS layer too.
+TEST_OS_SRC = src/os_compat.c src/os_win.c src/util.c

 .PHONY: all clean check demo sashidown-lib FORCE cmake-gcc cmake-clang asan install uninstall

@@ -184,26 +210,29 @@ cst-agent: $(OBJ) $(SD_DIR)/libsashidown.a
 src/tui.o: src/tui.c $(HDRS) $(SD_DIR)/include/sashidown.h
 	$(CC) $(CFLAGS) $(CPPFLAGS) -c -o $@ $<

-tests/test_ctl_frontier: tests/test_ctl_frontier.c src/ctl_parse.c include/ctl.h
-	$(CC) $(CFLAGS) $(CPPFLAGS) -o $@ tests/test_ctl_frontier.c src/ctl_parse.c
+tests/test_ctl_frontier: tests/test_ctl_frontier.c src/ctl_parse.c include/ctl.h $(TEST_OS_SRC)
+	$(CC) $(CFLAGS) $(CPPFLAGS) -o $@ tests/test_ctl_frontier.c src/ctl_parse.c $(TEST_OS_SRC) $(TEST_SYS_LIBS)

-tests/test_ctl_heredoc: tests/test_ctl_heredoc.c src/ctl_parse.c include/ctl.h
-	$(CC) $(CFLAGS) $(CPPFLAGS) -o $@ tests/test_ctl_heredoc.c src/ctl_parse.c
+tests/test_ctl_heredoc: tests/test_ctl_heredoc.c src/ctl_parse.c include/ctl.h $(TEST_OS_SRC)
+	$(CC) $(CFLAGS) $(CPPFLAGS) -o $@ tests/test_ctl_heredoc.c src/ctl_parse.c $(TEST_OS_SRC) $(TEST_SYS_LIBS)

-tests/test_json_baseline: tests/test_json_baseline.c src/json_baseline.c $(CJSON_DIR)/cJSON.c $(HDRS)
-	$(CC) $(CFLAGS) $(CPPFLAGS) -o $@ tests/test_json_baseline.c src/json_baseline.c $(CJSON_DIR)/cJSON.c -lm
+tests/test_json_baseline: tests/test_json_baseline.c src/json_baseline.c $(CJSON_DIR)/cJSON.c $(HDRS) $(TEST_OS_SRC)
+	$(CC) $(CFLAGS) $(CPPFLAGS) -o $@ tests/test_json_baseline.c src/json_baseline.c $(CJSON_DIR)/cJSON.c $(TEST_OS_SRC) $(TEST_SYS_LIBS)

-tests/test_replay_starvation: tests/test_replay_starvation.c src/ctl_parse.c src/json_baseline.c $(CJSON_DIR)/cJSON.c
-	$(CC) $(CFLAGS) $(CPPFLAGS) -o $@ tests/test_replay_starvation.c src/ctl_parse.c src/json_baseline.c $(CJSON_DIR)/cJSON.c -lm
+tests/test_replay_starvation: tests/test_replay_starvation.c src/ctl_parse.c src/json_baseline.c $(CJSON_DIR)/cJSON.c $(TEST_OS_SRC)
+	$(CC) $(CFLAGS) $(CPPFLAGS) -o $@ tests/test_replay_starvation.c src/ctl_parse.c src/json_baseline.c $(CJSON_DIR)/cJSON.c $(TEST_OS_SRC) $(TEST_SYS_LIBS)

-tests/test_replay_real: tests/test_replay_real.c src/ctl_parse.c src/json_baseline.c $(CJSON_DIR)/cJSON.c
-	$(CC) $(CFLAGS) $(CPPFLAGS) -o $@ tests/test_replay_real.c src/ctl_parse.c src/json_baseline.c $(CJSON_DIR)/cJSON.c -lm
+tests/test_replay_real: tests/test_replay_real.c src/ctl_parse.c src/json_baseline.c $(CJSON_DIR)/cJSON.c $(TEST_OS_SRC)
+	$(CC) $(CFLAGS) $(CPPFLAGS) -o $@ tests/test_replay_real.c src/ctl_parse.c src/json_baseline.c $(CJSON_DIR)/cJSON.c $(TEST_OS_SRC) $(TEST_SYS_LIBS)

-tests/test_mm_tool: tests/test_mm_tool.c src/mm_tool.c include/mm_tool.h
-	$(CC) $(CFLAGS) $(CPPFLAGS) -o $@ tests/test_mm_tool.c src/mm_tool.c
+tests/test_mm_tool: tests/test_mm_tool.c src/mm_tool.c include/mm_tool.h $(TEST_OS_SRC)
+	$(CC) $(CFLAGS) $(CPPFLAGS) -o $@ tests/test_mm_tool.c src/mm_tool.c $(TEST_OS_SRC) $(TEST_SYS_LIBS)

-tests/test_textsel: tests/test_textsel.c src/textsel.c src/os_compat.c src/os_win.c src/util.c include/cst_textsel.h
-	$(CC) $(CFLAGS) $(CPPFLAGS) -o $@ tests/test_textsel.c src/textsel.c src/os_compat.c src/os_win.c src/util.c
+tests/test_textsel: tests/test_textsel.c src/textsel.c $(TEST_OS_SRC) include/cst_textsel.h
+	$(CC) $(CFLAGS) $(CPPFLAGS) -o $@ tests/test_textsel.c src/textsel.c $(TEST_OS_SRC) $(TEST_SYS_LIBS)
+
+tests/test_pdf_text: tests/test_pdf_text.c src/pdf_text.c include/cst_pdf.h $(TEST_OS_SRC)
+	$(CC) $(CFLAGS) $(CPPFLAGS) -o $@ tests/test_pdf_text.c src/pdf_text.c $(TEST_OS_SRC) $(TEST_SYS_LIBS)

 tests/test_markdown: tests/test_markdown.c $(SD_DIR)/libsashidown.a
 	$(CC) $(CFLAGS) $(CPPFLAGS) -o $@ tests/test_markdown.c $(LDFLAGS) -lsashidown -lm
@@ -27,6 +27,10 @@ cst-agent --version
 cd ~/my-project && cst-agent
 ```

+不想编译：[Release](https://git.whcself.xyz/WHC/CST-agent/releases) 里有静态链接的 Linux 单文件
+（`cst-agent-VERSION-linux-{x86_64,x86,aarch64}.tar.gz`，musl，任何发行版都能直接跑），解压后把 `cst-agent`
+放进 `PATH` 即可；每个文件旁边有 `.sha256`。
+
 Windows 10 1903+（PowerShell，不需要管理员）：

 ```powershell
@@ -34,7 +38,8 @@ irm https://git.whcself.xyz/WHC/CST-agent/raw/branch/main/scripts/install_window
 # 新开终端后：cd 到项目目录运行 cst-agent；升级用 cst-agent-update
 ```

-详见 [docs/README-windows.md](docs/README-windows.md)。
+也可以下载 Release 里的安装包 `cst-agent-VERSION-windows-x86_64-setup.exe`（按用户安装、自动加 PATH、
+开始菜单有 WebUI 入口），或便携版 zip。详见 [docs/README-windows.md](docs/README-windows.md)。

 一键脚本可以重复运行：已有仓库会 `git pull --ff-only`，代码有更新时干净重编，然后重新安装；
 同时装一个 `cst-agent-update` 命令（等同重跑脚本）。NixOS 上会自动进 `nix-shell` 编译。
@@ -49,7 +54,8 @@ irm https://git.whcself.xyz/WHC/CST-agent/raw/branch/main/scripts/install_window
 | Linux | [`scripts/bootstrap_linux.sh`](scripts/bootstrap_linux.sh) |
 | Termux | [`scripts/bootstrap_termux.sh`](scripts/bootstrap_termux.sh) |
 | Windows 安装 / 升级 | [`scripts/install_windows.ps1`](scripts/install_windows.ps1)（从 Release 安装），之后 `cst-agent-update` |
-| Windows 交叉构建（在 Linux 上） | [`scripts/bootstrap_windows.sh`](scripts/bootstrap_windows.sh) → `dist/windows-x86_64/` |
+| Linux 静态发行包（musl 交叉） | [`scripts/package_linux.sh`](scripts/package_linux.sh) `x86_64\|x86\|aarch64` → `dist/*.tar.gz` |
+| Windows 交叉构建（在 Linux 上） | [`scripts/bootstrap_windows.sh`](scripts/bootstrap_windows.sh) → `dist/windows-x86_64/`、zip 和安装包 |
 | Windows 原生构建（MSYS2 UCRT64） | [`scripts/bootstrap_msys2.sh`](scripts/bootstrap_msys2.sh) → `dist/windows-x86_64/` |

 ### 迷你验收工程（论文案例）
@@ -86,6 +92,7 @@ cd CST-agent
 make            # 依赖 pkg-config 找 libcurl / ncursesw（找不到时退回 ncurses）
 make check      # 单元测试 + 截断实验（含真实录制流）
 make asan       # AddressSanitizer + UBSan 构建并跑全部测试
+make CC=tcc AR=ar check   # TinyCC：整套构建 + 测试十几秒（CI 的校验就是这条）
 make install PREFIX=$HOME/.local   # 或 sudo make install（默认 /usr/local）
 ```

@@ -111,7 +118,7 @@ cmake -B build-hd -DCST_BUILD_TUI=OFF && cmake --build build-hd
 ./scripts/bootstrap_windows.sh
 ```

-依赖：`gcc` 或 `clang`、`libcurl`、TUI 需 `ncursesw`（Windows 版用打过补丁的 ncurses 6.6，脚本自动编译）。
+依赖：`gcc`、`clang` 或 `tcc`（0.9.27 mob 以上）、`libcurl`、TUI 需 `ncursesw`（Windows 版用打过补丁的 ncurses 6.6，脚本自动编译）。
 Sashidown（`vendor/sashidown`）与 cJSON（`vendor/cjson`，MIT）已随仓库提供。  
 API Key：TUI 里输入并按 provider 保存到配置目录的 `credentials`（权限 0600）；`/key` 更换，`--ask-key` 在命令行输入。
 `--api-key` 只用于本次进程、不落盘；`DEEPSEEK_API_KEY` 等环境变量和 `CST_API_KEY` 仅作为没有保存 key 时的兜底（方便 CI）。
@@ -120,7 +127,7 @@ API Key：TUI 里输入并按 provider 保存到配置目录的 `credentials`（

 | OS | Arch | 编译器 | Headless | TUI | Shell |
 |---|---|---|---|---|---|
-| Linux（glibc / musl） | x86_64 / aarch64 | GCC, Clang | yes | ncursesw | `/bin/sh` 或旁路 `busybox` |
+| Linux（glibc / musl） | x86_64 / x86 / aarch64 | GCC, Clang, TinyCC | yes | ncursesw | `/bin/sh` 或旁路 `busybox` |
 | WSL2 | x86_64 | 同 Linux | yes | ncursesw | `/bin/sh`（工作区建议放在 Linux 文件系统内） |
 | macOS | arm64 / x86_64 | Apple Clang | yes | ncurses（Homebrew） | `/bin/sh` |
 | Android Termux | aarch64 / … | Clang (Termux) | yes | ncurses（Termux 包） | `$PREFIX/bin/sh`；可旁路 busybox |
@@ -164,6 +171,18 @@ make && make install      # 装到 Termux 的 $PREFIX/bin，之后任意目录
 临时目录：依次试 `$TMPDIR`、`$PREFIX/tmp`、`/tmp`、`/var/tmp`、`~/.cache/cst-agent/tmp`，用第一个真能建文件的，
 并写回 `TMPDIR` 给 shell 子进程和沙箱用（Termux 的 `/tmp` 常常不存在或被 SELinux 拒绝）。

+### Linux 静态发行包
+
+```bash
+scripts/package_linux.sh x86_64     # 或 x86 / aarch64；CST_LINUX_CHECK=1 顺便跑 make check
+```
+
+用 Bootlin 的 musl 交叉 gcc（固定版本、校验 sha256）静态编译 mbedTLS、curl 8.22（只开 HTTP/HTTPS）和宽字符 ncurses，
+再以 `STATIC=1` 和 `-Werror` 编出 `cst-agent`，产物是 `dist/cst-agent-VERSION-linux-ARCH.tar.gz`（+ `.sha256`）。
+宿主需要 `cmake ninja make perl` 以及 `tic`/`infocmp` 和终端条目（Debian：`ncurses-bin ncurses-term`）。
+ncurses 会去常见目录找 terminfo，找不到时用编进去的 `xterm-256color`、`xterm`、`screen-256color`、`tmux-256color`、`linux`、`vt100`。
+CA 证书：`CST_CA_BUNDLE` → `SSL_CERT_FILE` → 各发行版的常见位置（Debian/Ubuntu、Fedora/RHEL、openSUSE、Alpine）。
+
 ### Windows 发行包

 见 [docs/README-windows.md](docs/README-windows.md)（安装、编码、沙箱、构建、已知限制）。
@@ -172,10 +191,17 @@ make && make install      # 装到 Termux 的 $PREFIX/bin，之后任意目录
 ./scripts/bootstrap_windows.sh   # Linux 交叉编译；Windows 上在 MSYS2 UCRT64 里用 scripts/bootstrap_msys2.sh
 ```

-产物：`dist/windows-x86_64/`（静态链接的 `cst-agent.exe` + `busybox.exe`，无其他 DLL）和
-`dist/cst-agent-VERSION-windows-x86_64.zip`。CI 在 `main` / `v*` tag 推送且交叉构建成功后会自动上传到
-Gitea Release `vVERSION`（同名资源会被替换）；仓库可设 `RELEASE_TOKEN`（`write:repository`），否则用任务自带 token。
-本地：`GITEA_TOKEN=... scripts/release_windows.sh`。
+产物：`dist/windows-x86_64/`（静态链接的 `cst-agent.exe` + `busybox.exe`，无其他 DLL）、
+`dist/cst-agent-VERSION-windows-x86_64.zip`，以及装了 `makensis` 时的安装包 `dist/cst-agent-VERSION-windows-x86_64-setup.exe`。
+
+### CI 与 Release
+
+CI 用 TinyCC 校验（构建、`make check`、WebUI 端到端），通过后用 gcc（`-Werror`）出发行包：
+Linux x86_64 / x86 / aarch64 静态包（x86 两种跑 `make check`，aarch64 用 qemu 冒烟），
+Windows x86_64 的 zip 和安装包（Debian 的 mingw-w64 UCRT gcc + NSIS）。
+推送到 `main` 或 `v*` tag 时每个任务把自己的产物上传到 Gitea Release `vVERSION`（同名资源会被替换）；
+仓库可设 `RELEASE_TOKEN`（`write:repository`），否则用任务自带 token。
+本地：`GITEA_TOKEN=... scripts/release.sh dist/cst-agent-*.tar.gz dist/cst-agent-*.zip dist/cst-agent-*-setup.exe`。

 环境变量：`CST_SHELL`、`CST_EXE_DIR`、`CST_AUTO_APPROVE` 等。

@@ -207,18 +233,20 @@ Gitea Release `vVERSION`（同名资源会被替换）；仓库可设 `RELEASE_T
 |---|---|
 | `CST_PROTOCOL` / `--protocol` | 工具调用协议（`native-eos`） |
 | `CST_TOOLS` / `--tools` | 工具集：`default`（全部）/ `minimal`（论文 4 件套） |
+| `CST_MODE` / `--mode` | 工作模式：`coding`（默认，代码口径）/ `general`（通用助手口径） |
 | `CST_EFFORT` / `--effort` | 推理强度 `low` / `medium` / `high` / `max` |
 | `CST_CONFIRM` / `--confirm` | 哪些类别要确认：`shell,write,net`（默认）或 `none` |
 | `CST_MAX_STEPS` / `--max-steps` | 每个用户回合最多调用模型几次（30） |
 | `CST_TOOL_OUT_CAP` | 单个工具输出发给模型的字节上限（16384，保留头 60% + 尾 40%） |
 | `CST_CTX_WINDOW` | 发给模型的最近非 system 消息条数（80） |
-| `CST_CTX_TOKENS` / `CST_AUTO_COMPACT=0` | 覆盖模型上下文窗口大小 / 关闭自动压缩（用到 80% 时自动 `/compact`） |
+| `CST_CTX_TOKENS` / `CST_AUTO_COMPACT=0` | 覆盖模型上下文窗口大小 / 关闭自动压缩（用到 75% 时自动 `/compact`） |
 | `CST_NO_PROJECT_NOTES=1` | 不把工作区的 `AGENTS.md` 注入 system prompt |
 | `NO_COLOR` / `CST_ASCII=1` | 关闭颜色 / 强制 ASCII 图形（非 UTF-8 终端自动退化） |
-| `CST_SHELL_TIMEOUT` | `run_shell` 超时秒数（120），超时整组进程被杀，exit=124 |
+| `CST_SHELL_TIMEOUT` | `run_shell` 超时秒数（默认不限），设了才在超时后整组杀进程，exit=124 |
 | `CST_METRICS=0` / `CST_METRICS_PATH` | 关闭 / 指定计时埋点 JSONL（默认 `.agent/metrics/<session>.jsonl`） |
 | `CST_RECORD_SSE=DIR` / `CST_REPLAY_SSE=DIR` | 录制 / 回放原始 SSE 流（`NNNN.sse` + `NNNN.req.json`） |
 | `CST_APPROVAL=ask\|auto\|yolo` / `CST_SANDBOX=off\|landlock\|seccomp` | 审批模式 / 关闭沙箱（回落到只读白名单规则）或只用指定沙箱 |
+| `CST_CA_BUNDLE` | HTTPS 用的 CA 证书文件（默认：`SSL_CERT_FILE`、libcurl 自带路径、常见发行版位置；Windows 用系统证书库） |
 | `CST_MODELS`、`CST_AUTO_APPROVE=1`、`CST_SHELL`、`CST_EXE_DIR` | 模型列表、自动批准（= yolo）、shell、旁路 busybox |

 ## Run
@@ -249,6 +277,7 @@ cst-agent --headless --prompt 'List files'
 | `-y` / `--yolo` | 等同 `--approval yolo`，只对本次运行有效 |
 | `-C LIST` / `--confirm LIST` | 需要确认的类别：`shell,write,net` 的任意组合或 `none` |
 | `-T PROFILE` / `--tools PROFILE` | `default` 或 `minimal` |
+| `-M MODE` / `--mode MODE` | 工作模式 `coding`（默认）/ `general`，也读 `CST_MODE`，并记进 `settings.json` |
 | `-P NAME` / `-m NAME` / `-e LEVEL` | 本次使用的 provider / 模型 / 推理强度（不传则用上次的） |
 | `-a` / `--ask-key` | 在命令行输入并保存当前 provider 的 key |

@@ -270,6 +299,8 @@ cst-agent --headless --prompt 'List files'
 | `delete_file` / `move_file` | 写 | 删除 / 移动（重命名）文件 |
 | `run_shell` | shell | 非交互命令，超时整组进程被杀，Esc 可中断 |
 | `web_fetch` | 网络 | 抓取 URL，HTML 转纯文本 |
+| `clipboard_read` | 读 | 读系统剪贴板里的文本（`max_bytes` 默认 64 KiB，超出截断并说明） |
+| `clipboard_write` | 写 | 把文本放进系统剪贴板（按写类权限确认；本机走 pbcopy / wl-copy / xclip / Win32 API） |
 | `todo_write` | 元 | 维护任务清单（TUI 右侧面板，Ctrl+T 开关） |
 | `ask_user` | 询问 | 向用户提问（可给选项）；headless 下返回错误让模型自行决定 |
 | `skill` | 元 | 加载一个 skill 的说明（`SKILL.md`）或其中的文件；有 skill 时才出现 |
@@ -281,6 +312,26 @@ cst-agent --headless --prompt 'List files'
 `--tools minimal` 只保留 `read_file` / `run_shell` / `write_file` / `apply_patch`（论文实验口径）。
 TUI 中 `/tools off NAME`、`/tools on NAME` 可临时开关单个工具。

+### 工作模式（coding / general）
+
+`--mode general`（或 TUI 里 `/mode general`、WebUI 运行设置里的「工作模式」、`CST_MODE=general`）把默认口径
+从"代码助手"换成"通用助手"：工作流从改代码/跑测试换成"先弄清要什么、再动手、给出处和不确定的地方"，
+工具说明从"仓库"改成"文件与数据"，并说明 `web_fetch` 可以拿外部资料；`/status` 第一行显示当前模式，
+切换后 history 里会留一条系统提示。默认仍是 `coding`，选择记在 `settings.json`；
+`--tools minimal`（论文口径）的提示词按字节冻结，不受模式影响。
+
+### 剪贴板与文档附件
+
+- **剪贴板**：`clipboard_read` / `clipboard_write` 两个工具（写剪贴板是不可逆发射，要确认）；
+  TUI 里 `/attach paste` 在剪贴板里没有图片时，把剪贴板的**文本**放进输入框（丢掉 CR 与控制字节，尊重输入长度上限）。
+- **附件**（`/attach <path>`、拖入终端、剪贴板图片都算）：图片按原样存进会话目录，发给模型时是 image part；
+  **PDF 在摄取时就抽出文本层**（零依赖解析器：Flate / ASCIIHex / ASCII85、对象流、CID 字体 ToUnicode、
+  内联图像跳过），存的是文本、`mime` 仍是 `application/pdf`；其余文件必须是 UTF-8 文本。
+  文档以 **text part 内联**发出，所以不受 `CST_NO_IMAGES` 或非视觉模型影响；图片照旧会被换成一行"未发送"说明。
+  上限：单个附件 ≤16 MB（文档文本 ≤4 MiB），每条消息最多 8 个。
+  加密 PDF、扫描件（无文本层）会被明确拒绝并说明原因；
+  `view_image` 仍只收图片，文本/PDF 走 `/attach`。
+
 ### 权限、撤销与项目说明

 - **审批模式**（Shift+Tab 循环切换，`/approval`，`--approval`，`CST_APPROVAL`）：
@@ -319,7 +370,7 @@ TUI 中 `/tools off NAME`、`/tools on NAME` 可临时开关单个工具。
  `sudo`、`rm`、解释器（`python3 -c …`）等危险前缀不会被建议成规则。
 - **撤销**：每轮写文件前自动快照到 `.agent/undo/`，`/undo` 恢复上一轮改动的全部文件（新建的文件会被删除）。
 - **AGENTS.md**：工作区根目录的 `AGENTS.md` 会注入 system prompt；`/init` 让 agent 自己生成一份。
- **自动压缩**：上一轮 prompt 用到模型上下文 80% 时，下一轮开始前自动总结历史（`CST_AUTO_COMPACT=0` 关闭）。
+- **自动压缩**：上一轮 prompt 用到模型上下文 75% 时，下一轮开始前自动总结历史（`CST_AUTO_COMPACT=0` 关闭）。

 ### 扩展：Skills、插件、MCP、子 agent

@@ -399,6 +450,7 @@ Thought（推理过程）默认折叠成一行“Thought · 3.2s”，生成中
 | `/provider [name]`、`/model [name]`、`/effort [level]` | 切换 provider（缺 key 时当场输入并用 `/models` 校验）、模型（列表优先显示 API 实际返回的模型，标 `(api)`）、推理强度；都会记住 |
 | `/key` | 输入或更换当前 provider 的 API key（遮罩输入并保存） |
 | `/protocol [name]` | 切换工具调用协议 |
+| `/mode [coding\|general]` | 工作模式（通用助手 / 代码助手口径） |
 | `/tools [on\|off NAME\|minimal\|default]` | 查看或开关工具 |
 | `/approval [ask\|auto\|yolo]` *、`/yolo` * | 审批模式；本次开关 yolo |
 | `/permissions [reset]` *、`/max-steps [N]` * | 权限规则、每轮步数 |
@@ -408,6 +460,8 @@ Thought（推理过程）默认折叠成一行“Thought · 3.2s”，生成中
 | `/diff [path]` * | `git status` + `git diff` |
 | `/compact` | 手动压缩上下文 |
 | `/export [file]` *、`/copy` * | 导出 Markdown 记录；复制上一条回复（OSC 52，需终端支持） |
+| `/attach [path\|paste\|list\|clear]` | 给下一条消息带附件：图片、PDF（自动取文本层）、文本文件；`paste` 优先取剪贴板图片，没有就把剪贴板文本放进输入框 |
+| `/web` * | 后台起一个 WebUI 进程（同 workspace、新会话，日志写到 `.agent/webui.log`） |
 | `/init` | 让 agent 写 `AGENTS.md` |
 | `/details` *、`/thinking [on\|off]`、`/keys` * | 展开全部 tool 输出；Thought 默认展开/折叠；快捷键列表 |
 | `/skills [reload\|list]`、`/skill <name> [任务]` | 浏览 skills；带着某个 skill 开始一轮 |
@@ -2,17 +2,25 @@
 #   ./scripts/build_third_party_mingw.sh curl
 #   cmake -B build-win -DCMAKE_TOOLCHAIN_FILE=cmake/mingw-w64-x86_64.cmake \
 #     -DCST_BUILD_TUI=OFF
+# $MINGW_HOST picks the tool prefix (default x86_64-w64-mingw32; Debian's UCRT
+# cross gcc is x86_64-w64-mingw32ucrt).
 set(CMAKE_SYSTEM_NAME Windows)
 set(CMAKE_SYSTEM_PROCESSOR x86_64)

-set(CMAKE_C_COMPILER x86_64-w64-mingw32-gcc)
-set(CMAKE_CXX_COMPILER x86_64-w64-mingw32-g++)
-set(CMAKE_RC_COMPILER x86_64-w64-mingw32-windres)
+if(DEFINED ENV{MINGW_HOST} AND NOT "$ENV{MINGW_HOST}" STREQUAL "")
+  set(_MINGW_HOST "$ENV{MINGW_HOST}")
+else()
+  set(_MINGW_HOST x86_64-w64-mingw32)
+endif()
+
+set(CMAKE_C_COMPILER ${_MINGW_HOST}-gcc)
+set(CMAKE_CXX_COMPILER ${_MINGW_HOST}-g++)
+set(CMAKE_RC_COMPILER ${_MINGW_HOST}-windres)

 get_filename_component(_AGENT_ROOT "${CMAKE_CURRENT_LIST_DIR}/.." ABSOLUTE)
 set(_MINGW_PREFIX "${_AGENT_ROOT}/third_party/prefix-mingw")
 list(APPEND CMAKE_PREFIX_PATH "${_MINGW_PREFIX}")
-set(CMAKE_FIND_ROOT_PATH "/usr/x86_64-w64-mingw32" "${_MINGW_PREFIX}")
+set(CMAKE_FIND_ROOT_PATH "/usr/${_MINGW_HOST}" "${_MINGW_PREFIX}")
 set(CMAKE_FIND_ROOT_PATH_MODE_PROGRAM NEVER)
 set(CMAKE_FIND_ROOT_PATH_MODE_LIBRARY BOTH)
 set(CMAKE_FIND_ROOT_PATH_MODE_INCLUDE BOTH)
@@ -5,6 +5,20 @@

 ## 安装

+### 安装包
+
+从 [Release](https://git.whcself.xyz/WHC/CST-agent/releases) 下载
+`cst-agent-VERSION-windows-x86_64-setup.exe` 运行即可：按当前用户安装到
+`%LOCALAPPDATA%\Programs\cst-agent`（不需要管理员），把这个目录加进**用户** PATH，
+开始菜单里放一个“CST-Agent WebUI”（在用户目录启动 `cst-agent --web --web-open`）。
+系统低于 Windows 10 1903 时安装包会直接拒绝。
+
+卸载：设置 → 应用 → CST-Agent，或运行安装目录里的 `uninstall.exe`；会去掉 PATH 里的这一项。
+API key 和设置在 `%APPDATA%\cst-agent`，卸载时保留。
+安装包装的文件和 zip 一样，所以之后也能用 `cst-agent-update` 升级，或者重新运行新版安装包覆盖。
+
+### PowerShell 脚本
+
 在 PowerShell 里（Windows PowerShell 5.1 或 PowerShell 7 都行）：

 ```powershell
@@ -123,7 +137,9 @@ windows-x86_64/

 ## 从源码构建

-两种方式产物一样：`dist/windows-x86_64/` 和 `dist/cst-agent-VERSION-windows-x86_64.zip`。
+两种方式产物一样：`dist/windows-x86_64/` 和 `dist/cst-agent-VERSION-windows-x86_64.zip`；
+装了 `makensis`（NSIS）时还有安装包 `dist/cst-agent-VERSION-windows-x86_64-setup.exe`
+（脚本在 `scripts/windows/cst-agent.nsi`）。
 第三方依赖（curl、ncurses、busybox-w32）都按固定版本从源码编译、静态链接。

 ### 在 Linux 上交叉编译
@@ -134,8 +150,16 @@ cd CST-agent
 scripts/bootstrap_windows.sh
 ```

-NixOS / 有 Nix 时自动进 `shell-win.nix`（`pkgsCross.ucrt64`）；否则需要 UCRT 版的
-`x86_64-w64-mingw32-gcc`（带 winpthreads）、cmake、ninja、git、perl。
+NixOS / 有 Nix 时自动进 `shell-win.nix`（`pkgsCross.ucrt64`，带 NSIS）；否则需要 UCRT 版的
+mingw-w64 gcc（带 winpthreads）、cmake、ninja、git、perl。工具前缀由 `MINGW_HOST` 指定，默认
+`x86_64-w64-mingw32`；Debian trixie 起的 `gcc-mingw-w64-ucrt64` 是 `x86_64-w64-mingw32ucrt`：
+
+```bash
+sudo apt install gcc-mingw-w64-ucrt64 binutils-mingw-w64-ucrt64 nsis cmake ninja-build perl
+MINGW_HOST=x86_64-w64-mingw32ucrt scripts/bootstrap_windows.sh
+```
+
+`CST_WIN_CFLAGS` 给 `cst-agent.exe` 追加编译选项（CI 用 `-Wall -Wextra -Werror`）。

 ### 在 Windows 上用 MSYS2 原生编译

@@ -159,17 +183,18 @@ scripts/bootstrap_msys2.sh

 ### 发布

-CI：推 `main` 或 `v*` tag 且 `windows-cross` 成功后，自动把 zip / `.sha256` 发到
+CI：推 `main` 或 `v*` tag 且 `windows` 任务成功后，自动把 zip、安装包和各自的 `.sha256` 发到
 Gitea Release `vVERSION`（版本来自 `include/cst_version.h`；同名资源替换）。仓库 Secrets
 可设 `RELEASE_TOKEN`（`write:repository`），没有则用 Actions 自带 token。

 本地：

 ```bash
-GITEA_TOKEN=... scripts/release_windows.sh --build
+scripts/bootstrap_windows.sh
+GITEA_TOKEN=... scripts/release.sh dist/cst-agent-*-windows-x86_64.zip dist/cst-agent-*-windows-x86_64-setup.exe
 ```

-创建（或复用）`vVERSION` Release，上传 zip 和 `.sha256`。token 只从环境变量读取，经 stdin
+创建（或复用）`vVERSION` Release，上传给出的文件和旁边的 `.sha256`。token 只从环境变量读取，经 stdin
 交给 curl，不会出现在命令行参数、磁盘或输出里。发版前先 bump `cst_version.h`；推
 `vX.Y.Z` tag 时必须与头文件一致。

@@ -21,11 +21,11 @@ The Makefile picks `ncursesw`, then `ncurses`, through pkg-config. On macOS it
 adds Homebrew's keg-only `ncurses`/`curl` to `PKG_CONFIG_PATH`; without
 Homebrew it falls back to the system `-lncurses` (wide-character capable).

-Verified so far: NixOS x86_64 with gcc and clang (`-Werror`, `make check`,
-`make asan`, ctest), and every source compiled `-fsyntax-only` with
-`x86_64-unknown-linux-musl-gcc` without warnings. Full musl, macOS and WSL
-builds run in CI (`.gitea/workflows/ci.yml`; the macOS job needs a runner and
-the `HAVE_MACOS_RUNNER` variable) or by hand with the checklist below.
+Verified so far: NixOS x86_64 with gcc, clang and TinyCC (`-Werror` except
+tcc, `make check`, `make asan`, ctest). CI (`.gitea/workflows/ci.yml`) checks
+with TinyCC on Debian and builds the static musl releases for x86_64, x86 and
+aarch64 with gcc `-Werror` (`make check` on both x86 targets, a qemu smoke
+test on aarch64). macOS and WSL are checked by hand with the list below.

 ## Files

@@ -73,7 +73,7 @@ export OPENAI_MODEL='your-model'
 | `CST_MAX_STEPS` | 每个用户回合最多调用模型几次（默认 30） |
 | `CST_TOOL_OUT_CAP` | 单个工具输出发给模型的字节上限（默认 16384） |
 | `CST_CTX_WINDOW` | 发给模型的最近非 system 消息条数（默认 80） |
-| `CST_SHELL_TIMEOUT` | `run_shell` 超时秒数（默认 120） |
+| `CST_SHELL_TIMEOUT` | `run_shell` 超时秒数（默认不限，Ctrl+C / 停止按钮仍可随时中断） |
 | `CST_METRICS=0` / `CST_METRICS_PATH` | 关闭 / 指定计时埋点 JSONL |
 | `CST_RECORD_SSE` / `CST_REPLAY_SSE` | 录制 / 回放原始 SSE 流目录；`CST_REPLAY_CHUNK` 控制回放分块字节数 |
 | `CST_MODELS` | `/model` 可选列表，`:` 或 `,` 分隔 |
@@ -47,7 +47,9 @@ webui: token 3f9a…  (append #token=3f9a… to the URL)

 | 场景 | 行为 |
 |---|---|
-| 默认 `127.0.0.1` | 不需要 token；`Host` 必须是 loopback（防 DNS rebinding） |
+| 默认 `127.0.0.1` | 不需要 token；`Host` 必须是 loopback（防 DNS rebinding）；带 `Origin` 的请求必须来自本页面自己（`http://<Host>`），否则 401 —— 防止任意网页跨站 POST 驱动 agent |
+| 静态资源 | 页面本身与 `app.js`/`style.css` 不需要 token；`/api/attach/<rel>` 图片仍要（页面会把 token 写进 `SameSite=Strict` cookie） |
+| 请求解析 | 鉴权与大小检查在读请求体**之前**完成；`Content-Length` 必须是纯数字且只有一个值；请求头整体 15 s 内收完 |
 | `--web-host 0.0.0.0` | 没给 `--web-token` 时自动生成随机 token 并打印；所有 `/api/*` 都要 token |
 | token 传递 | `Authorization: Bearer <t>`、`Cookie: cst_token=<t>` 或 `?token=<t>`（`EventSource` 不能设 header，所以查询参数是必需的一条路） |
 | 资源 | 只有白名单路由；路径里的 `..`/编码 NUL 一律 400；静态资源来自二进制内嵌（或 `--web-dev DIR`） |
@@ -70,7 +72,10 @@ GET /api/events?since=<seq>      →  text/event-stream (chunked)

 * 每条事件都有 `id: <seq>`（单调递增）与 `data: {"type":…,"seq":…, …}`。
 * 断线重连带上 `since=<最后收到的 seq>`（浏览器用 `Last-Event-ID` 也行），
-  服务端保留最近 2000 条，超出后从可用的最老一条开始。
+  服务端保留最近 2000 条，超出后从可用的最老一条开始；比当前 seq 还大（服务端重启过）
+  就当作新连接。
+* 新连接（不带 `since`）只收之后的事件；但若有尚未回答的 `permission`/`ask_user`，
+  会从那条事件开始补发，刷新页面不会丢掉正在等的问题。
 * 心跳是 SSE 注释 `: ping`（每秒一次，客户端忽略）。
 * `state` 事件在状态变化时整份重发，客户端可以直接替换本地状态。

@@ -83,10 +88,10 @@ GET /api/events?since=<seq>      →  text/event-stream (chunked)
 | `assistant_html` | `text` | **已渲染且已消毒的 HTML 片段**，按顺序追加 |
 | `assistant_delta` | `text` | 原始文本增量（逐字节真相，做复制/导出时用） |
 | `think_delta` | `text` | 推理过程增量（Thought） |
-| `tool` | `id`,`name`,`args`,`phase`,`detail` | `phase` = `open`/`ready`/`run`/`done` |
+| `tool` | `id`,`name`,`args`,`phase`,`detail` | `phase` = `open`/`ready`/`run`/`progress`/`done`；`progress` 是子代理的最近几步（`detail` 形如 `explore · grep foo · read_file x`），可以出现多次 |
 | `tool_result` | `id`,`name`,`out`,`exit`,`bytes`,`truncated` | 完整输出被截到 32 KiB，`truncated=true` |
 | `todo` | `text` | `todo_write` 替换后的清单（一行一项） |
-| `permission` | `id`,`tool`,`detail`,`preview`,`sandbox` | 需要回答（见下） |
+| `permission` | `id`,`tool`,`detail`,`preview`,`sandbox`,`timeout` | 需要回答（见下）；`timeout` 秒后按拒绝处理 |
 | `permission_done` | `id`,`answer` | 问题已结束（超时/取消也会发） |
 | `ask_user` | `id`,`question`,`options[]` | 需要回答 |
 | `session` | `id` | 切换/新建了会话，客户端应重新拉 `/api/session` |
@@ -95,30 +100,80 @@ GET /api/events?since=<seq>      →  text/event-stream (chunked)
 `state` 字段：`workspace` `session` `provider` `model` `effort` `approval`
 (`ask|auto|yolo`) `plan_mode` `protocol_tools` `status_busy` `todo`
 `context_tokens` `undo_available` `pending_images[]` `usage{requests,prompt_tokens,
-completion_tokens,cost,priced,currency}`。
+completion_tokens,cached_tokens,cache_miss_tokens,cost,priced,currency}`
+`turn_usage{…}`（上一轮，字段同 `usage`）`api_base` `key_set` `key_from`
+（`saved` / 环境变量名 / `session` / 空，**从不含 key 本身**）
+`model_info{label,context,vision,reasoning,source,price?}` `effort_applies`
+`sees_images`（1 / 0 / -1 未知）`max_steps` `tool_profile` `perm_timeout`
+`show_thinking`（思考过程默认展开，存在 `settings.json`，和 TUI 的 `/thinking` 是同一个开关）。
+`cached_tokens` / `cache_miss_tokens` 来自上游的 prompt cache 统计，前端用它算命中率。
+
+`think_delta` / `assistant_delta` 的 `text` 总是完整的 UTF-8：上游在多字节字符中间
+断开时，残缺的尾巴留到下一块再发。

 ### 请求

 | 方法 | 路径 | 说明 |
 |---|---|---|
 | GET | `/api/state` | 上面那份快照 |
-| GET | `/api/session?from=N` | 会话消息（`{messages:[…]}`，形状与 `.agent/sessions/*.jsonl` 一致；assistant 消息额外带 `html`） |
+| GET | `/api/session?from=N` | 会话消息（`{messages:[…]}`，形状与 `.agent/sessions/*.jsonl` 一致，另加的字段见下） |
 | GET | `/api/sessions` | `{current, sessions:[{id,title,messages,mtime}]}` |
 | POST | `/api/sessions/new` | `{}` → `{id}` |
 | POST | `/api/sessions/switch` | `{"id":"…"}` |
+| POST | `/api/sessions/rename` | `{"id":"…"?,"title":"…"}`（不带 `id` = 当前会话）；标题按字符截到 120 字节，空串 = 恢复默认（第一句提问）；没这个会话 → 404 |
+| POST | `/api/sessions/delete` | `{"id":"…"}` → `{ok,current}`；删的是当前会话时先开一个新会话（并发 `session` 事件）；没这个会话 → 404 |
+| POST | `/api/retry` | `{}` = 重新生成最后一轮；`{"text":"…"}` = 把最后一句提问换成新文字再跑（编辑重发）。先把最后一轮从会话里退掉，再开新一轮 → 202 `{ok,text}`；没有可重来的轮次 / 轮次在跑 → 409；空串或非字符串 → 400 |
+| GET | `/api/export` | 当前会话的 Markdown（`text/markdown`，与 TUI `/export` 同一格式） |
+| GET | `/api/diff[?path=p]` | 工作区的真实 diff：`{git,path,files:[{status,path}],diff,truncated}`。`files` 来自 `git status --short`（含未跟踪），`diff` 是 `git diff`（未暂存部分，限 512 KiB）；不是 git 仓库 / 没装 git → `git:false`；10 s 超时 |
+| GET | `/api/settings` | 运行设置：`protocol`,`protocols[]`,`max_steps`,`tool_profile`,`profiles[]`,`confirm{shell,write,net}`,`approval`,`tools[{name,perm,on,desc,…}]`,`sandbox{kind,active,labels?}` |
+| POST | `/api/settings` | `{protocol?,max_steps?,tool_profile?,tools?:{名字:bool},confirm?:{shell?,write?,net?}}`；整体校验后才生效（任何一项不对 → 400，什么都不改），成功后发 `state` |
+| GET | `/api/permissions` | 已保存的"总是允许"规则：`{text,tools[],shell_prefixes[]}`（来自 `<workspace>/.agent/permissions.json`） |
+| DELETE | `/api/permissions` | 清空这些规则 |
+| POST | `/api/init` | `{}` → 202 `{ok,text}`：开一轮，让 agent 分析仓库写 `AGENTS.md`（与 TUI `/init` 同一段提示词） |
+| POST | `/api/skill` | `{"name":"…","task":"…"?}` → 202 `{ok,text}`：把 skill 正文内联进这条消息再开一轮（同 TUI `/skill`）；没有这个 skill → 404；缺 `name` / 超过 63 字节 / `task` 不是字符串 → 400 |
+| GET | `/api/extensions` | `{enabled,trusted,report,task_tool,mcp:[{name,source,command,state,tools,error}],plugins:[{name,source,desc,tools,mcp,skills,error}],skills:[{name,source,desc}],roles:[{name,source,desc,model,read_only}],runs:[{file,role,desc,status,steps,tools,ms}]}`；`state` = `starting/ready/failed/exited/off`，`runs` 是当前会话的子代理运行 |
+| POST | `/api/extensions/reload` | 停掉 MCP 服务器和插件，重新读取它们和 skills → `{ok,trusted,report}` |
+| POST | `/api/extensions/trust` | `{"on":true\|false}`：信任（或取消）当前工作区——决定它的 `.agent/plugins`、`.agent/mcp.json` 是否加载；写进 `settings.json` 后重新加载，回复同上 |
+| POST | `/api/workspace` | `{"path":"/abs/dir"}` —— 换工作目录（权限规则/埋点/AGENTS.md 跟着换，扩展按新目录重新加载），并**在新目录里开一个新会话**（会话是按目录存的）；返回 `{id}`，坏路径 → 400 |
 | POST | `/api/turn` | `{"text":"…"}` → 202 `{ok:true}`；已有轮次在跑时 **409** |
-| POST | `/api/cancel` | `{}` —— 取消当前轮（不会杀进程） |
+| POST | `/api/cancel` | `{}` —— 取消当前轮（不会杀进程）；挂起的权限问题按拒绝、提问按忽略处理 |
 | POST | `/api/permission` | `{"id":N,"answer":0\|1\|2\|3}` → deny / once / always-tool / always-prefix |
 | POST | `/api/ask_user` | `{"id":N,"text":"…"}` 或 `{"id":N,"dismiss":true}` |
-| POST | `/api/config` | `{model?,provider?,effort?,approval?,plan_mode?}` |
+| POST | `/api/config` | `{model?,effort?,approval?,plan_mode?,show_thinking?}`；先整体校验再生效，model/effort/approval（`yolo` 除外）写进 `settings.json`；当前模型不吃 effort 时回 `warning`；带 `provider` → 400（改用 `/api/provider`） |
+| GET | `/api/providers` | `{current, providers:[{name,api_base,model,has_key,key_from,key_env,source,current}]}` |
+| GET | `/api/models[?refresh=1]` | 当前提供商的模型：`{provider,current,api_base,key_set,api_count,models:[{id,api,label,context,vision,reasoning,source,price?}]}`；先用缓存，过期/缺失/`refresh=1` 时向 `/v1/models` 拉；key 被拒 → 401（仍带内置模型和 `auth:true`），其它失败放在 `api_error` |
+| POST | `/api/provider` | `{"name":"…","key":"…"?}`：切换提供商。缺 key → 400 `{need_key:true}`；key 被拒 → 401；没这个提供商 → 404；成功 → `{ok,provider,model,warning?}`，key 验证后存进 `credentials`（0600） |
+| POST | `/api/key` | `{"key":"…"}`：换当前提供商的 key，验证与回复同上 |
 | POST | `/api/undo` | 回滚最近一轮的文件改动，回 200 + 报告文本 |
 | POST | `/api/compact` | 手动压缩上下文 |
-| POST | `/api/attach?name=x.png` | 请求体是图片原始字节（≤16 MiB），进入"待发送"列表 |
+| POST | `/api/attach?name=x.png` | 请求体是图片原始字节（≤32 MiB），进入"待发送"列表 |
 | DELETE | `/api/attach` | 清空待发送图片 |
 | GET | `/api/attach/<rel>` | 读回已存进会话的图片（`rel` 由消息的 `attach[].rel` 给出） |
 | POST | `/api/quit` | 关闭进程 |

 未命中路由 → `404 {"error":"not found"}`；需要 token 而没给 → `401`。
+轮次运行中（或另一个修改正在进行）时，`sessions/new`、`sessions/switch`、`sessions/delete`、
+`workspace`、`undo`、`compact`、`attach`（上传/清空）、`provider`、`key`、`retry`、`diff`、
+`POST /api/settings`、`DELETE /api/permissions`、`init`、`skill`、`extensions/reload`、
+`extensions/trust` 以及改 `model/effort` 的 `config` 一律 **409**；
+`approval`/`plan_mode`/`show_thinking` 可以中途改，`sessions/rename`、`export` 与
+`GET /api/extensions` 在轮次进行中也能用。
+
+`/api/session` 的每条消息在会话文件的字段之外还带：
+
+| 字段 | 出现在 | 说明 |
+|---|---|---|
+| `html` | assistant | 已消毒的 Markdown 渲染结果（文本协议时只渲染说明文字，不含调用帧） |
+| `text` | 文本协议的 assistant | 去掉调用帧后的说明文字（`content` 仍是原文） |
+| `calls` | 带工具调用的 assistant | `[{id,name,args}]`：原生协议来自 `tool_calls`，文本协议从 `content` 解析 |
+| `tool_result` | 工具结果 | `true`；原生协议 role 是 `tool`，文本协议 role 是 `user`（凭 `tool_call_id` 区分） |
+| `out`,`exit`,`truncated` | 工具结果 | 去掉 `[tool …]` / `result … end` 外框后的输出（截到 32 KiB，按字符边界）与退出码 |
+| `note` | 引擎插入的提示 | `true`；如中途切换计划模式时的 `[cst] …` 说明，不是用户说的话 |
+
+客户端据此在刷新后重建"本轮工具调用"列表：`calls` 建行，`tool_call_id` 相同的结果填进去。
+
+key 只出现在请求体里：回复、事件、日志都不会带它。`/api/provider` 与 `/api/key`
+跟其它 `/api/*` 一样受 loopback 同源检查或 token 保护。

 ### 客户端接入示例（10 行）

@@ -149,12 +204,70 @@ api("/api/turn", {"text": "hello"})           # 之后读 /api/events 收流
 Markdown 在**服务端**渲染（复用 TUI 用的 sashidown 流式 API），客户端只贴 HTML；
 因此不需要引入任何 JS 依赖，也顺带保证了"只发送已确认的 HTML"。

+界面本身：左栏可收（顶栏按钮 / `?side=0`），右栏可收（`?panel=0`），主题 `?theme=dark|light`；
+三个开关都会记进 `localStorage`，带查询参数时以参数为准。空状态只留一句提示，第一轮
+对话开始后自动隐去；右栏只给三件事——任务清单、本轮文件改动行数（`+N −M`）、上下文
+占用与 prompt cache 命中率。布局类在页面脚本之前就写进 `<body>`，避免首帧闪一下再收起。
+
+对话列：正文、过程块、提示行和输入框共用同一条左边线，正文宽度上限 720px（`--content`）；
+左右两栏随屏宽在 `clamp()` 范围内变宽，超宽屏下中间列不会无限拉开。
+
+过程时间线：一段**连续**的思考和工具调用合成一个过程块，折叠时只有一行
+（"思考 4 次 · 工具 6 次"；运行中显示当前动作，比如"执行 run_shell…"），展开后是一条
+细竖线串起的时间线，每次思考、每个工具调用各占一行，思考只露第一行，点开看全文。
+出现可见的正文（或引擎提示、下一句提问）时过程块就此结束，后面的思考/工具另起一块——
+历史重建（`/api/session`）和实时事件流用的是同一条规则，所以刷新前后看到的一样。
+`/thinking on|off`（`show_thinking`）决定过程块和其中的思考默认展开还是收起，流式输出时也照此执行；
+`/details` 只管工具行是否展开。
+
+图片：提问附带的图片显示在气泡**上方**、靠右，一张保持原比例（最长 260px），多张是
+152px 的圆角方块；点开是大图预览（Esc 关闭，←/→ 切换）。从本页上传/粘贴/拖入的图片在
+输入框上方显示缩略图，发送后立刻出现在新提问上（本地预览，本轮结束后换成 `/api/attach/<rel>`）；
+从别的客户端加的待发图片仍然显示为文字条。
+
+### 命令与快捷键
+
+输入框以 `/` 开头时是命令，在页面里执行，不发给模型：打 `/` 弹出列表（↑↓ 选，Tab 补全，
+Enter 执行），`/model`、`/effort`、`/approval`、`/provider`、`/skill`、`/thinking`、`/details`
+还会补全参数。`/usr/bin 是什么` 这种第一个词里带 `/` 的照常发送；未知命令留在输入框里并提示，
+开头加个空格即可当普通消息发出。
+
+| 命令 | 作用 |
+|---|---|
+| `/new` `/sessions` `/rename [标题]` | 新建 / 会话列表 / 重命名 |
+| `/retry` `/edit` `/stop` `/undo` `/compact` | 重新生成 / 编辑重发 / 停止 / 撤销写入 / 压缩上下文 |
+| `/export` `/copy` `/diff [文件]` | 导出 Markdown / 复制最后一条回答 / 工作区 diff |
+| `/model [id]` `/effort [级别]` `/provider [名字]` `/key` | 模型、推理强度、提供商、API key |
+| `/approval [ask\|auto\|yolo]` `/plan` `/build` | 审批权限、计划/构建模式 |
+| `/settings`（`/protocol` `/tools` `/max-steps` `/permissions`） | 运行设置 |
+| `/init` `/skill 名字 [任务]` `/skills` `/extensions`（`/mcp` `/plugins` `/agents`） | 写 AGENTS.md、用 skill、扩展状态 |
+| `/thinking [on\|off]` `/details [on\|off]` | 思考过程默认展开（存 `settings.json`）、工具输出全部展开（存浏览器） |
+| `/theme` `/palette` `/help` | 主题、命令面板、帮助 |
+
+快捷键：Ctrl+P 命令面板，Ctrl+K 会话列表，Shift+Tab 循环审批（每次询问 → 自动 → 完全放行），
+F1 或 `?` 帮助，Esc 关闭弹窗 / 取消编辑 / 停止当前一轮，Ctrl+B / Ctrl+J 收起左栏 / 右栏。
+
+引擎的英文状态、错误和会话里的 `[cst]` 提示在页面上换成中文；一轮失败时（`turn_end.rc≠0`）
+显示可读的原因（比如"API key 无效或已过期（HTTP 401）"、"用满了 30 步上限"），不再是 `rc=-1`。
+
+TUI 里打 `/web` 会另起一个 WebUI 进程（`--web --workspace <当前目录> --new`），
+日志在 `<workspace>/.agent/webui.log`：TUI 不能被 `cst_webui_run()` 的阻塞循环占住，
+而且两个进程同时写一个会话文件会把它写坏，所以子进程总是开新会话。
+
+自动压缩的门槛是模型上下文窗口的 **75%**（`CST_CTX_BUDGET` 可覆盖）；到点后下一轮
+开始前先总结历史，而不是等到塞满。
+
 ## 测试

 ```bash
 make check                        # 含 tests/test_webui（HTTP/SSE/鉴权/消毒）
 sh scripts/e2e_webui.sh          # 用录制流跑完整轮次 + 权限往返，不需要 API key
-sh scripts/e2e_webui.sh tests/fixtures/sse/deepseek-ctl ctl
+                                  # （含 e2e_webui_settings.py：假 /v1/models、换提供商/key、中文思考流；
+                                  #  e2e_webui_sessions.py：重新生成/编辑重发/导出/重命名/删除/历史字段；
+                                  #  e2e_webui_workspace.py：diff、运行设置、权限规则；
+                                  #  e2e_webui_ext.py：/init、/skill、扩展、信任、思考开关）
+sh scripts/e2e_webui.sh tests/fixtures/sse/deepseek-ctl-heredoc ctl-heredoc
+                                  # （deepseek-ctl 录的是一次契约错误连发的失败运行，不能拿来跑这个）
 ```

 手工验收清单（浏览器）：
@@ -162,8 +275,11 @@ sh scripts/e2e_webui.sh tests/fixtures/sse/deepseek-ctl ctl
 1. 打开首页 → 顶部显示 workspace/模型/审批模式，右下状态"就绪"。
 2. 输入一句话 → 助手气泡逐字出现；工具卡片自动展开、结束后带 `exit=` 标记。
 3. 切到 `ask` 审批 → 写文件时弹窗；点"允许一次"后继续，点"拒绝"则该工具返回失败但对话继续。
-4. `会话` 按钮：新建/切换后历史正确重建（刷新页面也一样）。
-5. 拖入一张图片 → 底部出现待发送提示，发送后历史里能看到图。
+4. `会话` 按钮：新建/切换后历史正确重建（刷新页面也一样）：连续的思考和工具调用折叠成一行
+   过程块，展开是时间线，工具行点开能看到参数和输出；搜索、重命名、删除可用。
+   `/thinking off` 时流式输出的思考也保持收起。
+4a. 最后一句提问上有"编辑重发"，最后一个回答上有"重新生成"；右栏文件行点开是真实 diff。
+5. 拖入一张图片 → 输入框上方出现缩略图，发送后图片立刻出现在提问气泡上方，点开是大图预览。
 6. 关掉标签页再打开 → 事件流自动续传，不丢已完成的消息。

 ## 限制与已知问题
@@ -47,15 +47,18 @@ int cst_protocol_is_text(cst_protocol_t p);
 /* Eager protocols execute a call during the stream, at confirmation time. */
 int cst_protocol_is_eager(cst_protocol_t p);

-/* One image carried by a message (see cst_attach.h). The bytes live beside
- * the session; this is the metadata plus the name of the stored copy. */
+/* One file carried by a message (see cst_attach.h): an image, or a document
+ * whose stored payload is its text. The payload lives beside the session;
+ * this is the metadata plus the name of the stored copy. */
 typedef struct {
  char id[32];      /* unique inside the session ("img-1a2b3c4d") */
  char name[256];   /* display name (original file name) */
-  char mime[32];    /* image/png, image/jpeg, image/gif, image/webp, image/bmp */
+  char mime[32];    /* image/png, ...; application/pdf for a PDF whose text
+                     * layer the stored file holds; text/plain for text */
  char rel[256];    /* file name inside the session's attachment directory */
-  long long bytes;  /* stored size */
+  long long bytes;  /* stored size (the text, for documents) */
  int width, height; /* pixels, 0 when the header could not be read */
+  int pages;        /* PDF page count, 0 for anything else */
  char detail[8];   /* auto | low | high (wire hint, "" = auto) */
  char origin[32];  /* user | view_image | mcp (for display only) */
 } cst_attachment_t;
@@ -82,6 +85,7 @@ typedef struct {
  char model[128];
  char provider[32];
  char effort[16];      /* low | medium | high | max */
+  char mode[16];        /* coding | general (CST_DEFAULT_MODE) */
  cst_protocol_t protocol;
  int max_steps;        /* model calls per user turn */
  size_t tool_out_cap;  /* bytes of one tool output sent to the model */
@@ -114,6 +118,14 @@ typedef enum {
 const char *cst_approval_name(int a);
 int cst_approval_parse(const char *s, int *out);

+/* Working mode: how the system prompt frames the agent. "coding" (default)
+ * is code framed; "general" is for tasks that are not about a codebase
+ * (documents, data, research, everyday files). Orthogonal to plan mode. */
+#define CST_DEFAULT_MODE "coding"
+int cst_mode_valid(const char *m);
+/* Canonical name of m, or the default when m is unknown / empty. */
+const char *cst_mode_name(const char *m);
+
 typedef struct cst_session cst_session_t;
 typedef struct cst_agent cst_agent_t;

@@ -130,6 +142,12 @@ const cst_message_t *cst_session_get(const cst_session_t *s, int i);
 const char *cst_session_path(const cst_session_t *s);
 int cst_session_clear(cst_session_t *s, int keep_system);
 int cst_session_rewrite(cst_session_t *s, const cst_message_t *msgs, int n);
+/* Tool output: role=tool, or a text-protocol user message answering a call. */
+int cst_message_is_tool_result(const cst_message_t *m);
+/* The transcript as Markdown, malloc'd (NULL on OOM). Tool output and call
+ * arguments are cut at 4000 bytes. Takes the session lock. */
+char *cst_session_export_md(cst_session_t *s, const char *title,
+                            const char *sid, const char *model);
 /* Readers on another thread must hold the lock while using pointers returned
 * by cst_session_get (append may reallocate). Append locks internally. */
 void cst_session_lock(cst_session_t *s);
@@ -272,6 +290,8 @@ int cst_llm_list_models(const char *api_base, const char *api_key,
 char *cst_llm_build_body(const cst_llm_request_t *req);
 /* Reset the CST_RECORD_SSE / CST_REPLAY_SSE sequence counter. */
 void cst_llm_trace_reset(void);
+/* Set CURLOPT_CAINFO on a CURL * when a CA bundle file is known. */
+void cst_curl_apply_ca(void *curl);

 /* on_permission answers */
 #define CST_PERM_DENY 0
@@ -321,6 +341,12 @@ cst_agent_config_t *cst_agent_config_mut(cst_agent_t *a);
 void cst_agent_request_cancel(cst_agent_t *a);
 int cst_agent_new_session(cst_agent_t *a, char *id_out, size_t id_sz);
 int cst_agent_switch_session(cst_agent_t *a, const char *id);
+/* Re-root the agent at another directory. Per-workspace state (permission
+ * rules, metrics, project notes) follows, the system prompt is rebuilt and the
+ * token budget is re-measured. The session is not touched: the caller decides
+ * whether to keep it or start a new one. */
+int cst_agent_set_workspace(cst_agent_t *a, const char *path, char *err,
+                            size_t err_sz);
 int cst_agent_compact(cst_agent_t *a, char *err, size_t err_sz);
 /* Bare id of the current session (file name without .jsonl). */
 void cst_agent_session_id(const cst_agent_t *a, char *out, size_t out_sz);
@@ -392,13 +418,14 @@ int cst_creds_resolve(const char *provider, char *out, size_t out_sz,
 /* Hidden console prompt (no echo). */
 int cst_creds_prompt_console(const char *provider, char *out, size_t out_sz);

-/* Last used provider, model per provider and effort, in
+/* Last used provider, model per provider, effort and working mode, in
 * <config>/settings.json. */
 #define CST_SETTINGS_MAX_MODELS 16
 #define CST_SETTINGS_MAX_TRUSTED 32
 typedef struct {
  char provider[32];
  char effort[16];
+  char mode[16];     /* coding | general, "" = not chosen yet */
  char approval[8]; /* "" = not chosen yet */
  int show_thinking; /* TUI: Thought blocks start expanded */
  /* workspaces whose .agent/plugins and .agent/mcp.json may run */
@@ -413,11 +440,13 @@ typedef struct {
 int cst_settings_load(cst_settings_t *s);
 int cst_settings_save(const cst_settings_t *s);
 const char *cst_settings_model(const cst_settings_t *s, const char *provider);
-/* Record cfg's provider, model and effort as the last used ones. */
+/* Record cfg's provider, model, effort and mode as the last used ones. */
 int cst_settings_remember(const cst_agent_config_t *cfg);
 /* Record an approval mode the user picked (CST_APPROVAL_*). */
 int cst_settings_remember_approval(int approval);
 int cst_settings_remember_thinking(int show_thinking);
+/* Record the working mode the user picked (coding | general). */
+int cst_settings_remember_mode(const char *mode);
 /* Workspace trust for project plugins / MCP servers (by real path). */
 int cst_settings_trusted(const char *workspace);
 int cst_settings_trust(const char *workspace, int on);
@@ -9,16 +9,24 @@
 extern "C" {
 #endif

-/* --- image attachments ---------------------------------------------------
- * An attachment is a validated copy of an image file, stored beside the
- * session so the bytes a model is shown stay the bytes that were picked.
- * Only the metadata travels in the JSONL; the bytes are encoded into the
- * request when it is built. */
+/* --- attachments ---------------------------------------------------------
+ * An attachment is a validated copy of a file, stored beside the session so
+ * what a model is shown stays what was picked. Images are stored as they are
+ * and encoded into the request when it is built; a PDF is reduced to its text
+ * layer and a text file is stored as is, both of which travel inline as text
+ * (so a text-only model still gets them). Only the metadata goes in the
+ * JSONL. */

-/* Supported: PNG, JPEG, GIF, WebP, BMP (static images). */
-#define CST_ATTACH_MAX_BYTES (16 * 1024 * 1024) /* one stored image */
+/* Supported: PNG, JPEG, GIF, WebP, BMP (static images), PDF, UTF-8 text. */
+#define CST_ATTACH_MAX_BYTES (16 * 1024 * 1024) /* one picked file */
 #define CST_ATTACH_MAX_TOTAL (32 * 1024 * 1024) /* per request */
 #define CST_ATTACH_MAX_COUNT 8                  /* per message */
+/* Text stored and inlined per document (a PDF also caps its text here). */
+#define CST_ATTACH_TEXT_MAX (4 * 1024 * 1024)
+
+/* Images go on the wire as image parts; everything else is inlined as text. */
+int cst_attach_is_image(const cst_attachment_t *a);
+int cst_attach_is_text(const cst_attachment_t *a);

 /* Magic-byte test: 0 and a canonical mime when data holds a supported
 * image, -1 otherwise (an extension alone never decides). */
@@ -35,23 +43,25 @@ int cst_attach_dir(const char *session_path, char *out, size_t out_sz);
 int cst_attach_full_path(const char *session_path, const cst_attachment_t *a,
                         char *out, size_t out_sz);

-/* Read src_path, verify it is a supported image within limits, and copy it
- * into the session's attachment directory. Returns a malloc'd entry the
- * caller owns (cst_attach_free), or NULL with err filled. */
+/* Read src_path and copy it into the session's attachment directory: an image
+ * as it is, a PDF as its extracted text, anything else only when it is text.
+ * Returns a malloc'd entry the caller owns (cst_attach_free), or NULL with
+ * err filled. */
 cst_attachment_t *cst_attach_import_file(const char *session_path,
                                         const char *src_path, char *err,
                                         size_t err_sz);
-/* Store already-read bytes (clipboard, downloads); name is display-only. */
+/* Store already-read image bytes (clipboard, downloads); name is
+ * display-only. */
 cst_attachment_t *cst_attach_import_bytes(const char *session_path,
                                          const char *name, const char *data,
                                          size_t n, char *err, size_t err_sz);

-/* Validate and read back a stored image, or NULL when it is gone. */
+/* Validate and read back a stored payload, or NULL when it is gone. */
 char *cst_attach_read(const char *session_path, const cst_attachment_t *a,
                      size_t *out_len, char *err, size_t err_sz);
-/* "data:image/png;base64,...." for the wire; NULL when the file is gone. */
+/* "data:image/png;base64,...." for the wire; images only, NULL otherwise. */
 char *cst_attach_data_url(const char *session_path, const cst_attachment_t *a);
-/* "1920x1080 PNG 412 KB" style one-liner for prompts and the UI. */
+/* "1920x1080 png 412 KB" / "pdf 1.2 MB, 12 pages" one-liner for prompts. */
 void cst_attach_describe(const cst_attachment_t *a, char *out, size_t out_sz);

 void cst_attach_free(cst_attachment_t *a);
@@ -13,6 +13,20 @@ extern "C" {
 * ("no image on the clipboard", "install wl-clipboard", ...). */
 int cst_clipboard_image(char **data, size_t *n, char *err, size_t err_sz);

+/* Cap on text moved through the clipboard (read and write). */
+#define CST_CLIPBOARD_TEXT_MAX (1024 * 1024)
+
+/* UTF-8 text on the system clipboard: Wayland (wl-paste), X11 (xclip),
+ * macOS (pbpaste), Windows (CF_UNICODETEXT). Returns 0 with *data malloc'd
+ * and NUL-terminated, or -1 with err filled in ("no text on the clipboard",
+ * "reading the clipboard needs wl-clipboard", ...). */
+int cst_clipboard_text(char **data, size_t *n, char *err, size_t err_sz);
+
+/* Replace the clipboard with UTF-8 text: wl-copy / xclip / pbcopy, or the
+ * Windows clipboard API. Returns 0 on success, else -1 with err. */
+int cst_clipboard_set_text(const char *text, size_t n, char *err,
+                           size_t err_sz);
+
 /* Windows CF_DIB bytes (BITMAPINFOHEADER + pixels) as an image file: a PNG
 * for uncompressed 24/32-bit bitmaps, else a BMP. NULL when malformed. */
 char *cst_dib_to_image(const void *dib, size_t sz, size_t *n);
@@ -32,6 +32,7 @@ typedef struct {
  const char *host;
  const char *content_type;
  const char *user_agent;
+  const char *last_event_id; /* an EventSource reconnecting */
  int keep_alive;
 } cst_http_req_t;

@@ -72,8 +73,10 @@ cst_http_server_t *cst_http_start(const cst_http_opts_t *o, char *err,
 int cst_http_port(const cst_http_server_t *s);
 /* Non-zero once stop() was called: long-lived handlers must return. */
 int cst_http_stopping(const cst_http_server_t *s);
-/* Close the listener and every connection, then wait for the threads. */
-void cst_http_stop(cst_http_server_t *s);
+/* Close the listener and every connection, then wait (30 s at most) for the
+ * handler threads. Returns how many still run: their ud must stay alive. */
+int cst_http_stop(cst_http_server_t *s);
+/* Does nothing while handlers still run on s. */
 void cst_http_free(cst_http_server_t *s);

 /* --- inside a handler --- */
@@ -0,0 +1,27 @@
+#ifndef CST_PDF_H
+#define CST_PDF_H
+
+#include <stddef.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/* Text layer of a PDF, in page order, with no external dependency (no
+ * poppler, no zlib): classic objects and object streams, FlateDecode plus
+ * ASCIIHex/ASCII85, PNG/TIFF predictors, Tj/TJ/'/" text with per-font
+ * ToUnicode CMaps and a WinAnsi / Latin-1 fallback. Scanned pages have no
+ * text layer and are reported as such.
+ * Returns a malloc'd NUL-terminated UTF-8 string, or NULL with err filled in
+ * ("PDF is encrypted ...", "no extractable text ..."). *pages (optional) gets
+ * the page count; the result is truncated at CST_PDF_TEXT_MAX bytes. */
+char *cst_pdf_text(const void *data, size_t n, int *pages, char *err,
+                   size_t err_sz);
+
+#define CST_PDF_TEXT_MAX (4u * 1024u * 1024u)
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
@@ -0,0 +1,79 @@
+#ifndef CST_PROVIDER_SWITCH_H
+#define CST_PROVIDER_SWITCH_H
+
+/* Model lists, API keys and provider switches, shared by the TUI and the
+ * WebUI so both front ends validate and persist the same way.
+ *
+ * A key is checked with GET /models: a 401/403 (CST_MODELS_AUTH) rejects
+ * it, any other failure (offline, no /models route) lets it through
+ * unverified. A key that passes is saved to <config>/credentials. Keys are
+ * never copied into err. */
+
+#include "cst_agent.h"
+#include "cst_models.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define CST_PS_MAX_MODELS 512
+#define CST_PS_UNKNOWN (-1)   /* no such provider */
+#define CST_PS_NEED_KEY (-10) /* no key given, none saved or in the env */
+/* Non-negative results are flags: */
+#define CST_PS_UNVERIFIED 1 /* /models failed for another reason; err says */
+#define CST_PS_UNSAVED 2    /* the credentials file could not be written */
+
+typedef struct {
+  char id[128];
+  int api; /* listed by the provider's /models (else built-in only) */
+  cst_model_info_t info;
+} cst_ps_model_t;
+
+/* GET /models; a good answer refreshes models-cache/<provider>.json. The
+ * count, CST_MODELS_AUTH, or another negative value with err set. */
+int cst_ps_fetch(const char *provider, const char *api_base, const char *key,
+                 char ids[][128], int max, char *err, size_t err_sz);
+/* The cached list: count, or -1; *fresh when younger than the TTL. */
+int cst_ps_cached(const char *provider, char ids[][128], int max, int *fresh);
+
+/* Base URL for provider: cfg's own when it is the current provider (it may
+ * be overridden), else the preset's; "" when unknown. */
+const char *cst_ps_base(const cst_agent_config_t *cfg, const char *provider);
+
+/* The model picker list for cfg's provider: the API's ids (the cache while
+ * it is fresh, else a fetch; refresh always fetches) followed by the
+ * built-in names it did not list. *api_rc gets the fetch result (count,
+ * CST_MODELS_AUTH, -1 with err) or the cached count. Returns the count. */
+int cst_ps_models(const cst_agent_config_t *cfg, int refresh,
+                  cst_ps_model_t *out, int max, int *api_rc, char *err,
+                  size_t err_sz);
+
+/* Checks key against /models and saves it: CST_MODELS_AUTH (err set), or
+ * the CST_PS_* flags. When ids is given it gets the fetched list and *nids
+ * the raw fetch result. */
+int cst_ps_verify_save(const char *provider, const char *api_base,
+                       const char *key, char ids[][128], int max, int *nids,
+                       char *err, size_t err_sz);
+
+/* Point cfg at provider p with key: base URL, the remembered model (else
+ * the preset's), settings saved. */
+void cst_ps_apply_provider(cst_agent_config_t *cfg, const cst_provider_t *p,
+                           const char *key);
+
+/* The whole switch without a prompt: key NULL/"" uses the saved or env
+ * key (else CST_PS_NEED_KEY); a given key is verified and saved first. A
+ * preset with no model takes the first one the API lists. The CST_PS_*
+ * flags, or CST_PS_UNKNOWN / CST_PS_NEED_KEY / CST_MODELS_AUTH. */
+int cst_ps_switch_provider(cst_agent_config_t *cfg, const char *name,
+                           const char *key, char *err, size_t err_sz);
+
+/* Replace the current provider's key: verified, saved, put in cfg. Same
+ * results as cst_ps_verify_save. */
+int cst_ps_set_key(cst_agent_config_t *cfg, const char *key, char *err,
+                   size_t err_sz);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif
@@ -45,6 +45,8 @@ char *cst_skill_read(const char *workspace, const char *name,
 * task. NULL when the skill does not exist. malloc'd. */
 char *cst_skill_prompt(const char *workspace, const char *name,
                       const char *task);
+/* User message for `/init`: have the agent write AGENTS.md. */
+const char *cst_init_prompt(void);
 /* Register the `skill` tool (idempotent). */
 int cst_skills_register_tool(void);

@@ -27,6 +27,17 @@ size_t cst_utf8_floor(const char *s, size_t n);
 char *cst_utf8_sanitize(const char *s, size_t n, size_t *out_len);
 int cst_utf8_valid(const char *s, size_t n);

+/* Re-chunks a byte stream so no piece ends inside a UTF-8 sequence: the
+ * incomplete tail (at most 3 bytes) waits in the hold for the next feed. */
+typedef struct {
+  char b[4];
+  size_t n;
+} cst_utf8_hold_t;
+/* Returns the complete prefix of hold+s[0..n) (malloc'd, NUL-terminated,
+ * *out_len may be 0), or NULL on OOM. */
+char *cst_utf8_hold_feed(cst_utf8_hold_t *h, const char *s, size_t n,
+                         size_t *out_len);
+
 /* Text files may start with a UTF-8 BOM (Windows editors): its length (0/3),
 * and an in-place strip of the NUL-terminated s (*len updated when given). */
 size_t cst_bom_len(const char *s, size_t n);
@@ -2,8 +2,8 @@
 #define CST_VERSION_H

 #define CST_VERSION_MAJOR 0
-#define CST_VERSION_MINOR 7
-#define CST_VERSION_PATCH 8
+#define CST_VERSION_MINOR 8
+#define CST_VERSION_PATCH 6

 #define CST_VERSION_STR_(x) #x
 #define CST_VERSION_STR(x) CST_VERSION_STR_(x)
@@ -17,6 +17,10 @@
 extern "C" {
 #endif

+/* What --web listens on when --web-port/CST_WEB_PORT says nothing (the
+ * README and --help quote this number, so keep them in sync). */
+#define CST_WEBUI_DEFAULT_PORT 8787
+
 typedef struct {
  const char *host;  /* NULL = 127.0.0.1; "0.0.0.0" = every interface */
  int port;          /* 0 = pick a free port */
@@ -114,9 +114,10 @@ int os_win_get_label(const char *path, char *out, size_t n);
 int os_clipboard_set(const char *utf8, size_t n);
 #endif

-/* Child processes get stdin=/dev/null, their own process group, and are
- * killed after the timeout (CST_SHELL_TIMEOUT seconds, default 120) or when
- * *flag becomes non-zero. Exit code 124 = timeout, 130 = cancelled. */
+/* Child processes get stdin=/dev/null and their own process group. They run
+ * with no time limit unless one is asked for (CST_SHELL_TIMEOUT seconds, or
+ * os_set_run_timeout()); either way a non-zero *flag cancels them.
+ * Exit code 124 = killed by that timeout, 130 = cancelled. */
 void os_set_cancel_flag(volatile int *flag);
 volatile int *os_get_cancel_flag(void);
 void os_set_run_timeout(int seconds); /* this thread; 0 = default */
@@ -1,8 +1,9 @@
 #!/usr/bin/env bash
 # Linux host: cross-build the Windows x86_64 release (mingw-w64, UCRT, static).
 #   scripts/bootstrap_windows.sh [repo-dir]
-# With Nix the toolchain comes from shell-win.nix; otherwise x86_64-w64-mingw32-gcc
-# (UCRT flavour, with winpthreads), cmake, ninja, git and perl must be installed.
+# With Nix the toolchain comes from shell-win.nix; otherwise $MINGW_HOST-gcc
+# (default x86_64-w64-mingw32; UCRT flavour, with winpthreads), cmake, ninja,
+# git and perl must be installed. CST_WIN_CFLAGS adds C flags (CI: -Werror).
 # Output: dist/windows-x86_64/ and dist/cst-agent-VERSION-windows-x86_64.zip
 set -euo pipefail

@@ -21,15 +22,17 @@ if [[ ! -f "$DIR/CMakeLists.txt" ]]; then
 fi
 cd "$DIR"

-if ! command -v x86_64-w64-mingw32-gcc >/dev/null && [[ -z "${IN_NIX_SHELL:-}" ]] &&
+H="${MINGW_HOST:-x86_64-w64-mingw32}"
+if ! command -v "$H-gcc" >/dev/null && [[ -z "${IN_NIX_SHELL:-}" ]] &&
  command -v nix-shell >/dev/null; then
  exec nix-shell shell-win.nix --run "$(printf '%q ' "$0" "$PWD")"
 fi

 need() { command -v "$1" >/dev/null || { echo "missing $1 (mingw-w64 UCRT gcc, cmake, ninja, git, perl)" >&2; exit 1; }; }
-need git; need cmake; need ninja; need perl; need x86_64-w64-mingw32-gcc
-if ! printf '#include <stdio.h>\n' | x86_64-w64-mingw32-gcc -dM -E - | grep -qw _UCRT; then
-  echo "x86_64-w64-mingw32-gcc targets msvcrt; the UTF-8 manifest needs a UCRT toolchain" >&2
+export MINGW_HOST="$H"
+need git; need cmake; need ninja; need perl; need "$H-gcc"
+if ! printf '#include <stdio.h>\n' | "$H-gcc" -dM -E - | grep -qw _UCRT; then
+  echo "$H-gcc targets msvcrt; the UTF-8 manifest needs a UCRT toolchain" >&2
  exit 1
 fi

@@ -40,7 +43,8 @@ fi
 echo "==> cmake (Windows, TUI)"
 cmake -S . -B build-win-tui -G Ninja \
  -DCMAKE_TOOLCHAIN_FILE=cmake/mingw-w64-x86_64.cmake \
-  -DCMAKE_BUILD_TYPE=Release -DCST_BUILD_TUI=ON
+  -DCMAKE_BUILD_TYPE=Release -DCST_BUILD_TUI=ON \
+  -DCMAKE_C_FLAGS="${CST_WIN_CFLAGS:-}"
 # Package only needs the agent binary; skip host-unrunnable Win32 tests.
 cmake --build build-win-tui --target cst-agent

@@ -0,0 +1,164 @@
+#!/usr/bin/env bash
+# Build pinned static deps for a Linux release into third_party/prefix-linux-ARCH:
+# mbedTLS, libcurl (HTTP(S) only, mbedTLS) and wide ncurses.
+#   scripts/build_third_party_linux.sh x86_64|x86|aarch64 [all|mbedtls|curl|ncurses]
+# Uses LINUX_HOST-gcc from PATH, else fetches it via ci_linux_toolchain.sh.
+# Host tools: cmake ninja make perl, and tic/infocmp with the fallback
+# terminal entries (Debian: ncurses-bin ncurses-term).
+set -euo pipefail
+ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+ARCH="${1:?usage: $0 x86_64|x86|aarch64 [all|mbedtls|curl|ncurses]}"
+TP="$ROOT/third_party"
+PREFIX="$TP/prefix-linux-$ARCH"
+SRC="$TP/src"
+WORK="$TP/build-linux-$ARCH"
+JOBS="${JOBS:-$(nproc 2>/dev/null || echo 4)}"
+
+MBEDTLS_VER=3.6.7
+MBEDTLS_SHA=a7e8bcbec0e6f761b4af24f25677626b35f762f68eef79c08677a363212d11f6
+CURL_VER=8.22.0
+CURL_SHA=f7ef3ae8a22e521f289803fe93543eb64c329b58aa73a9e224dfd915a2a5f4f7
+NCURSES_VER=6.6
+NCURSES_SHA=355b4cbbed880b0381a04c46617b7656e362585d52e9cf84a67e2009b749ff11
+# Compiled in, so the TUI starts even where no terminfo database is found.
+NCURSES_FALLBACKS=xterm-256color,xterm,screen-256color,tmux-256color,linux,vt100
+
+if [[ -z "${LINUX_HOST:-}" ]] || ! command -v "$LINUX_HOST-gcc" >/dev/null; then
+  read -r LINUX_HOST bindir < <(bash "$ROOT/scripts/ci_linux_toolchain.sh" "$ARCH")
+  export PATH="$bindir:$PATH"
+fi
+T="$LINUX_HOST-"
+need() { command -v "$1" >/dev/null || { echo "missing: $1" >&2; exit 1; }; }
+need "${T}gcc"
+need cmake
+need ninja
+need make
+need tic
+need infocmp
+mkdir -p "$PREFIX" "$SRC" "$WORK"
+
+fetch() { # file sha url...
+  local file="$1" sha="$2" u
+  shift 2
+  if [[ -f "$SRC/$file" ]] && echo "$sha  $SRC/$file" | sha256sum -c --quiet - 2>/dev/null; then
+    return 0
+  fi
+  for u in "$@"; do
+    echo "  try $u"
+    if curl -fsSL --retry 2 --connect-timeout 15 --max-time 600 -o "$SRC/$file.part" "$u" &&
+        echo "$sha  $SRC/$file.part" | sha256sum -c --quiet -; then
+      mv "$SRC/$file.part" "$SRC/$file"
+      return 0
+    fi
+  done
+  rm -f "$SRC/$file.part"
+  echo "download failed: $file" >&2
+  exit 1
+}
+
+unpack() { # file dir
+  rm -rf "${WORK:?}/$2"
+  tar -C "$WORK" -xf "$SRC/$1"
+}
+
+gh() { # path -> GitHub URL candidates (CN proxies first)
+  printf '%s\n' "https://ghfast.top/https://github.com/$1" \
+    "https://ghproxy.net/https://github.com/$1" "https://github.com/$1"
+}
+
+cmake_cross() { # src build args...
+  local s="$1" b="$2"
+  shift 2
+  cmake -S "$s" -B "$b" -G Ninja \
+    -DCMAKE_SYSTEM_NAME=Linux \
+    -DCMAKE_C_COMPILER="${T}gcc" -DCMAKE_AR="$(command -v "${T}ar")" \
+    -DCMAKE_RANLIB="$(command -v "${T}ranlib")" \
+    -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX="$PREFIX" \
+    -DCMAKE_INSTALL_LIBDIR=lib -DCMAKE_PREFIX_PATH="$PREFIX" \
+    -DCMAKE_FIND_ROOT_PATH="$PREFIX" \
+    -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+    -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+    -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=ONLY \
+    -DCMAKE_POSITION_INDEPENDENT_CODE=OFF \
+    "$@"
+  cmake --build "$b" -j"$JOBS"
+  cmake --install "$b"
+}
+
+build_mbedtls() {
+  local f="mbedtls-$MBEDTLS_VER.tar.bz2"
+  mapfile -t urls < <(gh "Mbed-TLS/mbedtls/releases/download/mbedtls-$MBEDTLS_VER/$f")
+  fetch "$f" "$MBEDTLS_SHA" "${urls[@]}"
+  unpack "$f" "mbedtls-$MBEDTLS_VER"
+  cmake_cross "$WORK/mbedtls-$MBEDTLS_VER" "$WORK/mbedtls-$MBEDTLS_VER/build" \
+    -DENABLE_TESTING=OFF -DENABLE_PROGRAMS=OFF -DGEN_FILES=OFF \
+    -DUSE_STATIC_MBEDTLS_LIBRARY=ON -DUSE_SHARED_MBEDTLS_LIBRARY=OFF \
+    -DMBEDTLS_FATAL_WARNINGS=OFF
+  echo "mbedTLS $MBEDTLS_VER -> $PREFIX"
+}
+
+build_curl() {
+  local f="curl-$CURL_VER.tar.xz"
+  mapfile -t urls < <(gh "curl/curl/releases/download/curl-${CURL_VER//./_}/$f")
+  fetch "$f" "$CURL_SHA" "https://curl.se/download/$f" "${urls[@]}"
+  unpack "$f" "curl-$CURL_VER"
+  # Cross builds skip CA detection; this is only libcurl's default, the
+  # program probes the usual distro paths when it is missing.
+  cmake_cross "$WORK/curl-$CURL_VER" "$WORK/curl-$CURL_VER/build" \
+    -DBUILD_SHARED_LIBS=OFF -DBUILD_STATIC_LIBS=ON \
+    -DCURL_USE_MBEDTLS=ON -DCURL_USE_OPENSSL=OFF \
+    -DMBEDTLS_INCLUDE_DIR="$PREFIX/include" \
+    -DMBEDTLS_LIBRARY="$PREFIX/lib/libmbedtls.a" \
+    -DMBEDX509_LIBRARY="$PREFIX/lib/libmbedx509.a" \
+    -DMBEDCRYPTO_LIBRARY="$PREFIX/lib/libmbedcrypto.a" \
+    -DCURL_CA_BUNDLE=/etc/ssl/certs/ca-certificates.crt -DCURL_CA_PATH=none \
+    -DHTTP_ONLY=ON -DCURL_DISABLE_LDAP=ON -DCURL_USE_LIBPSL=OFF \
+    -DCURL_USE_LIBSSH2=OFF -DCURL_USE_LIBSSH=OFF -DUSE_LIBIDN2=OFF \
+    -DUSE_NGHTTP2=OFF -DCURL_ZLIB=OFF -DCURL_BROTLI=OFF -DCURL_ZSTD=OFF \
+    -DCURL_USE_GSSAPI=OFF -DENABLE_THREADED_RESOLVER=ON \
+    -DBUILD_CURL_EXE=OFF -DBUILD_TESTING=OFF -DBUILD_EXAMPLES=OFF \
+    -DBUILD_LIBCURL_DOCS=OFF -DBUILD_MISC_DOCS=OFF -DENABLE_CURL_MANUAL=OFF
+  echo "curl $CURL_VER -> $PREFIX"
+}
+
+build_ncurses() {
+  local f="ncurses-$NCURSES_VER.tar.gz"
+  fetch "$f" "$NCURSES_SHA" \
+    "https://invisible-island.net/archives/ncurses/$f" \
+    "https://mirrors.tuna.tsinghua.edu.cn/gnu/ncurses/$f" \
+    "https://ftp.gnu.org/gnu/ncurses/$f"
+  unpack "$f" "ncurses-$NCURSES_VER"
+  local b="$WORK/ncurses-$NCURSES_VER/build"
+  mkdir -p "$b"
+  cd "$b"
+  CC="${T}gcc" AR="${T}ar" RANLIB="${T}ranlib" STRIP="${T}strip" \
+    ../configure --host="$LINUX_HOST" --prefix="$PREFIX" \
+    --without-ada --without-cxx --without-cxx-binding --without-manpages \
+    --without-progs --without-tests --without-debug --without-shared \
+    --with-normal --enable-widec --disable-db-install \
+    --with-default-terminfo-dir=/usr/share/terminfo \
+    --with-terminfo-dirs=/etc/terminfo:/lib/terminfo:/usr/share/terminfo:/usr/lib/terminfo \
+    --with-fallbacks="$NCURSES_FALLBACKS" \
+    --with-tic-path="$(command -v tic)" --with-infocmp-path="$(command -v infocmp)" \
+    --enable-pc-files --with-pkg-config-libdir="$PREFIX/lib/pkgconfig" \
+    --with-build-cc="${BUILD_CC:-cc}" >/dev/null
+  make -j"$JOBS" libs >/dev/null
+  make install.libs >/dev/null
+  echo "ncurses $NCURSES_VER -> $PREFIX"
+}
+
+case "${2:-all}" in
+mbedtls) build_mbedtls ;;
+curl) build_curl ;;
+ncurses) build_ncurses ;;
+all)
+  build_mbedtls
+  build_curl
+  build_ncurses
+  ;;
+*)
+  echo "usage: $0 ARCH [all|mbedtls|curl|ncurses]" >&2
+  exit 1
+  ;;
+esac
+echo "DONE $ARCH"
@@ -106,7 +106,10 @@ build_ncurses() {
  rm -rf "$src/build-mingw"
  mkdir -p "$src/build-mingw"
  cd "$src/build-mingw"
-  ../configure --host="${HOST:-x86_64-w64-mingw32}" --prefix="$PREFIX" \
+  # --host stays canonical (config.sub may not know Debian's "mingw32ucrt");
+  # the tools themselves come from $HOST.
+  CC="${T}gcc" AR="${T}ar" RANLIB="${T}ranlib" STRIP="${T}strip" WINDRES="${T}windres" \
+    ../configure --host=x86_64-w64-mingw32 --prefix="$PREFIX" \
    --without-ada --without-cxx --without-cxx-binding --without-manpages \
    --without-progs --without-tests --without-debug --without-shared \
    --enable-widec --enable-sp-funcs --enable-term-driver \
@@ -0,0 +1,87 @@
+#!/usr/bin/env bash
+# Fetch a musl cross gcc for a static Linux release build.
+#   scripts/ci_linux_toolchain.sh x86_64|x86|aarch64
+# Prints "HOST BINDIR" on stdout (e.g. "x86_64-buildroot-linux-musl
+# /opt/x86-64--musl--stable-2025.08-1/bin"); progress goes to stderr.
+# Bootlin's pinned stable release first, musl.cc as a fallback.
+set -euo pipefail
+
+ARCH="${1:?usage: $0 x86_64|x86|aarch64}"
+BOOTLIN_VER="${CST_BOOTLIN_VER:-2025.08-1}"
+case "$ARCH" in
+x86_64)
+  bl=x86-64 bl_host=x86_64-buildroot-linux-musl mc=x86_64-linux-musl
+  sha=09fca3aa89540f1b01b5f4210d488cbeb00f522044c53e9989b1dd8a38076912
+  ;;
+x86)
+  bl=x86-i686 bl_host=i686-buildroot-linux-musl mc=i686-linux-musl
+  sha=3a340ebc386a057f5ddb24d8f4bbde07d2c013c99cebc02ac38565e2849bfda0
+  ;;
+aarch64)
+  bl=aarch64 bl_host=aarch64-buildroot-linux-musl mc=aarch64-linux-musl
+  sha=defba831ffa1175236f137069333e21ed46d4d19feb5080a90cf248b6fc2cb08
+  ;;
+*)
+  echo "unknown arch: $ARCH" >&2
+  exit 1
+  ;;
+esac
+[[ "$BOOTLIN_VER" == 2025.08-1 ]] || sha="" # other releases: no pinned hash
+
+if [[ -n "${CST_TOOLCHAIN_ROOT:-}" ]]; then
+  ROOT_DIR="$CST_TOOLCHAIN_ROOT"
+elif [[ -w /opt || "$(id -u)" -eq 0 ]]; then
+  ROOT_DIR=/opt
+else
+  ROOT_DIR="$HOME/.local/toolchains"
+fi
+mkdir -p "$ROOT_DIR"
+
+fetch() { # out url...
+  local out="$1" u
+  shift
+  for u in "$@"; do
+    echo "  try $u" >&2
+    if curl -fsSL --retry 2 --connect-timeout 15 --max-time 900 -o "$out" "$u"; then
+      return 0
+    fi
+  done
+  return 1
+}
+
+emit() { # host bindir
+  "$2/$1-gcc" -dumpmachine >&2
+  [[ -n "${GITHUB_PATH:-}" ]] && echo "$2" >>"$GITHUB_PATH"
+  echo "$1 $2"
+}
+
+name="$bl--musl--stable-$BOOTLIN_VER"
+dir="$ROOT_DIR/$name"
+if [[ -x "$dir/bin/$bl_host-gcc" ]]; then
+  emit "$bl_host" "$dir/bin"
+  exit 0
+fi
+
+echo "==> toolchain $name" >&2
+tmp="$ROOT_DIR/$name.tar.xz"
+if fetch "$tmp" "https://toolchains.bootlin.com/downloads/releases/toolchains/$bl/tarballs/$name.tar.xz" &&
+    { [[ -z "$sha" ]] || echo "$sha  $tmp" | sha256sum -c --quiet - >&2; }; then
+  tar -C "$ROOT_DIR" -xJf "$tmp"
+  rm -f "$tmp"
+  # Bootlin toolchains are relocatable but must be told once where they are.
+  [[ -x "$dir/relocate-sdk.sh" ]] && "$dir/relocate-sdk.sh" >&2
+  emit "$bl_host" "$dir/bin"
+  exit 0
+fi
+rm -f "$tmp"
+
+echo "==> bootlin unavailable, falling back to musl.cc" >&2
+mdir="$ROOT_DIR/$mc-cross"
+if [[ ! -x "$mdir/bin/$mc-gcc" ]]; then
+  tgz="$ROOT_DIR/$mc-cross.tgz"
+  fetch "$tgz" "https://musl.cc/$mc-cross.tgz" "https://more.musl.cc/x86_64-linux-musl/$mc-cross.tgz" ||
+    { echo "no musl toolchain for $ARCH" >&2; exit 1; }
+  tar -C "$ROOT_DIR" -xzf "$tgz"
+  rm -f "$tgz"
+fi
+emit "$mc" "$mdir/bin"
@@ -0,0 +1,47 @@
+#!/usr/bin/env bash
+# CI helper: build the pinned TinyCC used for the verification job.
+# Distro tcc packages are too old (no <stdatomic.h>), so this builds the same
+# mob snapshot nixpkgs ships. Installs to $CST_TCC_PREFIX (default ~/.local/tcc)
+# and, under Actions, puts its bin/ on $GITHUB_PATH.
+set -euo pipefail
+
+COMMIT="${CST_TCC_COMMIT:-f6385c05308f715bdd2c06336801193a21d69b50}"
+PREFIX="${CST_TCC_PREFIX:-$HOME/.local/tcc}"
+WORK="${TMPDIR:-/tmp}/tcc-build"
+
+if [[ -x "$PREFIX/bin/tcc" ]]; then
+  echo "==> tcc already at $PREFIX"
+else
+  rm -rf "$WORK"
+  mkdir -p "$WORK"
+  TGZ="$WORK/tinycc.tar.gz"
+  # repo.or.cz is upstream but often unreachable from CN runners; the GitHub
+  # mirror carries the same commits.
+  urls=(
+    "https://ghfast.top/https://github.com/TinyCC/tinycc/archive/${COMMIT}.tar.gz"
+    "https://ghproxy.net/https://github.com/TinyCC/tinycc/archive/${COMMIT}.tar.gz"
+    "https://github.com/TinyCC/tinycc/archive/${COMMIT}.tar.gz"
+    "https://repo.or.cz/tinycc.git/snapshot/${COMMIT}.tar.gz"
+  )
+  ok=0
+  for u in "${urls[@]}"; do
+    echo "  try $u"
+    if curl -fsSL --retry 2 --connect-timeout 15 --max-time 300 -o "$TGZ" "$u" &&
+        tar -tzf "$TGZ" 2>/dev/null | sed -n 1p | grep -q "${COMMIT:0:7}"; then
+      ok=1
+      break
+    fi
+  done
+  [[ $ok -eq 1 ]] || { echo "failed to download tinycc $COMMIT" >&2; exit 1; }
+  tar -C "$WORK" -xzf "$TGZ"
+  src=$(find "$WORK" -mindepth 1 -maxdepth 1 -type d -name 'tinycc*' | head -1)
+  cd "$src"
+  ./configure --prefix="$PREFIX" --cc="${BUILD_CC:-cc}"
+  make -j"$(nproc)"
+  make install
+fi
+
+"$PREFIX/bin/tcc" -v
+if [[ -n "${GITHUB_PATH:-}" ]]; then
+  echo "$PREFIX/bin" >>"$GITHUB_PATH"
+fi
@@ -1,64 +1,26 @@
 #!/usr/bin/env bash
-# CI helper: MinGW-w64 UCRT toolchain via llvm-mingw (no Nix / no /nix).
-# Downloads a pinned release, puts it on PATH, then runs bootstrap_windows.sh.
+# CI helper: Windows x86_64 release with Debian's mingw-w64 UCRT gcc.
+#   apt-get install gcc-mingw-w64-ucrt64 binutils-mingw-w64-ucrt64 nsis   (trixie+)
+# Checks the toolchain, then runs bootstrap_windows.sh, which builds the
+# third-party libs, cst-agent.exe, the zip and the NSIS installer.
 set -euo pipefail
 ROOT="$(cd "$(dirname "$0")/.." && pwd)"
 cd "$ROOT"

-VER="${CST_LLVM_MINGW_VER:-20241030}"
-NAME="llvm-mingw-${VER}-ucrt-ubuntu-20.04-x86_64"
-TGZ="/tmp/${NAME}.tar.xz"
-# Prefer /opt when writable (root CI); else unpack under $HOME.
-if [[ -w /opt || "$(id -u)" -eq 0 ]]; then
-  DIR="/opt/${NAME}"
-  PARENT=/opt
-else
-  DIR="${HOME}/${NAME}"
-  PARENT="$HOME"
-fi
-
-if [[ ! -x "$DIR/bin/x86_64-w64-mingw32-gcc" ]]; then
-  echo "==> download $NAME"
-  # Prefer CN-reachable GitHub proxies; raw github.com is often unreachable.
-  urls=(
-    "https://ghfast.top/https://github.com/mstorsjo/llvm-mingw/releases/download/${VER}/${NAME}.tar.xz"
-    "https://ghproxy.net/https://github.com/mstorsjo/llvm-mingw/releases/download/${VER}/${NAME}.tar.xz"
-    "https://gitclone.com/github.com/mstorsjo/llvm-mingw/releases/download/${VER}/${NAME}.tar.xz"
-    "https://github.com/mstorsjo/llvm-mingw/releases/download/${VER}/${NAME}.tar.xz"
-  )
-  ok=0
-  for u in "${urls[@]}"; do
-    echo "  try $u"
-    if curl -fL --retry 2 --connect-timeout 15 --max-time 600 \
-        --retry-delay 2 -o "$TGZ" "$u"; then
-      # Reject tiny error pages that some proxies return with HTTP 200.
-      sz=$(wc -c <"$TGZ")
-      if [[ "$sz" -gt 10000000 ]]; then
-        ok=1
-        break
-      fi
-      echo "  too small ($sz bytes), next mirror"
-    fi
-  done
-  [[ $ok -eq 1 ]] || { echo "failed to download llvm-mingw" >&2; exit 1; }
-  rm -rf "$DIR"
-  mkdir -p "$PARENT"
-  tar -C "$PARENT" -xf "$TGZ"
-fi
-
-export PATH="$DIR/bin:$PATH"
-export MINGW_HOST=x86_64-w64-mingw32
+export MINGW_HOST="${MINGW_HOST:-x86_64-w64-mingw32ucrt}"
 export BUILD_CC="${BUILD_CC:-cc}"
+export CST_REQUIRE_NSIS=1

 need() { command -v "$1" >/dev/null || { echo "missing $1" >&2; exit 1; }; }
-need x86_64-w64-mingw32-gcc
-need x86_64-w64-mingw32-windres
-need cmake; need ninja; need perl; need zip; need git; need make
+need "$MINGW_HOST-gcc"
+need "$MINGW_HOST-windres"
+need cmake; need ninja; need perl; need zip; need git; need make; need makensis

-if ! printf '#include <stdio.h>\n' | x86_64-w64-mingw32-gcc -dM -E - | grep -qw _UCRT; then
-  echo "toolchain is not UCRT" >&2
+if ! printf '#include <stdio.h>\n' | "$MINGW_HOST-gcc" -dM -E - | grep -qw _UCRT; then
+  echo "$MINGW_HOST-gcc is not UCRT" >&2
  exit 1
 fi
+"$MINGW_HOST-gcc" --version | head -1

 # Do not set a global insteadOf: a single broken proxy (e.g. gitclone 502)
 # would take down every third_party clone. build_third_party_mingw.sh tries
@@ -47,6 +47,19 @@ BASE="http://127.0.0.1:$PORT"
 echo "   port $PORT"

 fail() { echo "FAIL: $1"; echo "--- events"; cat "$EVENTS" 2>/dev/null | head -40; exit 1; }
+# SIGTERM must shut the server down cleanly, even mid-turn (128+ = a crash)
+stop_server() {
+  kill "$1" 2>/dev/null || return 0
+  n=0
+  while kill -0 "$1" 2>/dev/null && [ $n -lt 200 ]; do n=$((n + 1)); sleep 0.2; done
+  if kill -0 "$1" 2>/dev/null; then
+    kill -9 "$1" 2>/dev/null || true
+    fail "$2 hung on SIGTERM"
+  fi
+  st=0
+  wait "$1" 2>/dev/null || st=$?
+  [ "$st" -lt 128 ] || fail "$2 died with status $st on SIGTERM"
+}

 # 1. the shell page and the state endpoint
 curl -fsS --max-time 5 "$BASE/" | grep -q "<title>CST-Agent" || fail "GET / is not the shell page"
@@ -160,7 +173,7 @@ fi
 # 7. the interactive approval path: a fresh server in ask mode must raise a
 # permission event, and answering it must let the turn continue.
 echo "== permission round trip (-A ask)"
-kill "$SRV_PID" 2>/dev/null || true   # the first server's replay is spent
+stop_server "$SRV_PID" "the first server"   # its replay is spent
 rm -rf "$WORK/pw"; mkdir -p "$WORK/pw"
 P_EVENTS="$WORK/permission-events.txt"
 P_LOG="$WORK/permission-server.log"   # its own log: the port parse must not
@@ -206,7 +219,7 @@ else
    i=$((i + 1)); sleep 0.2
  done
  grep -q "event: permission_done" "$P_EVENTS" || fail "the answer was not acknowledged"
-  kill "$P_PID" 2>/dev/null || true
+  stop_server "$P_PID" "the ask-mode server"
  kill "$P_CURL" 2>/dev/null || true
  echo "   answered request $PID_REQ with 'allow once'"
 fi
@@ -260,4 +273,27 @@ done
 kill -0 "$T_PID" 2> /dev/null && fail "the server ignored /api/quit"
 wait "$T_PID" 2>/dev/null || true

+# 9. ask_user waits for the answer, and the HTTP hardening checks (both
+# need a scripted model: python3)
+if command -v python3 > /dev/null; then
+  echo "== ask_user round trip"
+  CST_AGENT="$BIN" python3 "$ROOT/scripts/e2e_webui_ask.py" > "$WORK/ask.txt" 2>&1 ||
+    { cat "$WORK/ask.txt"; fail "the ask_user round trip"; }
+  echo "== HTTP hardening"
+  CST_AGENT="$BIN" python3 "$ROOT/scripts/e2e_webui_http.py" > "$WORK/http.txt" 2>&1 ||
+    { cat "$WORK/http.txt"; fail "the HTTP hardening checks"; }
+  echo "== settings and streaming"
+  CST_AGENT="$BIN" python3 "$ROOT/scripts/e2e_webui_settings.py" > "$WORK/settings.txt" 2>&1 ||
+    { cat "$WORK/settings.txt"; fail "the settings checks"; }
+  echo "== sessions, retry and export"
+  CST_AGENT="$BIN" python3 "$ROOT/scripts/e2e_webui_sessions.py" > "$WORK/sessions.txt" 2>&1 ||
+    { cat "$WORK/sessions.txt"; fail "the session checks"; }
+  echo "== diff, run settings and permission rules"
+  CST_AGENT="$BIN" python3 "$ROOT/scripts/e2e_webui_workspace.py" > "$WORK/workspace.txt" 2>&1 ||
+    { cat "$WORK/workspace.txt"; fail "the workspace checks"; }
+  echo "== /init, /skill, extensions and the thinking switch"
+  CST_AGENT="$BIN" python3 "$ROOT/scripts/e2e_webui_ext.py" > "$WORK/ext.txt" 2>&1 ||
+    { cat "$WORK/ext.txt"; fail "the extension checks"; }
+fi
+
 echo "== OK: (events: $(grep -c '^event: ' "$EVENTS"))"
@@ -0,0 +1,164 @@
+"""WebUI ask_user round trip against a fake OpenAI-compatible API.
+
+  python3 scripts/e2e_webui_ask.py      (CST_AGENT= picks the binary)
+
+Checks: the turn really waits for the answer, a page opened while the
+question is open still receives it, and Last-Event-ID resumes the stream.
+"""
+import json, os, shutil, subprocess, sys, tempfile, threading, time, urllib.request
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+
+BIN = os.environ.get("CST_AGENT", os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "cst-agent"))
+REQS = []
+
+
+class H(BaseHTTPRequestHandler):
+    def log_message(self, *a):
+        pass
+
+    def do_GET(self):
+        body = json.dumps({"data": [{"id": "deepseek-flash"}]}).encode()
+        self.send_response(200)
+        self.send_header("Content-Type", "application/json")
+        self.send_header("Content-Length", str(len(body)))
+        self.end_headers()
+        self.wfile.write(body)
+
+    def do_POST(self):
+        n = int(self.headers.get("Content-Length", 0))
+        REQS.append(json.loads(self.rfile.read(n)))
+        self.send_response(200)
+        self.send_header("Content-Type", "text/event-stream")
+        self.end_headers()
+        base = {"id": "x", "object": "chat.completion.chunk", "model": "deepseek-flash"}
+
+        def ev(o):
+            self.wfile.write(b"data: " + json.dumps({**base, **o}).encode() + b"\n\n")
+            self.wfile.flush()
+        if len(REQS) == 1:
+            args = json.dumps({"question": "Pick a colour?", "options": "red\nblue"})
+            ev({"choices": [{"index": 0, "delta": {"role": "assistant", "tool_calls": [
+                {"index": 0, "id": "call_1", "type": "function",
+                 "function": {"name": "ask_user", "arguments": args}}]}}]})
+            ev({"choices": [{"index": 0, "delta": {}, "finish_reason": "tool_calls"}]})
+        else:
+            ev({"choices": [{"index": 0, "delta": {"role": "assistant", "content": "done"}}]})
+            ev({"choices": [{"index": 0, "delta": {}, "finish_reason": "stop"}]})
+        self.wfile.write(b"data: [DONE]\n\n")
+        self.wfile.flush()
+
+
+fails = 0
+
+
+def check(ok, msg):
+    global fails
+    print(("PASS " if ok else "FAIL ") + msg)
+    fails += 0 if ok else 1
+
+
+class Sse:
+    def __init__(self, base, last_id=None):
+        self.lines = []
+        req = urllib.request.Request(base + "/api/events")
+        if last_id is not None:
+            req.add_header("Last-Event-ID", str(last_id))
+        self.r = urllib.request.urlopen(req, timeout=60)
+        threading.Thread(target=self.pump, daemon=True).start()
+
+    def pump(self):
+        try:
+            for raw in self.r:
+                self.lines.append(raw.decode())
+        except Exception:
+            pass
+
+    def events(self, name):
+        out = []
+        for i, l in enumerate(self.lines):
+            if l.strip() == "event: " + name and i + 1 < len(self.lines):
+                out.append(json.loads(self.lines[i + 1][6:]))
+        return out
+
+    def ids(self):
+        return [int(l[4:]) for l in self.lines if l.startswith("id: ")]
+
+    def close(self):
+        try:
+            self.r.close()
+        except Exception:
+            pass
+
+
+def wait(fn, t=15):
+    end = time.time() + t
+    while time.time() < end:
+        v = fn()
+        if v:
+            return v
+        time.sleep(0.1)
+    return None
+
+
+def post(base, path, obj):
+    req = urllib.request.Request(base + path, data=json.dumps(obj).encode(),
+                                 headers={"Content-Type": "application/json"})
+    try:
+        with urllib.request.urlopen(req, timeout=5) as r:
+            return r.status
+    except urllib.error.HTTPError as e:
+        return e.code
+
+
+api = ThreadingHTTPServer(("127.0.0.1", 0), H)
+threading.Thread(target=api.serve_forever, daemon=True).start()
+ws = tempfile.mkdtemp(prefix="cst-ask-")
+log = open(ws + "/server.log", "w+")
+p = subprocess.Popen([BIN, "--web", "--web-port", "0", "--workspace", ws, "--new",
+                      "--api-key", "sk-x", "--provider", "deepseek", "--model", "deepseek-flash",
+                      "--api-base", "http://127.0.0.1:%d/v1" % api.server_address[1],
+                      "--protocol", "native-eos", "--approval", "auto"],
+                     stdout=log, stderr=subprocess.STDOUT)
+try:
+    def port():
+        log.seek(0)
+        for l in log.read().splitlines():
+            if l.startswith("webui: http://127.0.0.1:"):
+                return l.split(":")[3].split("/")[0]
+    pt = wait(port)
+    base = "http://127.0.0.1:%s" % pt
+    a = Sse(base)
+    time.sleep(0.3)
+    check(post(base, "/api/turn", {"text": "ask me"}) == 202, "turn accepted")
+    ask = wait(lambda: a.events("ask_user"))
+    check(bool(ask), "ask_user event raised")
+    time.sleep(1.5)
+    check(len(REQS) == 1, "the turn waits for the answer (requests so far: %d)" % len(REQS))
+    # a page opened now must still see the open question
+    b = Sse(base)
+    check(bool(wait(lambda: b.events("ask_user"), 5)), "a late page gets the open question")
+    # reconnect with Last-Event-ID: nothing before that id is resent
+    last = a.ids()[-1]
+    c = Sse(base, last_id=last - 1)
+    got = wait(lambda: c.ids(), 5)
+    check(bool(got) and got[0] == last, "Last-Event-ID resumes after that id (%s, want %d)" % (got[:1] if got else got, last))
+    if ask:
+        check(post(base, "/api/ask_user", {"id": ask[0]["id"], "text": "blue"}) == 200, "answer accepted")
+    end = wait(lambda: a.events("turn_end"), 20)
+    check(bool(end) and end[0].get("rc") == 0, "turn ended rc=0")
+    tool_msgs = [m for m in (REQS[1]["messages"] if len(REQS) > 1 else []) if m.get("role") == "tool"]
+    check(bool(tool_msgs) and "blue" in json.dumps(tool_msgs), "the answer reached the model: %s" % json.dumps(tool_msgs)[:200])
+    for s in (a, b, c):
+        s.close()
+finally:
+    p.terminate()
+    try:
+        p.wait(20)
+    except subprocess.TimeoutExpired:
+        p.kill()
+        check(False, "server hung on SIGTERM")
+    check(p.returncode is not None and p.returncode < 128 and p.returncode >= 0, "clean exit (%s)" % p.returncode)
+    log.close()
+    shutil.rmtree(ws, ignore_errors=True)
+print("FAILED:", fails)
+sys.exit(1 if fails else 0)
@@ -0,0 +1,227 @@
+"""WebUI /init, /skill, extensions and the thinking switch against a fake API.
+
+  python3 scripts/e2e_webui_ext.py      (CST_AGENT= picks the binary)
+
+Checks: show_thinking is validated, lands in the state and in settings.json;
+/api/skill inlines the skill and 404s an unknown one; /api/init sends the
+AGENTS.md prompt; /api/extensions lists skills and the built-in roles;
+trust and reload work and write settings.json; the changing calls are a 409
+while a turn runs, the listing is not. HOME points at a temp dir.
+"""
+import json, os, shutil, subprocess, sys, tempfile, threading, time, urllib.request
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+
+BIN = os.environ.get("CST_AGENT", os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "cst-agent"))
+SEEN = []   # the full last user message of every request
+
+
+class Api(BaseHTTPRequestHandler):
+    """Answers "好的"; a prompt with "slow" takes 2 s."""
+    def log_message(self, *a):
+        pass
+
+    def do_GET(self):
+        body = json.dumps({"data": [{"id": "deepseek-flash"}]}).encode()
+        self.send_response(200)
+        self.send_header("Content-Type", "application/json")
+        self.send_header("Content-Length", str(len(body)))
+        self.end_headers()
+        self.wfile.write(body)
+
+    def do_POST(self):
+        req = json.loads(self.rfile.read(int(self.headers.get("Content-Length", 0))) or b"{}")
+        users = [m.get("content") for m in req.get("messages", []) if m.get("role") == "user"]
+        last = users[-1] if users and isinstance(users[-1], str) else ""
+        SEEN.append(last)
+        if "slow" in last:
+            time.sleep(2)
+        self.send_response(200)
+        self.send_header("Content-Type", "text/event-stream")
+        self.end_headers()
+        for o in ({"delta": {"role": "assistant", "content": "好的"}},
+                  {"delta": {}, "finish_reason": "stop"}):
+            chunk = {"id": "x", "object": "chat.completion.chunk", "model": "deepseek-flash",
+                     "choices": [dict(index=0, **o)]}
+            self.wfile.write(b"data: " + json.dumps(chunk, ensure_ascii=False).encode() + b"\n\n")
+            self.wfile.flush()
+        self.wfile.write(b"data: [DONE]\n\n")
+        self.wfile.flush()
+
+
+fails = 0
+
+
+def check(ok, msg):
+    global fails
+    print(("PASS " if ok else "FAIL ") + msg)
+    fails += 0 if ok else 1
+
+
+def wait(fn, t=15):
+    end = time.time() + t
+    while time.time() < end:
+        v = fn()
+        if v:
+            return v
+        time.sleep(0.1)
+    return None
+
+
+def call(base, path, obj=None, method=None):
+    data = None if obj is None else json.dumps(obj).encode()
+    req = urllib.request.Request(base + path, data=data, method=method or ("POST" if data else "GET"),
+                                 headers={"Content-Type": "application/json"})
+    try:
+        with urllib.request.urlopen(req, timeout=30) as r:
+            return r.status, json.loads(r.read() or b"{}")
+    except urllib.error.HTTPError as e:
+        try:
+            return e.code, json.loads(e.read() or b"{}")
+        except ValueError:
+            return e.code, {}
+
+
+def skill(ws, name, desc, body):
+    d = os.path.join(ws, ".agent", "skills", name)
+    os.makedirs(d, exist_ok=True)
+    with open(os.path.join(d, "SKILL.md"), "w", encoding="utf-8") as f:
+        f.write("---\nname: %s\ndescription: %s\n---\n\n%s\n" % (name, desc, body))
+
+
+api = ThreadingHTTPServer(("127.0.0.1", 0), Api)
+threading.Thread(target=api.serve_forever, daemon=True).start()
+ws = tempfile.mkdtemp(prefix="cst-ext-")
+home = tempfile.mkdtemp(prefix="cst-home-")
+skill(ws, "demo", "演示用的 skill", "DEMO-INSTRUCTIONS: 先说你好。")
+env = dict(os.environ, HOME=home, XDG_CONFIG_HOME=home + "/.config",
+           XDG_DATA_HOME=home + "/.local/share", XDG_CACHE_HOME=home + "/.cache")
+for k in ("DEEPSEEK_API_KEY", "OPENAI_API_KEY", "CST_API_KEY"):
+    env.pop(k, None)
+log = open(ws + "/server.log", "w+")
+proc = subprocess.Popen([BIN, "--web", "--web-port", "0", "--workspace", ws, "--new", "--api-key", "sk-x",
+                         "--provider", "deepseek", "--model", "deepseek-flash",
+                         "--api-base", "http://127.0.0.1:%d/v1" % api.server_address[1],
+                         "--protocol", "native-eos", "--approval", "auto"],
+                        stdout=log, stderr=subprocess.STDOUT, env=env)
+
+
+def port():
+    log.seek(0)
+    for l in log.read().splitlines():
+        if l.startswith("webui: http://127.0.0.1:"):
+            return l.split(":")[3].split("/")[0]
+
+
+def settings():
+    try:
+        with open(home + "/.config/cst-agent/settings.json", encoding="utf-8") as f:
+            return json.load(f)
+    except OSError:
+        return {}
+
+
+def busy(base):
+    return call(base, "/api/state")[1].get("status_busy")
+
+
+def turn(base, path, body):
+    code, d = call(base, path, body)
+    if code == 202:
+        time.sleep(0.3)
+        wait(lambda: not busy(base), 20)
+    return code, d
+
+
+try:
+    base = "http://127.0.0.1:%s" % wait(port)
+
+    # 1. the thinking switch
+    st = call(base, "/api/state")[1]
+    check(st.get("show_thinking") is False, "show_thinking starts off")
+    check(call(base, "/api/config", {"show_thinking": "yes"})[0] == 400, "show_thinking must be a bool")
+    code, _ = call(base, "/api/config", {"show_thinking": True})
+    check(code == 200 and call(base, "/api/state")[1].get("show_thinking") is True, "show_thinking on")
+    check(settings().get("show_thinking") is True, "show_thinking is saved in settings.json")
+    call(base, "/api/config", {"show_thinking": False})
+    check(settings() and not settings().get("show_thinking"), "and cleared again (%s)" % settings())
+
+    # 2. the extensions listing
+    code, ex = call(base, "/api/extensions")
+    check(code == 200 and ex.get("enabled") is True and ex.get("trusted") is False,
+          "extensions: enabled, not trusted (%s)" % code)
+    names = [s["name"] for s in ex.get("skills", [])]
+    check(names == ["demo"] and ex["skills"][0].get("desc") == "演示用的 skill"
+          and ex["skills"][0].get("source") == "workspace", "the workspace skill is listed (%s)" % ex.get("skills"))
+    roles = [r["name"] for r in ex.get("roles", [])]
+    check("explore" in roles and all(r.get("source") for r in ex["roles"]), "built-in roles are listed (%s)" % roles)
+    check(ex.get("mcp") == [] and ex.get("plugins") == [] and ex.get("runs") == [], "no MCP, plugins or runs")
+    check(ex.get("task_tool") is True and "1 skill" in ex.get("report", ""), "task tool on, report %r" % ex.get("report"))
+
+    # 3. /skill
+    code, d = turn(base, "/api/skill", {"name": "demo", "task": "做点事"})
+    check(code == 202 and "DEMO-INSTRUCTIONS" in d.get("text", "") and d["text"].endswith("做点事"),
+          "skill turn accepted (%s)" % code)
+    check(SEEN and "DEMO-INSTRUCTIONS" in SEEN[-1] and "做点事" in SEEN[-1], "the model saw the skill and the task")
+    code, d = turn(base, "/api/skill", {"name": "demo"})
+    check(code == 202 and SEEN[-1].endswith("Apply this skill now."), "a skill without a task")
+    check(call(base, "/api/skill", {"name": "nope"})[0] == 404, "unknown skill: 404")
+    check(call(base, "/api/skill", {"name": "../demo"})[0] == 404, "skill ../demo: 404")
+    check(call(base, "/api/skill", {})[0] == 400, "skill without a name: 400")
+    check(call(base, "/api/skill", {"name": "x" * 100})[0] == 400, "a long skill name: 400")
+    check(call(base, "/api/skill", {"name": "demo", "task": 5})[0] == 400, "a non-string task: 400")
+
+    # 4. /init
+    code, d = turn(base, "/api/init", {})
+    check(code == 202 and "AGENTS.md" in d.get("text", ""), "init accepted (%s)" % code)
+    check("AGENTS.md" in SEEN[-1], "the model saw the AGENTS.md prompt")
+
+    # 5. trust and reload
+    code, d = call(base, "/api/extensions/trust", {"on": True})
+    check(code == 200 and d.get("trusted") is True, "trust the workspace (%s %s)" % (code, d))
+    tw = settings().get("trusted_workspaces", [])
+    check(any(os.path.realpath(ws) == os.path.realpath(p) for p in tw), "settings.json lists it (%s)" % tw)
+    check(call(base, "/api/extensions")[1].get("trusted") is True, "the listing says trusted")
+    code, d = call(base, "/api/extensions/trust", {"on": False})
+    check(code == 200 and d.get("trusted") is False and not settings().get("trusted_workspaces"),
+          "untrust (%s)" % settings().get("trusted_workspaces"))
+    check(call(base, "/api/extensions/trust", {"on": "yes"})[0] == 400, "trust needs a bool")
+    skill(ws, "zeta", "第二个", "ZETA")
+    code, d = call(base, "/api/extensions/reload", {})
+    check(code == 200 and "2 skills" in d.get("report", ""), "reload sees the new skill (%s)" % d)
+    check([s["name"] for s in call(base, "/api/extensions")[1]["skills"]] == ["demo", "zeta"], "both listed")
+
+    # 6. while a turn runs
+    call(base, "/api/turn", {"text": "slow 慢一点"})
+    time.sleep(0.4)
+    check(call(base, "/api/extensions/reload", {})[0] == 409, "reload during a turn: 409")
+    check(call(base, "/api/extensions/trust", {"on": True})[0] == 409, "trust during a turn: 409")
+    check(call(base, "/api/skill", {"name": "demo"})[0] == 409, "skill during a turn: 409")
+    check(call(base, "/api/init", {})[0] == 409, "init during a turn: 409")
+    check(call(base, "/api/extensions")[0] == 200, "the listing works during a turn")
+    check(call(base, "/api/config", {"show_thinking": True})[0] == 200, "the thinking switch works during a turn")
+    wait(lambda: not busy(base), 20)
+    check(not settings().get("trusted_workspaces"), "the refused trust changed nothing")
+
+    # 7. another workspace brings its own skills
+    ws2 = tempfile.mkdtemp(prefix="cst-ext2-")
+    skill(ws2, "other", "另一个工作区的", "OTHER")
+    code, _ = call(base, "/api/workspace", {"path": ws2})
+    ex = call(base, "/api/extensions")[1]
+    check(code == 200 and [s["name"] for s in ex.get("skills", [])] == ["other"]
+          and ex.get("report") == "1 skill", "a workspace switch reloads the extensions (%s)" % ex.get("report"))
+    code, d = turn(base, "/api/skill", {"name": "other"})
+    check(code == 202 and "OTHER" in SEEN[-1], "and its skill runs")
+    shutil.rmtree(ws2, ignore_errors=True)
+finally:
+    proc.terminate()
+    try:
+        proc.wait(20)
+    except subprocess.TimeoutExpired:
+        proc.kill()
+        check(False, "server hung on SIGTERM")
+    log.close()
+    shutil.rmtree(ws, ignore_errors=True)
+    shutil.rmtree(home, ignore_errors=True)
+    api.shutdown()
+print("FAILED:", fails)
+sys.exit(1 if fails else 0)
@@ -0,0 +1,234 @@
+"""WebUI HTTP hardening checks over raw sockets, against a scripted model.
+
+  python3 scripts/e2e_webui_http.py      (CST_AGENT= picks the binary)
+
+Cross-site requests, request parsing (Content-Length, NUL, long lines,
+keep-alive), auth before the body, changes refused while a turn runs,
+cancel waking a pending question, token mode and shutdown.
+"""
+import json, os, re, shutil, socket, subprocess, sys, tempfile, threading, time
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+
+BIN = os.environ.get("CST_AGENT", os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "cst-agent"))
+REQS = []
+
+
+class Api(BaseHTTPRequestHandler):
+    """First call: ask_user. Later calls: a plain answer."""
+
+    def log_message(self, *a):
+        pass
+
+    def do_GET(self):
+        body = json.dumps({"data": [{"id": "deepseek-flash"}]}).encode()
+        self.send_response(200)
+        self.send_header("Content-Type", "application/json")
+        self.send_header("Content-Length", str(len(body)))
+        self.end_headers()
+        self.wfile.write(body)
+
+    def do_POST(self):
+        n = int(self.headers.get("Content-Length", 0))
+        REQS.append(json.loads(self.rfile.read(n)))
+        self.send_response(200)
+        self.send_header("Content-Type", "text/event-stream")
+        self.end_headers()
+        base = {"id": "x", "object": "chat.completion.chunk", "model": "deepseek-flash"}
+
+        def ev(o):
+            self.wfile.write(b"data: " + json.dumps({**base, **o}).encode() + b"\n\n")
+            self.wfile.flush()
+        if len(REQS) == 1:
+            args = json.dumps({"question": "Pick?", "options": "a\nb"})
+            ev({"choices": [{"index": 0, "delta": {"role": "assistant", "tool_calls": [
+                {"index": 0, "id": "call_1", "type": "function",
+                 "function": {"name": "ask_user", "arguments": args}}]}}]})
+            ev({"choices": [{"index": 0, "delta": {}, "finish_reason": "tool_calls"}]})
+        else:
+            ev({"choices": [{"index": 0, "delta": {"role": "assistant", "content": "ok"}}]})
+            ev({"choices": [{"index": 0, "delta": {}, "finish_reason": "stop"}]})
+        self.wfile.write(b"data: [DONE]\n\n")
+        self.wfile.flush()
+
+
+fails = 0
+
+
+def check(ok, msg):
+    global fails
+    print(("PASS " if ok else "FAIL ") + msg)
+    fails += 0 if ok else 1
+
+
+def raw(port, data, timeout=5, read_all=True):
+    """Send bytes, return everything read until EOF/timeout."""
+    s = socket.create_connection(("127.0.0.1", port), timeout=timeout)
+    s.sendall(data)
+    out = b""
+    try:
+        while True:
+            b = s.recv(65536)
+            if not b:
+                break
+            out += b
+            if not read_all:
+                break
+    except socket.timeout:
+        pass
+    s.close()
+    return out.decode("latin1")
+
+
+def status(resp):
+    # a pipelined response starts right after the previous body, not on a new line
+    return [int(m) for m in re.findall(r"HTTP/1\.1 (\d{3}) ", resp)]
+
+
+def req(method, path, headers=None, body=b"", host=None):
+    h = {"Host": host or "127.0.0.1:%d" % PORT, "Connection": "close"}
+    h.update(headers or {})
+    if body or method == "POST":
+        h["Content-Length"] = str(len(body))
+    head = "%s %s HTTP/1.1\r\n" % (method, path) + "".join("%s: %s\r\n" % kv for kv in h.items()) + "\r\n"
+    return head.encode() + body
+
+
+def start(ws, *extra):
+    log = open(os.path.join(ws, "server.log"), "w+")
+    p = subprocess.Popen([BIN, "--web", "--web-port", "0", "--workspace", ws, "--new",
+                          "--api-key", "sk-x", "--provider", "deepseek", "--model", "deepseek-flash",
+                          "--api-base", "http://127.0.0.1:%d/v1" % api.server_address[1],
+                          "--protocol", "native-eos", "--approval", "auto", *extra],
+                         stdout=log, stderr=subprocess.STDOUT)
+    end = time.time() + 10
+    while time.time() < end:
+        log.seek(0)
+        for l in log.read().splitlines():
+            if l.startswith("webui: http://127.0.0.1:"):
+                return p, int(l.split(":")[3].split("/")[0]), log
+        time.sleep(0.1)
+    raise SystemExit("server did not start")
+
+
+def stop(p, name):
+    p.terminate()
+    try:
+        p.wait(20)
+    except subprocess.TimeoutExpired:
+        p.kill()
+        check(False, name + " hung on SIGTERM")
+        return
+    check(0 <= p.returncode < 128, "%s exits cleanly (%s)" % (name, p.returncode))
+
+
+api = ThreadingHTTPServer(("127.0.0.1", 0), Api)
+threading.Thread(target=api.serve_forever, daemon=True).start()
+root = tempfile.mkdtemp(prefix="cst-http-")
+try:
+    ws = os.path.join(root, "a")
+    os.makedirs(ws)
+    p, PORT, log = start(ws)
+    me = "http://127.0.0.1:%d" % PORT
+    cfg = b'{"approval":"ask"}'
+    # 1. cross-site requests
+    r = raw(PORT, req("POST", "/api/config", {"Origin": "http://evil.example", "Content-Type": "text/plain"}, cfg))
+    check(status(r)[:1] == [401], "a cross-site POST is refused (%s)" % status(r))
+    r = raw(PORT, req("POST", "/api/config", {"Origin": "null"}, cfg))
+    check(status(r)[:1] == [401], "Origin: null is refused")
+    r = raw(PORT, req("POST", "/api/config", {"Origin": me, "Content-Type": "application/json"}, b'{"approval":"auto"}'))
+    check(status(r)[:1] == [200], "our own page may POST (%s)" % status(r))
+    r = raw(PORT, req("POST", "/api/config", {}, b'{"approval":"auto"}'))
+    check(status(r)[:1] == [200], "a client without Origin (curl) may POST")
+    r = raw(PORT, req("GET", "/api/attach/img-0000.png", host="evil.example"))
+    check(status(r)[:1] == [401], "session images need the Host check too (%s)" % status(r))
+    r = raw(PORT, req("GET", "/app.js"))
+    check(status(r)[:1] == [200], "assets are served")
+    # 2. request parsing
+    r = raw(PORT, req("GET", "/api/state%00x"))
+    check(status(r)[:1] == [400], "an encoded NUL in the path is a 400 (%s)" % status(r))
+    smuggle = b"GET /api/sessions HTTP/1.1\r\nHost: 127.0.0.1\r\n\r\n"
+    r = raw(PORT, ("POST /api/cancel HTTP/1.1\r\nHost: 127.0.0.1:%d\r\nContent-Length: abc\r\n\r\n" % PORT).encode() + smuggle)
+    check(status(r) == [400], "Content-Length: abc is a 400, nothing smuggled (%s)" % status(r))
+    r = raw(PORT, ("POST /api/cancel HTTP/1.1\r\nHost: 127.0.0.1:%d\r\nContent-Length: 0\r\nContent-Length: %d\r\n\r\n"
+                   % (PORT, len(smuggle))).encode() + smuggle)
+    check(status(r) == [400], "two different Content-Lengths are a 400 (%s)" % status(r))
+    r = raw(PORT, ("GET /%s HTTP/1.1\r\nHost: x\r\n\r\n" % ("a" * 9000)).encode())
+    check(status(r)[:1] == [400], "an over-long request line is a 400 (%s)" % status(r))
+    r = raw(PORT, req("GET", "/api/state?x=http://a/b"))
+    check(status(r)[:1] == [200], "'://' in the query leaves the path alone (%s)" % status(r))
+    r = raw(PORT, req("GET", "/api/state", {"Connection": "Close"}), timeout=3)
+    check("Connection: close" in r, "Connection: Close (any case) closes")
+    ka = ("GET /api/state HTTP/1.1\r\nHost: 127.0.0.1:%d\r\n\r\n" % PORT).encode()
+    big = ("POST /api/turn HTTP/1.1\r\nHost: 127.0.0.1:%d\r\nContent-Length: 5000000\r\n\r\n" % PORT).encode()
+    r = raw(PORT, ka + big)
+    check(status(r) == [200, 413], "a 413 after a keep-alive request still arrives (%s)" % status(r))
+    # 3. a pending question: changes are refused, approval may change, cancel wakes it
+    events = []
+
+    def sse():
+        s = socket.create_connection(("127.0.0.1", PORT))
+        s.sendall(("GET /api/events HTTP/1.1\r\nHost: 127.0.0.1:%d\r\n\r\n" % PORT).encode())
+        buf = b""
+        s.settimeout(60)
+        try:
+            while True:
+                b = s.recv(65536)
+                if not b:
+                    break
+                buf += b
+                events[:] = [l[7:] for l in buf.decode("latin1").split("\n") if l.startswith("event: ")]
+        except Exception:
+            pass
+    threading.Thread(target=sse, daemon=True).start()
+    time.sleep(0.4)
+    r = raw(PORT, req("POST", "/api/turn", {"Content-Type": "application/json"}, b'{"text":"go"}'))
+    check(status(r)[:1] == [202], "turn accepted")
+    end = time.time() + 15
+    while time.time() < end and "ask_user" not in events:
+        time.sleep(0.1)
+    check("ask_user" in events, "the question is open")
+    for path, body in (("/api/sessions/new", b""), ("/api/workspace", json.dumps({"path": root}).encode()),
+                       ("/api/compact", b""), ("/api/undo", b""), ("/api/config", b'{"model":"x"}')):
+        r = raw(PORT, req("POST", path, {"Content-Type": "application/json"}, body))
+        check(status(r)[:1] == [409], "%s is refused mid-turn (%s)" % (path, status(r)))
+    r = raw(PORT, req("POST", "/api/config", {"Content-Type": "application/json"}, b'{"approval":"yolo"}'))
+    check(status(r)[:1] == [200], "approval may change mid-turn")
+    r = raw(PORT, req("POST", "/api/cancel", {"Content-Type": "application/json"}, b"{}"))
+    end = time.time() + 10
+    while time.time() < end and "turn_end" not in events:
+        time.sleep(0.1)
+    check("turn_end" in events, "cancel ends a turn parked on a question")
+    r = raw(PORT, req("POST", "/api/sessions/new", {"Content-Type": "application/json"}))
+    check(status(r)[:1] == [200], "after the turn, a new session is fine (%s)" % status(r))
+    stop(p, "the server")
+
+    # 4. token mode: the page loads, the API and images do not without it
+    ws2 = os.path.join(root, "b")
+    os.makedirs(ws2)
+    p, PORT, log = start(ws2, "--web-token", "tok-x")
+    r = raw(PORT, req("GET", "/app.js"))
+    check(status(r)[:1] == [200], "token mode: app.js loads without the token (%s)" % status(r))
+    r = raw(PORT, req("GET", "/api/state"))
+    check(status(r)[:1] == [401], "token mode: the API needs it")
+    r = raw(PORT, req("GET", "/api/attach/img-0000.png"))
+    check(status(r)[:1] == [401], "token mode: images need it")
+    r = raw(PORT, req("GET", "/api/state", {"Cookie": "cst_token=tok-x"}))
+    check(status(r)[:1] == [200], "token mode: the cookie works")
+    t0 = time.time()
+    r = raw(PORT, ("POST /api/attach HTTP/1.1\r\nHost: 127.0.0.1:%d\r\nContent-Length: 33554431\r\n"
+                   "Expect: 100-continue\r\n\r\n" % PORT).encode(), timeout=5)
+    check(status(r) == [401] and time.time() - t0 < 3,
+          "no token, no body read: 401 at once (%s, %.1fs)" % (status(r), time.time() - t0))
+    stop(p, "the token server")
+
+    # 5. option checks
+    rc = subprocess.run([BIN, "--web", "--web-port", "70000", "--workspace", ws2],
+                        capture_output=True, timeout=10).returncode
+    check(rc != 0, "--web-port 70000 is refused")
+    rc = subprocess.run([BIN, "--web", "--web-token", "x" * 80, "--web-port", "0", "--workspace", ws2,
+                         "--api-key", "k"], capture_output=True, timeout=10).returncode
+    check(rc != 0, "an over-long --web-token is refused")
+finally:
+    shutil.rmtree(root, ignore_errors=True)
+print("FAILED:", fails)
+sys.exit(1 if fails else 0)
@@ -0,0 +1,327 @@
+"""WebUI session management against a fake OpenAI-compatible API.
+
+  python3 scripts/e2e_webui_sessions.py      (CST_AGENT= picks the binary)
+
+Checks: retry runs the last prompt again and edit-and-resend replaces it
+(the model sees the new text, the session keeps one copy); retry is a 409
+while a turn runs and when there is nothing to retry; export returns the
+transcript as Markdown; rename, clear and delete work on the store, and
+deleting the open session moves to a new one; history messages carry the
+note / tool_result flags. HOME points at a temp dir.
+"""
+import json, os, shutil, subprocess, sys, tempfile, threading, time, urllib.request
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+
+BIN = os.environ.get("CST_AGENT", os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "cst-agent"))
+SEEN = []
+
+
+def last_prompt(msgs):
+    """The prompts are one line; a note before one shares its user message."""
+    for m in reversed(msgs):
+        c = m.get("content")
+        if m.get("role") == "user" and isinstance(c, str):
+            return c.rstrip("\n").split("\n")[-1]
+    return ""
+
+
+class Api(BaseHTTPRequestHandler):
+    """Answers "回答：<last prompt>"; a prompt with "slow" takes 2 s."""
+    def log_message(self, *a):
+        pass
+
+    def do_GET(self):
+        body = json.dumps({"data": [{"id": "deepseek-flash"}]}).encode()
+        self.send_response(200)
+        self.send_header("Content-Type", "application/json")
+        self.send_header("Content-Length", str(len(body)))
+        self.end_headers()
+        self.wfile.write(body)
+
+    def do_POST(self):
+        req = json.loads(self.rfile.read(int(self.headers.get("Content-Length", 0))) or b"{}")
+        msgs = req.get("messages", [])
+        p = last_prompt(msgs)
+        tail = msgs[-1] if msgs else {}
+        tc = tail.get("content") if isinstance(tail.get("content"), str) else ""
+        answered = tail.get("role") == "tool" or tc.startswith("result ")
+        if not answered:
+            SEEN.append(p)
+        if "slow" in p:
+            time.sleep(2)
+        self.send_response(200)
+        self.send_header("Content-Type", "text/event-stream")
+        self.end_headers()
+        if answered:
+            chunks = ({"delta": {"role": "assistant", "content": "读完了"}},
+                      {"delta": {}, "finish_reason": "stop"})
+        elif p.startswith("native"):
+            args = json.dumps({"path": "hello.txt"})
+            chunks = ({"delta": {"role": "assistant", "tool_calls": [
+                          {"index": 0, "id": "c1", "type": "function",
+                           "function": {"name": "read_file", "arguments": args}}]}},
+                      {"delta": {}, "finish_reason": "tool_calls"})
+        elif p.startswith("ctl"):
+            chunks = ({"delta": {"role": "assistant",
+                                 "content": "我来读。\ncall t1 read_file\narg path 9\nhello.txtend\n"}},
+                      {"delta": {}, "finish_reason": "stop"})
+        else:
+            chunks = ({"delta": {"role": "assistant", "content": "回答：" + p}},
+                      {"delta": {}, "finish_reason": "stop"})
+        for o in chunks:
+            chunk = {"id": "x", "object": "chat.completion.chunk", "model": "deepseek-flash",
+                     "choices": [dict(index=0, **o)]}
+            self.wfile.write(b"data: " + json.dumps(chunk, ensure_ascii=False).encode() + b"\n\n")
+            self.wfile.flush()
+        self.wfile.write(b"data: [DONE]\n\n")
+        self.wfile.flush()
+
+
+fails = 0
+
+
+def check(ok, msg):
+    global fails
+    print(("PASS " if ok else "FAIL ") + msg)
+    fails += 0 if ok else 1
+
+
+def wait(fn, t=15):
+    end = time.time() + t
+    while time.time() < end:
+        v = fn()
+        if v:
+            return v
+        time.sleep(0.1)
+    return None
+
+
+def call(base, path, obj=None, method=None, raw_reply=False):
+    data = None if obj is None else json.dumps(obj).encode()
+    req = urllib.request.Request(base + path, data=data, method=method or ("POST" if data else "GET"),
+                                 headers={"Content-Type": "application/json"})
+    try:
+        with urllib.request.urlopen(req, timeout=20) as r:
+            raw = r.read()
+            if raw_reply:
+                return r.status, r.headers.get("Content-Type", ""), raw
+            return r.status, json.loads(raw or b"{}")
+    except urllib.error.HTTPError as e:
+        raw = e.read()
+        if raw_reply:
+            return e.code, e.headers.get("Content-Type", ""), raw
+        try:
+            return e.code, json.loads(raw or b"{}")
+        except ValueError:
+            return e.code, {}
+
+
+class Sse:
+    def __init__(self, base):
+        self.lines = []
+        self.r = urllib.request.urlopen(base + "/api/events", timeout=60)
+        threading.Thread(target=self.pump, daemon=True).start()
+
+    def pump(self):
+        try:
+            for raw in self.r:
+                self.lines.append(raw)
+        except Exception:
+            pass
+
+    def count(self, name):
+        return sum(1 for l in self.lines if l.strip() == b"event: " + name.encode())
+
+    def close(self):
+        try:
+            self.r.close()
+        except Exception:
+            pass
+
+
+api = ThreadingHTTPServer(("127.0.0.1", 0), Api)
+threading.Thread(target=api.serve_forever, daemon=True).start()
+ws = tempfile.mkdtemp(prefix="cst-sess-")
+home = tempfile.mkdtemp(prefix="cst-home-")
+env = dict(os.environ, HOME=home, XDG_CONFIG_HOME=home + "/.config",
+           XDG_DATA_HOME=home + "/.local/share", XDG_CACHE_HOME=home + "/.cache")
+for k in ("DEEPSEEK_API_KEY", "OPENAI_API_KEY", "CST_API_KEY"):
+    env.pop(k, None)
+log = open(ws + "/server.log", "w+")
+proc = subprocess.Popen([BIN, "--web", "--web-port", "0", "--workspace", ws, "--new", "--api-key", "sk-x",
+                         "--provider", "deepseek", "--model", "deepseek-flash",
+                         "--api-base", "http://127.0.0.1:%d/v1" % api.server_address[1],
+                         "--protocol", "native-eos", "--approval", "auto"],
+                        stdout=log, stderr=subprocess.STDOUT, env=env)
+
+
+def port():
+    log.seek(0)
+    for l in log.read().splitlines():
+        if l.startswith("webui: http://127.0.0.1:"):
+            return l.split(":")[3].split("/")[0]
+
+
+def messages(base):
+    return call(base, "/api/session")[1].get("messages", [])
+
+
+def prompts(base):
+    return [m["content"] for m in messages(base)
+            if m["role"] == "user" and not m.get("note") and not m.get("tool_result")]
+
+
+def answers(base):
+    return [m["content"] for m in messages(base) if m["role"] == "assistant"]
+
+
+def turn(base, sse, path, body, want=202):
+    n = sse.count("turn_end")
+    code, d = call(base, path, body)
+    if code == want == 202:
+        wait(lambda: sse.count("turn_end") > n, 20)
+    return code, d
+
+
+try:
+    base = "http://127.0.0.1:%s" % wait(port)
+    sse = Sse(base)
+    time.sleep(0.3)
+
+    # 1. retry with nothing to retry, then a first turn
+    code, d = call(base, "/api/retry", {})
+    check(code == 409 and "nothing to retry" in d.get("error", ""), "retry on an empty session: 409 (%s)" % code)
+    turn(base, sse, "/api/turn", {"text": "第一句"})
+    check(prompts(base) == ["第一句"] and answers(base) == ["回答：第一句"], "first turn stored")
+
+    # 2. retry: same prompt again, one copy in the session
+    code, d = turn(base, sse, "/api/retry", {})
+    check(code == 202 and d.get("text") == "第一句", "retry accepted (%s %s)" % (code, d))
+    check(SEEN[-1] == "第一句", "the model saw the same prompt again")
+    check(prompts(base) == ["第一句"] and answers(base) == ["回答：第一句"],
+          "after retry: one prompt, one answer (%s)" % prompts(base))
+
+    # 3. edit and resend: the new text replaces the old
+    code, d = turn(base, sse, "/api/retry", {"text": "改过的问题"})
+    check(code == 202 and d.get("text") == "改过的问题", "edit-and-resend accepted")
+    check(SEEN[-1] == "改过的问题", "the model saw the edited prompt")
+    check(prompts(base) == ["改过的问题"] and answers(base) == ["回答：改过的问题"],
+          "after edit: %s / %s" % (prompts(base), answers(base)))
+    check(call(base, "/api/retry", {"text": ""})[0] == 400, "an empty edit is a 400")
+    check(call(base, "/api/retry", {"text": 5})[0] == 400, "a non-string edit is a 400")
+
+    # 4. a second turn, then retry only touches the last one
+    turn(base, sse, "/api/turn", {"text": "第二句"})
+    turn(base, sse, "/api/retry", {})
+    check(prompts(base) == ["改过的问题", "第二句"] and len(answers(base)) == 2,
+          "retry keeps earlier turns: %s" % prompts(base))
+
+    # 5. retry while a turn runs is a 409
+    n = sse.count("turn_end")
+    call(base, "/api/turn", {"text": "slow 请慢一点"})
+    code, d = call(base, "/api/retry", {})
+    check(code == 409, "retry during a turn: 409 (%s)" % code)
+    wait(lambda: sse.count("turn_end") > n, 20)
+
+    # 6. a note (plan mode changed mid-session) is flagged, and retry skips it
+    call(base, "/api/config", {"plan_mode": True})
+    turn(base, sse, "/api/turn", {"text": "计划一下"})
+    ms = messages(base)
+    notes = [m for m in ms if m.get("note")]
+    check(len(notes) >= 1 and all(m["role"] == "user" for m in notes),
+          "a mid-session note is flagged note:true (%d)" % len(notes))
+    turn(base, sse, "/api/retry", {})
+    check(SEEN[-1] == "计划一下" and prompts(base)[-1] == "计划一下", "retry after a note reruns the prompt")
+    call(base, "/api/config", {"plan_mode": False})
+
+    # 7. export
+    code, ctype, raw = call(base, "/api/export", raw_reply=True)
+    md = raw.decode("utf-8", "replace")
+    check(code == 200 and ctype.startswith("text/markdown"), "export: %s %s" % (code, ctype))
+    check(md.startswith("# ") and "## User" in md and "改过的问题" in md and "## Assistant" in md
+          and "回答：第二句" in md and "第一句" not in md, "export has the transcript")
+
+    # 8. rename and clear
+    st = call(base, "/api/state")[1]
+    cur = st.get("session")
+    code, _ = call(base, "/api/sessions/rename", {"title": "我的会话"})
+    check(code == 200, "rename the open session")
+    ss = call(base, "/api/sessions")[1]
+    check(any(s["id"] == cur and s["title"] == "我的会话" for s in ss.get("sessions", [])),
+          "the list shows the new title")
+    code, _, raw = call(base, "/api/export", raw_reply=True)
+    check(raw.decode("utf-8", "replace").startswith("# 我的会话"), "export uses the title")
+    long = "长" * 100
+    call(base, "/api/sessions/rename", {"id": cur, "title": long})
+    t = [s["title"] for s in call(base, "/api/sessions")[1]["sessions"] if s["id"] == cur][0]
+    check(0 < len(t.encode()) <= 120 and t == "长" * len(t), "a long title is cut on a character (%d bytes)"
+          % len(t.encode()))
+    call(base, "/api/sessions/rename", {"title": ""})
+    t = [s["title"] for s in call(base, "/api/sessions")[1]["sessions"] if s["id"] == cur][0]
+    check(t == "改过的问题", "an empty title falls back to the first prompt (%r)" % t)
+    check(call(base, "/api/sessions/rename", {"id": "nope", "title": "x"})[0] == 404, "rename unknown: 404")
+    check(call(base, "/api/sessions/rename", {"id": "../x", "title": "x"})[0] == 404, "rename ../x: 404")
+    check(call(base, "/api/sessions/rename", {})[0] == 400, "rename without a title: 400")
+
+    # 9. delete another session, then the open one
+    code, d = call(base, "/api/sessions/new", {})
+    second = d.get("id")
+    check(code == 200 and second and second != cur, "a second session")
+    code, d = call(base, "/api/sessions/delete", {"id": cur})
+    check(code == 200 and d.get("current") == second, "delete a closed session (%s %s)" % (code, d))
+    ids = [s["id"] for s in call(base, "/api/sessions")[1]["sessions"]]
+    check(cur not in ids and not os.path.exists(os.path.join(ws, ".agent", "sessions", cur + ".jsonl")),
+          "it is gone from the list and the disk")
+    n = sse.count("session")
+    code, d = call(base, "/api/sessions/delete", {"id": second})
+    check(code == 200 and d.get("current") and d["current"] != second, "delete the open session (%s)" % d)
+    check(call(base, "/api/state")[1].get("session") == d.get("current"), "the state moved to the new one")
+    check(wait(lambda: sse.count("session") > n, 5), "a session event was sent")
+    check(call(base, "/api/sessions/delete", {"id": "nope"})[0] == 404, "delete unknown: 404")
+    check(call(base, "/api/sessions/delete", {})[0] == 400, "delete without an id: 400")
+
+    # 10. history: calls and results come back apart, for native and ctl
+    with open(os.path.join(ws, "hello.txt"), "w", encoding="utf-8") as f:
+        f.write("你好世界\n")
+    turn(base, sse, "/api/turn", {"text": "native 读文件"})
+    ms = messages(base)
+    a = [m for m in ms if m["role"] == "assistant" and m.get("calls")]
+    r = [m for m in ms if m.get("tool_result")]
+    check(len(a) == 1 and a[0]["calls"][0].get("name") == "read_file"
+          and a[0]["calls"][0].get("args", {}).get("path") == "hello.txt" and a[0]["calls"][0].get("id") == "c1",
+          "native: the call is listed with its args (%s)" % (a[0].get("calls") if a else ms))
+    check(len(r) == 1 and r[0]["role"] == "tool" and r[0].get("tool_call_id") == "c1"
+          and "你好世界" in r[0].get("out", "") and not r[0]["out"].startswith("[tool ")
+          and r[0].get("exit") == 0 and r[0].get("truncated") is False,
+          "native: the result carries out/exit without the header (%s)" % r)
+    check(answers(base)[-1] == "读完了", "native: the turn finished")
+
+    check(call(base, "/api/settings", {"protocol": "ctl"})[0] == 200, "switch to the ctl protocol")
+    turn(base, sse, "/api/turn", {"text": "ctl 读文件"})
+    ms = messages(base)
+    a = [m for m in ms if m["role"] == "assistant" and m.get("calls") and m["calls"][0].get("id") == "t1"]
+    r = [m for m in ms if m.get("tool_result") and m.get("tool_call_id") == "t1"]
+    check(len(a) == 1 and a[0].get("text", "").strip() == "我来读。" and "call t1" in a[0]["content"]
+          and a[0]["calls"][0].get("args", {}).get("path") == "hello.txt"
+          and "call t1" not in a[0].get("html", ""),
+          "ctl: prose and calls are split (%s)" % (a[0] if a else ms[-4:]))
+    check(len(r) == 1 and r[0]["role"] == "user" and "你好世界" in r[0].get("out", "")
+          and not r[0]["out"].startswith("result ") and not r[0]["out"].endswith("end")
+          and r[0].get("exit") == 0,
+          "ctl: the result is flagged with out/exit (%s)" % r)
+    check(r and r[0]["content"] not in prompts(base), "ctl: the result is not a prompt")
+    call(base, "/api/settings", {"protocol": "native-eos"})
+    sse.close()
+finally:
+    proc.terminate()
+    try:
+        proc.wait(20)
+    except subprocess.TimeoutExpired:
+        proc.kill()
+        check(False, "server hung on SIGTERM")
+    log.close()
+    shutil.rmtree(ws, ignore_errors=True)
+    shutil.rmtree(home, ignore_errors=True)
+    api.shutdown()
+print("FAILED:", fails)
+sys.exit(1 if fails else 0)
@@ -0,0 +1,313 @@
+"""WebUI settings and streaming against fake OpenAI-compatible APIs.
+
+  python3 scripts/e2e_webui_settings.py      (CST_AGENT= picks the binary)
+
+Checks: Chinese reasoning streams as whole characters in few events; the
+provider list, the model list with capabilities, config validation and
+persistence; a wrong key is rejected, a good one switches the provider
+(requests then go to the new base URL) and is saved; and all of it
+survives a restart. HOME and XDG_CONFIG_HOME point at a temp dir, so no
+real credentials or settings are read or written.
+"""
+import json, os, shutil, stat, subprocess, sys, tempfile, threading, time, urllib.request
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+
+BIN = os.environ.get("CST_AGENT", os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "cst-agent"))
+THINK = ["好的，我先看看工作区里", "有哪些文件，然后用中文", "回答这个问题。😀 结束"]
+ANSWER = "这是中文回答。"
+GOOD_KEYS = {"sk-good", "sk-good2"}
+B_REQS = []
+
+
+def sse_reply(h, model, think, answer):
+    h.send_response(200)
+    h.send_header("Content-Type", "text/event-stream")
+    h.end_headers()
+    base = {"id": "x", "object": "chat.completion.chunk", "model": model}
+
+    def ev(o):
+        h.wfile.write(b"data: " + json.dumps({**base, **o}, ensure_ascii=False).encode() + b"\n\n")
+        h.wfile.flush()
+    for part in think:
+        ev({"choices": [{"index": 0, "delta": {"role": "assistant", "reasoning_content": part}}]})
+    ev({"choices": [{"index": 0, "delta": {"content": answer}}]})
+    ev({"choices": [{"index": 0, "delta": {}, "finish_reason": "stop"}]})
+    h.wfile.write(b"data: [DONE]\n\n")
+    h.wfile.flush()
+
+
+def send_json(h, code, obj):
+    body = json.dumps(obj).encode()
+    h.send_response(code)
+    h.send_header("Content-Type", "application/json")
+    h.send_header("Content-Length", str(len(body)))
+    h.end_headers()
+    h.wfile.write(body)
+
+
+class A(BaseHTTPRequestHandler):
+    """The starting provider: any key, one model, Chinese reasoning."""
+    def log_message(self, *a):
+        pass
+
+    def do_GET(self):
+        send_json(self, 200, {"data": [{"id": "deepseek-flash"}]})
+
+    def do_POST(self):
+        self.rfile.read(int(self.headers.get("Content-Length", 0)))
+        sse_reply(self, "deepseek-flash", THINK, ANSWER)
+
+
+class B(BaseHTTPRequestHandler):
+    """The provider switched to: /models checks the key."""
+    def log_message(self, *a):
+        pass
+
+    def key_ok(self):
+        auth = self.headers.get("Authorization", "")
+        return auth.startswith("Bearer ") and auth[7:] in GOOD_KEYS
+
+    def do_GET(self):
+        if not self.key_ok():
+            return send_json(self, 401, {"error": {"message": "invalid api key"}})
+        send_json(self, 200, {"data": [{"id": "fake-vis"}, {"id": "fake-text"}]})
+
+    def do_POST(self):
+        body = json.loads(self.rfile.read(int(self.headers.get("Content-Length", 0))) or b"{}")
+        B_REQS.append({"auth": self.headers.get("Authorization", ""), "model": body.get("model")})
+        if not self.key_ok():
+            return send_json(self, 401, {"error": {"message": "invalid api key"}})
+        sse_reply(self, body.get("model") or "fake-vis", [], "B 的回答")
+
+
+fails = 0
+
+
+def check(ok, msg):
+    global fails
+    print(("PASS " if ok else "FAIL ") + msg)
+    fails += 0 if ok else 1
+
+
+class Sse:
+    def __init__(self, base):
+        self.lines = []
+        self.r = urllib.request.urlopen(base + "/api/events", timeout=60)
+        threading.Thread(target=self.pump, daemon=True).start()
+
+    def pump(self):
+        try:
+            for raw in self.r:
+                self.lines.append(raw)
+        except Exception:
+            pass
+
+    def events(self, name):
+        out = []
+        for i, l in enumerate(self.lines):
+            if l.strip() == b"event: " + name.encode() and i + 1 < len(self.lines):
+                out.append(self.lines[i + 1][6:])
+        return out
+
+    def close(self):
+        try:
+            self.r.close()
+        except Exception:
+            pass
+
+
+def wait(fn, t=15):
+    end = time.time() + t
+    while time.time() < end:
+        v = fn()
+        if v:
+            return v
+        time.sleep(0.1)
+    return None
+
+
+def call(base, path, obj=None, method=None):
+    data = None if obj is None else json.dumps(obj).encode()
+    req = urllib.request.Request(base + path, data=data, method=method or ("POST" if data else "GET"),
+                                 headers={"Content-Type": "application/json"})
+    try:
+        with urllib.request.urlopen(req, timeout=20) as r:
+            raw = r.read()
+            return r.status, json.loads(raw or b"{}"), raw
+    except urllib.error.HTTPError as e:
+        raw = e.read()
+        try:
+            return e.code, json.loads(raw or b"{}"), raw
+        except ValueError:
+            return e.code, {}, raw
+
+
+def serve(handler):
+    s = ThreadingHTTPServer(("127.0.0.1", 0), handler)
+    threading.Thread(target=s.serve_forever, daemon=True).start()
+    return s
+
+
+api_a, api_b = serve(A), serve(B)
+ws = tempfile.mkdtemp(prefix="cst-set-")
+home = tempfile.mkdtemp(prefix="cst-home-")
+cfgdir = home + "/.config/cst-agent"
+os.makedirs(cfgdir)
+with open(cfgdir + "/models.json", "w") as f:
+    json.dump({"providers": {"fakeb": {"api_base": "http://127.0.0.1:%d/v1" % api_b.server_address[1],
+                                       "key_env": "FAKEB_API_KEY", "default_model": ""}},
+               "models": [{"match": "fake-vis", "vision": True, "context": 200000},
+                          {"match": "fake-text", "vision": False}]}, f)
+env = dict(os.environ, HOME=home, XDG_CONFIG_HOME=home + "/.config",
+           XDG_DATA_HOME=home + "/.local/share", XDG_CACHE_HOME=home + "/.cache")
+for k in ("DEEPSEEK_API_KEY", "FAKEB_API_KEY", "OPENAI_API_KEY", "CST_API_KEY"):
+    env.pop(k, None)
+procs = []
+
+
+def start(args, tag):
+    log = open("%s/%s.log" % (ws, tag), "w+")
+    p = subprocess.Popen([BIN, "--web", "--web-port", "0", "--workspace", ws] + args,
+                         stdout=log, stderr=subprocess.STDOUT, env=env)
+    procs.append((p, log))
+
+    def port():
+        log.seek(0)
+        for l in log.read().splitlines():
+            if l.startswith("webui: http://127.0.0.1:"):
+                return l.split(":")[3].split("/")[0]
+    pt = wait(port)
+    return p, "http://127.0.0.1:%s" % pt
+
+
+def stop(p):
+    p.terminate()
+    try:
+        p.wait(20)
+    except subprocess.TimeoutExpired:
+        p.kill()
+        check(False, "server hung on SIGTERM")
+    check(p.returncode is not None and 0 <= p.returncode < 128, "clean exit (%s)" % p.returncode)
+
+
+def leaked(raw):
+    return any(k.encode() in raw for k in GOOD_KEYS | {"sk-bad", "sk-x"})
+
+
+try:
+    p, base = start(["--new", "--api-key", "sk-x", "--provider", "deepseek", "--model", "deepseek-flash",
+                     "--api-base", "http://127.0.0.1:%d/v1" % api_a.server_address[1],
+                     "--protocol", "native-eos", "--approval", "auto"], "first")
+    a = Sse(base)
+    time.sleep(0.3)
+
+    # 1. Chinese reasoning: whole characters, a handful of events
+    code, _, _ = call(base, "/api/turn", {"text": "用中文想一想"})
+    check(code == 202, "turn accepted")
+    end = wait(lambda: a.events("turn_end"), 20)
+    check(bool(end) and json.loads(end[0]).get("rc") == 0, "turn ended rc=0")
+    raw = a.events("think_delta")
+    ok = True
+    for r in raw:
+        try:
+            r.decode("utf-8")
+        except UnicodeDecodeError:
+            ok = False
+    check(ok, "every think_delta frame is valid UTF-8")
+    text = "".join(json.loads(r.decode("utf-8", "replace"))["text"] for r in raw)
+    want = "".join(THINK)
+    check(text == want and "\ufffd" not in text, "the thinking text round trips: %r" % text[:60])
+    check(0 < len(raw) <= 2 * len(THINK) + 2,
+          "think_delta events: %d for %d bytes" % (len(raw), len(want.encode())))
+    ans = "".join(json.loads(r)["text"] for r in a.events("assistant_delta"))
+    check(ans == ANSWER, "the answer round trips: %r" % ans)
+
+    # 2. providers and models
+    code, d, rawp = call(base, "/api/providers")
+    names = {x["name"]: x for x in d.get("providers", [])}
+    check(code == 200 and "deepseek" in names and "fakeb" in names, "provider list: %s" % sorted(names))
+    check(names.get("deepseek", {}).get("current") and names["deepseek"].get("has_key")
+          and names["deepseek"].get("key_from") == "session", "deepseek: current, key for this run")
+    check(names.get("fakeb", {}).get("has_key") is False and names["fakeb"].get("source") == "user",
+          "fakeb: no key yet, from models.json")
+    check(not leaked(rawp), "no key in /api/providers")
+    code, d, _ = call(base, "/api/models")
+    ids = [m["id"] for m in d.get("models", [])]
+    first = (d.get("models") or [{}])[0]
+    check(code == 200 and "deepseek-flash" in ids and first.get("api") is True
+          and "context" in first and "label" in first, "model list: %s" % ids[:5])
+
+    # 3. config: validation and persistence
+    check(call(base, "/api/config", {"effort": "bogus"})[0] == 400, "a bad effort is a 400")
+    check(call(base, "/api/config", {"provider": "fakeb"})[0] == 400, "provider via /api/config is a 400")
+    check(call(base, "/api/config", {"model": ""})[0] == 400, "an empty model is a 400")
+    code, d, _ = call(base, "/api/config", {"model": "deepseek-v4-pro", "effort": "high"})
+    check(code == 200 and d.get("ok"), "model + effort accepted")
+    code, s, _ = call(base, "/api/state")
+    check(s.get("model") == "deepseek-v4-pro" and s.get("effort") == "high", "the state has them")
+    check(s.get("effort_applies") is True and s.get("key_set") is True and "model_info" in s
+          and s.get("perm_timeout", 0) > 0 and "turn_usage" in s, "the state has the new fields")
+    st = open(cfgdir + "/settings.json").read() if os.path.exists(cfgdir + "/settings.json") else ""
+    check("deepseek-v4-pro" in st and "high" in st, "settings.json remembers model and effort")
+
+    # 4. switching provider: no key, a wrong key, a good key
+    code, d, _ = call(base, "/api/provider", {"name": "fakeb"})
+    check(code == 400 and d.get("need_key"), "fakeb without a key: need_key (%s)" % code)
+    code, d, rawp = call(base, "/api/provider", {"name": "fakeb", "key": "sk-bad"})
+    check(code == 401 and "拒绝" in d.get("error", ""), "a wrong key is rejected (%s %s)" % (code, d.get("error")))
+    check(not leaked(rawp), "no key in the rejection")
+    check(call(base, "/api/state")[1].get("provider") == "deepseek", "still on deepseek")
+    check(call(base, "/api/provider", {"name": "nope"})[0] == 404, "an unknown provider is a 404")
+    code, d, rawp = call(base, "/api/provider", {"name": "fakeb", "key": "sk-good"})
+    check(code == 200 and d.get("provider") == "fakeb" and d.get("model") == "fake-vis",
+          "switched to fakeb, first listed model picked (%s %s)" % (code, d))
+    check(not leaked(rawp), "no key in the switch reply")
+    creds = cfgdir + "/credentials"
+    cr = open(creds).read() if os.path.exists(creds) else ""
+    check("fakeb.api_key=sk-good" in cr, "the key was saved to credentials")
+    if os.name == "posix" and os.path.exists(creds):
+        check(stat.S_IMODE(os.stat(creds).st_mode) == 0o600, "credentials is 0600")
+    code, d, _ = call(base, "/api/models")
+    ms = {m["id"]: m for m in d.get("models", [])}
+    check("fake-vis" in ms and ms["fake-vis"].get("vision") is True and ms["fake-vis"].get("context") == 200000
+          and ms.get("fake-text", {}).get("vision") is False, "fakeb models with capabilities")
+    code, d, _ = call(base, "/api/config", {"effort": "low"})
+    check(code == 200 and d.get("warning"), "effort on a model without it: saved with a warning")
+
+    # 5. a turn now goes to the new base URL with the new key
+    n0 = len(B_REQS)
+    code, _, _ = call(base, "/api/turn", {"text": "hi B"})
+    check(code == 202, "turn on fakeb accepted")
+    wait(lambda: len(a.events("turn_end")) >= 2, 20)
+    check(len(B_REQS) > n0 and B_REQS[-1]["auth"] == "Bearer sk-good" and B_REQS[-1]["model"] == "fake-vis",
+          "the request reached fakeb: %s" % (B_REQS[-1:] or "none"))
+
+    # 6. replacing the key
+    code, d, _ = call(base, "/api/key", {"key": "sk-bad"})
+    check(code == 401, "a wrong replacement key is rejected (%s)" % code)
+    code, d, _ = call(base, "/api/key", {"key": "sk-good2"})
+    check(code == 200 and d.get("ok"), "the replacement key is accepted")
+    check("fakeb.api_key=sk-good2" in open(creds).read(), "the replacement key was saved")
+    check(call(base, "/api/key", {})[0] == 400, "an empty key body is a 400")
+    a.close()
+    stop(p)
+
+    # 7. the next start remembers provider, model and key
+    p, base = start(["--new"], "second")
+    code, s, raws = call(base, "/api/state")
+    check(s.get("provider") == "fakeb" and s.get("model") == "fake-vis" and s.get("key_set") is True
+          and s.get("key_from") == "saved", "restart: %s/%s key %s" % (s.get("provider"), s.get("model"), s.get("key_from")))
+    check(not leaked(raws), "no key in /api/state")
+    stop(p)
+
+    logs = b"".join(open(lg.name, "rb").read() for _, lg in procs)
+    check(not leaked(logs.replace(b"sk-x", b"")) , "no key in the server logs")
+finally:
+    for p, log in procs:
+        if p.poll() is None:
+            p.kill()
+        log.close()
+    shutil.rmtree(ws, ignore_errors=True)
+    shutil.rmtree(home, ignore_errors=True)
+print("FAILED:", fails)
+sys.exit(1 if fails else 0)
@@ -0,0 +1,206 @@
+"""WebUI workspace views: git diff, run settings, saved permission rules.
+
+  python3 scripts/e2e_webui_workspace.py     (CST_AGENT= picks the binary)
+
+Checks: /api/diff lists git status and the working-tree diff (per file too),
+an option-looking path is only ever a path, and a directory without git
+says so; /api/settings reports protocol, tools, confirmations and the
+sandbox, rejects a bad body without applying any of it, applies a good one,
+and is a 409 while a turn runs; /api/permissions shows the saved rules and
+DELETE clears them. HOME points at a temp dir.
+"""
+import json, os, shutil, subprocess, sys, tempfile, threading, time, urllib.request
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+
+BIN = os.environ.get("CST_AGENT", os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "cst-agent"))
+
+
+class Api(BaseHTTPRequestHandler):
+    """A slow plain answer, so a turn is running for a while."""
+    def log_message(self, *a):
+        pass
+
+    def do_POST(self):
+        self.rfile.read(int(self.headers.get("Content-Length", 0)))
+        time.sleep(2)
+        self.send_response(200)
+        self.send_header("Content-Type", "text/event-stream")
+        self.end_headers()
+        for o in ({"delta": {"role": "assistant", "content": "好"}}, {"delta": {}, "finish_reason": "stop"}):
+            chunk = {"id": "x", "object": "chat.completion.chunk", "model": "m", "choices": [dict(index=0, **o)]}
+            self.wfile.write(b"data: " + json.dumps(chunk, ensure_ascii=False).encode() + b"\n\n")
+        self.wfile.write(b"data: [DONE]\n\n")
+
+
+fails = 0
+
+
+def check(ok, msg):
+    global fails
+    print(("PASS " if ok else "FAIL ") + msg)
+    fails += 0 if ok else 1
+
+
+def wait(fn, t=15):
+    end = time.time() + t
+    while time.time() < end:
+        v = fn()
+        if v:
+            return v
+        time.sleep(0.1)
+    return None
+
+
+def call(base, path, obj=None, method=None):
+    data = None if obj is None else json.dumps(obj).encode()
+    req = urllib.request.Request(base + path, data=data, method=method or ("POST" if data else "GET"),
+                                 headers={"Content-Type": "application/json"})
+    try:
+        with urllib.request.urlopen(req, timeout=30) as r:
+            return r.status, json.loads(r.read() or b"{}")
+    except urllib.error.HTTPError as e:
+        raw = e.read()
+        try:
+            return e.code, json.loads(raw or b"{}")
+        except ValueError:
+            return e.code, {}
+
+
+def git(*args):
+    return subprocess.run(["git", "-C", ws] + list(args), capture_output=True, text=True)
+
+
+api = ThreadingHTTPServer(("127.0.0.1", 0), Api)
+threading.Thread(target=api.serve_forever, daemon=True).start()
+ws = tempfile.mkdtemp(prefix="cst-ws-")
+plain = tempfile.mkdtemp(prefix="cst-nogit-")
+home = tempfile.mkdtemp(prefix="cst-home-")
+env = dict(os.environ, HOME=home, XDG_CONFIG_HOME=home + "/.config",
+           XDG_DATA_HOME=home + "/.local/share", XDG_CACHE_HOME=home + "/.cache",
+           GIT_CONFIG_NOSYSTEM="1")
+for k in ("DEEPSEEK_API_KEY", "OPENAI_API_KEY", "CST_API_KEY"):
+    env.pop(k, None)
+have_git = shutil.which("git") is not None
+if have_git:
+    git("init", "-q")
+    git("config", "user.email", "e2e@example.com")
+    git("config", "user.name", "e2e")
+    open(ws + "/a.txt", "w").write("one\n")
+    git("add", "a.txt")
+    git("commit", "-q", "-m", "first")
+    open(ws + "/a.txt", "w").write("one\n第二行\n")
+    open(ws + "/new.txt", "w").write("new\n")
+os.makedirs(ws + "/.agent", exist_ok=True)
+json.dump({"tools": ["write_file"], "shell_prefixes": ["make"]}, open(ws + "/.agent/permissions.json", "w"))
+
+log = open(home + "/server.log", "w+")
+proc = subprocess.Popen([BIN, "--web", "--web-port", "0", "--workspace", ws, "--new", "--api-key", "sk-x",
+                         "--provider", "deepseek", "--model", "deepseek-flash",
+                         "--api-base", "http://127.0.0.1:%d/v1" % api.server_address[1],
+                         "--protocol", "native-eos", "--approval", "auto"],
+                        stdout=log, stderr=subprocess.STDOUT, env=env)
+
+
+def port():
+    log.seek(0)
+    for l in log.read().splitlines():
+        if l.startswith("webui: http://127.0.0.1:"):
+            return l.split(":")[3].split("/")[0]
+
+
+try:
+    base = "http://127.0.0.1:%s" % wait(port)
+
+    # 1. diff
+    if have_git:
+        code, d = call(base, "/api/diff")
+        files = {f["path"]: f["status"] for f in d.get("files", [])}
+        check(code == 200 and d.get("git") is True, "diff: git repo seen")
+        check(files.get("a.txt") == " M" and files.get("new.txt") == "??", "status lists both files: %s" % files)
+        check("+第二行" in d.get("diff", "") and "a.txt" in d.get("diff", ""), "the diff has the change")
+        code, d = call(base, "/api/diff?path=a.txt")
+        check(code == 200 and "+第二行" in d.get("diff", "") and d.get("path") == "a.txt", "diff of one file")
+        code, d = call(base, "/api/diff?path=new.txt")
+        check(code == 200 and d.get("diff") == "", "an untracked file has no diff")
+        evil = ws + "/pwned"
+        code, d = call(base, "/api/diff?path=" + urllib.request.quote("--output=" + evil))
+        check(code == 200 and not os.path.exists(evil), "an option-looking path stays a path")
+    else:
+        print("SKIP diff checks: no git")
+
+    # 2. settings: what is there
+    code, s = call(base, "/api/settings")
+    names = [t["name"] for t in s.get("tools", [])]
+    check(code == 200 and s.get("protocol") == "native-eos" and len(s.get("protocols", [])) == 5,
+          "settings: protocol %s of %s" % (s.get("protocol"), s.get("protocols")))
+    check("read_file" in names and all("perm" in t and "on" in t for t in s["tools"]), "tool list: %d tools" % len(names))
+    check(set(s.get("confirm", {})) == {"shell", "write", "net"}, "confirm flags present")
+    check(isinstance(s.get("sandbox", {}).get("kind"), str), "sandbox kind: %s" % s.get("sandbox"))
+    victim = next((t["name"] for t in s["tools"] if t["perm"] == "net"), names[-1])
+
+    # 3. a bad body changes nothing
+    for body, why in (({"protocol": "nope"}, "unknown protocol"), ({"max_steps": 0}, "max_steps 0"),
+                      ({"max_steps": "5"}, "max_steps as text"), ({"tool_profile": "huge"}, "unknown profile"),
+                      ({"tools": {"nope": False}}, "unknown tool"), ({"tools": {victim: "off"}}, "tool value not bool"),
+                      ({"confirm": {"disk": True}}, "unknown confirm key"),
+                      ({"max_steps": 9, "protocol": "nope"}, "mixed good and bad")):
+        check(call(base, "/api/settings", body)[0] == 400, "400 for %s" % why)
+    check(call(base, "/api/settings")[1].get("max_steps") == s.get("max_steps"), "nothing applied from a bad body")
+
+    # 4. a good body applies
+    code, d = call(base, "/api/settings", {"protocol": "ctl", "max_steps": 7, "tool_profile": "minimal",
+                                            "tools": {victim: False}, "confirm": {"shell": True, "net": False}})
+    check(code == 200 and d.get("ok"), "settings accepted (%s %s)" % (code, d))
+    code, s2 = call(base, "/api/settings")
+    t2 = {t["name"]: t for t in s2.get("tools", [])}
+    check(s2.get("protocol") == "ctl" and s2.get("max_steps") == 7 and s2.get("tool_profile") == "minimal",
+          "protocol/steps/profile applied")
+    check(t2.get(victim, {}).get("off") is True and t2[victim]["on"] is False, "%s switched off" % victim)
+    check(s2["confirm"]["shell"] is True and s2["confirm"]["net"] is False, "confirm flags applied")
+    st = call(base, "/api/state")[1]
+    check(st.get("protocol_tools") == "ctl" and st.get("max_steps") == 7 and st.get("tool_profile") == "minimal",
+          "the state shows them")
+    call(base, "/api/settings", {"protocol": "native-eos", "tool_profile": "default", "tools": {victim: True}})
+    t3 = {t["name"]: t for t in call(base, "/api/settings")[1]["tools"]}
+    check(t3[victim]["off"] is False and t3[victim]["on"] is True, "%s back on" % victim)
+
+    # 5. permissions
+    code, p = call(base, "/api/permissions")
+    check(code == 200 and p.get("tools") == ["write_file"] and p.get("shell_prefixes") == ["make"]
+          and "write_file" in p.get("rules", ""), "saved rules: %s %s" % (p.get("tools"), p.get("shell_prefixes")))
+    code, _ = call(base, "/api/permissions", method="DELETE")
+    check(code == 200, "DELETE /api/permissions")
+    p = call(base, "/api/permissions")[1]
+    check(p.get("tools") == [] and p.get("shell_prefixes") == [], "the rules are gone")
+    saved = json.load(open(ws + "/.agent/permissions.json"))
+    check(saved.get("tools") == [] and saved.get("shell_prefixes") == [], "and gone from the file")
+
+    # 6. while a turn runs: settings and the rule reset are 409, reads still work
+    code, _ = call(base, "/api/turn", {"text": "慢慢来"})
+    check(code == 202, "a slow turn started")
+    check(call(base, "/api/settings", {"max_steps": 3})[0] == 409, "settings during a turn: 409")
+    check(call(base, "/api/permissions", method="DELETE")[0] == 409, "rule reset during a turn: 409")
+    check(call(base, "/api/settings")[0] == 200 and call(base, "/api/permissions")[0] == 200,
+          "reads during a turn: 200")
+    if have_git:
+        check(call(base, "/api/diff")[0] == 200, "diff during a turn: 200")
+    wait(lambda: not call(base, "/api/state")[1].get("status_busy"), 20)
+
+    # 7. a directory without git
+    if have_git:
+        code, _ = call(base, "/api/workspace", {"path": plain})
+        code, d = call(base, "/api/diff")
+        check(code == 200 and d.get("git") is False and "diff" not in d, "no git repo: git=false")
+finally:
+    proc.terminate()
+    try:
+        proc.wait(20)
+    except subprocess.TimeoutExpired:
+        proc.kill()
+        check(False, "server hung on SIGTERM")
+    log.close()
+    for d in (ws, plain, home):
+        shutil.rmtree(d, ignore_errors=True)
+    api.shutdown()
+print("FAILED:", fails)
+sys.exit(1 if fails else 0)
@@ -0,0 +1,51 @@
+#!/usr/bin/env bash
+# Static musl build of cst-agent for one architecture, packed as
+# dist/cst-agent-VERSION-linux-ARCH.tar.gz (+ .sha256).
+#   scripts/package_linux.sh x86_64|x86|aarch64
+# CST_LINUX_CHECK=1 also runs `make check` (only where the binaries can run).
+# Builds in the source tree, so the tree is cleaned first.
+set -euo pipefail
+ROOT="$(cd "$(dirname "$0")/.." && pwd)"
+ARCH="${1:?usage: $0 x86_64|x86|aarch64}"
+PREFIX="$ROOT/third_party/prefix-linux-$ARCH"
+cd "$ROOT"
+
+read -r LINUX_HOST bindir < <(bash scripts/ci_linux_toolchain.sh "$ARCH")
+export PATH="$bindir:$PATH" LINUX_HOST
+T="$LINUX_HOST-"
+
+if [[ ! -f "$PREFIX/lib/libcurl.a" || ! -f "$PREFIX/lib/libncursesw.a" ]]; then
+  bash scripts/build_third_party_linux.sh "$ARCH"
+fi
+
+# Only the static prefix: no host .pc file may leak into the link.
+export PKG_CONFIG_LIBDIR="$PREFIX/lib/pkgconfig"
+unset PKG_CONFIG_PATH CC CFLAGS LDFLAGS
+mk=(make CC="${T}gcc" AR="${T}ar" STATIC=1
+  CFLAGS="-std=c11 -Wall -Wextra -O2 -g -Werror")
+
+make clean >/dev/null
+make -C vendor/sashidown clean >/dev/null 2>&1 || true
+"${mk[@]}" -j"$(nproc)" cst-agent
+if "${T}readelf" -l cst-agent | grep -q 'Requesting program interpreter'; then
+  echo "cst-agent is dynamically linked" >&2
+  exit 1
+fi
+if [[ "${CST_LINUX_CHECK:-0}" == 1 ]]; then
+  "${mk[@]}" check
+fi
+
+ver="$(sed -n 's/^#define CST_VERSION_\(MAJOR\|MINOR\|PATCH\) \([0-9]*\)$/\2/p' include/cst_version.h | paste -sd.)"
+name="cst-agent-$ver-linux-$ARCH"
+rel="dist/$name"
+rm -rf "$rel"
+mkdir -p "$rel"
+cp -f cst-agent "$rel/cst-agent"
+"${T}strip" "$rel/cst-agent"
+cp -f README.md "$rel/README.md"
+echo "$ver" >"$rel/VERSION"
+tgz="dist/$name.tar.gz"
+tar -C dist -czf "$tgz" "$name"
+(cd dist && sha256sum "$name.tar.gz" >"$name.tar.gz.sha256")
+ls -la "$rel"
+echo "$ROOT/$tgz"
@@ -36,3 +36,18 @@ rm -f "$zip"
 (cd "$ROOT/dist" && sha256sum "$(basename "$zip")" >"$(basename "$zip").sha256")
 ls -la "$REL"
 echo "$zip"
+
+# NSIS installer from the same folder (CST_REQUIRE_NSIS=1: fail without makensis).
+setup="$ROOT/dist/cst-agent-$ver-windows-x86_64-setup.exe"
+if command -v makensis >/dev/null; then
+  rm -f "$setup"
+  makensis -V2 -WX -INPUTCHARSET UTF8 -DVERSION="$ver" -DSRCDIR="$REL" \
+    -DOUTFILE="$setup" "$ROOT/scripts/windows/cst-agent.nsi"
+  (cd "$ROOT/dist" && sha256sum "$(basename "$setup")" >"$(basename "$setup").sha256")
+  echo "$setup"
+elif [[ "${CST_REQUIRE_NSIS:-0}" == 1 ]]; then
+  echo "makensis not found" >&2
+  exit 1
+else
+  echo "makensis not found; skipping the installer" >&2
+fi
@@ -1,10 +1,12 @@
 #!/usr/bin/env bash
-# Publish dist/cst-agent-VERSION-windows-x86_64.zip (+ .sha256) to the Gitea
-# release vVERSION, creating the release (and its tag at HEAD) when missing.
-#   GITEA_TOKEN=... scripts/release_windows.sh [--build]
-# --build runs scripts/bootstrap_windows.sh first. The token is only read from
-# the environment and handed to curl on stdin: it never appears in argv, on
-# disk or in the output. Assets of the same name are replaced.
+# Upload release files to the Gitea release vVERSION, creating the release
+# (and its tag at HEAD) when missing.
+#   GITEA_TOKEN=... scripts/release.sh FILE...
+# Each FILE's FILE.sha256 is uploaded too when it exists; assets of the same
+# name are replaced. Several CI jobs may call this at once: the first creates
+# the release, the others find it. The token is only read from the
+# environment and handed to curl on stdin: it never appears in argv, on disk
+# or in the output.
 #
 # CI: set GITEA_TOKEN (or GITHUB_TOKEN), CST_GITEA, CST_REPO. The remote-branch
 # check is skipped under GITEA_ACTIONS / GITHUB_ACTIONS / CI.
@@ -19,16 +21,20 @@ TOKEN="${GITEA_TOKEN:-${GITHUB_TOKEN:-}}"
  echo "set GITEA_TOKEN (a Gitea token with write:repository)" >&2
  exit 1
 }
+[[ $# -gt 0 ]] || { echo "usage: $0 FILE..." >&2; exit 1; }
 command -v python3 >/dev/null || { echo "missing python3" >&2; exit 1; }
+
+files=()
+for f in "$@"; do
+  [[ -f "$f" ]] || { echo "missing $f" >&2; exit 1; }
+  f="$(cd "$(dirname "$f")" && pwd)/$(basename "$f")"
+  files+=("$f")
+  [[ "$f" != *.sha256 && -f "$f.sha256" ]] && files+=("$f.sha256")
+done
 cd "$ROOT"
-[[ "${1:-}" == --build ]] && scripts/bootstrap_windows.sh "$ROOT"

 ver="$(sed -n 's/^#define CST_VERSION_\(MAJOR\|MINOR\|PATCH\) \([0-9]*\)$/\2/p' include/cst_version.h | paste -sd.)"
 tag="v$ver"
-zip="dist/cst-agent-$ver-windows-x86_64.zip"
-for f in "$zip" "$zip.sha256"; do
-  [[ -f "$f" ]] || { echo "missing $f (scripts/package_windows.sh, or --build)" >&2; exit 1; }
-done

 # If this job was triggered by a version tag, it must match cst_version.h.
 if [[ "${GITHUB_REF:-}" == refs/tags/v* ]]; then
@@ -51,6 +57,12 @@ json() { # CODE [ARGS...]: d = the JSON on stdin, ARGS in sys.argv[1:]
  python3 -c "import json,sys; d=json.load(sys.stdin); $code" "$@"
 }

+BODY="Linux (static, musl, any distro): cst-agent-$ver-linux-{x86_64,x86,aarch64}.tar.gz -- extract and run ./cst-agent, or copy it onto PATH.
+
+Windows x86_64: cst-agent-$ver-windows-x86_64-setup.exe installs per user (no admin rights) and adds PATH; cst-agent-$ver-windows-x86_64.zip is the portable build (install_windows.ps1 inside).
+
+Every file has a .sha256 next to it."
+
 if rel="$(api GET "/releases/tags/$tag" 2>/dev/null)"; then
  echo "==> release $tag exists"
 else
@@ -62,12 +74,22 @@ else
    fi
  fi
  echo "==> creating release $tag at $head"
-  body="$(python3 -c 'import json,sys; print(json.dumps({"tag_name": sys.argv[1], "target_commitish": sys.argv[2], "name": "CST-Agent " + sys.argv[1], "body": "Windows x86_64: unzip, or install with install_windows.ps1 from the zip.\\n\\nArtifacts: `cst-agent-VERSION-windows-x86_64.zip` + `.sha256`."}))' "$tag" "$head")"
-  rel="$(api POST /releases -H 'Content-Type: application/json' -d "$body")"
+  body="$(python3 -c 'import json,sys; print(json.dumps({"tag_name": sys.argv[1], "target_commitish": sys.argv[2], "name": "CST-Agent " + sys.argv[1], "body": sys.argv[3]}))' "$tag" "$head" "$BODY")"
+  if ! rel="$(api POST /releases -H 'Content-Type: application/json' -d "$body" 2>/dev/null)"; then
+    # Another job created it first (409).
+    rel=""
+    for _ in 1 2 3 4 5 6; do
+      sleep 5
+      rel="$(api GET "/releases/tags/$tag" 2>/dev/null)" && break
+      rel=""
+    done
+    [[ -n "$rel" ]] || { echo "could not create or find release $tag" >&2; exit 1; }
+    echo "==> release $tag created by another job"
+  fi
 fi
 id="$(json 'print(d["id"])' <<<"$rel")"

-for f in "$zip" "$zip.sha256"; do
+for f in "${files[@]}"; do
  name="$(basename "$f")"
  old="$(json "print(' '.join(str(a['id']) for a in d.get('assets') or [] if a['name'] == sys.argv[1]))" "$name" <<<"$rel")"
  for aid in $old; do
@@ -0,0 +1,65 @@
+"""TUI /provider and /key against a fake API whose /models checks the key.
+
+  nix-shell -p python3Packages.pyte --run "python3 scripts/tui_keys/provider.py"
+
+A wrong key is asked for again, a good one is saved and switches the
+provider; /key replaces it. Same code path as the WebUI (provider_switch.c).
+"""
+import json, os, sys, threading
+from http.server import BaseHTTPRequestHandler, ThreadingHTTPServer
+sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
+from harness import Term
+from keys import check, base
+
+GOOD = {"sk-good", "sk-good2"}
+
+
+class B(BaseHTTPRequestHandler):
+    def log_message(self, *a):
+        pass
+
+    def do_GET(self):
+        auth = self.headers.get("Authorization", "")
+        ok = auth.startswith("Bearer ") and auth[7:] in GOOD
+        body = json.dumps({"data": [{"id": "fake-one"}]} if ok else {"error": {"message": "bad key"}}).encode()
+        self.send_response(200 if ok else 401)
+        self.send_header("Content-Type", "application/json")
+        self.send_header("Content-Length", str(len(body)))
+        self.end_headers()
+        self.wfile.write(body)
+
+
+srv = ThreadingHTTPServer(("127.0.0.1", 0), B)
+threading.Thread(target=srv.serve_forever, daemon=True).start()
+
+t = Term(args=("-b", base, "-k", "sk-test"))
+cfg = os.path.join(t.home, ".config", "cst-agent")
+os.makedirs(cfg, exist_ok=True)
+t.close()
+with open(os.path.join(cfg, "models.json"), "w") as f:
+    json.dump({"providers": {"fakeb": {"api_base": "http://127.0.0.1:%d/v1" % srv.server_address[1],
+                                       "key_env": "FAKEB_API_KEY", "default_model": ""}}}, f)
+t = Term(args=("-b", base, "-k", "sk-test"), home=t.home, env={"FAKEB_API_KEY": ""})
+creds = os.path.join(cfg, "credentials")
+
+
+def screen():
+    return "\n".join(t.lines())
+
+
+t.send("/provider fakeb\r", 0.8)
+check("a key prompt for fakeb", "API key for fakeb" in screen(), t)
+t.send("sk-bad\r", 1.5)
+check("a wrong key is asked for again", "rejected that key" in screen(), t)
+check("the wrong key is not saved", not os.path.exists(creds) or "sk-bad" not in open(creds).read(), t)
+t.send("sk-good\r", 2.0)
+check("switched to fakeb", "provider → fakeb" in screen(), t)
+check("the good key is saved", os.path.exists(creds) and "fakeb.api_key=sk-good" in open(creds).read(), t)
+check("no key on screen", "sk-good" not in screen() and "sk-bad" not in screen(), t)
+t.send("/key\r", 0.8)
+t.send("sk-good2\r", 2.0)
+check("/key replaces it", "fakeb.api_key=sk-good2" in open(creds).read(), t)
+t.close()
+from keys import FAILS
+print("FAILED:", len(FAILS))
+sys.exit(min(len(FAILS), 100))
@@ -0,0 +1,151 @@
+; CST-Agent per-user installer (no admin rights).
+;   makensis -INPUTCHARSET UTF8 -DVERSION=0.8.5 -DSRCDIR=dist/windows-x86_64 -DOUTFILE=setup.exe cst-agent.nsi
+; Installs the same files as the zip into %LOCALAPPDATA%\Programs\cst-agent,
+; adds that folder to the user PATH and registers an uninstaller under HKCU.
+Unicode true
+SetCompressor /SOLID lzma
+RequestExecutionLevel user
+
+!ifndef VERSION
+  !error "pass -DVERSION=x.y.z"
+!endif
+!ifndef SRCDIR
+  !error "pass -DSRCDIR=<dist/windows-x86_64>"
+!endif
+!ifndef OUTFILE
+  !define OUTFILE "cst-agent-${VERSION}-windows-x86_64-setup.exe"
+!endif
+
+!define APP "CST-Agent"
+!define UNINST_KEY "Software\Microsoft\Windows\CurrentVersion\Uninstall\cst-agent"
+
+!include "MUI2.nsh"
+!include "LogicLib.nsh"
+!include "x64.nsh"
+
+Name "${APP} ${VERSION}"
+OutFile "${OUTFILE}"
+InstallDir "$LOCALAPPDATA\Programs\cst-agent"
+InstallDirRegKey HKCU "${UNINST_KEY}" "InstallLocation"
+BrandingText "${APP} ${VERSION}"
+
+VIProductVersion "${VERSION}.0"
+VIAddVersionKey /LANG=1033 "ProductName" "${APP}"
+VIAddVersionKey /LANG=1033 "ProductVersion" "${VERSION}"
+VIAddVersionKey /LANG=1033 "FileVersion" "${VERSION}"
+VIAddVersionKey /LANG=1033 "FileDescription" "${APP} installer"
+VIAddVersionKey /LANG=1033 "LegalCopyright" "WHC"
+
+!define MUI_ABORTWARNING
+!define MUI_FINISHPAGE_RUN
+!define MUI_FINISHPAGE_RUN_TEXT "$(RunWebUI)"
+!define MUI_FINISHPAGE_RUN_FUNCTION LaunchWebUI
+
+!insertmacro MUI_PAGE_DIRECTORY
+!insertmacro MUI_PAGE_INSTFILES
+!insertmacro MUI_PAGE_FINISH
+!insertmacro MUI_UNPAGE_CONFIRM
+!insertmacro MUI_UNPAGE_INSTFILES
+
+!insertmacro MUI_LANGUAGE "English"
+!insertmacro MUI_LANGUAGE "SimpChinese"
+
+LangString NeedX64 ${LANG_ENGLISH} "${APP} needs 64-bit Windows."
+LangString NeedX64 ${LANG_SIMPCHINESE} "${APP} 需要 64 位 Windows。"
+LangString NeedBuild ${LANG_ENGLISH} "${APP} needs Windows 10 1903 (build 18362) or newer; this is build $0."
+LangString NeedBuild ${LANG_SIMPCHINESE} "${APP} 需要 Windows 10 1903（内部版本 18362）或更新版本，当前为 $0。"
+LangString RunWebUI ${LANG_ENGLISH} "Open the ${APP} WebUI"
+LangString RunWebUI ${LANG_SIMPCHINESE} "打开 ${APP} WebUI"
+LangString PathFail ${LANG_ENGLISH} "Could not add $INSTDIR to the user PATH; add it by hand."
+LangString PathFail ${LANG_SIMPCHINESE} "未能把 $INSTDIR 加入用户 PATH，请手动添加。"
+
+; PowerShell edits the user PATH: NSIS strings are capped at 1024 chars, and
+; .NET broadcasts the change so new terminals see it. The folder is passed in
+; the environment, so quotes in it cannot break the command.
+!macro EditUserPath ACTION
+  System::Call 'Kernel32::SetEnvironmentVariable(t "CST_INSTDIR", t "$INSTDIR")'
+  nsExec::ExecToLog `powershell -NoProfile -ExecutionPolicy Bypass -Command "$$d = $$env:CST_INSTDIR.TrimEnd('\'); $$p = [Environment]::GetEnvironmentVariable('Path', 'User'); $$parts = @([string]$$p -split ';' | Where-Object { $$_ -and $$_.TrimEnd('\') -ine $$d }); if ('${ACTION}' -eq 'add') { $$parts += $$d }; [Environment]::SetEnvironmentVariable('Path', ($$parts -join ';'), 'User')"`
+  Pop $0
+!macroend
+
+; A running exe cannot be overwritten but can be renamed out of the way.
+!macro MoveAside NAME
+  Delete "$INSTDIR\${NAME}.old"
+  ClearErrors
+  Delete "$INSTDIR\${NAME}"
+  ${If} ${Errors}
+    Rename "$INSTDIR\${NAME}" "$INSTDIR\${NAME}.old"
+  ${EndIf}
+!macroend
+
+Function .onInit
+  ${IfNot} ${RunningX64}
+    MessageBox MB_ICONSTOP "$(NeedX64)"
+    Abort
+  ${EndIf}
+  SetRegView 64
+  ReadRegStr $0 HKLM "SOFTWARE\Microsoft\Windows NT\CurrentVersion" "CurrentBuildNumber"
+  ${If} $0 < 18362
+    MessageBox MB_ICONSTOP "$(NeedBuild)"
+    Abort
+  ${EndIf}
+FunctionEnd
+
+Function LaunchWebUI
+  SetOutPath "$PROFILE"
+  Exec '"$INSTDIR\cst-agent.exe" --web --web-open'
+FunctionEnd
+
+Section "Install"
+  SetShellVarContext current
+  SetOutPath "$INSTDIR"
+  !insertmacro MoveAside "cst-agent.exe"
+  !insertmacro MoveAside "busybox.exe"
+  File "${SRCDIR}/cst-agent.exe"
+  File "${SRCDIR}/busybox.exe"
+  File "${SRCDIR}/LICENSE.busybox"
+  File "${SRCDIR}/README.md"
+  File "${SRCDIR}/cst-agent-update.cmd"
+  File "${SRCDIR}/install_windows.ps1"
+  File "${SRCDIR}/VERSION"
+  WriteUninstaller "$INSTDIR\uninstall.exe"
+
+  !insertmacro EditUserPath "add"
+  ${If} $0 != 0
+    DetailPrint "$(PathFail)"
+  ${EndIf}
+
+  ; The WebUI's workspace is the shortcut's start folder.
+  SetOutPath "$PROFILE"
+  CreateShortCut "$SMPROGRAMS\${APP} WebUI.lnk" "$INSTDIR\cst-agent.exe" "--web --web-open" "$INSTDIR\cst-agent.exe" 0
+  SetOutPath "$INSTDIR"
+
+  WriteRegStr HKCU "${UNINST_KEY}" "DisplayName" "${APP}"
+  WriteRegStr HKCU "${UNINST_KEY}" "DisplayVersion" "${VERSION}"
+  WriteRegStr HKCU "${UNINST_KEY}" "Publisher" "WHC"
+  WriteRegStr HKCU "${UNINST_KEY}" "DisplayIcon" "$INSTDIR\cst-agent.exe"
+  WriteRegStr HKCU "${UNINST_KEY}" "InstallLocation" "$INSTDIR"
+  WriteRegStr HKCU "${UNINST_KEY}" "UninstallString" '"$INSTDIR\uninstall.exe"'
+  WriteRegStr HKCU "${UNINST_KEY}" "QuietUninstallString" '"$INSTDIR\uninstall.exe" /S'
+  WriteRegStr HKCU "${UNINST_KEY}" "URLInfoAbout" "https://git.whcself.xyz/WHC/CST-agent"
+  WriteRegDWORD HKCU "${UNINST_KEY}" "NoModify" 1
+  WriteRegDWORD HKCU "${UNINST_KEY}" "NoRepair" 1
+SectionEnd
+
+Section "Uninstall"
+  SetShellVarContext current
+  !insertmacro EditUserPath "remove"
+  Delete "$SMPROGRAMS\${APP} WebUI.lnk"
+  Delete "$INSTDIR\cst-agent.exe"
+  Delete "$INSTDIR\cst-agent.exe.old"
+  Delete "$INSTDIR\busybox.exe"
+  Delete "$INSTDIR\busybox.exe.old"
+  Delete "$INSTDIR\LICENSE.busybox"
+  Delete "$INSTDIR\README.md"
+  Delete "$INSTDIR\cst-agent-update.cmd"
+  Delete "$INSTDIR\install_windows.ps1"
+  Delete "$INSTDIR\VERSION"
+  Delete "$INSTDIR\uninstall.exe"
+  RMDir "$INSTDIR"
+  DeleteRegKey HKCU "${UNINST_KEY}"
+SectionEnd
@@ -15,6 +15,7 @@ win.mkShell {
    pkg-config
    zip
    curl
+    nsis
  ];
  buildInputs = [ win.windows.pthreads ];
  # mkShell exports CC/AR for the target; the build scripts call the prefixed tools directly.
@@ -177,8 +177,62 @@ static int tool_on(const cst_agent_config_t *cfg, const char *name) {
  return cst_tool_active(cfg, cst_tool_find(name));
 }

+/* /mode general (cst_mode_name covers "" and unknown values). */
+static int general_mode(const cst_agent_config_t *cfg) {
+  return cfg && strcmp(cst_mode_name(cfg->mode), "general") == 0;
+}
+
+/* Workflow for /mode general: gather, then do the work, then check it. */
+static void put_workflow_general(cst_buf_t *b, const cst_agent_config_t *cfg) {
+  cst_buf_puts(b, "\nWorkflow:\n1) Gather what you need");
+  const char *explore[] = {"list_dir", "glob", "grep"};
+  int n = 0;
+  for (int i = 0; i < 3; i++) {
+    if (tool_on(cfg, explore[i]))
+      cst_buf_printf(b, "%s%s", n++ ? ", " : " (", explore[i]);
+  }
+  if (n)
+    cst_buf_puts(b, ")");
+  if (tool_on(cfg, "read_file"))
+    cst_buf_puts(b, "; read big files in slices with read_file offset/limit, "
+                    "and use the documents the user attached");
+  cst_buf_puts(b, ".\n2) Do the work");
+  n = 0;
+  if (tool_on(cfg, "write_file")) {
+    cst_buf_puts(b, ": write_file for new files");
+    n++;
+  }
+  if (tool_on(cfg, "edit_file"))
+    cst_buf_printf(b, "%sedit_file for small exact replacements",
+                   n++ ? ", " : ": ");
+  if (tool_on(cfg, "apply_patch"))
+    cst_buf_printf(b, "%sapply_patch for multi-hunk diffs to existing files",
+                   n++ ? ", " : ": ");
+  if (tool_on(cfg, "run_shell"))
+    cst_buf_printf(b, "%srun_shell for commands, computations, data "
+                      "processing and other CLIs",
+                   n++ ? ", " : ": ");
+  cst_buf_puts(b, ".\n");
+  if (tool_on(cfg, "run_shell"))
+    cst_buf_puts(b, "3) Check the result before saying it works (re-read it, "
+                    "run it, or verify the numbers); fix and re-run until it "
+                    "holds.\n");
+  if (tool_on(cfg, "todo_write"))
+    cst_buf_puts(b, "- For work with several steps keep a todo list with "
+                    "todo_write and update it as you go.\n");
+  if (tool_on(cfg, "ask_user"))
+    cst_buf_puts(b, "- Use ask_user only when a decision is truly the user's; "
+                    "otherwise choose sensibly and say so.\n");
+  cst_buf_puts(b, "Finish with a short plain-text summary of the result.\n"
+                  "Never invent other tool names.\n");
+}
+
 /* Workflow section for the default tool profile; names only active tools. */
 static void put_workflow(cst_buf_t *b, const cst_agent_config_t *cfg) {
+  if (general_mode(cfg)) {
+    put_workflow_general(b, cfg);
+    return;
+  }
  cst_buf_puts(b, "\nWorkflow:\n1) Inspect only what you need");
  const char *explore[] = {"list_dir", "glob", "grep"};
  int n = 0;
@@ -235,6 +289,25 @@ static const char DEF_INTRO[] =
    "could not verify, say so.\n"
    "- Keep going until the task is done or truly blocked.\n\n";

+/* Same role, framed for /mode general: the workspace may hold documents,
+ * data or anything else, not just code. */
+static const char DEF_INTRO_GENERAL[] =
+    "You are CST-Agent, a general-purpose assistant working in the user's "
+    "local workspace. All tool paths are relative to the workspace root and "
+    "shell commands run there; the material you work on may be code, "
+    "documents, data or anything else.\n\n"
+    "Principles:\n"
+    "- Reply in the user's language. Be concise: lead with the result, no "
+    "filler.\n"
+    "- Understand before answering: read the files, documents and sources the "
+    "task refers to instead of guessing; never invent content, numbers or "
+    "tool results.\n"
+    "- Do what was asked, at the scope asked: leave unrelated files, data and "
+    "material alone.\n"
+    "- Check your work when you can (re-read it, run it, cross-check the "
+    "source); if you could not verify something, say so.\n"
+    "- Keep going until the task is done or truly blocked.\n\n";
+
 static void put_tool_use(cst_buf_t *b, const cst_agent_config_t *cfg) {
  cst_buf_puts(b, "\nTool use:\n");
  if (tool_on(cfg, "run_shell") &&
@@ -248,15 +321,26 @@ static void put_tool_use(cst_buf_t *b, const cst_agent_config_t *cfg) {
    cst_buf_puts(b, "- Read a file before editing it and again after a failed "
                    "edit. The `N|` prefixes in read_file output are not part "
                    "of the file.\n");
+  if (general_mode(cfg) && tool_on(cfg, "web_fetch"))
+    cst_buf_puts(b, "- Use web_fetch for pages the user points you to; say so "
+                    "when a page could not be fetched instead of guessing its "
+                    "content.\n");
  cst_buf_puts(b, "- When a call fails, read the error and change approach "
                  "instead of repeating the same call.\n");
 }

 static void put_safety(cst_buf_t *b, const cst_agent_config_t *cfg) {
-  cst_buf_puts(b, "\nSafety:\n"
-                  "- Do not commit, push or rewrite git history unless asked. "
-                  "Do not delete or overwrite files you did not create unless "
-                  "the task requires it. Never print secrets or API keys.\n");
+  if (general_mode(cfg))
+    cst_buf_puts(b, "\nSafety:\n"
+                    "- Do not delete or overwrite files or data you did not "
+                    "create unless the task requires it, and do not send the "
+                    "user's content anywhere the task did not ask for. Never "
+                    "print secrets or API keys.\n");
+  else
+    cst_buf_puts(b, "\nSafety:\n"
+                    "- Do not commit, push or rewrite git history unless asked. "
+                    "Do not delete or overwrite files you did not create unless "
+                    "the task requires it. Never print secrets or API keys.\n");
  if (tool_on(cfg, "run_shell"))
    cst_buf_puts(b, "- run_shell may run in a sandbox (writes only inside the "
                    "workspace and temp dir; the network works). If a command is "
@@ -365,7 +449,9 @@ char *cst_agent_build_prompt(const cst_agent_config_t *cfg, cst_protocol_t p,
                             const char *project_notes) {
  int minimal = !cfg || cfg->tool_profile == CST_TOOLS_MINIMAL;
  cst_buf_t b = {0};
-  cst_buf_puts(&b, minimal ? PROMPT_BASE : DEF_INTRO);
+  cst_buf_puts(&b, minimal ? PROMPT_BASE
+                           : general_mode(cfg) ? DEF_INTRO_GENERAL
+                                               : DEF_INTRO);
  if (cst_protocol_is_native(p))
    cst_buf_puts(&b, minimal ? SEC_NATIVE
                             : "Call tools through the native "
@@ -1282,13 +1368,22 @@ static void on_mm(const ctl_event_t *ev, void *ud) {
 }

 static void feed_visible(turn_t *t, const char *data, size_t len) {
-  if (!len || t->stop)
-    return;
-  cst_buf_append(&t->raw, data, len);
-  if (t->ctl)
-    ctl_parser_feed(t->ctl, data, len);
-  else if (t->mm)
-    mm_scan_feed(t->mm, data, len);
+  /* cut after each '\n' and '>': a call only becomes ready there, and
+   * raw_mark must not take in what follows it in the same chunk */
+  while (len && !t->stop) {
+    size_t n = 0;
+    while (n < len && data[n] != '\n' && data[n] != '>')
+      n++;
+    if (n < len)
+      n++;
+    cst_buf_append(&t->raw, data, n);
+    if (t->ctl)
+      ctl_parser_feed(t->ctl, data, n);
+    else if (t->mm)
+      mm_scan_feed(t->mm, data, n);
+    data += n;
+    len -= n;
+  }
 }

 static void think_out(turn_t *t, const char *s, size_t n) {
@@ -1306,10 +1401,16 @@ static void on_delta(const char *data, size_t len, void *ud) {
    const char *want = t->in_think ? "</think>" : "<think>";
    size_t wl = strlen(want);
    if (t->taglen == 0 && c != '<') {
+      /* the whole run up to the next '<' in one call: byte by byte, a
+       * front end would get every CJK character in three invalid pieces */
+      size_t j = i + 1;
+      while (j < len && data[j] != '<')
+        j++;
      if (t->in_think)
-        think_out(t, &c, 1);
+        think_out(t, data + i, j - i);
      else
-        feed_visible(t, &c, 1);
+        feed_visible(t, data + i, j - i);
+      i = j - 1;
      continue;
    }
    t->tagbuf[t->taglen++] = c;
@@ -2226,11 +2327,12 @@ int cst_agent_undo_available(const cst_agent_t *a) {
 * later changes reach the model as persisted notes appended to the history,
 * which keeps the cached prefix intact. */

-#define PROMPT_VERSION 2
+#define PROMPT_VERSION 3

 typedef struct {
  int plan;
  char approval[96];
+  char mode[16];
  char branch[128];
  char day[16];
  char notes[24]; /* hash of AGENTS.md, "" = none */
@@ -2240,6 +2342,7 @@ static void state_now(const cst_agent_t *a, pstate_t *s) {
  memset(s, 0, sizeof(*s));
  s->plan = a->cfg.plan_mode;
  snprintf(s->approval, sizeof(s->approval), "%s", approval_desc(&a->cfg));
+  snprintf(s->mode, sizeof(s->mode), "%s", cst_mode_name(a->cfg.mode));
  git_branch(a->cfg.workspace_root, s->branch, sizeof(s->branch));
  today(s->day, sizeof(s->day));
  if (a->notes && a->notes[0])
@@ -2252,6 +2355,7 @@ static cJSON *state_json(const pstate_t *s) {
  cJSON *o = cJSON_CreateObject();
  cJSON_AddNumberToObject(o, "plan", s->plan);
  cJSON_AddStringToObject(o, "approval", s->approval);
+  cJSON_AddStringToObject(o, "mode", s->mode);
  cJSON_AddStringToObject(o, "branch", s->branch);
  cJSON_AddStringToObject(o, "day", s->day);
  cJSON_AddStringToObject(o, "notes", s->notes);
@@ -2267,6 +2371,7 @@ static void state_parse(const cJSON *o, pstate_t *s) {
  if (cJSON_IsString(v))                                                       \
    snprintf(s->field, sizeof(s->field), "%s", v->valuestring);
  STATE_STR("approval", approval)
+  STATE_STR("mode", mode)
  STATE_STR("branch", branch)
  STATE_STR("day", day)
  STATE_STR("notes", notes)
@@ -2278,6 +2383,7 @@ static char *system_meta(const cst_agent_t *a, const pstate_t *s) {
  cJSON_AddStringToObject(o, "protocol", cst_protocol_name(a->cfg.protocol));
  cJSON_AddNumberToObject(o, "pv", PROMPT_VERSION);
  cJSON_AddNumberToObject(o, "tools", a->cfg.tool_profile);
+  cJSON_AddStringToObject(o, "mode", cst_mode_name(a->cfg.mode));
  cJSON_AddItemToObject(o, "state", state_json(s));
  char *out = cJSON_PrintUnformatted(o);
  cJSON_Delete(o);
@@ -2297,10 +2403,13 @@ static int system_current(const cst_message_t *m, const cst_agent_config_t *c) {
  const cJSON *pv = cJSON_GetObjectItemCaseSensitive(meta, "pv");
  const cJSON *pr = cJSON_GetObjectItemCaseSensitive(meta, "protocol");
  const cJSON *tp = cJSON_GetObjectItemCaseSensitive(meta, "tools");
+  const cJSON *md = cJSON_GetObjectItemCaseSensitive(meta, "mode");
  int ok = cJSON_IsNumber(pv) && pv->valueint == PROMPT_VERSION &&
           cJSON_IsString(pr) &&
           strcmp(pr->valuestring, cst_protocol_name(c->protocol)) == 0 &&
           cJSON_IsNumber(tp) && tp->valueint == c->tool_profile;
+  ok = ok && cJSON_IsString(md) &&
+       strcmp(md->valuestring, cst_mode_name(c->mode)) == 0;
  cJSON_Delete(meta);
  return ok;
 }
@@ -2378,6 +2487,14 @@ static void sync_notes(cst_agent_t *a) {
                       "is available again.\n");
  if (strcmp(now.approval, was.approval) != 0)
    cst_buf_printf(&b, "[cst] Approval mode is now %s.\n", now.approval);
+  if (strcmp(now.mode, was.mode) != 0)
+    cst_buf_printf(&b,
+                   "[cst] Working mode is now %s%s. Earlier instructions about "
+                   "how to frame the work no longer apply.\n",
+                   now.mode,
+                   strcmp(now.mode, "general") == 0
+                       ? " (general: documents, data, research)"
+                       : " (coding: repo, build, tests)");
  if (strcmp(now.branch, was.branch) != 0)
    cst_buf_printf(&b, "[cst] The git branch is now %s.\n",
                   now.branch[0] ? now.branch : "(none)");
@@ -2558,9 +2675,9 @@ static int can_continue(const cst_agent_t *a, const turn_t *t,
  return st->http_status > 0 && st->http_status < 400 && !st->client_stopped;
 }

-/* Prompt size at which the session is compacted: CST_CTX_BUDGET, else 60% of
- * the context window capped at 256K tokens. Even a cached prompt is billed
- * on every step, so huge windows are not worth filling. */
+/* Prompt size at which the session is compacted: CST_CTX_BUDGET, else 75% of
+ * the context window, capped at 256K tokens. Even a cached prompt is billed on
+ * every step, so a 1M-token window is not worth filling. */
 static long ctx_budget(const cst_agent_t *a) {
  const char *e = getenv("CST_CTX_BUDGET");
  if (e && atol(e) > 0)
@@ -2568,7 +2685,7 @@ static long ctx_budget(const cst_agent_t *a) {
  long win = cst_model_context_tokens(a->cfg.model);
  if (win <= 0)
    return 0;
-  long b = win / 10 * 6;
+  long b = win / 4 * 3;
  return b > 262144 ? 262144 : b;
 }

@@ -2933,13 +3050,18 @@ int cst_agent_new_session(cst_agent_t *a, char *id_out, size_t id_sz) {
    return -1;
  if (cst_agent_store_ensure(a->cfg.workspace_root) != 0)
    return -1;
-  char id[128];
-  if (cst_agent_store_new_id(id, sizeof(id)) != 0)
-    return -1;
-  char path[4096];
-  if (cst_agent_store_path_for_id(a->cfg.workspace_root, id, path,
-                                  sizeof(path)) != 0)
-    return -1;
+  char id[128], path[4096];
+  /* ids are per second: never reopen (and append to) one that exists */
+  for (int tries = 0;; tries++) {
+    if (tries == 16 || cst_agent_store_new_id(id, sizeof(id)) != 0 ||
+        cst_agent_store_path_for_id(a->cfg.workspace_root, id, path,
+                                    sizeof(path)) != 0)
+      return -1;
+    FILE *fp = fopen(path, "rb");
+    if (!fp)
+      break;
+    fclose(fp);
+  }
  if (agent_reopen_session(a, path) != 0)
    return -1;
  cst_agent_store_set_current(a->cfg.workspace_root, id);
@@ -2948,6 +3070,42 @@ int cst_agent_new_session(cst_agent_t *a, char *id_out, size_t id_sz) {
  return 0;
 }

+int cst_agent_set_workspace(cst_agent_t *a, const char *path, char *err,
+                            size_t err_sz) {
+  char dummy[8];
+  if (!err || !err_sz) {
+    err = dummy;
+    err_sz = sizeof(dummy);
+  }
+  err[0] = 0;
+  if (!a || !path || !path[0]) {
+    snprintf(err, err_sz, "empty path");
+    return -1;
+  }
+  char abs[sizeof(a->cfg.workspace_root)];
+  if (os_realpath(path, abs, sizeof(abs)) != 0) {
+    snprintf(err, err_sz, "no such directory: %s", path);
+    return -1;
+  }
+  /* a file, or a tree we may not write to, fails here instead of halfway
+   * through the next turn */
+  if (cst_agent_store_ensure(abs) != 0) {
+    snprintf(err, err_sz, "cannot use as a workspace: %s", abs);
+    return -1;
+  }
+  snprintf(a->cfg.workspace_root, sizeof(a->cfg.workspace_root), "%s", abs);
+  /* everything that is kept per workspace follows the new root */
+  cst_perms_close(a->perms);
+  a->perms = cst_perms_open(a->cfg.workspace_root);
+  cst_metrics_close(a->metrics);
+  a->metrics = cst_metrics_open(a->cfg.workspace_root,
+                                cst_session_path(a->session));
+  cst_agent_reload_notes(a);
+  ensure_system(a);          /* the prompt names the directory */
+  a->last_prompt_tokens = 0; /* the old budget was measured elsewhere */
+  return 0;
+}
+
 int cst_agent_switch_session(cst_agent_t *a, const char *id) {
  if (!a || !id || !id[0])
    return -1;
@@ -57,7 +57,8 @@ static void make_id(char *out, size_t n) {
  time_t now = time(NULL);
  struct tm tm;
  os_localtime(&now, &tm);
-  unsigned r = (unsigned)(os_getpid() ^ (unsigned)now);
+  static unsigned seq;
+  unsigned r = (unsigned)(os_getpid() ^ (unsigned)now) + 0x9e37u * seq++;
  snprintf(out, n, "%04d%02d%02d-%02d%02d%02d-%04x", tm.tm_year + 1900,
           tm.tm_mon + 1, tm.tm_mday, tm.tm_hour, tm.tm_min, tm.tm_sec,
           r & 0xffff);
@@ -4,6 +4,7 @@

 #include "cJSON.h"
 #include "cst_attach.h"
+#include "cst_pdf.h"
 #include "cst_util.h"
 #include "os_compat.h"

@@ -102,6 +103,9 @@ static int le16(const unsigned char *p) {
 }

 static const char *mime_of_ext(const char *mime) {
+  if (strcmp(mime, "application/pdf") == 0 ||
+      strcmp(mime, "text/plain") == 0)
+    return "txt"; /* a document is stored as the text it was reduced to */
  if (strcmp(mime, "image/png") == 0)
    return "png";
  if (strcmp(mime, "image/jpeg") == 0)
@@ -352,23 +356,73 @@ static void make_id(const char *name, const char *data, size_t n, char *out,
  snprintf(out, out_sz, "img-%08x", (unsigned)(h & 0xFFFFFFFFu));
 }

+/* n bytes, NUL-terminated: a text payload is handed around as a C string. */
+static char *memdup_z(const char *data, size_t n) {
+  char *copy = malloc(n + 1);
+  if (!copy)
+    return NULL;
+  memcpy(copy, data, n);
+  copy[n] = 0;
+  return copy;
+}
+
+/* What a picked file is, plus the payload to store for it: an image stays as
+ * it is, a PDF is reduced to its text layer (the mime still says PDF, the
+ * stored file holds text), anything else must be UTF-8 text. The caller frees
+ * the result; *out_len is its length. */
+static char *prepare(const char *data, size_t n, char *mime, size_t mime_sz,
+                     int *pages, size_t *out_len, char *err, size_t err_sz) {
+  *pages = 0;
+  *out_len = n;
+  if (cst_attach_sniff(data, n, mime, mime_sz) == 0) {
+    if (n > CST_ATTACH_MAX_BYTES) {
+      snprintf(err, err_sz, "image is %.1f MB; the limit is %d MB",
+               (double)n / (1024 * 1024), (int)(CST_ATTACH_MAX_BYTES >> 20));
+      return NULL;
+    }
+    char *copy = memdup_z(data, n);
+    if (!copy)
+      snprintf(err, err_sz, "out of memory");
+    return copy;
+  }
+  if (n >= 4 && memcmp(data, "%PDF", 4) == 0) {
+    int pg = 0;
+    char perr[256];
+    char *text = cst_pdf_text(data, n, &pg, perr, sizeof(perr));
+    if (!text) {
+      snprintf(err, err_sz, "cannot read the PDF: %.200s", perr);
+      return NULL;
+    }
+    snprintf(mime, mime_sz, "application/pdf");
+    *pages = pg;
+    *out_len = strlen(text);
+    return text;
+  }
+  if (n > CST_ATTACH_TEXT_MAX) {
+    snprintf(err, err_sz, "text is %.1f MB; the limit is %d MB",
+             (double)n / (1024 * 1024), (int)(CST_ATTACH_TEXT_MAX >> 20));
+    return NULL;
+  }
+  if (memchr(data, 0, n) || !cst_utf8_valid(data, n)) {
+    snprintf(err, err_sz,
+             "not an image (PNG, JPEG, GIF, WebP or BMP), a PDF or UTF-8 "
+             "text; the content decides, not the extension");
+    return NULL;
+  }
+  snprintf(mime, mime_sz, "text/plain");
+  char *copy = memdup_z(data, n);
+  if (!copy)
+    snprintf(err, err_sz, "out of memory");
+  return copy;
+}
+
+/* Write one payload whose kind is already known. */
 static cst_attachment_t *store(const char *session_path, const char *name,
-                               const char *data, size_t n, const char *origin,
-                               char *err, size_t err_sz) {
+                               const char *data, size_t n, const char *mime,
+                               int pages, const char *origin, char *err,
+                               size_t err_sz) {
  if (!name || !name[0])
    name = "image";
-  char mime[32];
-  if (cst_attach_sniff(data, n, mime, sizeof(mime)) != 0) {
-    snprintf(err, err_sz,
-             "not a supported image (PNG, JPEG, GIF, WebP or BMP; "
-             "checked by content, not by extension)");
-    return NULL;
-  }
-  if (n > CST_ATTACH_MAX_BYTES) {
-    snprintf(err, err_sz, "image is %.1f MB; the limit is %d MB",
-             (double)n / (1024 * 1024), (int)(CST_ATTACH_MAX_BYTES >> 20));
-    return NULL;
-  }
  char dir[OS_PATH_MAX];
  if (cst_attach_dir(session_path, dir, sizeof(dir)) != 0) {
    snprintf(err, err_sz, "cannot create the attachment directory");
@@ -386,7 +440,9 @@ static cst_attachment_t *store(const char *session_path, const char *name,
  snprintf(a->origin, sizeof(a->origin), "%s", origin ? origin : "user");
  snprintf(a->detail, sizeof(a->detail), "auto");
  a->bytes = (long long)n;
-  cst_attach_dimensions(data, n, &a->width, &a->height);
+  a->pages = pages;
+  if (cst_attach_is_image(a))
+    cst_attach_dimensions(data, n, &a->width, &a->height);

  char full[OS_PATH_MAX];
  if (cst_attach_full_path(session_path, a, full, sizeof(full)) != 0) {
@@ -419,14 +475,26 @@ cst_attachment_t *cst_attach_import_bytes(const char *session_path,
  }
  err[0] = 0;
  if (!session_path || !session_path[0]) {
-    snprintf(err, err_sz, "no session to store the image in");
+    snprintf(err, err_sz, "no session to store the file in");
    return NULL;
  }
  if (!data || !n) {
-    snprintf(err, err_sz, "image is empty");
+    snprintf(err, err_sz, "the file is empty");
    return NULL;
  }
-  return store(session_path, name, data, n, "user", err, err_sz);
+  char mime[32];
+  if (cst_attach_sniff(data, n, mime, sizeof(mime)) != 0) {
+    snprintf(err, err_sz,
+             "not a supported image (PNG, JPEG, GIF, WebP or BMP; "
+             "checked by content, not by extension)");
+    return NULL;
+  }
+  if (n > CST_ATTACH_MAX_BYTES) {
+    snprintf(err, err_sz, "image is %.1f MB; the limit is %d MB",
+             (double)n / (1024 * 1024), (int)(CST_ATTACH_MAX_BYTES >> 20));
+    return NULL;
+  }
+  return store(session_path, name, data, n, mime, 0, "user", err, err_sz);
 }

 cst_attachment_t *cst_attach_import_file(const char *session_path,
@@ -449,11 +517,30 @@ cst_attachment_t *cst_attach_import_file(const char *session_path,
  for (const char *p = src_path; *p; p++)
    if (*p == '/' || *p == '\\')
      leaf = p + 1;
-  cst_attachment_t *a = store(session_path, leaf, data, n, "user", err, err_sz);
+  char mime[32];
+  int pages = 0;
+  size_t pn = 0;
+  char *payload =
+      prepare(data, n, mime, sizeof(mime), &pages, &pn, err, err_sz);
  free(data);
+  if (!payload)
+    return NULL;
+  cst_attachment_t *a =
+      store(session_path, leaf, payload, pn, mime, pages, "user", err, err_sz);
+  free(payload);
  return a;
 }

+/* Images are opaque and travel as image parts; everything else is text the
+ * store holds and the request inlines. */
+int cst_attach_is_image(const cst_attachment_t *a) {
+  return a && strncmp(a->mime, "image/", 6) == 0;
+}
+
+int cst_attach_is_text(const cst_attachment_t *a) {
+  return a && !cst_attach_is_image(a);
+}
+
 char *cst_attach_read(const char *session_path, const cst_attachment_t *a,
                      size_t *out_len, char *err, size_t err_sz) {
  char dummy[8];
@@ -470,11 +557,12 @@ char *cst_attach_read(const char *session_path, const cst_attachment_t *a,
  char *data = read_whole(full, out_len, CST_ATTACH_MAX_BYTES);
  if (!data) {
    snprintf(err, err_sz,
-             "%.200s is missing; the image was deleted from the session "
-             "store",
+             "%.200s is missing; it was deleted from the session store",
             a->name);
    return NULL;
  }
+  if (!cst_attach_is_image(a)) /* a document is text, nothing to verify */
+    return data;
  char mime[32];
  if (cst_attach_sniff(data, *out_len, mime, sizeof(mime)) != 0) {
    snprintf(err, err_sz, "%.200s in the session store is no longer an image",
@@ -486,6 +574,8 @@ char *cst_attach_read(const char *session_path, const cst_attachment_t *a,
 }

 char *cst_attach_data_url(const char *session_path, const cst_attachment_t *a) {
+  if (!cst_attach_is_image(a)) /* a document travels as text, not as a URL */
+    return NULL;
  char err[256];
  size_t n = 0;
  char *data = cst_attach_read(session_path, a, &n, err, sizeof(err));
@@ -509,10 +599,21 @@ void cst_attach_describe(const cst_attachment_t *a, char *out, size_t out_sz) {
    snprintf(out, out_sz, "(none)");
    return;
  }
-  const char *kind = strncmp(a->mime, "image/", 6) == 0 ? a->mime + 6 : a->mime;
+  const char *kind = a->mime;
  char dim[32] = "";
-  if (a->width > 0 && a->height > 0)
-    snprintf(dim, sizeof(dim), "%dx%d ", a->width, a->height);
+  if (cst_attach_is_image(a)) {
+    if (strncmp(a->mime, "image/", 6) == 0)
+      kind = a->mime + 6;
+    if (a->width > 0 && a->height > 0)
+      snprintf(dim, sizeof(dim), "%dx%d ", a->width, a->height);
+  } else if (strcmp(a->mime, "application/pdf") == 0) {
+    kind = "pdf";
+    if (a->pages > 0)
+      snprintf(dim, sizeof(dim), "%d page%s ", a->pages,
+               a->pages == 1 ? "" : "s");
+  } else if (strcmp(a->mime, "text/plain") == 0 || !a->mime[0]) {
+    kind = "text";
+  }
  out[0] = 0;
  if (a->bytes >= 1024 * 1024)
    snprintf(out, out_sz, "%s%s %.1f MB", dim, kind,
@@ -580,6 +681,8 @@ char *cst_attachments_to_json(const cst_attachment_t *list, int n) {
      cJSON_AddNumberToObject(o, "width", a->width);
      cJSON_AddNumberToObject(o, "height", a->height);
    }
+    if (a->pages > 0)
+      cJSON_AddNumberToObject(o, "pages", a->pages);
    if (a->detail[0] && strcmp(a->detail, "auto") != 0)
      cJSON_AddStringToObject(o, "detail", a->detail);
    if (a->origin[0])
@@ -628,6 +731,8 @@ cst_attachment_t *cst_attachments_from_json(const void *array_json, int *n_out)
    a->width = cJSON_IsNumber(v) ? v->valueint : 0;
    v = cJSON_GetObjectItemCaseSensitive(it, "height");
    a->height = cJSON_IsNumber(v) ? v->valueint : 0;
+    v = cJSON_GetObjectItemCaseSensitive(it, "pages");
+    a->pages = cJSON_IsNumber(v) ? v->valueint : 0;
    v = cJSON_GetObjectItemCaseSensitive(it, "detail");
    snprintf(a->detail, sizeof(a->detail), "%s",
             cJSON_IsString(v) ? v->valuestring : "auto");
@@ -390,6 +390,175 @@ int cst_clipboard_image(char **data, size_t *n, char *err, size_t err_sz) {

 #endif

+/* --- text clipboard ------------------------------------------------------ */
+
+#ifdef _WIN32
+
+int cst_clipboard_text(char **data, size_t *n, char *err, size_t err_sz) {
+  *data = NULL;
+  *n = 0;
+  if (!IsClipboardFormatAvailable(CF_UNICODETEXT) || !OpenClipboard(NULL)) {
+    set_err(err, err_sz, "no text on the clipboard");
+    return -1;
+  }
+  HANDLE h = GetClipboardData(CF_UNICODETEXT);
+  const wchar_t *w = h ? GlobalLock(h) : NULL;
+  if (!w) {
+    if (h)
+      GlobalUnlock(h);
+    CloseClipboard();
+    set_err(err, err_sz, "no text on the clipboard");
+    return -1;
+  }
+  char *utf8 = os_wide_to_utf8(w);
+  size_t len = utf8 ? strlen(utf8) : 0;
+  GlobalUnlock(h);
+  CloseClipboard();
+  if (!utf8 || !len) {
+    free(utf8);
+    set_err(err, err_sz, "no text on the clipboard");
+    return -1;
+  }
+  if (len > CST_CLIPBOARD_TEXT_MAX) {
+    free(utf8);
+    set_err(err, err_sz, "clipboard text is too large");
+    return -1;
+  }
+  *data = utf8;
+  *n = len;
+  return 0;
+}
+
+int cst_clipboard_set_text(const char *text, size_t n, char *err,
+                           size_t err_sz) {
+  if (!text || os_clipboard_set(text, n) != 0) {
+    set_err(err, err_sz, "could not put text on the clipboard");
+    return -1;
+  }
+  return 0;
+}
+
+#else
+
+/* stdout of a clipboard reader, capped and NUL-terminated. */
+static int clip_read(const char *cmd, const char *tool, char **data, size_t *n,
+                     char *err, size_t err_sz) {
+  os_run_result_t r;
+  memset(&r, 0, sizeof(r));
+  if (os_run_shell_c(NULL, cmd, CST_CLIPBOARD_TEXT_MAX + 1, &r) != 0) {
+    os_run_result_free(&r);
+    set_err(err, err_sz, "could not read the clipboard");
+    return -1;
+  }
+  if (r.exit_code != 0 || r.out_len == 0) {
+    int missing = r.exit_code == 127;
+    os_run_result_free(&r);
+    if (missing) {
+      char msg[96];
+      snprintf(msg, sizeof(msg), "reading the clipboard needs %s", tool);
+      set_err(err, err_sz, msg);
+    } else {
+      set_err(err, err_sz, "no text on the clipboard");
+    }
+    return -1;
+  }
+  if (r.out_len > CST_CLIPBOARD_TEXT_MAX) {
+    os_run_result_free(&r);
+    set_err(err, err_sz, "clipboard text is too large");
+    return -1;
+  }
+  char *copy = malloc(r.out_len + 1);
+  if (!copy) {
+    os_run_result_free(&r);
+    set_err(err, err_sz, "out of memory");
+    return -1;
+  }
+  memcpy(copy, r.out, r.out_len);
+  copy[r.out_len] = 0;
+  os_run_result_free(&r);
+  *data = copy;
+  *n = r.out_len;
+  return 0;
+}
+
+int cst_clipboard_text(char **data, size_t *n, char *err, size_t err_sz) {
+  *data = NULL;
+  *n = 0;
+#ifdef __APPLE__
+  return clip_read("pbpaste 2>/dev/null", "pbpaste", data, n, err, err_sz);
+#else
+  const char *wl = getenv("WAYLAND_DISPLAY");
+  if (wl && *wl)
+    return clip_read("wl-paste --type text/plain --no-newline 2>/dev/null",
+                     "wl-clipboard (wl-paste)", data, n, err, err_sz);
+  const char *x = getenv("DISPLAY");
+  if (x && *x)
+    return clip_read("xclip -selection clipboard -o 2>/dev/null", "xclip",
+                     data, n, err, err_sz);
+  set_err(err, err_sz, "no clipboard here (no Wayland or X11 display)");
+  return -1;
+#endif
+}
+
+int cst_clipboard_set_text(const char *text, size_t n, char *err,
+                           size_t err_sz) {
+  if (!text)
+    return -1;
+  if (n > CST_CLIPBOARD_TEXT_MAX) {
+    set_err(err, err_sz, "text is too large for the clipboard");
+    return -1;
+  }
+  const char *fmt = NULL, *tool = NULL;
+#ifdef __APPLE__
+  fmt = "pbcopy < '%s'";
+  tool = "pbcopy";
+#else
+  if (getenv("WAYLAND_DISPLAY")) {
+    fmt = "wl-copy < '%s'";
+    tool = "wl-clipboard (wl-copy)";
+  } else if (getenv("DISPLAY")) {
+    fmt = "xclip -selection clipboard -i < '%s'";
+    tool = "xclip";
+  }
+#endif
+  if (!fmt) {
+    set_err(err, err_sz, "no clipboard here (no Wayland or X11 display)");
+    return -1;
+  }
+  char tmp[OS_PATH_MAX];
+  int fd = os_mkstemp(tmp, sizeof(tmp), "cst-clip");
+  if (fd < 0) {
+    set_err(err, err_sz, "could not create a temporary file");
+    return -1;
+  }
+  FILE *fp = fdopen(fd, "wb");
+  int ok = fp && fwrite(text, 1, n, fp) == n && fclose(fp) == 0;
+  if (!ok) {
+    if (fp)
+      fclose(fp);
+    remove(tmp);
+    set_err(err, err_sz, "could not write the text");
+    return -1;
+  }
+  char cmd[OS_PATH_MAX + 64];
+  snprintf(cmd, sizeof(cmd), fmt, tmp);
+  os_run_result_t r;
+  memset(&r, 0, sizeof(r));
+  int rc = os_run_shell_c(NULL, cmd, 4096, &r);
+  int code = r.exit_code;
+  os_run_result_free(&r);
+  remove(tmp);
+  if (rc != 0 || code != 0) {
+    char msg[96];
+    snprintf(msg, sizeof(msg), "writing the clipboard needs %s", tool);
+    set_err(err, err_sz, msg);
+    return -1;
+  }
+  return 0;
+}
+
+#endif
+
 static int hexval(int c) {
  if (c >= '0' && c <= '9')
    return c - '0';
@@ -4,8 +4,11 @@

 #include "cst_http.h"
 #include "cst_util.h"
+#include "os_compat.h"

+#include <errno.h>
 #include <pthread.h>
+#include <stdint.h>
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
@@ -44,6 +47,9 @@ struct cst_http_conn {
  int sent;
  int streaming;
  int failed;
+  int nreq;            /* requests served on this connection */
+  int expect_continue; /* the client waits for "100 Continue" */
+  int keep_alive_req;  /* what the client asked for */
  cst_http_conn_t *prev, *next;
 };

@@ -72,10 +78,15 @@ static const char *status_text(int code) {

 /* --- reading -------------------------------------------------------------- */

-static int fill(cst_http_conn_t *c, int timeout_ms) {
+/* deadline: os_now_ms() value; the whole read must be done by then, so a
+ * client trickling one byte at a time cannot hold the thread forever. */
+static int fill(cst_http_conn_t *c, double deadline) {
  if (c->rpos < c->rlen)
    return 1;
-  long r = os_net_recv(c->sock, c->rbuf, sizeof(c->rbuf), timeout_ms);
+  double left = deadline - os_now_ms();
+  if (left <= 0)
+    return -2;
+  long r = os_net_recv(c->sock, c->rbuf, sizeof(c->rbuf), (int)left + 1);
  if (r <= 0)
    return (int)r; /* 0 eof, -1 error, -2 timeout */
  c->rpos = 0;
@@ -83,10 +94,12 @@ static int fill(cst_http_conn_t *c, int timeout_ms) {
  return 1;
 }

-static int rd_line(cst_http_conn_t *c, char *out, size_t cap, int timeout_ms) {
+#define RD_TOO_LONG (-3)
+
+static int rd_line(cst_http_conn_t *c, char *out, size_t cap, double deadline) {
  size_t n = 0;
  for (;;) {
-    int rc = fill(c, timeout_ms);
+    int rc = fill(c, deadline);
    if (rc <= 0)
      return rc;
    while (c->rpos < c->rlen) {
@@ -98,7 +111,7 @@ static int rd_line(cst_http_conn_t *c, char *out, size_t cap, int timeout_ms) {
        return 1;
      }
      if (n + 2 >= cap)
-        return -1;
+        return RD_TOO_LONG;
      out[n++] = ch;
    }
  }
@@ -132,7 +145,8 @@ static char *rd_body(cst_http_conn_t *c, size_t n, int timeout_ms) {
  return buf;
 }

-static void url_decode(const char *in, char *out, size_t cap, int plus_space) {
+/* -1 when the input encodes a NUL (out then holds what came before it). */
+static int url_decode(const char *in, char *out, size_t cap, int plus_space) {
  size_t o = 0;
  for (const char *p = in; *p && o + 1 < cap; p++) {
    if (*p == '%' && p[1] && p[2]) {
@@ -146,8 +160,10 @@ static void url_decode(const char *in, char *out, size_t cap, int plus_space) {
                                            : -1;
      if (hi >= 0 && lo >= 0) {
        char ch = (char)((hi << 4) | lo);
-        if (ch == 0)
-          return; /* a decoded NUL would truncate the path */
+        if (ch == 0) {
+          out[o] = 0;
+          return -1;
+        }
        out[o++] = ch;
        p += 2;
        continue;
@@ -156,6 +172,23 @@ static void url_decode(const char *in, char *out, size_t cap, int plus_space) {
    out[o++] = (*p == '+' && plus_space) ? ' ' : *p;
  }
  out[o] = 0;
+  return 0;
+}
+
+/* Comma-separated header tokens, compared without case ("Close"). */
+static int has_token(const char *val, const char *tok) {
+  size_t n = strlen(tok);
+  for (const char *p = val; p && *p;) {
+    while (*p == ' ' || *p == '\t' || *p == ',')
+      p++;
+    size_t i = 0;
+    while (i < n && p[i] && (p[i] | 0x20) == (tok[i] | 0x20))
+      i++;
+    if (i == n && (p[i] == 0 || p[i] == ',' || p[i] == ' ' || p[i] == '\t'))
+      return 1;
+    p = strchr(p, ',');
+  }
+  return 0;
 }

 static int key_is(const char *line, const char *key) {
@@ -189,12 +222,16 @@ static int read_request(cst_http_conn_t *c, size_t *want_body) {
  cst_http_server_t *s = c->srv;
  char line[HTTP_LINE_MAX];
  cst_http_req_t *r = &c->req;
-  int first = c->rpos >= c->rlen; /* nothing buffered: this is a new client */
+  int first = c->nreq == 0; /* a keep-alive pause gets the idle timeout */
  memset(r, 0, sizeof(*r));
  c->hdrs_used = 0;
+  c->expect_continue = 0;
  *want_body = 0;
  int rc = rd_line(c, line, sizeof(line),
-                   first ? s->o.read_timeout_ms : s->o.idle_timeout_ms);
+                   os_now_ms() + (first ? s->o.read_timeout_ms
+                                        : s->o.idle_timeout_ms));
+  if (rc == RD_TOO_LONG)
+    return -1;
  if (rc <= 0)
    return 0;
  char *sp1 = strchr(line, ' ');
@@ -214,9 +251,10 @@ static int read_request(cst_http_conn_t *c, size_t *want_body) {
  r->keep_alive = strncmp(version, "HTTP/1.", 7) == 0 &&
                  version[7] >= '1' && version[7] <= '9';
  /* absolute-form ("http://host/path") is legal: keep the path */
-  char *scheme = strstr(target, "://");
-  if (scheme)
-    target = strchr(scheme + 3, '/');
+  if (target[0] != '/') {
+    char *scheme = strstr(target, "://");
+    target = scheme ? strchr(scheme + 3, '/') : NULL;
+  }
  if (!target || target[0] != '/')
    return -1;
  {
@@ -227,7 +265,8 @@ static int read_request(cst_http_conn_t *c, size_t *want_body) {
      *q = 0;
      snprintf(r->query, sizeof(r->query), "%s", q + 1);
    }
-    url_decode(raw, r->path, sizeof(r->path), 0);
+    if (url_decode(raw, r->path, sizeof(r->path), 0) != 0)
+      return -1;
    size_t plen = strlen(r->path);
    if (plen == 0 || strstr(r->path, ".."))
      return -1;
@@ -235,8 +274,9 @@ static int read_request(cst_http_conn_t *c, size_t *want_body) {
  size_t hdr_bytes = 0;
  size_t content_length = 0;
  int have_length = 0;
+  double deadline = os_now_ms() + s->o.read_timeout_ms;
  for (;;) {
-    rc = rd_line(c, line, sizeof(line), s->o.read_timeout_ms);
+    rc = rd_line(c, line, sizeof(line), deadline);
    if (rc <= 0)
      return -1;
    if (!line[0])
@@ -251,18 +291,29 @@ static int read_request(cst_http_conn_t *c, size_t *want_body) {
    while (*val == ' ' || *val == '\t')
      val++;
    if (key_is(line, "content-length")) {
-      content_length = (size_t)strtoul(val, NULL, 10);
+      /* digits only, and one value: "5x" or two lengths would let the
+       * rest of the body run as a second request behind a proxy */
+      char *end = NULL;
+      errno = 0;
+      unsigned long long v = strtoull(val, &end, 10);
+      while (end && (*end == ' ' || *end == '\t'))
+        end++;
+      if (*val < '0' || *val > '9' || !end || *end || errno == ERANGE ||
+          v > (unsigned long long)SIZE_MAX ||
+          (have_length && content_length != (size_t)v))
+        return -1;
+      content_length = (size_t)v;
      have_length = 1;
    } else if (key_is(line, "transfer-encoding")) {
      return -1; /* chunked request bodies are not supported */
    } else if (key_is(line, "connection")) {
-      if (strstr(val, "close"))
+      if (has_token(val, "close"))
        r->keep_alive = 0;
-      else if (strstr(val, "keep-alive"))
+      else if (has_token(val, "keep-alive"))
        r->keep_alive = 1;
    } else if (key_is(line, "expect")) {
-      const char *cont = "HTTP/1.1 100 Continue\r\n\r\n";
-      os_net_send_all(c->sock, cont, strlen(cont), 5000);
+      /* answered once auth and the size cap have passed */
+      c->expect_continue = has_token(val, "100-continue");
    } else if (key_is(line, "authorization")) {
      r->authorization = store_str(c, val);
    } else if (key_is(line, "cookie")) {
@@ -275,6 +326,8 @@ static int read_request(cst_http_conn_t *c, size_t *want_body) {
      r->content_type = store_str(c, val);
    } else if (key_is(line, "user-agent")) {
      r->user_agent = store_str(c, val);
+    } else if (key_is(line, "last-event-id")) {
+      r->last_event_id = store_str(c, val);
    }
  }
  if (have_length)
@@ -429,24 +482,30 @@ int cst_http_query(const cst_http_req_t *req, const char *key, char *out,

 /* --- connections ---------------------------------------------------------- */

-static void dispatch(cst_http_server_t *s, cst_http_conn_t *c) {
-  const cst_http_route_t *route = NULL;
+static const cst_http_route_t *find_route(const cst_http_server_t *s,
+                                          const cst_http_req_t *req) {
  const cst_http_route_t *fallback = NULL;
  for (int i = 0; i < s->nroutes; i++) {
    const cst_http_route_t *r = &s->routes[i];
-    if (strcmp(r->method, "*") != 0 && strcmp(r->method, c->req.method) != 0)
+    if (strcmp(r->method, "*") != 0 && strcmp(r->method, req->method) != 0)
      continue;
    if (strcmp(r->path, "*") == 0) {
      fallback = r;
      continue;
    }
-    if (strcmp(r->path, c->req.path) == 0) {
-      route = r;
-      break;
-    }
+    if (strcmp(r->path, req->path) == 0)
+      return r;
  }
-  if (!route)
-    route = fallback;
+  return fallback;
+}
+
+/* Route, auth and the size cap are settled before the body is read: an
+ * unauthenticated client must not make us buffer 32 MiB. */
+static void serve(cst_http_server_t *s, cst_http_conn_t *c, size_t want) {
+  const cst_http_route_t *route = find_route(s, &c->req);
+  size_t cap = route && route->max_body ? route->max_body : s->o.max_body;
+  if (want > 0)
+    c->req.keep_alive = 0; /* until the body is read, the stream is ours */
  if (!route) {
    cst_http_send_err(c, 404, "not found");
    return;
@@ -455,6 +514,25 @@ static void dispatch(cst_http_server_t *s, cst_http_conn_t *c) {
    cst_http_send_err(c, 401, "unauthorized");
    return;
  }
+  if (want > cap) {
+    cst_http_send_err(c, 413, "request body too large");
+    return;
+  }
+  if (want > 0) {
+    if (c->expect_continue) {
+      const char *cont = "HTTP/1.1 100 Continue\r\n\r\n";
+      os_net_send_all(c->sock, cont, strlen(cont), 5000);
+    }
+    char *body = rd_body(c, want, s->o.read_timeout_ms);
+    if (!body) {
+      c->failed = 1;
+      return;
+    }
+    c->body = body;
+    c->req.body = body;
+    c->req.body_len = want;
+    c->req.keep_alive = c->keep_alive_req;
+  }
  int rc = route->fn(c, route->ud);
  if (!c->sent)
    cst_http_send_err(c, rc == 0 ? 500 : 404,
@@ -477,32 +555,14 @@ static void *conn_main(void *arg) {
    }
    if (rc == 0)
      break;
-    size_t cap = s->o.max_body;
-    for (int i = 0; i < s->nroutes; i++) {
-      const cst_http_route_t *r = &s->routes[i];
-      if (r->max_body && strcmp(r->method, c->req.method) == 0 &&
-          strcmp(r->path, c->req.path) == 0)
-        cap = r->max_body;
-    }
-    if (want > cap) {
-      c->req.keep_alive = 0;
-      cst_http_send_err(c, 413, "request body too large");
-      break;
-    }
-    if (want > 0) {
-      char *body = rd_body(c, want, s->o.read_timeout_ms);
-      if (!body)
-        break;
-      c->body = body;
-      c->req.body = body;
-      c->req.body_len = want;
-    }
+    c->nreq++;
    c->sent = 0;
    c->streaming = 0;
-    dispatch(s, c);
+    c->keep_alive_req = c->req.keep_alive;
+    serve(s, c, want);
    if (s->o.on_log) {
      char line[512];
-      snprintf(line, sizeof(line), "%s %s -> %d", c->req.method, c->req.path,
+      snprintf(line, sizeof(line), "%s %.480s -> %d", c->req.method, c->req.path,
               c->failed ? 499 : (c->sent ? 200 : 500));
      s->o.on_log(line, s->o.log_ud);
    }
@@ -512,11 +572,12 @@ static void *conn_main(void *arg) {
    c->req.body_len = 0;
    c->req.authorization = c->req.cookie = c->req.origin = NULL;
    c->req.host = c->req.content_type = c->req.user_agent = NULL;
+    c->req.last_event_id = NULL;
    if (!c->req.keep_alive || s->stopping || c->failed)
      break;
  }
-  os_net_shutdown(c->sock);
-  os_net_close(c->sock);
+  /* unlink first: stop() shuts down every listed socket, and a closed fd
+   * number may already belong to someone else */
  pthread_mutex_lock(&s->mu);
  if (c->prev)
    c->prev->next = c->next;
@@ -524,10 +585,14 @@ static void *conn_main(void *arg) {
    s->conns_first = c->next;
  if (c->next)
    c->next->prev = c->prev;
+  pthread_mutex_unlock(&s->mu);
+  os_net_shutdown(c->sock);
+  os_net_close(c->sock);
+  free(c);
+  pthread_mutex_lock(&s->mu);
  s->nconn--;
  pthread_cond_broadcast(&s->cv);
  pthread_mutex_unlock(&s->mu);
-  free(c);
  return NULL;
 }

@@ -536,8 +601,22 @@ static void *accept_main(void *arg) {
  while (!s->stopping) {
    int timed_out = 0;
    os_socket_t sock = os_net_accept(s->listener, 500, &timed_out);
-    if (sock == OS_SOCKET_BAD)
+    if (sock == OS_SOCKET_BAD) {
+      if (!timed_out) { /* EMFILE and friends: the listener stays readable */
+        struct timespec ts;
+        clock_gettime(CLOCK_REALTIME, &ts);
+        ts.tv_nsec += 100000000;
+        if (ts.tv_nsec >= 1000000000) {
+          ts.tv_sec++;
+          ts.tv_nsec -= 1000000000;
+        }
+        pthread_mutex_lock(&s->mu);
+        if (!s->stopping)
+          pthread_cond_timedwait(&s->cv, &s->mu, &ts);
+        pthread_mutex_unlock(&s->mu);
+      }
      continue;
+    }
    if (s->stopping) {
      os_net_close(sock);
      break;
@@ -644,9 +723,9 @@ int cst_http_stopping(const cst_http_server_t *s) {
  return s ? s->stopping : 1;
 }

-void cst_http_stop(cst_http_server_t *s) {
+int cst_http_stop(cst_http_server_t *s) {
  if (!s)
-    return;
+    return 0;
  pthread_mutex_lock(&s->mu);
  s->stopping = 1;
  for (cst_http_conn_t *c = s->conns_first; c; c = c->next)
@@ -670,13 +749,20 @@ void cst_http_stop(cst_http_server_t *s) {
    }
    pthread_cond_timedwait(&s->cv, &s->mu, &ts);
  }
+  int left = s->nconn;
  pthread_mutex_unlock(&s->mu);
  os_net_close(s->listener);
+  return left;
 }

 void cst_http_free(cst_http_server_t *s) {
  if (!s)
    return;
+  pthread_mutex_lock(&s->mu);
+  int left = s->nconn;
+  pthread_mutex_unlock(&s->mu);
+  if (left > 0)
+    return; /* a handler still runs on it (see stop()): leak, not crash */
  pthread_mutex_destroy(&s->mu);
  pthread_cond_destroy(&s->cv);
  free(s->routes);
@@ -22,43 +22,75 @@
 #define ERRBODY_MAX 4096
 #define LLM_MAX_RETRIES 3

-/* Point libcurl at a CA bundle on Windows.
+static int ca_file_ok(const char *path) {
+  if (!path || !path[0])
+    return 0;
+  FILE *f = fopen(path, "rb");
+  if (!f)
+    return 0;
+  fclose(f);
+  return 1;
+}
+
+/* Point libcurl at a CA bundle.
 *
- * An OpenSSL-based libcurl keeps the path it was built with -- MSYS2's is
- * "/mingw64/etc/ssl/certs/ca-bundle.crt", which resolves to nothing once the
- * binary runs outside an MSYS2 shell, and that build ignores CURL_CA_BUNDLE.
- * Every HTTPS request then dies with CURLE_SSL_CACERT_BADFILE (77).
+ * libcurl keeps the bundle path it was built with. On Windows an OpenSSL
+ * build carries MSYS2's "/mingw64/etc/ssl/certs/ca-bundle.crt"; the static
+ * Linux release is built on one distro and runs on all of them, whose bundles
+ * live in different places. A path that does not exist fails every HTTPS
+ * request with CURLE_SSL_CACERT_BADFILE (77).
 *
- * Order: $CST_CA_BUNDLE, then ca-bundle.crt beside the executable. If neither
- * exists nothing is set, so a libcurl linked against Schannel (the Windows
- * certificate store) behaves exactly as before. */
-static void llm_apply_ca_bundle(CURL *curl) {
-#if defined(_WIN32)
+ * Order: $CST_CA_BUNDLE; on Windows ca-bundle.crt beside the executable (if
+ * absent nothing is set, so Schannel uses the certificate store); elsewhere
+ * $SSL_CERT_FILE, libcurl's own default if that file exists, then the usual
+ * distro locations. Decided once per process. */
+void cst_curl_apply_ca(void *curl) {
  static char ca[OS_PATH_MAX];
-  static int checked;
-  if (!checked) {
-    checked = 1;
+  static atomic_int state; /* 0 unchecked, 1 checking, 2 done */
+  int expect = 0;
+  if (atomic_compare_exchange_strong(&state, &expect, 1)) {
    const char *env = getenv("CST_CA_BUNDLE");
-    if (env && env[0]) {
+    if (ca_file_ok(env)) {
      snprintf(ca, sizeof(ca), "%s", env);
    } else {
-      char dir[OS_PATH_MAX];
-      if (os_exe_dir(dir, sizeof(dir)) == 0)
-        os_path_join(ca, sizeof(ca), dir, "ca-bundle.crt");
-    }
-    if (ca[0]) {
-      FILE *f = fopen(ca, "rb");
-      if (f)
-        fclose(f);
+#if defined(_WIN32)
+      char dir[OS_PATH_MAX], p[OS_PATH_MAX];
+      if (os_exe_dir(dir, sizeof(dir)) == 0 &&
+          os_path_join(p, sizeof(p), dir, "ca-bundle.crt") == 0 &&
+          ca_file_ok(p))
+        snprintf(ca, sizeof(ca), "%s", p);
+#else
+      static const char *const known[] = {
+          "/etc/ssl/certs/ca-certificates.crt", /* Debian, Ubuntu, Arch, Alpine */
+          "/etc/pki/tls/certs/ca-bundle.crt",   /* Fedora, RHEL */
+          "/etc/ssl/cert.pem",                  /* macOS, BSD, Alpine */
+          "/etc/ssl/ca-bundle.pem",             /* openSUSE */
+          "/etc/pki/ca-trust/extracted/pem/tls-ca-bundle.pem",
+      };
+      const char *pick = NULL;
+      const char *sslenv = getenv("SSL_CERT_FILE");
+      const curl_version_info_data *vi = curl_version_info(CURLVERSION_NOW);
+      if (ca_file_ok(sslenv))
+        pick = sslenv;
+      else if (vi && ca_file_ok(vi->cainfo))
+        pick = NULL; /* libcurl's built-in default already works */
      else
-        ca[0] = '\0';
+        for (size_t i = 0; i < sizeof(known) / sizeof(known[0]); i++)
+          if (ca_file_ok(known[i])) {
+            pick = known[i];
+            break;
+          }
+      if (pick)
+        snprintf(ca, sizeof(ca), "%s", pick);
+#endif
    }
+    atomic_store(&state, 2);
+  } else {
+    while (atomic_load(&state) != 2)
+      ;
  }
  if (ca[0])
-    curl_easy_setopt(curl, CURLOPT_CAINFO, ca);
-#else
-  (void)curl;
-#endif
+    curl_easy_setopt((CURL *)curl, CURLOPT_CAINFO, ca);
 }

 #include "cst_attach.h"
@@ -223,28 +255,78 @@ static cJSON *image_part(const cst_llm_request_t *req,
  return p;
 }

-/* Append every attachment of m as image parts (or a note saying why not). */
-static void parts_add_images(const cst_llm_request_t *req, cJSON *parts,
-                             const cst_message_t *m, size_t *budget) {
+/* One stored document as a text part: its text, framed so the model knows
+ * where it came from. Documents are plain text on the wire, so a text-only
+ * model and CST_NO_IMAGES both still get them. */
+static cJSON *document_part(const cst_llm_request_t *req,
+                            const cst_attachment_t *a, size_t *budget,
+                            char *why, size_t why_sz) {
+  why[0] = 0;
+  if (!req->session_path || !req->session_path[0]) {
+    snprintf(why, why_sz, "there is no session store to read it from");
+    return NULL;
+  }
+  char err[256];
+  size_t n = 0;
+  char *text = cst_attach_read(req->session_path, a, &n, err, sizeof(err));
+  if (!text) {
+    snprintf(why, why_sz, "%s", err);
+    return NULL;
+  }
+  int cut = 0;
+  if (n > *budget) {
+    n = *budget;
+    cut = 1;
+  }
+  *budget -= n;
+  char desc[128];
+  cst_attach_describe(a, desc, sizeof(desc));
+  cst_buf_t b = {0};
+  cst_buf_printf(&b, "[document \"%.120s\" (%s)%s]\n", a->name, desc,
+                 cut ? " -- only the start is included" : "");
+  cst_buf_append(&b, text, n);
+  cst_buf_puts(&b, cut ? "\n[-- document truncated --]"
+                       : "\n[-- end of document --]");
+  free(text);
+  cJSON *p = cJSON_CreateObject();
+  cJSON_AddStringToObject(p, "type", "text");
+  cJSON_AddStringToObject(p, "text", b.buf ? b.buf : "");
+  cst_buf_free(&b);
+  return p;
+}
+
+/* Append every attachment of m: images as image parts (or a note saying why
+ * not), documents inline as text. */
+static void parts_add_attachments(const cst_llm_request_t *req, cJSON *parts,
+                                  const cst_message_t *m, size_t *budget,
+                                  size_t *doc_budget, const char *img_off) {
  int n = m->nattach;
  if (n > CST_ATTACH_MAX_COUNT) {
    char note[160];
-    snprintf(note, sizeof(note), "[%d of %d images sent]", CST_ATTACH_MAX_COUNT,
-             n);
+    snprintf(note, sizeof(note),
+             "[only the first %d of %d attachments are sent]",
+             CST_ATTACH_MAX_COUNT, n);
    parts_add_text(parts, cJSON_CreateString(note));
    n = CST_ATTACH_MAX_COUNT;
  }
  for (int i = 0; i < n; i++) {
    const cst_attachment_t *a = &m->attach[i];
+    int is_img = cst_attach_is_image(a);
    char why[256];
-    cJSON *p = image_part(req, a, budget, why, sizeof(why));
+    cJSON *p = NULL;
+    if (is_img && img_off)
+      snprintf(why, sizeof(why), "%s", img_off);
+    else if (is_img)
+      p = image_part(req, a, budget, why, sizeof(why));
+    else
+      p = document_part(req, a, doc_budget, why, sizeof(why));
    if (p) {
      cJSON_AddItemToArray(parts, p);
      continue;
    }
    char note[512];
-    snprintf(note, sizeof(note), "[image \"%.120s\" not sent: %.200s]",
-             a->name, why);
+    snprintf(note, sizeof(note), "[%s \"%.120s\" not sent: %.200s]",
+             is_img ? "image" : "document", a->name, why);
    parts_add_text(parts, cJSON_CreateString(note));
  }
 }
@@ -252,8 +334,10 @@ static void parts_add_images(const cst_llm_request_t *req, cJSON *parts,
 /* Images a tool result produced are deferred: an assistant tool_call must be
 * answered by tool messages only, so they ride in one extra user message
 * after the run of tool messages ends. *pending collects it. */
-static void defer_images(const cst_llm_request_t *req, const cst_message_t *m,
-                         cJSON **pending, size_t *budget) {
+static void defer_attachments(const cst_llm_request_t *req,
+                              const cst_message_t *m, cJSON **pending,
+                              size_t *budget, size_t *doc_budget,
+                              const char *img_off) {
  cJSON *p = *pending;
  if (!p) {
    p = cJSON_CreateObject();
@@ -261,15 +345,18 @@ static void defer_images(const cst_llm_request_t *req, const cst_message_t *m,
    cJSON_AddArrayToObject(p, "content");
    *pending = p;
  }
-  parts_add_images(req, cJSON_GetObjectItemCaseSensitive(p, "content"), m,
-                   budget);
+  parts_add_attachments(req, cJSON_GetObjectItemCaseSensitive(p, "content"), m,
+                        budget, doc_budget, img_off);
 }

-/* m's text plus one line per image it would have carried (malloc'd). */
+/* m's text plus one line per image it would have carried (malloc'd); a
+ * document is not affected (it is text and is sent as such). */
 static char *image_notes(const cst_message_t *m, const char *why) {
  cst_buf_t b = {0};
  cst_buf_puts(&b, m->content ? m->content : "");
  for (int i = 0; i < m->nattach; i++) {
+    if (!cst_attach_is_image(&m->attach[i]))
+      continue;
    char d[128];
    cst_attach_describe(&m->attach[i], d, sizeof(d));
    cst_buf_printf(&b, "%s[image \"%.120s\" (%s) not sent: %s]",
@@ -362,6 +449,7 @@ char *cst_llm_build_body(const cst_llm_request_t *req) {
  cJSON *prev_user = NULL; /* merge consecutive user messages */
  cJSON *pending = NULL;   /* images from tool results, waiting for a break */
  size_t img_budget = CST_ATTACH_MAX_TOTAL;
+  size_t doc_budget = CST_ATTACH_MAX_TOTAL;
  const char *noimg = getenv("CST_NO_IMAGES");
  const char *img_off =
      noimg && noimg[0] && noimg[0] != '0'
@@ -375,15 +463,26 @@ char *cst_llm_build_body(const cst_llm_request_t *req) {
    cst_role_t role = m->role;
    if (role == CST_ROLE_TOOL && !m->tool_call_id[0])
      role = CST_ROLE_USER;
-    int has_img = m->attach && m->nattach > 0;
+    int n_att = m->attach ? m->nattach : 0;
+    int has_img = 0, has_doc = 0;
+    for (int k = 0; k < n_att; k++) {
+      if (cst_attach_is_image(&m->attach[k]))
+        has_img = 1;
+      else
+        has_doc = 1;
+    }
    char *noted = NULL;
    if (has_img && img_off) {
-      /* plain text: text-only APIs may refuse content arrays */
-      noted = image_notes(m, img_off);
-      if (noted)
-        mm.content = noted;
      has_img = 0;
+      if (!has_doc) {
+        /* plain text: text-only APIs may refuse content arrays */
+        noted = image_notes(m, img_off);
+        if (noted)
+          mm.content = noted;
+      }
    }
+    /* a document needs a content array even when no image can go out */
+    int att_parts = has_img || has_doc;

    /* Every tool_call must be answered by tool messages, so tool-produced
     * images only leave the queue when the run of tool results is over. */
@@ -399,7 +498,7 @@ char *cst_llm_build_body(const cst_llm_request_t *req) {
      cJSON *pc = prev_user
                      ? cJSON_GetObjectItemCaseSensitive(prev_user, "content")
                      : NULL;
-      if (!has_img && !cJSON_IsArray(pc)) {
+      if (!att_parts && !cJSON_IsArray(pc)) {
        cJSON *c = clean_string(m->content, 0);
        if (prev_user) {
          cst_buf_t b = {0};
@@ -431,8 +530,8 @@ char *cst_llm_build_body(const cst_llm_request_t *req) {
      }
      cJSON *parts = content_parts(o);
      parts_add_text(parts, clean_string(m->content, 0));
-      if (has_img)
-        parts_add_images(req, parts, m, &img_budget);
+      if (att_parts)
+        parts_add_attachments(req, parts, m, &img_budget, &doc_budget, img_off);
      free(noted);
      continue;
    }
@@ -466,8 +565,8 @@ char *cst_llm_build_body(const cst_llm_request_t *req) {
        cJSON_AddTrueToObject(o, "prefix");
    }
    cJSON_AddItemToArray(msgs, o);
-    if (role == CST_ROLE_TOOL && has_img)
-      defer_images(req, m, &pending, &img_budget);
+    if (role == CST_ROLE_TOOL && n_att)
+      defer_attachments(req, m, &pending, &img_budget, &doc_budget, img_off);
    free(noted);
  }
  if (pending)
@@ -988,7 +1087,7 @@ int cst_llm_chat_stream(const cst_llm_request_t *req, char *err,
      ret = -1;
      break;
    }
-    llm_apply_ca_bundle(curl);
+    cst_curl_apply_ca(curl);
    sse_reset(&s);
    s.curl = curl;
    if (rec_dir && rec_dir[0]) {
@@ -1233,7 +1332,7 @@ int cst_llm_list_models_meta(const char *api_base, const char *api_key,
    snprintf(err, err_sz, "curl init failed");
    return -1;
  }
-  llm_apply_ca_bundle(curl);
+  cst_curl_apply_ca(curl);
  char url[1024];
  size_t bl = strlen(api_base);
  while (bl && api_base[bl - 1] == '/')
@@ -26,6 +26,8 @@ static void usage(const char *argv0) {
          "                        (last used)\n"
          "  -m, --model NAME      model (last used for the provider)\n"
          "  -e, --effort LEVEL    low | medium | high | max (last used)\n"
+          "  -M, --mode NAME       coding (default, code framed) | general\n"
+          "                        (documents, data, research; last used)\n"
          "  -w, --workspace DIR   project directory (default: current)\n"
          "  -c, --continue        reopen the last session (default)\n"
          "  -N, --new             start a new session\n"
@@ -43,7 +45,7 @@ static void usage(const char *argv0) {
          "  -n, --max-steps N     model calls per turn (30)\n"
          "  -H, --headless -p TEXT  run one prompt without the TUI\n"
          "      --web             serve the UI over HTTP on 127.0.0.1:8787\n"
-          "      --web-port N      port (0 = pick a free one)\n"
+          "      --web-port N      port (default 8787, 0 = pick a free one)\n"
          "      --web-host ADDR   bind address (default 127.0.0.1; a\n"
          "                        non-loopback address requires a token)\n"
          "      --web-token T     require ?token=T / Authorization: Bearer T\n"
@@ -59,13 +61,13 @@ static void usage(const char *argv0) {
          "API keys: the TUI asks for a missing key and saves it per provider\n"
          "  in the config dir (credentials, mode 0600); /key changes it.\n"
          "  Fallbacks: <provider env> (DEEPSEEK_API_KEY, ...), CST_API_KEY.\n"
-          "Settings: provider, models and effort are remembered in\n"
+          "Settings: provider, models, effort and mode are remembered in\n"
          "  settings.json next to the credentials.\n"
-          "Env:  CST_PROVIDER CST_MODEL CST_EFFORT CST_PROTOCOL CST_TOOLS\n"
-          "      CST_CONFIRM CST_MAX_STEPS CST_TOOL_OUT_CAP CST_CTX_WINDOW\n"
-          "      CST_CTX_TOKENS CST_AUTO_COMPACT=0 CST_SHELL_TIMEOUT\n"
-          "      CST_METRICS=0 CST_RECORD_SSE CST_REPLAY_SSE CST_MODELS\n"
-          "      CST_APPROVAL=ask|auto|yolo CST_AUTO_APPROVE=1 (=yolo)\n"
+          "Env:  CST_PROVIDER CST_MODEL CST_EFFORT CST_MODE CST_PROTOCOL\n"
+          "      CST_TOOLS CST_CONFIRM CST_MAX_STEPS CST_TOOL_OUT_CAP\n"
+          "      CST_CTX_WINDOW CST_CTX_TOKENS CST_AUTO_COMPACT=0\n"
+          "      CST_SHELL_TIMEOUT CST_METRICS=0 CST_RECORD_SSE CST_REPLAY_SSE\n"
+          "      CST_MODELS CST_APPROVAL=ask|auto|yolo CST_AUTO_APPROVE=1 (=yolo)\n"
          "      CST_SANDBOX=off CST_SHELL CST_NO_PROJECT_NOTES=1 NO_COLOR\n"
          "TUI: type /help inside for commands and keys\n",
          argv0);
@@ -255,6 +257,16 @@ static int demo_ctl_impl(const char *workspace) {
  return 0;
 }

+/* 0..65535, digits only: atoi turned "70000" into port 4464, "abc" into 0. */
+static int parse_port(const char *s, int *out) {
+  char *end = NULL;
+  long v = strtol(s, &end, 10);
+  if (!s[0] || *end || v < 0 || v > 65535)
+    return -1;
+  *out = (int)v;
+  return 0;
+}
+
 int main(int argc, char **argv) {
  if (os_console_init() != 0)
    fprintf(stderr, "warning: the process code page is not UTF-8 (Windows 10 "
@@ -332,9 +344,12 @@ int main(int argc, char **argv) {
  int web_mode = 0;
  cst_webui_opts_t web;
  memset(&web, 0, sizeof(web));
+  web.port = CST_WEBUI_DEFAULT_PORT; /* 0 would mean "pick a free one" */
  const char *env = getenv("CST_WEB_PORT");
-  if (env && env[0])
-    web.port = atoi(env);
+  if (env && env[0] && parse_port(env, &web.port) != 0) {
+    fprintf(stderr, "CST_WEB_PORT=%s is not a port (0-65535)\n", env);
+    return 1;
+  }
  web.host = getenv("CST_WEB_HOST");
  web.token = getenv("CST_WEB_TOKEN");
  web.root = getenv("CST_WEB_ROOT");
@@ -361,6 +376,7 @@ int main(int argc, char **argv) {
      {"approval", required_argument, 0, 'A'},
      {"confirm", required_argument, 0, 'C'},
      {"effort", required_argument, 0, 'e'},
+      {"mode", required_argument, 0, 'M'},
      {"help", no_argument, 0, 'h'},
      {"version", no_argument, 0, 'V'},
      {"list-models", optional_argument, 0, 0x100},
@@ -372,8 +388,8 @@ int main(int argc, char **argv) {
      {"web-open", no_argument, 0, 0x106},
      {0, 0, 0, 0}};
  int c;
-  while ((c = getopt_long(argc, argv, "w:s:b:k:m:P:r:n:T:Hp:dahVNcR:LyA:C:e:", opts, NULL)) !=
-         -1) {
+  while ((c = getopt_long(argc, argv, "w:s:b:k:m:P:r:n:T:Hp:dahVNcR:LyA:C:e:M:",
+                          opts, NULL)) != -1) {
    switch (c) {
    case 'w':
      snprintf(cfg.workspace_root, sizeof(cfg.workspace_root), "%s", optarg);
@@ -465,6 +481,14 @@ int main(int argc, char **argv) {
      }
      snprintf(cfg.effort, sizeof(cfg.effort), "%s", optarg);
      break;
+    case 'M':
+      if (!cst_mode_valid(optarg)) {
+        fprintf(stderr, "bad --mode %s (coding | general)\n", optarg);
+        curl_global_cleanup();
+        return 1;
+      }
+      snprintf(cfg.mode, sizeof(cfg.mode), "%s", optarg);
+      break;
    case 'h':
      usage(argv[0]);
      curl_global_cleanup();
@@ -481,7 +505,11 @@ int main(int argc, char **argv) {
      web_mode = 1;
      break;
    case 0x102:
-      web.port = atoi(optarg);
+      if (parse_port(optarg, &web.port) != 0) {
+        fprintf(stderr, "--web-port %s is not a port (0-65535)\n", optarg);
+        curl_global_cleanup();
+        return 1;
+      }
      web_mode = 1;
      break;
    case 0x103:
@@ -540,6 +568,12 @@ int main(int argc, char **argv) {
    snprintf(cfg.effort, sizeof(cfg.effort), "%s",
             cst_effort_valid(e) ? e : CST_DEFAULT_EFFORT);
  }
+  if (!cfg.mode[0]) {
+    const char *m = getenv("CST_MODE");
+    if (!cst_mode_valid(m))
+      m = saved.mode;
+    snprintf(cfg.mode, sizeof(cfg.mode), "%s", cst_mode_name(m));
+  }
  if (!cfg.api_base[0]) {
    const char *env_base = getenv("CST_API_BASE");
    if (!env_base || !env_base[0])
@@ -12,6 +12,10 @@
 #ifndef WIN32_LEAN_AND_MEAN
 #define WIN32_LEAN_AND_MEAN
 #endif
+#if !defined(_WIN32_WINNT) || _WIN32_WINNT < 0x0600
+#undef _WIN32_WINNT
+#define _WIN32_WINNT 0x0600 /* WSAPoll */
+#endif
 #include <winsock2.h>
 #include <ws2tcpip.h>
 #else
@@ -21,8 +25,8 @@
 #include <netdb.h>
 #include <netinet/in.h>
 #include <netinet/tcp.h>
+#include <poll.h>
 #include <signal.h>
-#include <sys/select.h>
 #include <sys/socket.h>
 #include <unistd.h>
 #endif
@@ -33,11 +37,6 @@

 #define NET_SEND_TIMEOUT_MS 20000

-static void set_err(char *err, size_t err_sz, const char *what) {
-  if (err && err_sz)
-    snprintf(err, err_sz, "%s", what);
-}
-
 #ifdef _WIN32
 static void set_err_win(char *err, size_t err_sz, const char *what) {
  if (err && err_sz)
@@ -45,7 +44,6 @@ static void set_err_win(char *err, size_t err_sz, const char *what) {
 }
 #define SET_ERR(e, n, w) set_err_win(e, n, w)
 #define NET_FD(s) ((SOCKET)(s))
-#define NET_NFDS(s) (0) /* Winsock ignores nfds */
 static void net_close_raw(os_socket_t s) { closesocket((SOCKET)s); }
 static int net_last_would_block(void) {
  int e = WSAGetLastError();
@@ -56,9 +54,12 @@ static int net_last_conn_reset(void) {
  return e == WSAECONNRESET || e == WSAECONNABORTED || e == WSAENOTSOCK;
 }
 #else
+static void set_err(char *err, size_t err_sz, const char *what) {
+  if (err && err_sz)
+    snprintf(err, err_sz, "%s", what);
+}
 #define SET_ERR(e, n, w) set_err(e, n, w)
 #define NET_FD(s) ((int)(s))
-#define NET_NFDS(s) ((int)((s) + 1))
 static void net_close_raw(os_socket_t s) { close(s); }
 static int net_last_would_block(void) { return errno == EINTR; }
 static int net_last_conn_reset(void) {
@@ -84,15 +85,22 @@ void os_net_cleanup(void) {
 #endif
 }

+/* poll, not select: FD_SET past FD_SETSIZE writes out of bounds, and a
+ * failed connect shows up as an error event, not as writable, on Windows. */
 static int wait_io(os_socket_t s, int want_write, int timeout_ms) {
-  fd_set fds;
-  FD_ZERO(&fds);
-  FD_SET(NET_FD(s), &fds);
-  struct timeval tv;
-  tv.tv_sec = timeout_ms / 1000;
-  tv.tv_usec = (timeout_ms % 1000) * 1000;
-  int rc = select(NET_NFDS(s), want_write ? NULL : &fds,
-                  want_write ? &fds : NULL, NULL, timeout_ms < 0 ? NULL : &tv);
+#ifdef _WIN32
+  WSAPOLLFD p;
+  p.fd = NET_FD(s);
+  p.events = want_write ? POLLWRNORM : POLLRDNORM;
+  p.revents = 0;
+  int rc = WSAPoll(&p, 1, timeout_ms < 0 ? -1 : timeout_ms);
+#else
+  struct pollfd p;
+  p.fd = NET_FD(s);
+  p.events = want_write ? POLLOUT : POLLIN;
+  p.revents = 0;
+  int rc = poll(&p, 1, timeout_ms < 0 ? -1 : timeout_ms);
+#endif
  if (rc > 0)
    return 1;
  if (rc == 0)
@@ -137,8 +145,15 @@ os_socket_t os_net_listen(const char *host, int port, char *err, size_t err_sz)
    return OS_SOCKET_BAD;
  }
  int one = 1;
+#ifdef _WIN32
+  /* SO_REUSEADDR on Windows lets another process bind the same port and
+   * take over our connections */
+  setsockopt(NET_FD(s), SOL_SOCKET, SO_EXCLUSIVEADDRUSE, (const char *)&one,
+             sizeof(one));
+#else
  setsockopt(NET_FD(s), SOL_SOCKET, SO_REUSEADDR, (const char *)&one,
             sizeof(one));
+#endif
 #ifdef IPV6_V6ONLY
  if (sa.ss_family == AF_INET6)
    setsockopt(NET_FD(s), IPPROTO_IPV6, IPV6_V6ONLY, (const char *)&one,
@@ -13,6 +13,7 @@
 #include "os_internal.h"

 #include <errno.h>
+#include <stdint.h>
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
@@ -503,23 +504,78 @@ int os_resolve_shell(char *out, size_t out_sz) {
 #endif
 }

-/* Per thread: parallel tool runs and sub-agents each watch their own flag. */
-static _Thread_local volatile int *g_run_cancel;
-static _Thread_local int g_run_timeout_s;
+/* Per thread: parallel tool runs and sub-agents each watch their own flag.
+ * Thread keys rather than _Thread_local, which tcc does not implement; each
+ * value fits in the key's pointer slot, so nothing is allocated. */
+#ifdef _WIN32
+static INIT_ONCE g_run_once = INIT_ONCE_STATIC_INIT;
+static DWORD g_key_cancel = TLS_OUT_OF_INDEXES, g_key_timeout = TLS_OUT_OF_INDEXES;
+static BOOL CALLBACK run_keys_init(PINIT_ONCE once, PVOID param, PVOID *ctx) {
+  (void)once;
+  (void)param;
+  (void)ctx;
+  g_key_cancel = TlsAlloc();
+  g_key_timeout = TlsAlloc();
+  return TRUE;
+}
+static void *run_key_get(DWORD *key) {
+  InitOnceExecuteOnce(&g_run_once, run_keys_init, NULL, NULL);
+  if (*key == TLS_OUT_OF_INDEXES)
+    return NULL;
+  DWORD err = GetLastError(); /* TlsGetValue clears it */
+  void *v = TlsGetValue(*key);
+  SetLastError(err);
+  return v;
+}
+static void run_key_set(DWORD *key, void *v) {
+  InitOnceExecuteOnce(&g_run_once, run_keys_init, NULL, NULL);
+  if (*key != TLS_OUT_OF_INDEXES)
+    TlsSetValue(*key, v);
+}
+#else
+static pthread_once_t g_run_once = PTHREAD_ONCE_INIT;
+static pthread_key_t g_key_cancel, g_key_timeout;
+static int g_run_keys_ok;
+static void run_keys_init(void) {
+  g_run_keys_ok = pthread_key_create(&g_key_cancel, NULL) == 0 &&
+                  pthread_key_create(&g_key_timeout, NULL) == 0;
+}
+static void *run_key_get(pthread_key_t *key) {
+  pthread_once(&g_run_once, run_keys_init);
+  return g_run_keys_ok ? pthread_getspecific(*key) : NULL;
+}
+static void run_key_set(pthread_key_t *key, void *v) {
+  pthread_once(&g_run_once, run_keys_init);
+  if (g_run_keys_ok)
+    pthread_setspecific(*key, v);
+}
+#endif

-void os_set_cancel_flag(volatile int *flag) { g_run_cancel = flag; }
+static volatile int *run_cancel(void) {
+  return (volatile int *)run_key_get(&g_key_cancel);
+}
+static int run_timeout(void) {
+  return (int)(intptr_t)run_key_get(&g_key_timeout);
+}

-void os_set_run_timeout(int seconds) { g_run_timeout_s = seconds; }
+void os_set_cancel_flag(volatile int *flag) {
+  run_key_set(&g_key_cancel, (void *)(uintptr_t)flag);
+}

-int os_get_run_timeout(void) { return g_run_timeout_s; }
+void os_set_run_timeout(int seconds) {
+  run_key_set(&g_key_timeout, (void *)(intptr_t)seconds);
+}
+
+int os_get_run_timeout(void) { return run_timeout(); }

 int os_run_timeout_effective(void) {
-  if (g_run_timeout_s > 0)
-    return g_run_timeout_s;
+  int t = run_timeout();
+  if (t > 0)
+    return t;
  const char *e = getenv("CST_SHELL_TIMEOUT");
  if (e && atoi(e) > 0)
    return atoi(e);
-  return 120;
+  return 0; /* no limit: a build, a test suite or a long clone must finish */
 }

 double os_now_ms(void) {
@@ -662,7 +718,8 @@ int os_run_capture_ex(const char *cwd, const char *const argv[],
  int truncated = 0, killed = 0;
  int status = 0, reaped = 0, bg_killed = 0;
  double reaped_at = 0, killed_at = 0;
-  double deadline = os_now_ms() + os_run_timeout_effective() * 1000.0;
+  int timeout_s = os_run_timeout_effective();
+  double deadline = timeout_s > 0 ? os_now_ms() + timeout_s * 1000.0 : 0;
  for (;;) {
    struct pollfd pfd = {pipefd[0], POLLIN, 0};
    int pr = poll(&pfd, 1, 100);
@@ -692,8 +749,8 @@ int os_run_capture_ex(const char *cwd, const char *const argv[],
    }
    /* Checked on every pass, so a chatty command still stops promptly. */
    if (!reaped && !killed &&
-        ((g_run_cancel && *g_run_cancel) || now > deadline)) {
-      killed = (g_run_cancel && *g_run_cancel) ? 2 : 1;
+        (os_cancel_requested() || (deadline > 0 && now > deadline))) {
+      killed = os_cancel_requested() ? 2 : 1;
      killed_at = now;
      kill(-pid, SIGTERM);
    }
@@ -899,9 +956,12 @@ void os_run_parallel(os_task_fn fn, void *const args[], int n) {
  free(tasks);
 }

-int os_cancel_requested(void) { return g_run_cancel && *g_run_cancel; }
+int os_cancel_requested(void) {
+  volatile int *c = run_cancel();
+  return c && *c;
+}

-volatile int *os_get_cancel_flag(void) { return g_run_cancel; }
+volatile int *os_get_cancel_flag(void) { return run_cancel(); }

 #ifndef _WIN32
 int os_proc_spawn(const char *cwd, const char *const argv[],
@@ -601,7 +601,10 @@ int os_run_capture_token(const char *cwd, const char *const argv[],
  buf[0] = 0;
  int truncated = 0, killed = 0, reaped = 0, bg_killed = 0;
  double reaped_at = 0, killed_at = 0;
-  double deadline = os_now_ms() + os_run_timeout_effective() * 1000.0;
+  /* no timeout unless one was asked for: a deadline far in the future keeps
+   * every `now > deadline` check working without a second code path */
+  int timeout_s = os_run_timeout_effective();
+  double deadline = timeout_s > 0 ? os_now_ms() + timeout_s * 1000.0 : 1e18;
  int idle = 0; /* empty polls in a row: wait longer once output stalls */
  for (;;) {
    DWORD avail = 0;
@@ -0,0 +1,193 @@
+#include "cst_provider_switch.h"
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+int cst_ps_fetch(const char *provider, const char *api_base, const char *key,
+                 char ids[][128], int max, char *err, size_t err_sz) {
+  cst_model_meta_t *m = calloc((size_t)(max > 0 ? max : 1), sizeof(*m));
+  if (!m) {
+    snprintf(err, err_sz, "out of memory");
+    return -1;
+  }
+  int r = cst_llm_list_models_meta(api_base, key, m, max, err, err_sz);
+  if (r >= 0)
+    cst_models_cache_save(provider, api_base, m, r);
+  for (int i = 0; i < r; i++)
+    snprintf(ids[i], 128, "%s", m[i].id);
+  free(m);
+  return r;
+}
+
+int cst_ps_cached(const char *provider, char ids[][128], int max, int *fresh) {
+  *fresh = 0;
+  cst_model_meta_t *m = calloc((size_t)(max > 0 ? max : 1), sizeof(*m));
+  if (!m)
+    return -1;
+  long age = 0;
+  int n = cst_models_cache_load(provider, m, max, &age);
+  for (int i = 0; i < n; i++)
+    snprintf(ids[i], 128, "%s", m[i].id);
+  free(m);
+  *fresh = n > 0 && age >= 0 && age < CST_MODELS_CACHE_TTL;
+  return n > 0 ? n : -1;
+}
+
+const char *cst_ps_base(const cst_agent_config_t *cfg, const char *provider) {
+  if (cfg && strcmp(cfg->provider, provider) == 0)
+    return cfg->api_base;
+  const cst_provider_t *p = cst_provider_find(provider);
+  return p ? p->api_base : "";
+}
+
+int cst_ps_models(const cst_agent_config_t *cfg, int refresh,
+                  cst_ps_model_t *out, int max, int *api_rc, char *err,
+                  size_t err_sz) {
+  err[0] = 0;
+  *api_rc = -1;
+  char(*ids)[128] = calloc(CST_PS_MAX_MODELS, sizeof(*ids));
+  if (!ids) {
+    snprintf(err, err_sz, "out of memory");
+    return 0;
+  }
+  int fresh = 0;
+  int na = cst_ps_cached(cfg->provider, ids, CST_PS_MAX_MODELS, &fresh);
+  *api_rc = na;
+  if ((refresh || !fresh) && cfg->api_key[0]) {
+    char(*got)[128] = calloc(CST_PS_MAX_MODELS, sizeof(*got));
+    if (got) {
+      int r = cst_ps_fetch(cfg->provider, cfg->api_base, cfg->api_key, got,
+                           CST_PS_MAX_MODELS, err, err_sz);
+      *api_rc = r;
+      if (r >= 0) {
+        memcpy(ids, got, (size_t)r * sizeof(*ids));
+        na = r;
+      }
+      free(got);
+    }
+  }
+  int n = 0;
+  for (int i = 0; i < na && n < max; i++, n++) {
+    snprintf(out[n].id, sizeof(out[n].id), "%s", ids[i]);
+    out[n].api = 1;
+    cst_model_info(cfg->provider, ids[i], &out[n].info);
+  }
+  char builtin[32][128];
+  int nb = cst_models_list(cfg->model, builtin, 32);
+  for (int i = 0; i < nb && n < max; i++) {
+    int dup = 0;
+    for (int k = 0; k < na && !dup; k++)
+      dup = strcmp(ids[k], builtin[i]) == 0;
+    if (dup)
+      continue;
+    memcpy(out[n].id, builtin[i], sizeof(out[n].id));
+    out[n].id[sizeof(out[n].id) - 1] = 0;
+    out[n].api = 0;
+    cst_model_info(cfg->provider, builtin[i], &out[n].info);
+    n++;
+  }
+  free(ids);
+  return n;
+}
+
+int cst_ps_verify_save(const char *provider, const char *api_base,
+                       const char *key, char ids[][128], int max, int *nids,
+                       char *err, size_t err_sz) {
+  err[0] = 0;
+  char(*own)[128] = NULL;
+  if (!ids) {
+    own = calloc(CST_PS_MAX_MODELS, sizeof(*own));
+    if (!own) {
+      snprintf(err, err_sz, "out of memory");
+      return CST_PS_UNVERIFIED | CST_PS_UNSAVED;
+    }
+    ids = own;
+    max = CST_PS_MAX_MODELS;
+  }
+  int r = cst_ps_fetch(provider, api_base, key, ids, max, err, err_sz);
+  free(own);
+  if (nids)
+    *nids = r;
+  if (r == CST_MODELS_AUTH)
+    return CST_MODELS_AUTH;
+  int flags = r < 0 ? CST_PS_UNVERIFIED : 0;
+  if (r >= 0)
+    err[0] = 0;
+  if (cst_creds_save(provider, key) != 0)
+    flags |= CST_PS_UNSAVED;
+  return flags;
+}
+
+void cst_ps_apply_provider(cst_agent_config_t *cfg, const cst_provider_t *p,
+                           const char *key) {
+  cst_settings_t s;
+  cst_settings_load(&s);
+  const char *model = cst_settings_model(&s, p->name);
+  snprintf(cfg->provider, sizeof(cfg->provider), "%s", p->name);
+  snprintf(cfg->api_base, sizeof(cfg->api_base), "%s", p->api_base);
+  snprintf(cfg->model, sizeof(cfg->model), "%s", model ? model : p->model);
+  snprintf(cfg->api_key, sizeof(cfg->api_key), "%s", key);
+  cst_settings_remember(cfg);
+}
+
+int cst_ps_switch_provider(cst_agent_config_t *cfg, const char *name,
+                           const char *key, char *err, size_t err_sz) {
+  err[0] = 0;
+  const cst_provider_t *p = name ? cst_provider_find(name) : NULL;
+  if (!p) {
+    snprintf(err, err_sz, "unknown provider '%.64s'", name ? name : "");
+    return CST_PS_UNKNOWN;
+  }
+  char k[512];
+  int flags = 0;
+  char(*ids)[128] = calloc(CST_PS_MAX_MODELS, sizeof(*ids));
+  int nids = -1;
+  if (key && key[0]) {
+    snprintf(k, sizeof(k), "%s", key);
+    flags = cst_ps_verify_save(p->name, p->api_base, k, ids,
+                               ids ? CST_PS_MAX_MODELS : 0, &nids, err,
+                               err_sz);
+    if (flags < 0) {
+      memset(k, 0, sizeof(k));
+      free(ids);
+      return flags;
+    }
+  } else if (cst_creds_resolve(p->name, k, sizeof(k), NULL) != 0) {
+    free(ids);
+    snprintf(err, err_sz, "%s needs an API key", p->name);
+    return CST_PS_NEED_KEY;
+  }
+  cst_ps_apply_provider(cfg, p, k);
+  memset(k, 0, sizeof(k));
+  if (!cfg->model[0] && ids) {
+    /* a router preset: the first model the API lists */
+    int fresh = 0;
+    if (nids <= 0)
+      nids = cst_ps_cached(p->name, ids, CST_PS_MAX_MODELS, &fresh);
+    if (nids <= 0) {
+      char e2[256];
+      nids = cst_ps_fetch(p->name, cfg->api_base, cfg->api_key, ids,
+                          CST_PS_MAX_MODELS, e2, sizeof(e2));
+    }
+    if (nids > 0) {
+      snprintf(cfg->model, sizeof(cfg->model), "%s", ids[0]);
+      cst_settings_remember(cfg);
+    }
+  }
+  free(ids);
+  return flags;
+}
+
+int cst_ps_set_key(cst_agent_config_t *cfg, const char *key, char *err,
+                   size_t err_sz) {
+  if (!key || !key[0]) {
+    snprintf(err, err_sz, "empty key");
+    return CST_PS_NEED_KEY;
+  }
+  int r = cst_ps_verify_save(cfg->provider, cfg->api_base, key, NULL, 0, NULL,
+                             err, err_sz);
+  if (r >= 0)
+    snprintf(cfg->api_key, sizeof(cfg->api_key), "%s", key);
+  return r;
+}
@@ -8,6 +8,7 @@
 #include "cJSON.h"
 #include "cst_agent.h"
 #include "cst_attach.h"
+#include "cst_util.h"

 #include <pthread.h>
 #include <stdio.h>
@@ -409,3 +410,65 @@ int cst_session_rewrite(cst_session_t *s, const cst_message_t *msgs, int n) {
  free(copy);
  return rc;
 }
+
+int cst_message_is_tool_result(const cst_message_t *m) {
+  return m && (m->role == CST_ROLE_TOOL ||
+               (m->role == CST_ROLE_USER && m->tool_call_id[0]));
+}
+
+static void md_fence(cst_buf_t *b, const char *s, size_t max) {
+  size_t n = strlen(s);
+  if (n > max)
+    n = cst_utf8_floor(s, max);
+  cst_buf_puts(b, "````\n");
+  cst_buf_append(b, s, n);
+  cst_buf_puts(b, n < strlen(s) ? "\n…(truncated)\n````\n\n" : "\n````\n\n");
+}
+
+char *cst_session_export_md(cst_session_t *s, const char *title,
+                            const char *sid, const char *model) {
+  if (!s)
+    return NULL;
+  cst_buf_t b = {0};
+  cst_buf_printf(&b, "# %s\n\nsession `%s` · model `%s`\n\n",
+                 title ? title : "", sid ? sid : "", model ? model : "");
+  cst_session_lock(s);
+  for (int i = 0; i < cst_session_count(s); i++) {
+    const cst_message_t *m = cst_session_get(s, i);
+    const char *c = m->content ? m->content : "";
+    if (m->role == CST_ROLE_SYSTEM)
+      continue;
+    if (cst_message_is_tool_result(m)) {
+      cst_buf_printf(&b, "**Tool result** `%s`\n\n", m->tool_name);
+      md_fence(&b, c, 4000);
+    } else if (cst_message_is_note(m)) {
+      cst_buf_puts(&b, "> ");
+      cst_buf_append(&b, c, strcspn(c, "\n"));
+      cst_buf_puts(&b, "\n\n");
+    } else if (m->role == CST_ROLE_USER) {
+      cst_buf_puts(&b, "## User\n\n");
+      cst_buf_puts(&b, c);
+      cst_buf_puts(&b, "\n\n");
+      for (int k = 0; k < m->nattach; k++) {
+        char d[128];
+        cst_attach_describe(&m->attach[k], d, sizeof(d));
+        cst_buf_printf(&b, "- %s `%s` (%s)\n",
+                       cst_attach_is_image(&m->attach[k]) ? "image"
+                                                         : "document",
+                       m->attach[k].name, d);
+      }
+      if (m->nattach)
+        cst_buf_puts(&b, "\n");
+    } else {
+      cst_buf_puts(&b, "## Assistant\n\n");
+      cst_buf_puts(&b, c);
+      cst_buf_puts(&b, "\n\n");
+      if (m->tool_calls_json && m->tool_calls_json[0]) {
+        cst_buf_puts(&b, "**Tool calls**\n\n");
+        md_fence(&b, m->tool_calls_json, 4000);
+      }
+    }
+  }
+  cst_session_unlock(s);
+  return cst_buf_steal(&b);
+}
@@ -28,11 +28,17 @@ int cst_settings_load(cst_settings_t *s) {
  FILE *fp = fopen(path, "rb");
  if (!fp)
    return -1;
-  static _Thread_local char buf[65536];
-  size_t n = fread(buf, 1, sizeof(buf) - 1, fp);
+  enum { SETTINGS_MAX = 65536 };
+  char *buf = malloc(SETTINGS_MAX);
+  if (!buf) {
+    fclose(fp);
+    return -1;
+  }
+  size_t n = fread(buf, 1, SETTINGS_MAX - 1, fp);
  fclose(fp);
  buf[n] = 0;
  cJSON *root = cJSON_Parse(buf + cst_bom_len(buf, n));
+  free(buf);
  if (!cJSON_IsObject(root)) {
    cJSON_Delete(root);
    return -1;
@@ -41,6 +47,8 @@ int cst_settings_load(cst_settings_t *s) {
           cJSON_GetObjectItemCaseSensitive(root, "provider"));
  copy_str(s->effort, sizeof(s->effort),
           cJSON_GetObjectItemCaseSensitive(root, "effort"));
+  copy_str(s->mode, sizeof(s->mode),
+           cJSON_GetObjectItemCaseSensitive(root, "mode"));
  copy_str(s->approval, sizeof(s->approval),
           cJSON_GetObjectItemCaseSensitive(root, "approval"));
  s->show_thinking =
@@ -80,6 +88,8 @@ int cst_settings_save(const cst_settings_t *s) {
    cJSON_AddStringToObject(root, "provider", s->provider);
  if (s->effort[0])
    cJSON_AddStringToObject(root, "effort", s->effort);
+  if (s->mode[0])
+    cJSON_AddStringToObject(root, "mode", s->mode);
  if (s->approval[0])
    cJSON_AddStringToObject(root, "approval", s->approval);
  if (s->show_thinking)
@@ -129,6 +139,21 @@ int cst_approval_parse(const char *s, int *out) {
  return -1;
 }

+static const char *const MODE_NAMES[] = {"coding", "general"};
+
+int cst_mode_valid(const char *m) {
+  for (size_t i = 0; m && m[0] && i < sizeof(MODE_NAMES) / sizeof(*MODE_NAMES);
+       i++) {
+    if (strcmp(m, MODE_NAMES[i]) == 0)
+      return 1;
+  }
+  return 0;
+}
+
+const char *cst_mode_name(const char *m) {
+  return cst_mode_valid(m) ? m : CST_DEFAULT_MODE;
+}
+
 const char *cst_settings_model(const cst_settings_t *s, const char *provider) {
  for (int i = 0; provider && i < s->nmodels; i++) {
    if (strcmp(s->models[i].provider, provider) == 0 && s->models[i].model[0])
@@ -142,6 +167,8 @@ int cst_settings_remember(const cst_agent_config_t *cfg) {
  cst_settings_load(&s);
  snprintf(s.provider, sizeof(s.provider), "%s", cfg->provider);
  snprintf(s.effort, sizeof(s.effort), "%s", cfg->effort);
+  if (cfg->mode[0])
+    snprintf(s.mode, sizeof(s.mode), "%s", cfg->mode);
  int i = 0;
  while (i < s.nmodels && strcmp(s.models[i].provider, cfg->provider) != 0)
    i++;
@@ -170,6 +197,13 @@ int cst_settings_remember_thinking(int show_thinking) {
  return cst_settings_save(&s);
 }

+int cst_settings_remember_mode(const char *mode) {
+  cst_settings_t s;
+  cst_settings_load(&s);
+  snprintf(s.mode, sizeof(s.mode), "%s", cst_mode_name(mode));
+  return cst_settings_save(&s);
+}
+
 static void real_ws(const char *ws, char *out, size_t n) {
  if (!ws || os_realpath(ws, out, n) != 0)
    snprintf(out, n, "%s", ws ? ws : "");
@@ -488,6 +488,16 @@ char *cst_skill_read(const char *ws, const char *name, const char *file,
  return cst_buf_steal(&b);
 }

+const char *cst_init_prompt(void) {
+  return "Analyze this repository and write an AGENTS.md file at the "
+         "workspace root for future coding agents. Cover: what the project "
+         "is, how to build, test and run it (exact commands), the directory "
+         "layout, coding conventions you can infer from the code, and "
+         "anything surprising. Explore with list_dir/glob/grep/read_file "
+         "before writing. Keep it under 150 lines. If AGENTS.md already "
+         "exists, improve it rather than starting over.";
+}
+
 char *cst_skill_prompt(const char *ws, const char *name, const char *task) {
  int code = 0;
  char *body = cst_skill_read(ws, name, NULL, &code);
@@ -40,7 +40,7 @@ static const struct {
     "runs commands as needed.",
     "", 0},
    {"explore",
-     "Fast read-only explorer: searches and reads the codebase to answer a "
+     "Fast read-only explorer: searches and reads the workspace to answer a "
     "question (where is X, how does Y work, which files do Z). Several can "
     "run at once.",
     "You only read. Search broadly first (grep, glob, list_dir), then read "
@@ -49,11 +49,11 @@ static const struct {
     "you looked for and did not find.",
     1},
    {"plan",
-     "Read-only planner: investigates the code involved and returns a "
+     "Read-only planner: investigates the material involved and returns a "
     "step-by-step implementation plan.",
-     "You only read. Investigate the code involved, then return a concrete "
-     "plan: the files to change and what to change in each, risks and edge "
-     "cases, and how to verify. Sketch code only where it removes ambiguity.",
+     "You only read. Investigate what is involved, then return a concrete "
+     "plan: what to produce or change and where, risks and edge cases, and "
+     "how to check the result. Sketch code only where it removes ambiguity.",
     1},
 };
 #define NBUILTIN ((int)(sizeof(BUILTIN) / sizeof(BUILTIN[0])))
@@ -7,6 +7,7 @@
 #include <string.h>

 #include "cst_tools.h"
+#include "cst_clipboard.h"
 #include "cst_util.h"
 #include "tools_internal.h"

@@ -106,3 +107,70 @@ int cst_tool_ask_user(const cst_tool_ctx_t *ctx, const ctl_tool_call_t *c,
                           1);
  return tool_set_resultf(r, 0, "user answered: %s", ans);
 }
+
+/* The system clipboard: what the user copied elsewhere. Reading needs no
+ * permission (READ); writing replaces what the user has copied (WRITE, so
+ * the confirmation frontier applies). */
+int cst_tool_clipboard_read(const cst_tool_ctx_t *ctx, const ctl_tool_call_t *c,
+                            cst_tool_result_t *r) {
+  (void)ctx;
+  char err[192];
+  char *text = NULL;
+  size_t n = 0;
+  if (cst_clipboard_text(&text, &n, err, sizeof(err)) != 0)
+    return tool_set_result(r, err, 1);
+  long cap = tool_arg_long(c, "max_bytes", 65536);
+  if (cap > 0 && (size_t)cap < n)
+    n = (size_t)cap;
+  cst_buf_t b = {0};
+  cst_buf_printf(&b, "clipboard text (%zu bytes%s):\n", (size_t)n,
+                 (size_t)cap < (size_t)n ? ", truncated" : "");
+  cst_buf_append(&b, text, n);
+  if (b.len && b.buf[b.len - 1] != '\n')
+    cst_buf_puts(&b, "\n");
+  free(text);
+  return tool_take_buf(r, &b, 0);
+}
+
+int cst_tool_clipboard_write(const cst_tool_ctx_t *ctx,
+                             const ctl_tool_call_t *c, cst_tool_result_t *r) {
+  (void)ctx;
+  size_t n = 0;
+  const char *text = tool_arg_n(c, "text", &n);
+  if (!text)
+    return tool_set_result(r, "error: missing text", 1);
+  char err[192];
+  if (cst_clipboard_set_text(text, n, err, sizeof(err)) != 0)
+    return tool_set_result(r, err, 1);
+  return tool_set_resultf(r, 0, "copied %zu bytes to the clipboard", n);
+}
+
+/* Confirmation preview: what the clipboard would hold afterwards. Clipboard
+ * text has no diff, so this is the text itself, cut to a readable size. */
+#define CLIP_PREVIEW_MAX 2000
+int cst_preview_clipboard_write(const cst_tool_ctx_t *ctx,
+                                const ctl_tool_call_t *c, char **out) {
+  (void)ctx;
+  size_t n = 0;
+  const char *text = tool_arg_n(c, "text", &n);
+  if (!text) {
+    *out = strdup("error: missing text");
+    return -1;
+  }
+  int lines = 0;
+  for (size_t i = 0; i < n; i++)
+    lines += text[i] == '\n';
+  size_t cut = n;
+  if (cut > CLIP_PREVIEW_MAX)
+    cut = cst_utf8_floor(text, CLIP_PREVIEW_MAX);
+  cst_buf_t b = {0};
+  cst_buf_printf(&b, "clipboard: replace with %zu bytes, %d line%s\n", n,
+                 lines + (n && text[n - 1] != '\n'), lines == 0 ? "" : "s");
+  cst_buf_append(&b, text, cut);
+  if (b.len && b.buf[b.len - 1] != '\n')
+    cst_buf_puts(&b, "\n");
+  if (cut < n)
+    cst_buf_printf(&b, "… (%zu more bytes)\n", n - cut);
+  *out = cst_buf_steal(&b);
+  return 0;
+}
@@ -267,6 +267,14 @@ int cst_tool_view_image(const cst_tool_ctx_t *ctx, const ctl_tool_call_t *c,
  cst_attachment_t *a = cst_attach_import_file(session, full, err, sizeof(err));
  if (!a)
    return tool_set_resultf(r, 1, "error: cannot show %.200s: %s", path, err);
+  if (!cst_attach_is_image(a)) {
+    cst_attach_free(a);
+    return tool_set_resultf(
+        r, 1,
+        "error: %.200s is not an image; a text file or PDF reaches the model "
+        "through /attach <path> (a PDF becomes its text layer)",
+        path);
+  }
  snprintf(a->origin, sizeof(a->origin), "view_image");
  const char *d = tool_arg(c, "detail");
  if (d && (strcmp(d, "low") == 0 || strcmp(d, "high") == 0))
@@ -66,8 +66,9 @@ static const cst_tool_def_t TOOLS[] = {
      {"detail", S, 0, "low or high (provider hint; default auto)"}},
     CST_TP_READ, 0, cst_tool_view_image, NULL, NO_DYN},
    {"run_shell",
-     "Run a non-interactive shell command in the workspace root for builds, "
-     "tests, git and other CLIs (stdin is empty; long runs are killed). "
+     "Run a non-interactive shell command in the workspace root: builds, "
+     "tests, git, scripts, computations, data processing, other CLIs "
+     "(stdin is empty; long runs are killed). "
     "Returns stdout+stderr with colors stripped and long output cut to head "
     "+ tail. Prefer read_file/grep/glob/list_dir over cat/grep/find.",
     "Run a shell command in the workspace root (non-interactive, stdin is "
@@ -158,6 +159,19 @@ static const cst_tool_def_t TOOLS[] = {
     {{"question", S, 1, "the question"},
      {"options", S, 0, "suggested answers, one per line"}},
     CST_TP_ASK, 0, cst_tool_ask_user, NULL, NO_DYN},
+    {"clipboard_read",
+     "Read the text on the system clipboard — what the user copied outside "
+     "the workspace. Returns UTF-8 text.",
+     NULL,
+     {{"max_bytes", I, 0, "cap on the returned text (default 65536)"}},
+     CST_TP_READ, 0, cst_tool_clipboard_read, NULL, NO_DYN},
+    {"clipboard_write",
+     "Replace the system clipboard with text, so the user can paste it into "
+     "another app. Overwrites what they had copied.",
+     NULL,
+     {{"text", S, 1, "text to put on the clipboard"}},
+     CST_TP_WRITE, 0, cst_tool_clipboard_write, cst_preview_clipboard_write,
+     NO_DYN},
 };
 #define NTOOLS ((int)(sizeof(TOOLS) / sizeof(TOOLS[0])))

@@ -659,9 +659,10 @@ static int plan_patch(const char *ws, const char *patch, patch_plan_t *pl,
    pl->full[i] = malloc(OS_PATH_MAX);
    if (!pl->full[i] ||
        cst_sandbox_resolve(ws, target, pl->full[i], OS_PATH_MAX) != 0) {
-      if (dry && pl->full[i] && !pl->del[i] && is_new) {
+      if (dry && pl->full[i] && !pl->del[i] && is_new &&
+          strlen(target) < OS_PATH_MAX) {
        /* parent dirs do not exist yet; the write would create them */
-        snprintf(pl->full[i], OS_PATH_MAX, "%s", target);
+        memcpy(pl->full[i], target, strlen(target) + 1);
      } else {
        snprintf(err, err_sz, "error: path outside workspace: %.200s",
                 target);
@@ -46,6 +46,13 @@ int cst_tool_todo_write(const cst_tool_ctx_t *ctx, const ctl_tool_call_t *c,
                        cst_tool_result_t *r);
 int cst_tool_ask_user(const cst_tool_ctx_t *ctx, const ctl_tool_call_t *c,
                      cst_tool_result_t *r);
+/* Clipboard text (system clipboard, not the workspace). */
+int cst_tool_clipboard_read(const cst_tool_ctx_t *ctx, const ctl_tool_call_t *c,
+                            cst_tool_result_t *r);
+int cst_tool_clipboard_write(const cst_tool_ctx_t *ctx,
+                             const ctl_tool_call_t *c, cst_tool_result_t *r);
+int cst_preview_clipboard_write(const cst_tool_ctx_t *ctx,
+                                const ctl_tool_call_t *c, char **out);
 int cst_tool_web_fetch(const cst_tool_ctx_t *ctx, const ctl_tool_call_t *c,
                       cst_tool_result_t *r);
 int cst_tool_view_image(const cst_tool_ctx_t *ctx, const ctl_tool_call_t *c,
@@ -215,6 +215,7 @@ int cst_tool_web_fetch(const cst_tool_ctx_t *ctx, const ctl_tool_call_t *c,
  memset(&f, 0, sizeof(f));
  char errbuf[CURL_ERROR_SIZE];
  errbuf[0] = 0;
+  cst_curl_apply_ca(curl);
  curl_easy_setopt(curl, CURLOPT_URL, url);
  curl_easy_setopt(curl, CURLOPT_FOLLOWLOCATION, 1L);
  curl_easy_setopt(curl, CURLOPT_MAXREDIRS, 5L);
@@ -11,6 +11,7 @@
 #include "cst_mcp.h"
 #include "cst_models.h"
 #include "cst_plugins.h"
+#include "cst_provider_switch.h"
 #include "cst_skills.h"
 #include "cst_subagent.h"
 #include "cst_textsel.h"
@@ -151,6 +152,7 @@ static int cst_clock_gettime_mono(struct timespec *ts) {
 #ifdef NCURSES_VERSION
 static int g_kitty_on;

+#ifndef _WIN32
 /* Bind the reports and ask whether the terminal speaks the protocol; the
 * flags are pushed once the reply arrives (see tui_wget_wch). */
 static void kitty_init(void) {
@@ -181,6 +183,7 @@ static void kitty_restore(void) {
    g_kitty_on = 0;
  }
 }
+#endif /* !_WIN32 */
 #endif

 /* wget_wch() that turns kitty key reports back into legacy input. */
@@ -1041,10 +1044,7 @@ static int find_tool_block(tui_t *t, const char *id) {
  return -1;
 }

-static int msg_is_tool_result(const cst_message_t *m) {
-  return m && (m->role == CST_ROLE_TOOL ||
-               (m->role == CST_ROLE_USER && m->tool_call_id[0]));
-}
+#define msg_is_tool_result cst_message_is_tool_result

 /* Strip the "[tool …]" / "result <id> exit=N <n>" header and "end" trailer. */
 static void tool_result_view(const char *content, const char **body,
@@ -2354,6 +2354,7 @@ static const char *const TIPS[] = {
    "Ctrl+P opens the command palette",
    "Type @ to mention a file in your message",
    "/effort sets reasoning effort: low, medium, high, max",
+    "/mode general reframes the agent for non-code work",
    "/model lists the models your API key can use",
    "/undo reverts the file changes of the last turn",
    "/provider switches provider; keys are saved per provider",
@@ -2832,7 +2833,10 @@ static void reload_hist_from_session(tui_t *t) {
      for (int k = 0; k < m->nattach; k++) {
        char d[128];
        cst_attach_describe(&m->attach[k], d, sizeof(d));
-        cst_buf_printf(&ub, "\n[image %s · %s]", m->attach[k].name, d);
+        cst_buf_printf(&ub, "\n[%s %s · %s]",
+                       cst_attach_is_image(&m->attach[k]) ? "image"
+                                                          : "document",
+                       m->attach[k].name, d);
      }
      block_set_body(b, ub.buf ? ub.buf : "", ub.len);
      cst_buf_free(&ub);
@@ -3886,46 +3890,11 @@ static void am_store(const char *provider, int r, char ids[][128],
  pthread_mutex_unlock(&g_am_mu);
 }

-/* GET /models; a good answer also refreshes models-cache/<provider>.json. */
-static int fetch_models(const char *provider, const char *base,
-                        const char *key, char ids[][128], int max, char *err,
-                        size_t err_sz) {
-  cst_model_meta_t *m = calloc((size_t)max, sizeof(*m));
-  if (!m) {
-    snprintf(err, err_sz, "out of memory");
-    return -1;
-  }
-  int r = cst_llm_list_models_meta(base, key, m, max, err, err_sz);
-  if (r >= 0)
-    cst_models_cache_save(provider, base, m, r);
-  for (int i = 0; i < r; i++)
-    snprintf(ids[i], 128, "%s", m[i].id);
-  free(m);
-  return r;
-}
-
-/* The cached list for provider into ids: count, or -1; *fresh when it is
- * younger than CST_MODELS_CACHE_TTL. */
-static int cached_models(const char *provider, char ids[][128], int max,
-                         int *fresh) {
-  *fresh = 0;
-  cst_model_meta_t *m = calloc((size_t)max, sizeof(*m));
-  if (!m)
-    return -1;
-  long age = 0;
-  int n = cst_models_cache_load(provider, m, max, &age);
-  for (int i = 0; i < n; i++)
-    snprintf(ids[i], 128, "%s", m[i].id);
-  free(m);
-  *fresh = n > 0 && age >= 0 && age < CST_MODELS_CACHE_TTL;
-  return n > 0 ? n : -1;
-}
-
 static void *probe_thread(void *arg) {
  probe_arg_t *a = arg;
  static char ids[AM_MAX][128]; /* one probe runs per provider switch */
  char err[256];
-  int r = fetch_models(a->provider, a->base, a->key, ids, AM_MAX, err,
+  int r = cst_ps_fetch(a->provider, a->base, a->key, ids, AM_MAX, err,
                       sizeof(err));
  if (r >= 0 || r == CST_MODELS_AUTH || !a->have_cache)
    am_store(a->provider, r, ids, err);
@@ -3937,23 +3906,13 @@ static void *probe_thread(void *arg) {
  return NULL;
 }

-/* Base URL for a provider: the live config for the current one (it may be
- * overridden), else the preset. */
-static const char *provider_base(const tui_t *t, const char *provider) {
-  const cst_agent_config_t *cfg = cst_agent_config(t->agent);
-  if (cfg && strcmp(cfg->provider, provider) == 0)
-    return cfg->api_base;
-  const cst_provider_t *p = cst_provider_find(provider);
-  return p ? p->api_base : "";
-}
-
 static void start_probe(tui_t *t) {
  const cst_agent_config_t *cfg = cst_agent_config(t->agent);
  if (!cfg || !cfg->api_key[0] || g_am_threads > 0)
    return;
  static char ids[AM_MAX][128];
  int fresh = 0;
-  int nc = cached_models(cfg->provider, ids, AM_MAX, &fresh);
+  int nc = cst_ps_cached(cfg->provider, ids, AM_MAX, &fresh);
  pthread_mutex_lock(&g_am_mu);
  snprintf(g_am.provider, sizeof(g_am.provider), "%s", cfg->provider);
  pthread_mutex_unlock(&g_am_mu);
@@ -4039,13 +3998,14 @@ static int obtain_key(tui_t *t, const char *provider, char *out, size_t n,
    return 0;
  char dir[1024] = "the config directory";
  os_config_dir(dir, sizeof(dir));
-  const char *base = provider_base(t, provider);
+  const char *base = cst_ps_base(cst_agent_config(t->agent), provider);
  char title[128], hint[1300];
  snprintf(title, sizeof(title), "API key for %s", provider);
  snprintf(hint, sizeof(hint), "%.200s · saved to %.900s/credentials (only you "
                               "can read it)",
           base, dir);
  char note[300] = "";
+  int flags;
  for (;;) {
    if (secret_input_run(t, title, hint, out, n) != 0)
      return -1;
@@ -4053,8 +4013,10 @@ static int obtain_key(tui_t *t, const char *provider, char *out, size_t n,
    draw_status(t);
    static char ids[AM_MAX][128];
    char err[256];
-    int r = fetch_models(provider, base, out, ids, AM_MAX, err, sizeof(err));
-    if (r == CST_MODELS_AUTH) {
+    int r = -1;
+    flags = cst_ps_verify_save(provider, base, out, ids, AM_MAX, &r, err,
+                               sizeof(err));
+    if (flags == CST_MODELS_AUTH) {
      snprintf(hint, sizeof(hint), "%s rejected that key (%.200s) — paste "
                                   "another one",
               base, err);
@@ -4064,13 +4026,13 @@ static int obtain_key(tui_t *t, const char *provider, char *out, size_t n,
    snprintf(g_am.provider, sizeof(g_am.provider), "%s", provider);
    pthread_mutex_unlock(&g_am_mu);
    am_store(provider, r, ids, err);
-    if (r < 0)
+    if (flags & CST_PS_UNVERIFIED)
      snprintf(note, sizeof(note), "could not verify the key (%.200s); "
                                   "saved anyway",
               err);
    break;
  }
-  if (cst_creds_save(provider, out) != 0)
+  if (flags & CST_PS_UNSAVED)
    set_status(t, "warning: could not save the key; it is used for this run "
                  "only");
  else if (note[0])
@@ -4302,7 +4264,7 @@ static void cmd_model(tui_t *t, char *arg, int *quit) {
      set_status(t, "fetching the model list…");
      draw_status(t);
      char err[256];
-      na = fetch_models(cfg->provider, cfg->api_base, cfg->api_key, names,
+      na = cst_ps_fetch(cfg->provider, cfg->api_base, cfg->api_key, names,
                        AM_MAX, err, sizeof(err));
      pthread_mutex_lock(&g_am_mu);
      snprintf(g_am.provider, sizeof(g_am.provider), "%s", cfg->provider);
@@ -4403,6 +4365,41 @@ static void cmd_effort(tui_t *t, char *arg, int *quit) {
                     cfg->effort, cfg->provider, cfg->model);
 }

+static void cmd_mode(tui_t *t, char *arg, int *quit) {
+  (void)quit;
+  cst_agent_config_t *cfg = cst_agent_config_mut(t->agent);
+  static const char *const names[] = {"coding", "general"};
+  static const char *const desc[] = {
+      "code framed: repo, build, tests (default)",
+      "documents, data, research, everyday files"};
+  const char *want = NULL;
+  if (arg && arg[0]) {
+    if (!cst_mode_valid(arg)) {
+      push_system_line(t, "Unknown mode '%s' (coding | general)", arg);
+      return;
+    }
+    want = arg;
+  } else {
+    char items[2][128];
+    int cur_idx = 0;
+    for (int i = 0; i < 2; i++) {
+      snprintf(items[i], 128, "%-8s %s", names[i], desc[i]);
+      if (strcmp(names[i], cst_mode_name(cfg->mode)) == 0)
+        cur_idx = i;
+    }
+    int pick = picker_run(t, "Working mode", items, 2, cur_idx);
+    if (pick < 0)
+      return;
+    want = names[pick];
+  }
+  snprintf(cfg->mode, sizeof(cfg->mode), "%s", want);
+  cst_settings_remember_mode(cfg->mode);
+  push_system_line(t, "(mode → %s%s)", cfg->mode,
+                   strcmp(cfg->mode, "general") == 0
+                       ? "; the system prompt is rewritten on the next turn"
+                       : "");
+}
+
 static void cmd_provider(tui_t *t, char *arg, int *quit) {
  (void)quit;
  cst_agent_config_t *cfg = cst_agent_config_mut(t->agent);
@@ -4436,15 +4433,8 @@ static void cmd_provider(tui_t *t, char *arg, int *quit) {
    push_system_line(t, "(provider unchanged: %s needs an API key)", p->name);
    return;
  }
-  cst_settings_t s;
-  cst_settings_load(&s);
-  const char *model = cst_settings_model(&s, p->name);
-  snprintf(cfg->provider, sizeof(cfg->provider), "%s", p->name);
-  snprintf(cfg->api_base, sizeof(cfg->api_base), "%s", p->api_base);
-  snprintf(cfg->model, sizeof(cfg->model), "%s", model ? model : p->model);
-  snprintf(cfg->api_key, sizeof(cfg->api_key), "%s", key);
+  cst_ps_apply_provider(cfg, p, key);
  memset(key, 0, sizeof(key));
-  cst_settings_remember(cfg);
  push_system_line(t, "(provider → %s · %s · model %s)", p->name, p->api_base,
                   cfg->model);
  if (am_models(p->name, NULL, 0) < 0)
@@ -4688,14 +4678,15 @@ static void cmd_status(tui_t *t, char *arg, int *quit) {
                   : !same          ? "this run only (--api-key / CST_API_KEY)"
                   : strcmp(from, "saved") == 0 ? "saved (/key to replace)"
                                                : from);
-  push_system_line(t, "  protocol   %s · tools %s · max steps %d",
+  push_system_line(t, "  protocol   %s · tools %s · mode %s · max steps %d",
                   cst_protocol_name(cfg->protocol),
-                   cst_tool_profile_name(cfg->tool_profile), cfg->max_steps);
+                   cst_tool_profile_name(cfg->tool_profile),
+                   cst_mode_name(cfg->mode), cfg->max_steps);
  char appr[64];
  approval_label(t, appr, sizeof(appr));
  push_system_line(t, "  approval   %s — %s", appr,
                   approval_help(approval_mode(t)));
-  push_system_line(t, "  mode       %s",
+  push_system_line(t, "  plan       %s",
                   cfg->plan_mode ? "Plan (read-only tools)" : "Build");
  const char *notes = cst_agent_notes(t->agent);
  push_system_line(t, "  notes      %s",
@@ -4822,12 +4813,6 @@ static void cmd_diff(tui_t *t, char *arg, int *quit) {
  free(df.out);
 }

-static void md_fence(FILE *fp, const char *s, size_t max) {
-  size_t n = strlen(s);
-  fprintf(fp, "````\n%.*s%s\n````\n\n", (int)(n > max ? max : n), s,
-          n > max ? "\n…(truncated)" : "");
-}
-
 static void cmd_export(tui_t *t, char *arg, int *quit) {
  (void)quit;
  const cst_agent_config_t *cfg = cst_agent_config(t->agent);
@@ -4847,44 +4832,20 @@ static void cmd_export(tui_t *t, char *arg, int *quit) {
    *sl = 0;
    os_mkdir_p(dir, 0755);
  }
-  FILE *fp = fopen(path, "wb");
+  char *md = cst_session_export_md(cst_agent_session(t->agent), t->title, sid,
+                                   cfg->model);
+  FILE *fp = md ? fopen(path, "wb") : NULL;
  if (!fp) {
+    free(md);
    push_system_line(t, "Error: cannot write %s", path);
    return;
  }
-  fprintf(fp, "# %s\n\nsession `%s` · model `%s`\n\n", t->title, sid,
-          cfg->model);
-  cst_session_t *s = cst_agent_session(t->agent);
-  cst_session_lock(s);
-  for (int i = 0; i < cst_session_count(s); i++) {
-    const cst_message_t *m = cst_session_get(s, i);
-    const char *c = m->content ? m->content : "";
-    if (m->role == CST_ROLE_SYSTEM)
-      continue;
-    if (msg_is_tool_result(m)) {
-      fprintf(fp, "**Tool result** `%s`\n\n", m->tool_name);
-      md_fence(fp, c, 4000);
-    } else if (cst_message_is_note(m)) {
-      fprintf(fp, "> %.*s\n\n", (int)strcspn(c, "\n"), c);
-    } else if (m->role == CST_ROLE_USER) {
-      fprintf(fp, "## User\n\n%s\n\n", c);
-      for (int k = 0; k < m->nattach; k++) {
-        char d[128];
-        cst_attach_describe(&m->attach[k], d, sizeof(d));
-        fprintf(fp, "- image `%s` (%s)\n", m->attach[k].name, d);
-      }
-      if (m->nattach)
-        fputc('\n', fp);
-    } else {
-      fprintf(fp, "## Assistant\n\n%s\n\n", c);
-      if (m->tool_calls_json && m->tool_calls_json[0]) {
-        fprintf(fp, "**Tool calls**\n\n");
-        md_fence(fp, m->tool_calls_json, 4000);
-      }
-    }
+  fputs(md, fp);
+  free(md);
+  if (fclose(fp) != 0) {
+    push_system_line(t, "Error: cannot write %s", path);
+    return;
  }
-  cst_session_unlock(s);
-  fclose(fp);
  push_system_line(t, "(exported to %s)", path);
 }

@@ -4904,19 +4865,39 @@ static void report_attached(tui_t *t, int added) {
                     cst_agent_config(t->agent)->model);
 }

-/* /attach: stage images for the next user message. The bytes are copied into
- * the session's attachment directory now, so a picture picked from /tmp keeps
- * working after the file is gone. */
+/* /attach paste: the clipboard image if there is one, else its text goes
+ * into the composer (CR dropped, control bytes kept out, INPUT_CAP respected). */
+static int paste_text_into_input(tui_t *t, const char *text, size_t n) {
+  char *buf = malloc(n + 1);
+  if (!buf)
+    return 0;
+  size_t k = 0;
+  for (size_t i = 0; i < n && t->in_len + k + 1 < INPUT_CAP; i++) {
+    unsigned char c = (unsigned char)text[i];
+    if (c == '\r' || (c < 0x20 && c != '\n' && c != '\t'))
+      continue;
+    buf[k++] = (char)c;
+  }
+  if (k)
+    input_insert(t, buf, k);
+  free(buf);
+  return (int)k;
+}
+
+/* /attach: stage files for the next user message -- images, PDFs (reduced to
+ * their text) and text. The payload is copied into the session's attachment
+ * directory now, so a file picked from /tmp keeps working after it is gone. */
 static void cmd_attach(tui_t *t, char *arg, int *quit) {
  (void)quit;
  int n = cst_agent_pending_count(t->agent);
  if (!arg || !arg[0] || strcmp(arg, "list") == 0) {
    if (n == 0) {
-      push_system_line(t, "No images picked. Usage: /attach <path>");
+      push_system_line(t, "Nothing attached yet. Usage: /attach <path>, "
+                          "/attach paste, /attach list, /attach clear");
      return;
    }
-    push_system_line(t, "%d image%s with the next message (sent once):", n,
-                     n > 1 ? "s" : "");
+    push_system_line(t, "%d attachment%s with the next message (sent once):",
+                     n, n > 1 ? "s" : "");
    for (int i = 0; i < n; i++) {
      char line[256];
      if (cst_agent_pending_describe(t->agent, i, line, sizeof(line)) == 0)
@@ -4927,7 +4908,7 @@ static void cmd_attach(tui_t *t, char *arg, int *quit) {
  }
  if (strcmp(arg, "clear") == 0 || strcmp(arg, "none") == 0) {
    cst_agent_pending_clear(t->agent);
-    push_system_line(t, "Dropped the picked images.");
+    push_system_line(t, "Dropped the attachments.");
    return;
  }
  char err[256];
@@ -4935,7 +4916,24 @@ static void cmd_attach(tui_t *t, char *arg, int *quit) {
    char *img = NULL;
    size_t len = 0;
    if (cst_clipboard_image(&img, &len, err, sizeof(err)) != 0) {
-      push_system_line(t, "Error: %s", err);
+      /* no image: put the clipboard text in the composer instead */
+      char *text = NULL;
+      size_t tn = 0;
+      char terr[192];
+      if (cst_clipboard_text(&text, &tn, terr, sizeof(terr)) != 0) {
+        push_system_line(t, "Error: %s", terr);
+        return;
+      }
+      int added = paste_text_into_input(t, text, tn);
+      free(text);
+      if (added <= 0) {
+        push_system_line(t, "(no room in the composer for the clipboard "
+                            "text)");
+        return;
+      }
+      push_system_line(t, "(pasted %d bytes of clipboard text into the "
+                          "composer)", added);
+      layout_maybe_grow_input(t);
      return;
    }
    char mime[32], name[32];
@@ -4960,19 +4958,29 @@ static void cmd_attach(tui_t *t, char *arg, int *quit) {
  report_attached(t, 1);
 }

-static int file_is_image(const char *path) {
+/* A file worth attaching: an image, a PDF, or something that reads as text.
+ * Deciding here in the TUI only avoids turning a random dropped binary into
+ * a pasted path; the store checks the content for real. */
+static int file_is_attachable(const char *path) {
  FILE *f = fopen(path, "rb");
  if (!f)
    return 0;
-  unsigned char head[16];
+  unsigned char head[64];
  size_t got = fread(head, 1, sizeof(head), f);
  fclose(f);
  char mime[32];
-  return cst_attach_sniff(head, got, mime, sizeof(mime)) == 0;
+  if (cst_attach_sniff(head, got, mime, sizeof(mime)) == 0)
+    return 1;
+  if (got >= 4 && memcmp(head, "%PDF", 4) == 0)
+    return 1;
+  for (size_t i = 0; i < got; i++)
+    if (head[i] == 0 || head[i] < 0x09)
+      return 0;
+  return got > 0;
 }

-/* A paste that is nothing but image file paths (a file dropped on the
- * terminal) attaches those files instead of typing the paths. */
+/* A paste that is nothing but file paths (a file dropped on the terminal)
+ * attaches those files instead of typing the paths. */
 static void paste_image_paths(tui_t *t) {
  size_t from = t->paste_at, to = t->cur;
  if (to <= from || to - from != t->in_len - t->paste_len0 ||
@@ -4984,7 +4992,7 @@ static void paste_image_paths(tui_t *t) {
  if (n <= 0)
    return;
  for (int i = 0; i < n; i++)
-    if (!file_is_image(paths[i]))
+    if (!file_is_attachable(paths[i]))
      return;
  input_erase(t, from, to);
  int added = 0;
@@ -5061,14 +5069,7 @@ static void cmd_init(tui_t *t, char *arg, int *quit) {
  (void)arg;
  (void)quit;
  free(t->pending_prompt);
-  t->pending_prompt = xstrdup(
-      "Analyze this repository and write an AGENTS.md file at the workspace "
-      "root for future coding agents. Cover: what the project is, how to "
-      "build, test and run it (exact commands), the directory layout, coding "
-      "conventions you can infer from the code, and anything surprising. "
-      "Explore with list_dir/glob/grep/read_file before writing. Keep it "
-      "under 150 lines. If AGENTS.md already exists, improve it rather than "
-      "starting over.");
+  t->pending_prompt = xstrdup(cst_init_prompt());
 }

 static void cmd_compact(tui_t *t, char *arg, int *quit) {
@@ -5364,6 +5365,151 @@ static void cmd_quit(tui_t *t, char *arg, int *quit) {
  *quit = 1;
 }

+/* /web: the WebUI is a second front end for the same engine and is designed to
+ * run as its own process, so this starts one on the same workspace and leaves
+ * the TUI holding the terminal. The child gets a fresh session: two writers on
+ * one session file is the one thing that would corrupt it. */
+#ifndef _WIN32
+#include <fcntl.h>
+#include <sys/wait.h> /* waitpid: collect the WebUI started by the last /web */
+#endif
+#ifndef CST_WEBUI_DEFAULT_PORT
+#define CST_WEBUI_DEFAULT_PORT 8787 /* keep in sync with include/cst_webui.h */
+#endif
+static void cmd_web(tui_t *t, char *arg, int *quit) {
+  (void)arg;
+  (void)quit;
+#ifdef _WIN32
+  push_system_line(t, "open another terminal and run: cst-agent --web");
+#else
+  static pid_t spawned = 0;
+  if (spawned > 0) {          /* collect the one we started last time */
+    waitpid(spawned, NULL, WNOHANG);
+    spawned = 0;
+  }
+  char dir[4096], exe[4096 + 16];
+  if (os_exe_dir(dir, sizeof(dir)) != 0) {
+    push_system_line(t, "cannot locate this binary: run cst-agent --web yourself");
+    return;
+  }
+  snprintf(exe, sizeof(exe), "%s/cst-agent", dir);
+  const cst_agent_config_t *cfg = cst_agent_config(t->agent);
+  const char *ws = (cfg && cfg->workspace_root[0]) ? cfg->workspace_root : ".";
+  char logf[4096 + 32];
+  snprintf(logf, sizeof(logf), "%s/.agent", ws);
+  os_mkdir_p(logf, 0755);
+  snprintf(logf, sizeof(logf), "%s/.agent/webui.log", ws);
+  /* The WebUI talks to what this session talks to (-b, -m, a /provider
+   * switch...). The key goes through the environment, not argv (ps). */
+  const char *argv[20];
+  int na = 0;
+  argv[na++] = "cst-agent";
+  argv[na++] = "--web";
+  argv[na++] = "--workspace";
+  argv[na++] = ws;
+  argv[na++] = "--new";
+  if (cfg && cfg->provider[0])
+    argv[na++] = "--provider", argv[na++] = cfg->provider;
+  if (cfg && cfg->model[0])
+    argv[na++] = "--model", argv[na++] = cfg->model;
+  if (cfg && cfg->api_base[0])
+    argv[na++] = "--api-base", argv[na++] = cfg->api_base;
+  if (cfg && cfg->effort[0])
+    argv[na++] = "--effort", argv[na++] = cfg->effort;
+  if (cfg)
+    argv[na++] = "--approval", argv[na++] = cst_approval_name(cfg->approval);
+  argv[na] = NULL;
+  size_t nenv = 0;
+  while (environ[nenv])
+    nenv++;
+  char **envp = calloc(nenv + 2, sizeof(char *));
+  char keyvar[600] = "";
+  if (!envp) {
+    push_system_line(t, "out of memory");
+    return;
+  }
+  size_t ne = 0;
+  for (size_t i = 0; i < nenv; i++)
+    if (strncmp(environ[i], "CST_API_KEY=", 12) != 0)
+      envp[ne++] = environ[i];
+  if (cfg && cfg->api_key[0]) {
+    snprintf(keyvar, sizeof(keyvar), "CST_API_KEY=%s", cfg->api_key);
+    envp[ne++] = keyvar;
+  }
+  envp[ne] = NULL;
+  pid_t pid = fork();
+  if (pid < 0) {
+    free(envp);
+    push_system_line(t, "cannot start a second process");
+    return;
+  }
+  if (pid == 0) {
+    setsid();
+    /* never inherit the TUI's terminal: its output would land on our screen */
+    int in = open("/dev/null", O_RDONLY);
+    if (in >= 0 && in != STDIN_FILENO) {
+      dup2(in, STDIN_FILENO);
+      close(in);
+    }
+    int fd = open(logf, O_WRONLY | O_CREAT | O_TRUNC, 0600);
+    if (fd < 0)
+      fd = open("/dev/null", O_WRONLY);
+    if (fd >= 0) {
+      dup2(fd, STDOUT_FILENO);
+      dup2(fd, STDERR_FILENO);
+      if (fd > 2)
+        close(fd);
+    }
+    execve(exe, (char *const *)argv, envp);
+    _exit(127);
+  }
+  free(envp);
+  spawned = pid;
+  /* Wait (2 s at most) for the URL it prints once it listens: the port may
+   * come from CST_WEB_PORT, or it may not start at all. */
+  char url[512] = "";
+  int died = 0;
+  for (int i = 0; i < 40 && !url[0] && !died; i++) {
+    usleep(50000);
+    FILE *fp = fopen(logf, "r");
+    char line[512];
+    while (fp && fgets(line, sizeof(line), fp)) {
+      if (strncmp(line, "webui: http", 11) == 0) {
+        line[strcspn(line, "\r\n")] = 0;
+        snprintf(url, sizeof(url), "%s", line + 7);
+        break;
+      }
+    }
+    if (fp)
+      fclose(fp);
+    if (!url[0] && waitpid(pid, NULL, WNOHANG) == pid) {
+      died = 1;
+      spawned = 0;
+    }
+  }
+  /* its --new made the WebUI's session CURRENT; a plain `cst-agent` later
+   * should still resume this one */
+  char sid[128] = "";
+  cst_agent_session_id(t->agent, sid, sizeof(sid));
+  if (sid[0])
+    cst_agent_store_set_current(ws, sid);
+  char msg[4200];
+  if (died)
+    snprintf(msg, sizeof(msg), "the WebUI did not start (port busy?): see %s",
+             logf);
+  else if (url[0])
+    snprintf(msg, sizeof(msg), "WebUI started in the background: %s (new session)",
+             url);
+  else
+    snprintf(msg, sizeof(msg), "WebUI starting in the background "
+                               "(http://127.0.0.1:%d/ unless CST_WEB_PORT says "
+                               "otherwise)", CST_WEBUI_DEFAULT_PORT);
+  push_system_line(t, msg);
+  snprintf(msg, sizeof(msg), "its log: %s", logf);
+  push_system_line(t, msg);
+#endif
+}
+
 static const slash_cmd_t SLASH[] = {
    {"help", "?", "", "list commands and keys", 1, cmd_help},
    {"new", "clear", "", "start a new session", 0, cmd_new},
@@ -5382,6 +5528,8 @@ static const slash_cmd_t SLASH[] = {
     cmd_protocol},
    {"tools", "", "[on|off NAME|minimal]", "list or toggle tools", 0,
     cmd_tools},
+    {"mode", "", "[coding|general]", "coding or general (non-code) framing", 1,
+     cmd_mode},
    {"permissions", "perms", "[reset]", "show saved always-allow rules", 1,
     cmd_permissions},
    {"sandbox", "", "[reset-labels]", "shell sandbox and labeled folders", 1,
@@ -5405,8 +5553,8 @@ static const slash_cmd_t SLASH[] = {
    {"export", "", "[file]", "write the transcript as Markdown", 1,
     cmd_export},
    {"copy", "", "", "copy the last reply (OSC 52)", 1, cmd_copy},
-    {"attach", "image", "[path|paste|clear]",
-     "send image(s) with the next message",
+    {"attach", "image|file", "[path|paste|clear]",
+     "send files (image, PDF, text) or clipboard text with the next message",
     1, cmd_attach},
    {"init", "", "", "have the agent write AGENTS.md", 0, cmd_init},
    {"details", "", "", "expand/collapse all tool output", 1, cmd_details},
@@ -5422,6 +5570,8 @@ static const slash_cmd_t SLASH[] = {
     cmd_agents},
    {"skill", "", "<name> [task]", "run a task with a skill loaded", 0,
     cmd_skill},
+    {"web", "", "", "start the WebUI in the background (127.0.0.1:8787)", 1,
+     cmd_web},
    {"quit", "q exit", "", "quit", 1, cmd_quit},
 };

@@ -99,6 +99,26 @@ size_t cst_utf8_floor(const char *s, size_t n) {
  return i - 1; /* sequence incomplete: cut before lead byte */
 }

+char *cst_utf8_hold_feed(cst_utf8_hold_t *h, const char *s, size_t n,
+                         size_t *out_len) {
+  size_t have = h->n;
+  char *all = malloc(have + n + 1);
+  if (!all)
+    return NULL;
+  memcpy(all, h->b, have);
+  if (n)
+    memcpy(all + have, s, n);
+  size_t len = have + n;
+  size_t cut = cst_utf8_floor(all, len);
+  if (len - cut > sizeof(h->b))
+    cut = len;
+  h->n = len - cut;
+  memcpy(h->b, all + cut, h->n);
+  all[cut] = 0;
+  *out_len = cut;
+  return all;
+}
+
 static size_t utf8_seq_ok(const unsigned char *p, size_t left) {
  unsigned char c = p[0];
  size_t need;
@@ -64,7 +64,17 @@ static void attr_append(cst_buf_t *b, const char *name, size_t nlen,
  cst_buf_puts(b, " ");
  cst_buf_append(b, name, nlen);
  cst_buf_puts(b, "=\"");
-  cst_buf_append(b, val, vlen);
+  /* a single-quoted or bare value may hold '"': it must not end ours */
+  for (size_t i = 0; i < vlen; i++) {
+    if (val[i] == '"')
+      cst_buf_puts(b, "&quot;");
+    else if (val[i] == '<')
+      cst_buf_puts(b, "&lt;");
+    else if (val[i] == '>')
+      cst_buf_puts(b, "&gt;");
+    else
+      cst_buf_append(b, val + i, 1);
+  }
  cst_buf_puts(b, "\"");
 }

@@ -325,6 +325,100 @@ static void test_import_file(const char *root) {
         "missing file refused");
 }

+/* Documents: text is stored as text, a PDF as its text layer while the mime
+ * still says PDF, and neither is an image as far as the wire is concerned. */
+static void test_documents(const char *root) {
+  char session[1024], src[1024];
+  snprintf(session, sizeof(session), "%.900s/sessions/doc.jsonl", root);
+  char err[256] = "";
+
+  const char *text = "# Notes\n\nFive painters, one canvas.\n";
+  snprintf(src, sizeof(src), "%.900s/notes.md", root);
+  FILE *f = fopen(src, "wb");
+  EXPECT(f != NULL, "write the text file");
+  if (f) {
+    fwrite(text, 1, strlen(text), f);
+    fclose(f);
+  }
+  cst_attachment_t *a = cst_attach_import_file(session, src, err, sizeof(err));
+  EXPECT(a != NULL, "text file imported");
+  if (a) {
+    EXPECT(strcmp(a->mime, "text/plain") == 0, "text mime");
+    EXPECT(!cst_attach_is_image(a) && cst_attach_is_text(a), "not an image");
+    EXPECT(a->pages == 0, "no pages for text");
+    EXPECT(strstr(a->rel, ".txt") != NULL, "text stored as .txt");
+    EXPECT(a->bytes == (long long)strlen(text), "text size");
+    size_t n = 0;
+    char *back = cst_attach_read(session, a, &n, err, sizeof(err));
+    EXPECT(back && n == strlen(text) && memcmp(back, text, n) == 0,
+           "text reads back exactly");
+    free(back);
+    EXPECT(cst_attach_data_url(session, a) == NULL,
+           "a document has no data url");
+    char desc[128];
+    cst_attach_describe(a, desc, sizeof(desc));
+    EXPECT(strstr(desc, "text") != NULL, "describe says text");
+    cst_attach_free(a);
+  }
+
+  /* a one-page PDF with a text layer */
+  static const char PDF[] =
+      "%PDF-1.4\n"
+      "1 0 obj\n<< /Type /Catalog /Pages 2 0 R >>\nendobj\n"
+      "2 0 obj\n<< /Type /Pages /Kids [3 0 R] /Count 1 >>\nendobj\n"
+      "3 0 obj\n<< /Type /Page /Parent 2 0 R /Resources << /Font << /F1 4 0 R "
+      ">> >> /Contents 5 0 R >>\nendobj\n"
+      "4 0 obj\n<< /Type /Font /Subtype /Type1 /BaseFont /Helvetica >>\n"
+      "endobj\n"
+      "5 0 obj\n<< >>\nstream\n"
+      "BT /F1 12 Tf 72 720 Td (Attached paper abstract) Tj ET\n"
+      "endstream\nendobj\n"
+      "trailer\n<< /Root 1 0 R >>\n%%EOF\n";
+  snprintf(src, sizeof(src), "%.900s/paper.pdf", root);
+  f = fopen(src, "wb");
+  EXPECT(f != NULL, "write the PDF");
+  if (f) {
+    fwrite(PDF, 1, strlen(PDF), f);
+    fclose(f);
+  }
+  err[0] = 0;
+  a = cst_attach_import_file(session, src, err, sizeof(err));
+  EXPECT(a != NULL, "PDF imported");
+  if (a) {
+    EXPECT(strcmp(a->mime, "application/pdf") == 0, "the PDF keeps its mime");
+    EXPECT(a->pages == 1, "page count comes from the extractor");
+    EXPECT(strstr(a->rel, ".txt") != NULL, "the text layer is what is stored");
+    EXPECT(cst_attach_is_text(a), "documents are text on the wire");
+    char desc[128];
+    cst_attach_describe(a, desc, sizeof(desc));
+    EXPECT(strstr(desc, "pdf") != NULL && strstr(desc, "1 page") != NULL,
+           "describe names the PDF and its pages");
+    size_t n = 0;
+    char *back = cst_attach_read(session, a, &n, err, sizeof(err));
+    EXPECT(back && strstr(back, "Attached paper abstract") != NULL,
+           "the extracted text reads back");
+    free(back);
+    cst_attach_free(a);
+  }
+  err[0] = 0;
+  EXPECT(cst_attach_import_file(session, "/nonexistent/x.pdf", err,
+                                sizeof(err)) == NULL,
+         "a missing PDF is refused");
+
+  /* neither image, PDF nor text: a binary blob */
+  static const char binary[] = "\x01\x02\x03\x00\x04\x05";
+  snprintf(src, sizeof(src), "%.900s/blob.bin", root);
+  f = fopen(src, "wb");
+  if (f) {
+    fwrite(binary, 1, sizeof(binary), f);
+    fclose(f);
+  }
+  err[0] = 0;
+  EXPECT(cst_attach_import_file(session, src, err, sizeof(err)) == NULL,
+         "a binary file is refused");
+  EXPECT(strstr(err, "not an image") != NULL, "err says what was expected");
+}
+
 static void test_json_roundtrip(void) {
  cst_attachment_t list[3];
  memset(list, 0, sizeof(list));
@@ -344,6 +438,8 @@ static void test_json_roundtrip(void) {
  list[1].bytes = 4096;
  snprintf(list[1].detail, sizeof(list[1].detail), "high");
  snprintf(list[1].origin, sizeof(list[1].origin), "view_image");
+  snprintf(list[1].mime, sizeof(list[1].mime), "application/pdf");
+  list[1].pages = 12;

  snprintf(list[2].id, sizeof(list[2].id), "img-3");
  snprintf(list[2].name, sizeof(list[2].name), "c.webp");
@@ -365,6 +461,9 @@ static void test_json_roundtrip(void) {
    EXPECT(strcmp(back[1].detail, "high") == 0 &&
               strcmp(back[1].origin, "view_image") == 0,
           "detail and origin survive");
+    EXPECT(strcmp(back[1].mime, "application/pdf") == 0 && back[1].pages == 12,
+           "a document keeps its mime and page count");
+    EXPECT(back[2].pages == 0, "pages default to 0");
    EXPECT(strcmp(back[2].mime, "image/webp") == 0 && back[2].width == 0,
           "entry without dimensions");
    EXPECT(strcmp(back[0].detail, "auto") == 0, "missing detail defaults");
@@ -530,6 +629,7 @@ int main(void) {
  test_dib();
  test_store(root);
  test_import_file(root);
+  test_documents(root);
  test_json_roundtrip();
  test_lists();
  test_agent_pending(root);
@@ -106,11 +106,13 @@ static void test_settings(void) {
  snprintf(cfg.provider, sizeof(cfg.provider), "openai");
  snprintf(cfg.model, sizeof(cfg.model), "gpt-5");
  snprintf(cfg.effort, sizeof(cfg.effort), "max");
+  snprintf(cfg.mode, sizeof(cfg.mode), "general");
  CHECK(cst_settings_remember(&cfg) == 0);

  CHECK(cst_settings_load(&s) == 0);
  CHECK(strcmp(s.provider, "openai") == 0);
  CHECK(strcmp(s.effort, "max") == 0);
+  CHECK(strcmp(s.mode, "general") == 0);
  CHECK(strcmp(cst_settings_model(&s, "deepseek"), "deepseek-v4-pro") == 0);
  CHECK(strcmp(cst_settings_model(&s, "openai"), "gpt-5") == 0);
  CHECK(s.nmodels == 2);
@@ -145,6 +147,24 @@ static void test_effort(void) {
  free(body);
 }

+static void test_mode(void) {
+  CHECK(cst_mode_valid("coding") && cst_mode_valid("general"));
+  CHECK(!cst_mode_valid("code") && !cst_mode_valid(NULL));
+  CHECK(!cst_mode_valid(""));
+  CHECK(strcmp(cst_mode_name("general"), "general") == 0);
+  CHECK(strcmp(cst_mode_name(NULL), CST_DEFAULT_MODE) == 0);
+  CHECK(strcmp(cst_mode_name("nonsense"), "coding") == 0);
+
+  CHECK(cst_settings_remember_mode("general") == 0);
+  cst_settings_t s;
+  CHECK(cst_settings_load(&s) == 0);
+  CHECK(strcmp(s.mode, "general") == 0);
+  /* an unknown value is recorded as the default, not as itself */
+  CHECK(cst_settings_remember_mode("junk") == 0);
+  CHECK(cst_settings_load(&s) == 0);
+  CHECK(strcmp(s.mode, "coding") == 0);
+}
+
 int main(void) {
  CHECK(test_tmpdir(g_dir, sizeof(g_dir), "cst-config") == 0);
  setenv("XDG_CONFIG_HOME", g_dir, 1);
@@ -158,6 +178,7 @@ int main(void) {
  test_creds();
  test_settings();
  test_effort();
+  test_mode();

  CHECK(test_rmrf(g_dir) == 0);
  printf("test_config: OK\n");
@@ -406,9 +406,118 @@ static void test_images_on(void) {
  cst_attach_free(img);
 }

+/* Every text part of message i, joined (malloc'd, "" when there is none). */
+static char *text_of(const cJSON *root, int i) {
+  const cJSON *m =
+      cJSON_GetArrayItem(cJSON_GetObjectItemCaseSensitive(root, "messages"), i);
+  const cJSON *c = cJSON_GetObjectItemCaseSensitive(m, "content");
+  if (cJSON_IsString(c)) {
+    size_t n = strlen(c->valuestring);
+    char *out = malloc(n + 1);
+    if (out)
+      memcpy(out, c->valuestring, n + 1);
+    return out;
+  }
+  size_t len = 1;
+  const cJSON *p, *t, *s;
+  cJSON_ArrayForEach(p, c) {
+    t = cJSON_GetObjectItemCaseSensitive(p, "type");
+    s = cJSON_GetObjectItemCaseSensitive(p, "text");
+    if (cJSON_IsString(t) && strcmp(t->valuestring, "text") == 0 &&
+        cJSON_IsString(s))
+      len += strlen(s->valuestring) + 1;
+  }
+  char *out = calloc(1, len);
+  if (!out)
+    return NULL;
+  cJSON_ArrayForEach(p, c) {
+    t = cJSON_GetObjectItemCaseSensitive(p, "type");
+    s = cJSON_GetObjectItemCaseSensitive(p, "text");
+    if (cJSON_IsString(t) && strcmp(t->valuestring, "text") == 0 &&
+        cJSON_IsString(s)) {
+      strcat(out, s->valuestring);
+      strcat(out, "\n");
+    }
+  }
+  return out;
+}
+
+/* A document is inlined as text, so a text-only model and a message that
+ * also carries an image both keep it; the image is what becomes a note. */
+static void test_document_parts(void) {
+  char dir[1024], sess[1100], err[256], src[1200];
+  if (test_tmpdir(dir, sizeof(dir), "cst-body-doc") != 0) {
+    EXPECT(0, "scratch dir");
+    return;
+  }
+  snprintf(sess, sizeof(sess), "%s/s.jsonl", dir);
+  snprintf(src, sizeof(src), "%s/paper.pdf", dir);
+  static const char PDF[] =
+      "%PDF-1.4\n"
+      "1 0 obj\n<< /Type /Catalog /Pages 2 0 R >>\nendobj\n"
+      "2 0 obj\n<< /Type /Pages /Kids [3 0 R] /Count 1 >>\nendobj\n"
+      "3 0 obj\n<< /Type /Page /Parent 2 0 R /Resources << /Font << /F1 4 0 R "
+      ">> >> /Contents 5 0 R >>\nendobj\n"
+      "4 0 obj\n<< /Type /Font /Subtype /Type1 /BaseFont /Helvetica >>\n"
+      "endobj\n"
+      "5 0 obj\n<< >>\nstream\n"
+      "BT /F1 12 Tf 72 720 Td (Inlined paper text) Tj ET\n"
+      "endstream\nendobj\n"
+      "trailer\n<< /Root 1 0 R >>\n%%EOF\n";
+  FILE *f = fopen(src, "wb");
+  EXPECT(f != NULL, "write the PDF");
+  if (!f)
+    return;
+  fwrite(PDF, 1, strlen(PDF), f);
+  fclose(f);
+  cst_attachment_t *doc = cst_attach_import_file(sess, src, err, sizeof(err));
+  EXPECT(doc != NULL, "PDF stored");
+  if (!doc)
+    return;
+  static const unsigned char png[32] = {0x89, 'P', 'N', 'G', '\r', '\n', 0x1A,
+                                        '\n', 0, 0, 0, 13, 'I', 'H', 'D', 'R',
+                                        0, 0, 0, 2, 0, 0, 0, 2};
+  cst_attachment_t *img = cst_attach_import_bytes(
+      sess, "shot.png", (const char *)png, sizeof(png), err, sizeof(err));
+  EXPECT(img != NULL, "image stored");
+  if (!img) {
+    cst_attach_free(doc);
+    return;
+  }
+  cst_attachment_t both[2] = {*doc, *img};
+  cst_message_t m[2];
+  m[0] = msg(CST_ROLE_USER, "what does this say?");
+  m[0].attach = both;
+  m[0].nattach = 2;
+  m[1] = msg(CST_ROLE_USER, "thanks");
+  cst_llm_request_t r;
+  memset(&r, 0, sizeof(r));
+  r.model = "deepseek-flash";
+  r.msgs = m;
+  r.nmsgs = 2;
+  r.images_off = 1; /* a text-only model */
+  r.session_path = sess;
+  cJSON *j = build_req(&r);
+  char *txt = text_of(j, 0);
+  EXPECT(txt && strstr(txt, "[document \"paper.pdf\"") &&
+             strstr(txt, "Inlined paper text") &&
+             strstr(txt, "[-- end of document --]"),
+         "the PDF text is inlined even when images are off");
+  EXPECT(txt && strstr(txt, "[image \"shot.png\"") && strstr(txt, "not sent"),
+         "the image beside it is still only a note");
+  free(txt);
+  char *s = cJSON_PrintUnformatted(j);
+  EXPECT(s && !strstr(s, "image_url"), "no image part with images off");
+  free(s);
+  cJSON_Delete(j);
+  cst_attach_free(doc);
+  cst_attach_free(img);
+}
+
 int main(void) {
  test_images_off();
  test_images_on();
+  test_document_parts();
  test_reasoning_and_prefix();
  test_cache_hints();
  test_tools_snapshot();
@@ -8,6 +8,7 @@
 #include "cst_test_util.h"
 #include "os_compat.h"

+#include <math.h>
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
@@ -49,7 +50,9 @@ static void test_builtin(void) {
  cst_model_info("deepseek", "deepseek-chat", &mi);
  EXPECT(mi.vision == 1, "deepseek-chat is deepseek-flash");
  cst_model_info(NULL, "gpt-4o-mini-2024", &mi);
-  EXPECT(mi.price.miss == 0.15 && mi.context == 128000 && mi.vision == 1,
+  /* x87 (i686) compares the literal at long double precision: no == here */
+  EXPECT(fabs(mi.price.miss - 0.15) < 1e-12 && mi.context == 128000 &&
+             mi.vision == 1,
         "gpt-4o-mini over gpt-4o");
  cst_model_info(NULL, "Qwen/Qwen2.5-VL-72B-Instruct", &mi);
  EXPECT(mi.vision == 1 && mi.context == 131072, "glob *-vl* + qwen prefix");
@@ -0,0 +1,250 @@
+/* Text extraction from PDFs: hand-made fixtures for the shapes that matter
+ * (plain content, FlateDecode, CID fonts with ToUnicode CMaps, object
+ * streams) plus the two error paths (encrypted, not a PDF). */
+
+#include "cst_pdf.h"
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+static int failures = 0;
+
+#define CHECK(cond, what)                                                      \
+  do {                                                                         \
+    if (!(cond)) {                                                             \
+      fprintf(stderr, "FAIL %s:%d: %s\n", __FILE__, __LINE__, what);           \
+      failures++;                                                              \
+    }                                                                          \
+  } while (0)
+
+static char *extract(const char *pdf, int *pages, char *err, size_t err_sz) {
+  return cst_pdf_text(pdf, strlen(pdf), pages, err, err_sz);
+}
+
+static int has(const char *hay, const char *needle) {
+  return hay && strstr(hay, needle) != NULL;
+}
+
+/* --- 1. plain, uncompressed content, two lines via Td --------------- */
+
+static const char PLAIN[] =
+    "%PDF-1.4\n"
+    "1 0 obj\n<< /Type /Catalog /Pages 2 0 R >>\nendobj\n"
+    "2 0 obj\n<< /Type /Pages /Kids [3 0 R] /Count 1 >>\nendobj\n"
+    "3 0 obj\n<< /Type /Page /Parent 2 0 R /MediaBox [0 0 612 792] "
+    "/Resources << /Font << /F1 4 0 R >> >> /Contents 5 0 R >>\nendobj\n"
+    "4 0 obj\n<< /Type /Font /Subtype /Type1 /BaseFont /Helvetica >>\nendobj\n"
+    "5 0 obj\n<< >>\nstream\n"
+    "BT\n/F1 12 Tf\n72 720 Td\n(Line one) Tj\n0 -14 Td\n"
+    "(A \\(paren\\) and a \\\\ backslash) Tj\nET\n"
+    "endstream\nendobj\n"
+    "trailer\n<< /Root 1 0 R >>\n%%EOF\n";
+
+static void test_plain(void) {
+  char err[256] = "";
+  int pages = 0;
+  char *t = extract(PLAIN, &pages, err, sizeof(err));
+  CHECK(t != NULL, err);
+  CHECK(pages == 1, "one page");
+  CHECK(has(t, "Line one"), "first line");
+  CHECK(has(t, "A (paren) and a \\ backslash"), "escapes");
+  CHECK(!has(t, "\r"), "no CR left");
+  free(t);
+}
+
+/* --- 2. FlateDecode content stream ---------------------------------- */
+
+/* raw deflate (zlib, level 9, no header) of
+ * "BT /F1 12 Tf 72 720 Td (Hello, compressed world) Tj ET\n" */
+static const unsigned char FLATE[] = {
+    0x73, 0x0a, 0x51, 0xd0, 0x77, 0x33, 0x54, 0x30, 0x34, 0x52, 0x08, 0x49,
+    0x53, 0x30, 0x37, 0x02, 0x22, 0x03, 0x85, 0x90, 0x14, 0x05, 0x0d, 0x8f,
+    0xd4, 0x9c, 0x9c, 0x7c, 0x1d, 0x85, 0xe4, 0xfc, 0xdc, 0x82, 0xa2, 0xd4,
+    0xe2, 0xe2, 0xd4, 0x14, 0x85, 0xf2, 0xfc, 0xa2, 0x9c, 0x14, 0x4d, 0x85,
+    0x90, 0x2c, 0x05, 0xd7, 0x10, 0x2e, 0x00};
+
+static void test_flate(void) {
+  char pdf[2048];
+  int n = snprintf(pdf, sizeof(pdf),
+                   "%%PDF-1.4\n"
+                   "1 0 obj\n<< /Type /Catalog /Pages 2 0 R >>\nendobj\n"
+                   "2 0 obj\n<< /Type /Pages /Kids [3 0 R] /Count 1 >>\nendobj\n"
+                   "3 0 obj\n<< /Type /Page /Parent 2 0 R /Resources << /Font "
+                   "<< /F1 4 0 R >> >> /Contents 5 0 R >>\nendobj\n"
+                   "4 0 obj\n<< /Type /Font /Subtype /Type1 >>\nendobj\n"
+                   "5 0 obj\n<< /Filter /FlateDecode >>\nstream\n");
+  memcpy(pdf + n, FLATE, sizeof(FLATE));
+  n += (int)sizeof(FLATE);
+  n += snprintf(pdf + n, sizeof(pdf) - (size_t)n,
+                "\nendstream\nendobj\ntrailer\n<< /Root 1 0 R >>\n%%EOF\n");
+  char err[256] = "";
+  int pages = 0;
+  char *t = cst_pdf_text(pdf, (size_t)n, &pages, err, sizeof(err));
+  CHECK(t != NULL, err);
+  CHECK(has(t, "Hello, compressed world"), "inflated text");
+  free(t);
+}
+
+/* 2b. zlib-framed FlateDecode (RFC 1950), which is what real writers emit:
+ * zlib.compress(b"BT /F1 12 Tf 72 720 Td (Zlib framed hello) Tj ET\n"). */
+static const unsigned char ZLIB[] = {
+    0x78, 0xda, 0x73, 0x0a, 0x51, 0xd0, 0x77, 0x33, 0x54, 0x30, 0x34, 0x52,
+    0x08, 0x49, 0x53, 0x30, 0x37, 0x02, 0x22, 0x03, 0x85, 0x90, 0x14, 0x05,
+    0x8d, 0xa8, 0x9c, 0xcc, 0x24, 0x85, 0xb4, 0xa2, 0xc4, 0xdc, 0xd4, 0x14,
+    0x85, 0x8c, 0xd4, 0x9c, 0x9c, 0x7c, 0x4d, 0x85, 0x90, 0x2c, 0x05, 0xd7,
+    0x10, 0x2e, 0x00, 0x33, 0x0f, 0x0d, 0x3a};
+
+static void test_zlib_flate(void) {
+  char pdf[2048];
+  int n = snprintf(pdf, sizeof(pdf),
+                   "%%PDF-1.4\n"
+                   "1 0 obj\n<< /Type /Catalog /Pages 2 0 R >>\nendobj\n"
+                   "2 0 obj\n<< /Type /Pages /Kids [3 0 R] /Count 1 >>\nendobj\n"
+                   "3 0 obj\n<< /Type /Page /Parent 2 0 R /Resources << /Font "
+                   "<< /F1 4 0 R >> >> /Contents 5 0 R >>\nendobj\n"
+                   "4 0 obj\n<< /Type /Font /Subtype /Type1 >>\nendobj\n"
+                   "5 0 obj\n<< /Filter /FlateDecode >>\nstream\n");
+  memcpy(pdf + n, ZLIB, sizeof(ZLIB));
+  n += (int)sizeof(ZLIB);
+  n += snprintf(pdf + n, sizeof(pdf) - (size_t)n,
+                "\nendstream\nendobj\ntrailer\n<< /Root 1 0 R >>\n%%EOF\n");
+  char err[256] = "";
+  int pages = 0;
+  char *t = cst_pdf_text(pdf, (size_t)n, &pages, err, sizeof(err));
+  CHECK(t != NULL, err);
+  CHECK(has(t, "Zlib framed hello"), "zlib header skipped, text inflated");
+  free(t);
+}
+
+/* 2c. an inline image: its bytes are not text and must be skipped. */
+
+static const char INLINE_IMG[] =
+    "%PDF-1.4\n"
+    "1 0 obj\n<< /Type /Catalog /Pages 2 0 R >>\nendobj\n"
+    "2 0 obj\n<< /Type /Pages /Kids [3 0 R] /Count 1 >>\nendobj\n"
+    "3 0 obj\n<< /Type /Page /Parent 2 0 R /Resources << /Font << /F1 4 0 R "
+    ">> >> /Contents 5 0 R >>\nendobj\n"
+    "4 0 obj\n<< /Type /Font /Subtype /Type1 /BaseFont /Helvetica >>\nendobj\n"
+    "5 0 obj\n<< >>\nstream\n"
+    "BT /F1 12 Tf 72 720 Td (Before the image) Tj ET\n"
+    "q 20 0 0 20 72 700 cm\n"
+    "BI /W 8 /H 8 /CS /RGB /BPC 8 /F /AHx ID\n"
+    "00ff0000ff00ff00ff00ff00 8123<0001> EI Q\n"
+    "BT /F1 12 Tf 72 680 Td (After the image) Tj ET\n"
+    "endstream\nendobj\n"
+    "trailer\n<< /Root 1 0 R >>\n%%EOF\n";
+
+static void test_inline_image(void) {
+  char err[256] = "";
+  int pages = 0;
+  char *t = extract(INLINE_IMG, &pages, err, sizeof(err));
+  CHECK(t != NULL, err);
+  CHECK(has(t, "Before the image"), "text before the inline image");
+  CHECK(has(t, "After the image"), "text after it is still found");
+  CHECK(!has(t, "00ff0000"), "the image bytes were skipped");
+  CHECK(!has(t, "8123"), "its hex data was skipped too");
+  free(t);
+}
+
+/* --- 3. CID font with a ToUnicode CMap ------------------------------ */
+
+static const char CID[] =
+    "%PDF-1.5\n"
+    "1 0 obj\n<< /Type /Catalog /Pages 2 0 R >>\nendobj\n"
+    "2 0 obj\n<< /Type /Pages /Kids [3 0 R] /Count 1 >>\nendobj\n"
+    "3 0 obj\n<< /Type /Page /Parent 2 0 R /Resources << /Font << /F1 4 0 R "
+    ">> >> /Contents 5 0 R >>\nendobj\n"
+    "4 0 obj\n<< /Type /Font /Subtype /Type0 /Encoding /Identity-H "
+    "/ToUnicode 6 0 R >>\nendobj\n"
+    "5 0 obj\n<< >>\nstream\n"
+    "BT /F1 12 Tf 72 720 Td <0001000200030004> Tj ET\n"
+    "endstream\nendobj\n"
+    "6 0 obj\n<< >>\nstream\n"
+    "1 begincodespacerange\n<0000> <FFFF>\nendcodespacerange\n"
+    "2 beginbfchar\n<0001> <4F60>\n<0002> <597D>\nendbfchar\n"
+    "1 beginbfrange\n<0003> <0004> <0041>\nendbfrange\n"
+    "endcmap\nendstream\nendobj\n"
+    "trailer\n<< /Root 1 0 R >>\n%%EOF\n";
+
+static void test_cid(void) {
+  char err[256] = "";
+  int pages = 0;
+  char *t = extract(CID, &pages, err, sizeof(err));
+  CHECK(t != NULL, err);
+  CHECK(has(t, "\xe4\xbd\xa0\xe5\xa5\xbd"), "bfchar maps to 你好");
+  CHECK(has(t, "AB"), "bfrange walks the code points");
+  free(t);
+}
+
+/* --- 4. objects inside an object stream ----------------------------- */
+
+static void test_objstm(void) {
+  char pdf[4096];
+  const char *page = "<< /Type /Page /Parent 2 0 R /Resources << /Font << /F1 "
+                     "4 0 R >> >> /Contents 5 0 R >>";
+  const char *font = "<< /Type /Font /Subtype /Type1 /BaseFont /Helvetica >>";
+  /* the stream is "num1 off1 num2 off2" -- /N and /First are in the dict --
+   * and offsets count from /First, which is where the bodies start, so
+   * /First is the length of the pairs list itself */
+  char hdr[64];
+  int off4 = (int)strlen(page) + 1; /* object 4 follows the page */
+  int first = snprintf(hdr, sizeof(hdr), "3 0 4 %d ", off4);
+  int n = snprintf(pdf, sizeof(pdf),
+                   "%%PDF-1.5\n"
+                   "1 0 obj\n<< /Type /Catalog /Pages 2 0 R >>\nendobj\n"
+                   "2 0 obj\n<< /Type /Pages /Kids [3 0 R] /Count 1 >>\nendobj\n"
+                   "5 0 obj\n<< >>\nstream\n"
+                   "BT /F1 12 Tf 72 720 Td (Inside an object stream) Tj ET\n"
+                   "endstream\nendobj\n"
+                   "6 0 obj\n<< /Type /ObjStm /N 2 /First %d >>\nstream\n"
+                   "%s%s\n%s",
+                   first, hdr, page, font);
+  n += snprintf(pdf + n, sizeof(pdf) - (size_t)n,
+                "\nendstream\nendobj\n"
+                "trailer\n<< /Root 1 0 R >>\n%%EOF\n");
+  char err[256] = "";
+  int pages = 0;
+  char *t = cst_pdf_text(pdf, (size_t)n, &pages, err, sizeof(err));
+  CHECK(t != NULL, err);
+  CHECK(pages == 1, "one page through the object stream");
+  CHECK(has(t, "Inside an object stream"), "text of the page in the ObjStm");
+  free(t);
+}
+
+/* --- 5. error paths -------------------------------------------------- */
+
+static void test_errors(void) {
+  char err[256] = "";
+  char *t = extract("%PDF-1.7\ntrailer\n<< /Root 1 0 R /Encrypt 9 0 R >>\n",
+                    NULL, err, sizeof(err));
+  CHECK(t == NULL, "encrypted file is refused");
+  CHECK(has(err, "encrypted"), "says it is encrypted");
+
+  err[0] = 0;
+  t = extract("hello, not a pdf at all\n", NULL, err, sizeof(err));
+  CHECK(t == NULL, "non-PDF is refused");
+  CHECK(has(err, "not a PDF"), "says it is not a PDF");
+
+  err[0] = 0;
+  t = extract("%PDF-1.4\n1 0 obj\n<< >>\nendobj\n", NULL, err, sizeof(err));
+  CHECK(t == NULL, "a PDF without text is refused");
+  CHECK(has(err, "no extractable text") || has(err, "no readable pages"),
+        "explains the missing text");
+}
+
+int main(void) {
+  test_plain();
+  test_flate();
+  test_zlib_flate();
+  test_inline_image();
+  test_cid();
+  test_objstm();
+  test_errors();
+  if (failures) {
+    fprintf(stderr, "test_pdf_text: %d failure(s)\n", failures);
+    return 1;
+  }
+  printf("test_pdf_text: OK\n");
+  return 0;
+}
@@ -236,6 +236,16 @@ static void test_sanitize(void) {
  CHECK(strstr(html, "<a") == NULL); /* an unusable href drops the tag */
  free(html);

+  /* a quote inside a single-quoted or bare value cannot open a new attribute */
+  const char *quoted = "<a href='http://x\" onmouseover=\"alert(1)'>q</a>"
+                       "<a href=http://y\"onmouseover=alert(2)//>r</a>";
+  html = cst_web_md_sanitize(quoted, strlen(quoted));
+  CHECK(html != NULL);
+  CHECK(strstr(html, "\" onmouseover") == NULL);
+  CHECK(strstr(html, "\"onmouseover") == NULL);
+  CHECK(strstr(html, "&quot;") != NULL);
+  free(html);
+
  /* streaming: a javascript: link is dropped, a split tag never leaks */
  {
    cst_buf_free(&g_md_out);
@@ -270,8 +280,38 @@ static void test_sanitize(void) {
  }
 }

+/* Streamed deltas cut anywhere must reach the browser as whole characters. */
+static void test_utf8_hold(void) {
+  const char *src = "思考：先读 main.c，再改 😀 emoji。";
+  size_t len = strlen(src);
+  for (size_t step = 1; step <= 5; step++) {
+    cst_utf8_hold_t h = {0};
+    cst_buf_t out = {0};
+    for (size_t i = 0; i < len; i += step) {
+      size_t n = len - i < step ? len - i : step;
+      size_t got = 0;
+      char *piece = cst_utf8_hold_feed(&h, src + i, n, &got);
+      CHECK(piece != NULL);
+      CHECK(h.n < 4);
+      CHECK(cst_utf8_valid(piece, got));
+      CHECK(cst_buf_append(&out, piece, got) == 0);
+      free(piece);
+    }
+    CHECK(h.n == 0);
+    CHECK(out.len == len && memcmp(out.buf, src, len) == 0);
+    cst_buf_free(&out);
+  }
+  /* a stray continuation byte is not held forever */
+  cst_utf8_hold_t h = {0};
+  size_t got = 0;
+  char *piece = cst_utf8_hold_feed(&h, "\x80\x80\x80\x80", 4, &got);
+  CHECK(piece && got == 4 && h.n == 0);
+  free(piece);
+}
+
 int main(void) {
  CHECK(os_net_init() == 0);
+  test_utf8_hold();
  test_query();
  test_sanitize();
  test_http();
@@ -89,8 +89,11 @@ typedef struct {
    const unsigned char *json;
    size_t position;
 } error;
-/* cst-agent: per thread, parses run concurrently (sub-agents, MCP) */
-#if defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L
+/* cst-agent: per thread, parses run concurrently (sub-agents, MCP).
+ * tcc has no thread-local storage; cst-agent never reads cJSON_GetErrorPtr. */
+#if defined(__TINYC__)
+static error global_error = { NULL, 0 };
+#elif defined(__STDC_VERSION__) && __STDC_VERSION__ >= 201112L
 static _Thread_local error global_error = { NULL, 0 };
 #elif defined(__GNUC__)
 static __thread error global_error = { NULL, 0 };
@@ -1,41 +1,250 @@
 <!DOCTYPE html>
-<html lang="zh-CN">
+<html lang="zh-CN" data-theme="dark">
 <head>
 <meta charset="utf-8">
-<meta name="viewport" content="width=device-width, initial-scale=1">
+<meta name="viewport" content="width=device-width, initial-scale=1, viewport-fit=cover">
+<meta name="color-scheme" content="dark light">
 <title>CST-Agent</title>
 <link rel="stylesheet" href="style.css">
 </head>
 <body>
-<header id="bar">
-  <span class="tag" id="planmode">Build</span>
-  <span class="tag" id="approval">auto</span>
-  <span class="ws" id="ws" title="workspace"></span>
-  <span class="model" id="model" title="model">-</span>
-  <span class="muted" id="ctx"></span>
-  <span class="grow"></span>
-  <button id="btn-sessions" title="sessions">会话</button>
-  <button id="btn-model" title="model">模型</button>
-  <button id="btn-undo" title="undo">撤销</button>
-  <button id="btn-compact" title="compact">压缩</button>
-  <button id="btn-clear" title="clear view">清屏</button>
-</header>
+<!-- one icon set, drawn as a sprite so every glyph shares the same 24px grid,
+     stroke weight and round caps: mixed emoji/unicode never looks like a family -->
+<svg id="atlas" aria-hidden="true" hidden>
+  <symbol id="i-menu" viewBox="0 0 24 24"><path d="M4 7h16M4 12h16M4 17h16"/></symbol>
+  <symbol id="i-panel-left" viewBox="0 0 24 24"><rect x="3.5" y="5" width="17" height="14" rx="3"/><path d="M9.5 5v14"/></symbol>
+  <symbol id="i-panel-right" viewBox="0 0 24 24"><rect x="3.5" y="5" width="17" height="14" rx="3"/><path d="M14.5 5v14"/></symbol>
+  <symbol id="i-chevron" viewBox="0 0 24 24"><path d="M6 9.5l6 5 6-5"/></symbol>
+  <symbol id="i-plus" viewBox="0 0 24 24"><path d="M12 5v14M5 12h14"/></symbol>
+  <symbol id="i-refresh" viewBox="0 0 24 24"><path d="M20 12a8 8 0 1 1-2.6-5.9M20 4.5V9h-4.5"/></symbol>
+  <symbol id="i-settings" viewBox="0 0 24 24"><path d="M4 6.5h5M15 6.5h5M4 12h7M17 12h3M4 17.5h5M15 17.5h5"/><circle cx="11.5" cy="6.5" r="2"/><circle cx="15" cy="12" r="2"/><circle cx="11.5" cy="17.5" r="2"/></symbol>
+  <symbol id="i-moon" viewBox="0 0 24 24"><path d="M20 14.6A8.6 8.6 0 0 1 9.4 4 8.6 8.6 0 1 0 20 14.6z"/></symbol>
+  <symbol id="i-sun" viewBox="0 0 24 24"><circle cx="12" cy="12" r="4.2"/><path d="M12 2.5v2.2M12 19.3v2.2M2.5 12h2.2M19.3 12h2.2M5.2 5.2l1.6 1.6M17.2 17.2l1.6 1.6M18.8 5.2l-1.6 1.6M6.8 17.2l-1.6 1.6"/></symbol>
+  <symbol id="i-more" viewBox="0 0 24 24"><circle cx="5.5" cy="12" r="1.6" class="f"/><circle cx="12" cy="12" r="1.6" class="f"/><circle cx="18.5" cy="12" r="1.6" class="f"/></symbol>
+  <symbol id="i-send" viewBox="0 0 24 24"><path d="M12 19.5V5M5.8 11.2 12 5l6.2 6.2"/></symbol>
+  <symbol id="i-stop" viewBox="0 0 24 24"><rect x="7" y="7" width="10" height="10" rx="2.5" class="f"/></symbol>
+  <symbol id="i-image" viewBox="0 0 24 24"><rect x="3.5" y="5" width="17" height="14" rx="3"/><path d="M4 15.5 8.6 11l4 4 2.8-2.8L20.5 16"/><circle cx="9" cy="9.4" r="1.5"/></symbol>
+  <symbol id="i-clipboard" viewBox="0 0 24 24"><rect x="8.5" y="3.5" width="7" height="4" rx="1.6"/><path d="M15.5 5.5h3v15h-13v-15h3M9 12h6M9 15.5h4"/></symbol>
+  <symbol id="i-list" viewBox="0 0 24 24"><path d="M4 6.8h2M4 12h2M4 17.2h2M9 6.8h11M9 12h11M9 17.2h11"/></symbol>
+  <symbol id="i-copy" viewBox="0 0 24 24"><rect x="8.5" y="8.5" width="11" height="11" rx="2.6"/><path d="M15.5 8.5V7a2.5 2.5 0 0 0-2.5-2.5H6.5A2.5 2.5 0 0 0 4 7v6.5A2.5 2.5 0 0 0 6.5 16h1.5"/></symbol>
+  <symbol id="i-compress" viewBox="0 0 24 24"><path d="M8 3.5V9H3M16 20.5V15h5M3 9h4M16 15h5"/></symbol>
+  <symbol id="i-undo" viewBox="0 0 24 24"><path d="M9.5 5.5 4 11l5.5 5.5"/><path d="M4 11h10.2a5.2 5.2 0 0 1 0 10.4H8.5"/></symbol>
+  <symbol id="i-close" viewBox="0 0 24 24"><path d="M6.5 6.5l11 11M17.5 6.5l-11 11"/></symbol>
+  <symbol id="i-check" viewBox="0 0 24 24"><path d="M5 12.6l4.6 4.6L19 7.6"/></symbol>
+  <symbol id="i-shield" viewBox="0 0 24 24"><path d="M12 3.5 5 6.2v5.4c0 4.2 2.9 7.5 7 8.9 4.1-1.4 7-4.7 7-8.9V6.2z"/></symbol>
+  <symbol id="i-bolt" viewBox="0 0 24 24"><path d="M13.4 3 5.5 13.6H11l-.9 7.4L18.5 10H13z"/></symbol>
+  <symbol id="i-alert" viewBox="0 0 24 24"><path d="M12 4 2.8 20h18.4z"/><path d="M12 10v4.4M12 17.2v.2"/></symbol>
+  <symbol id="i-cpu" viewBox="0 0 24 24"><rect x="7" y="7" width="10" height="10" rx="2.6"/><path d="M10 3v4M14 3v4M10 17v4M14 17v4M3 10h4M3 14h4M17 10h4M17 14h4"/></symbol>
+  <symbol id="i-lv1" viewBox="0 0 24 24"><path d="M5 18.5V13h3.4v5.5z" class="f"/></symbol>
+  <symbol id="i-lv2" viewBox="0 0 24 24"><path d="M5 18.5V13h3.4v5.5zM10.3 18.5V9.6h3.4v8.9z" class="f"/></symbol>
+  <symbol id="i-lv3" viewBox="0 0 24 24"><path d="M5 18.5V13h3.4v5.5zM10.3 18.5V9.6h3.4v8.9zM15.6 18.5V6.4h3.4v12.1z" class="f"/></symbol>
+  <symbol id="i-lv4" viewBox="0 0 24 24"><path d="M4 18.5V13h2.6v5.5zM8.7 18.5V10h2.6v8.5zM13.4 18.5V6.8h2.6v11.7zM18.1 18.5V4h2.6v14.5z" class="f"/></symbol>
+  <symbol id="i-file" viewBox="0 0 24 24"><path d="M13.5 3.5H7A2.5 2.5 0 0 0 4.5 6v12A2.5 2.5 0 0 0 7 20.5h10A2.5 2.5 0 0 0 19.5 18V9.5z"/><path d="M13.5 3.5v6h6"/></symbol>
+  <symbol id="i-clock" viewBox="0 0 24 24"><circle cx="12" cy="12" r="8.5"/><path d="M12 7.2V12l3.2 2"/></symbol>
+  <symbol id="i-edit" viewBox="0 0 24 24"><path d="M14.8 5.2l4 4L9 19H5v-4z"/><path d="M12.6 7.4l4 4"/></symbol>
+  <symbol id="i-retry" viewBox="0 0 24 24"><path d="M4.5 12a7.5 7.5 0 0 1 13-5.1M19.5 12a7.5 7.5 0 0 1-13 5.1"/><path d="M17.8 3.6v3.6h-3.6M6.2 20.4v-3.6h3.6"/></symbol>
+  <symbol id="i-left" viewBox="0 0 24 24"><path d="M14.5 6l-6 6 6 6"/></symbol>
+  <symbol id="i-right" viewBox="0 0 24 24"><path d="M9.5 6l6 6-6 6"/></symbol>
+  <symbol id="i-folder" viewBox="0 0 24 24"><path d="M3.5 6.5A1.5 1.5 0 0 1 5 5h4l2 2.5h8A1.5 1.5 0 0 1 20.5 9v9A1.5 1.5 0 0 1 19 19.5H5A1.5 1.5 0 0 1 3.5 18z"/></symbol>
+</svg>

-<main id="log"></main>
-<aside id="todo" hidden></aside>
+<div id="app">

-<footer id="composer">
-  <div id="pending" hidden></div>
-  <div class="row">
-    <textarea id="input" rows="1" placeholder="输入消息，Enter 发送，Shift+Enter 换行"></textarea>
-    <button id="btn-send">发送</button>
-    <button id="btn-stop" hidden>停止</button>
-    <label class="att" title="附上图片"><input type="file" id="file" accept="image/*" hidden>+图</label>
-  </div>
-  <div class="hint"><span id="status">就绪</span><span class="grow"></span><span id="usage" class="muted"></span></div>
-</footer>
+  <!-- rail: identity, one primary action, sessions, identity again.
+       Everything else is a popover anchored where it belongs. -->
+  <aside id="side">
+   <div class="side-inner">
+    <div class="side-head">
+      <div class="brand">
+        <span class="brand-name">CST-Agent</span>
+      </div>
+      <button id="btn-collapse" class="icon" title="收起侧栏 (Ctrl+B)">
+        <svg class="i"><use href="#i-panel-left"/></svg>
+      </button>
+    </div>

+    <button id="btn-new" class="new-task">
+      <svg class="i"><use href="#i-plus"/></svg><span>新建任务</span>
+    </button>
+
+    <div class="side-sec">
+      <span>最近</span>
+      <button id="btn-refresh-sessions" class="icon tiny" title="刷新">
+        <svg class="i"><use href="#i-refresh"/></svg>
+      </button>
+    </div>
+    <input id="sess-search" class="sess-search" type="search" placeholder="搜索会话" autocomplete="off">
+    <div id="sessions-list" class="sessions"></div>
+
+    <div class="side-foot">
+      <div class="who">
+        <span class="avatar">C</span>
+        <span class="who-text">
+          <b id="foot-ws">workspace</b>
+          <i id="foot-session">-</i>
+        </span>
+      </div>
+      <button id="btn-settings" class="icon" title="设置">
+        <svg class="i"><use href="#i-settings"/></svg>
+      </button>
+    </div>
+   </div>
+  </aside>
+
+  <!-- main column -->
+  <main id="main">
+    <header id="topbar">
+      <button id="btn-side" class="icon" title="展开侧栏 (Ctrl+B)">
+        <svg class="i"><use href="#i-panel-left"/></svg>
+      </button>
+      <button id="btn-drawer" class="icon only-narrow" title="侧栏">
+        <svg class="i"><use href="#i-menu"/></svg>
+      </button>
+      <button id="chip-model" class="chip model" title="模型与推理强度">
+        <svg class="i"><use href="#i-cpu"/></svg><span class="lbl">-</span>
+        <svg class="i chev"><use href="#i-chevron"/></svg>
+      </button>
+      <span class="grow"></span>
+      <div class="seg" id="seg-mode" title="计划模式只读，不改文件">
+        <button data-mode="build">构建</button>
+        <button data-mode="plan">计划</button>
+      </div>
+      <button id="chip-approval" class="chip quiet" title="审批权限">
+        <svg class="i"><use href="#i-shield"/></svg><span class="lbl">auto</span>
+        <svg class="i chev"><use href="#i-chevron"/></svg>
+      </button>
+      <button id="btn-theme" class="icon" title="明暗切换">
+        <svg class="i only-dark"><use href="#i-moon"/></svg>
+        <svg class="i only-light"><use href="#i-sun"/></svg>
+      </button>
+      <button id="btn-panel" class="icon" title="运行详情 (Ctrl+J)">
+        <svg class="i"><use href="#i-panel-right"/></svg>
+      </button>
+      <button id="btn-menu" class="icon" title="更多">
+        <svg class="i"><use href="#i-more"/></svg>
+      </button>
+    </header>
+
+    <div id="scroll">
+      <section id="empty">
+        <p class="ready">准备好开始构建了吗</p>
+      </section>
+      <div id="log"></div>
+    </div>
+
+    <button id="to-bottom" class="round" hidden title="回到最新">
+      <svg class="i"><use href="#i-chevron"/></svg>
+    </button>
+
+    <footer id="composer">
+      <div id="pending" class="pending" hidden></div>
+      <div id="edit-bar" class="edit-bar" hidden>
+        <span>编辑上一条消息：发送后替换它并重新回答（文件改动保留，需要时先撤销）</span>
+        <span class="grow"></span>
+        <button id="btn-edit-cancel" class="link">取消</button>
+      </div>
+
+      <!-- permission gate: inline, the turn really is waiting on it -->
+      <div id="gate" class="gate" hidden>
+        <div class="gate-head">
+          <span class="gate-title" id="gate-title">需要确认</span>
+          <span class="grow"></span>
+          <span class="gate-timer" id="gate-timer"></span>
+        </div>
+        <pre id="gate-detail"></pre>
+        <div class="gate-keys" id="gate-keys"></div>
+        <div class="gate-row" id="gate-buttons"></div>
+      </div>
+
+      <div id="slash" class="slash" role="listbox" hidden></div>
+      <div class="box">
+        <textarea id="input" rows="1" placeholder="给 CST-Agent 派个任务…（/ 打开命令，F1 看快捷键）"></textarea>
+        <div class="box-row">
+          <button id="btn-plus" class="icon att" title="添加图片或计划模式">
+            <svg class="i"><use href="#i-plus"/></svg>
+          </button>
+          <button id="btn-undo" class="chip quiet" title="撤销上一次文件写入">
+            <svg class="i"><use href="#i-undo"/></svg><span>撤销</span>
+          </button>
+          <span class="grow"></span>
+          <button id="cost" class="cost" title="用量详情"></button>
+          <button id="btn-stop" class="icon stop" hidden title="停止">
+            <svg class="i"><use href="#i-stop"/></svg>
+          </button>
+          <button id="btn-send" class="round send" title="发送 (Enter)">
+            <svg class="i"><use href="#i-send"/></svg>
+          </button>
+        </div>
+      </div>
+      <div class="hint-line">
+        <span id="status">就绪</span>
+        <span class="grow"></span>
+        <span id="usage" class="quiet"></span>
+      </div>
+      <input type="file" id="file" accept="image/*" multiple hidden>
+    </footer>
+  </main>
+
+  <!-- right panel: the run, not the chat -->
+  <aside id="panel">
+   <div class="panel-inner">
+    <div class="panel-head">
+      <span>运行详情</span>
+      <span class="grow"></span>
+      <button id="btn-panel-close" class="icon tiny" title="关闭">
+        <svg class="i"><use href="#i-close"/></svg>
+      </button>
+    </div>
+
+    <section class="card" id="card-todo">
+      <h4>任务清单</h4>
+      <div id="todo-body" class="muted">模型还没有给出任务清单。</div>
+    </section>
+
+    <section class="card" id="card-files">
+      <h4>文件改动 <span class="badge" id="files-count">0</span>
+        <span class="grow"></span><button id="btn-diff" class="link" title="git status + git diff">查看 diff</button></h4>
+      <div id="files-body" class="muted">本轮还没有改动文件。</div>
+    </section>
+
+    <section class="card">
+      <h4>上下文</h4>
+      <div class="kv"><span>占用</span><b id="kv-ctx">0</b></div>
+      <div class="kv"><span>缓存命中</span><b id="kv-cache">-</b></div>
+    </section>
+   </div>
+  </aside>
+
+</div>
+
+<div id="backdrop"></div>
+<div id="pop" class="pop" hidden></div>
 <div id="modal" hidden><div id="modal-box"></div></div>
+<div id="lightbox" hidden>
+  <img id="lb-img" alt="">
+  <button id="lb-close" class="lb-btn" title="关闭 (Esc)"><svg class="i"><use href="#i-close"/></svg></button>
+  <button id="lb-prev" class="lb-btn" title="上一张 (←)"><svg class="i"><use href="#i-left"/></svg></button>
+  <button id="lb-next" class="lb-btn" title="下一张 (→)"><svg class="i"><use href="#i-right"/></svg></button>
+  <span id="lb-count"></span>
+</div>
+<div id="toast" class="toast" role="status" aria-live="polite" hidden></div>
+<script>
+/* Layout is decided here, before app.js runs and before the first paint: the
+ * rail must not flash open and then fold, and a broken script further down can
+ * no longer take the layout down with it. */
+(() => {
+  const q = new URLSearchParams(location.search);
+  const pick = (key) => {
+    const v = q.get(key);
+    if (v === '0' || v === '1') return v === '0';
+    return localStorage.getItem('cst.' + key) === '0';
+  };
+  if (pick('side')) document.body.classList.add('no-side');
+  if (pick('panel')) document.body.classList.add('no-panel');
+  const t = q.get('theme') || localStorage.getItem('cst.theme');
+  if (t) document.documentElement.dataset.theme = t;
+})();
+</script>
 <script src="app.js"></script>
 </body>
 </html>
Author	SHA1	Message	Date
WHC	e0b0eddf3a	v0.8.6: general mode, clipboard text, PDF text extraction ci / tcc-check (push) Successful in 59s Details ci / linux (aarch64, 0) (push) Successful in 19m11s Details ci / linux (x86, 1) (push) Successful in 11m31s Details ci / linux (x86_64, 1) (push) Successful in 11m11s Details ci / windows (push) Successful in 9m24s Details Working mode: --mode coding\|general (CST_MODE, settings.json, /mode, WebUI run settings). general swaps the default framing for a general assistant one -- workflow, safety notes, tool wording (run_shell, the subagents) -- and /status shows it; minimal (the paper profile) stays byte-frozen. The prompt state carries the mode so a switch leaves a system note in history. Clipboard: clipboard_read / clipboard_write tools (16 MiB cap read, default 64 KiB window; writing is a WRITE permission with a preview) on pbcopy / wl-copy / xclip / pbpaste / wl-paste / xclip and the Win32 clipboard. In the TUI, /attach paste falls back to clipboard text when there is no image on it. PDF: a dependency-free text extractor (src/pdf_text.c): inflate, Flate/ASCIIHex/ASCII85, PNG/TIFF predictors, layout-independent object scan, object streams, page tree walk with inherited resources, content lexer (Tj/TJ/'/" and the Td/TD/T*/Tm spacing heuristics), per-font ToUnicode CMaps and inline images skipped. Encrypted files and pages with no text layer are refused with the reason. Attachments: /attach (a path or files dropped on the terminal) now take images, PDFs and text. A PDF is stored as its text layer with the mime still saying application/pdf and its page count; anything else has to be UTF-8 text. Documents travel inline as text parts, so CST_NO_IMAGES and text-only models no longer drop them, while images keep their old path. Tests: test_pdf_text (plain, bare and zlib-framed Flate, inline image, CID + ToUnicode, object stream, error paths), documents in test_attach, inlined documents in test_llm_body, mode in test_config.	2026-10-11 11:09:39 +08:00
WHC	50abcf1a91	build: finish cross-platform release tooling ci / tcc-check (push) Successful in 1m9s Details ci / linux (aarch64, 0) (push) Failing after 58m44s Details ci / linux (x86, 1) (push) Successful in 13m47s Details ci / linux (x86_64, 1) (push) Failing after 1h0m2s Details ci / windows (push) Successful in 10m46s Details	2026-10-10 16:21:00 +08:00
WHC	9bf7707912	v0.8.5: WebUI process timeline, Gemini-style images, aligned layout ci / linux (clang) (push) Successful in 6m15s Details ci / linux (gcc) (push) Successful in 7m14s Details ci / musl (push) Successful in 7m14s Details ci / windows-cross (push) Successful in 11m4s Details Process timeline: a run of consecutive thinking and tool calls is one block that folds to a single line ("思考 4 次 · 工具 6 次"; while running it shows the current action) and opens into a compact timeline, one row per thought or call. Visible text, a note or the next prompt closes the block; history and the live stream follow the same rule. show_thinking now also applies while streaming instead of forcing thoughts open. Layout: the avatar gutter and the flex-wrap that dropped it under long answers are gone, so answer, process blocks, notes and composer share one left edge. The text column and composer are both 720px (--content); the rail and run panel grow with clamp() on wide screens. A rebuild after each turn no longer replays every entrance animation, and the turn time survives it. Message actions are icon buttons and only appear under text. Images: a prompt's images sit above the bubble, right-aligned (one keeps its shape, several are 152px rounded tiles) and open in a lightbox (Esc, arrow keys). Images uploaded from the page show as thumbnails over the composer and appear on the prompt as soon as it is sent. Also: ?theme= is no longer overridden by app.js, the user bubble is visible in the light theme, and the narrow top bar no longer wraps.	2026-10-10 14:54:36 +08:00
WHC	7e1dbd9afa	v0.8.4: WebUI sessions, diff, settings, slash commands and extensions ci / linux (clang) (push) Successful in 6m13s Details ci / musl (push) Canceled after 0s Details ci / windows-cross (push) Canceled after 0s Details ci / linux (gcc) (push) Canceled after 6m33s Details Sessions: rename, delete and search; /api/retry regenerates the last answer, editing a prompt resends from there; /api/export downloads the transcript as Markdown (cst_session_export_md). Session messages carry note/tool_result markers, so a reload rebuilds the tool steps with their arguments, output and exit status; subagent progress is shown. Diff and settings: /api/diff shows the real diff; /api/settings covers protocol, max_steps, tool switches and confirm; /api/permissions edits the rules; sandbox status is reported. Session rename no longer reads past a short title. Commands: slash commands with completion in the composer (/model, /provider, /new, /retry, /undo, /compact, /export, /diff, /init, /skill and more), a Ctrl+P palette, Shift+Tab cycles the approval mode, F1 or ? shows the keys. "Thinking expanded" is saved in the settings (show_thinking), "expand all tool output" in the browser. Extensions: GET /api/extensions lists MCP servers, plugins, skills, agent roles and subagent runs; POST /api/extensions/reload and /api/extensions/trust; a workspace switch reloads them. /api/init and /api/skill run on the server (cst_init_prompt shared with the TUI). Errors, statuses and [cst] notes are shown in Chinese with a readable reason instead of rc=; a failed turn's error survives the reload. Tests: scripts/e2e_webui_sessions.py, e2e_webui_workspace.py and e2e_webui_ext.py, run from e2e_webui.sh.	2026-10-10 13:57:17 +08:00
WHC	2d1052fd5f	v0.8.3: WebUI models, providers and keys; garbled thinking fixed ci / linux (clang) (push) Successful in 5m56s Details ci / linux (gcc) (push) Successful in 8m38s Details ci / musl (push) Successful in 7m22s Details ci / windows-cross (push) Successful in 9m21s Details Streaming: Chinese reasoning showed up garbled in the WebUI. The agent loop now feeds visible/thinking text in runs instead of byte by byte, and the WebUI holds an incomplete UTF-8 tail until the next chunk (cst_utf8_hold_feed), so every think_delta/assistant_delta is whole. src/provider_switch.c: model list fetch/cache, key verify-and-save and provider switching moved out of tui.c and shared by TUI and WebUI. WebUI API: - GET /api/providers, GET /api/models[?refresh=1] (401 on a rejected key, still listing built-ins), POST /api/provider, POST /api/key (key checked against /v1/models, saved to credentials 0600, never echoed in replies, events or logs) - /api/config validates everything before applying, persists model, effort and approval (not yolo), warns when effort does not apply; plan_mode no longer forces read-only - state: turn_usage, api_base, key_set/key_from, model_info, effort_applies, sees_images, max_steps, tool_profile, perm_timeout; permission events carry their timeout Frontend: searchable model picker with refresh and manual entry, provider picker and key dialog, effort/vision hints, toasts instead of silent failures, Chinese error text, real protocol in the settings and usage panels, approval countdown from the server's timeout. Tests: utf8 hold unit test, scripts/e2e_webui_settings.py (fake /v1/models, wrong key, switching, persistence, Chinese thinking) run from e2e_webui.sh, scripts/tui_keys/provider.py.	2026-10-09 15:54:53 +08:00
WHC	e238f5f37d	v0.8.2: build fixes, WebUI security and correctness ci / linux (clang) (push) Successful in 5m45s Details ci / linux (gcc) (push) Successful in 6m46s Details ci / musl (push) Successful in 7m9s Details ci / windows-cross (push) Successful in 11m52s Details Build: <fcntl.h> in tui.c, format-truncation in http_srv.c and /web. WebUI: - refuse cross-site requests (Origin must be the page itself) in the default no-token mode: any page could POST approval=yolo and a prompt - session/workspace/undo/compact/attach/model changes are 409 while a turn runs (use-after-free, messages landing in the wrong session) - ask_user returned "dismissed" at once instead of waiting - cancel and shutdown wake a turn parked on a permission or question; SIGTERM mid-turn no longer hangs or crashes - a page opened while a question is pending still gets it; Last-Event-ID resumes the stream; event seq matches the SSE id - token mode: page assets load without the token, images via cookie - --web-open works, --web-port/--web-token are validated - frontend: gate keys ignored while typing, clear-images button, reasoning no longer wiped while streaming HTTP server: auth and size cap before reading the body, strict Content-Length, NUL in url_decode, 413 on keep-alive, idle timeout, absolute-form only without a leading '/', case-insensitive Connection, header deadline, 400 for long request lines, unlink before close, poll() instead of select(), SO_EXCLUSIVEADDRUSE on Windows. Markdown sanitizer escapes quotes in attribute values. /web: passes this session's provider/model/base/effort/approval (key via the environment), keeps output off the TUI, shows the real URL and keeps CURRENT on the TUI's session. Tests: scripts/e2e_webui_ask.py, scripts/e2e_webui_http.py (run from e2e_webui.sh), sanitizer quote test.	2026-10-09 13:51:34 +08:00
WHC	adf3e9f024	v0.8.0: WebUI redesign, /web, workspace switch, 75% budget, no shell timeout ci / linux (clang) (push) Failing after 31s Details ci / linux (gcc) (push) Failing after 32s Details ci / musl (push) Failing after 39s Details ci / windows-cross (push) Canceled after 2m32s Details The front end is rebuilt on the design tokens the desktop app ships (surfaces, color-mix borders, alpha hovers, accent): a 30-icon inline SVG sprite, dark and light themes, layout switches as deep links (?side=0, ?panel=0, ?theme=light) resolved before the first paint. The empty page is one line and leaves with the first turn; the right rail keeps three cards - todo, file changes (+N -M) and context (tokens, prompt cache hit rate). Both rails animate in and out, and a collapsed rail widens its grid track to 0 instead of being removed, so the transcript never slides into a zero-width column. - src/webui.c: POST /api/workspace - re-root the agent at another directory (permission rules, metrics, AGENTS.md follow) and start a session there, as sessions live per directory; state also reports cached/cache_miss tokens - src/agent_loop.c: cst_agent_set_workspace(), rebuilt prompt and re-measured budget; the auto-compact threshold is 75% of the window (was 60%) - src/tui.c: /web forks a WebUI next to the TUI - separate process and a new session, because cst_webui_run() blocks and two writers corrupt one session - src/os_compat.c, src/os_win.c: run_shell runs without a time limit unless CST_SHELL_TIMEOUT asks for one, so builds and long clones finish - src/main.c, include/cst_webui.h: one CST_WEBUI_DEFAULT_PORT for --help, the docs and /web; Termux opens the browser through termux-open-url - README.md, docs/webui.md, docs/env_openai.md: /web, /api/workspace, the new state fields, the 75% threshold and the layout deep links tests/e2e_webui.sh green (permission round-trip, token paths, --web-dev); make check green.	2026-10-09 11:50:13 +08:00