From 257b4c93c6418ef0b425ce85432ceeac076038a0 Mon Sep 17 00:00:00 2001
From: bowen628 <bowen628@noreply.gitcode.com>
Date: Wed, 1 Apr 2026 18:07:51 +0800
Subject: [PATCH] feat(agentic): primary model vision gating and computer-use
 tool updates

- Resolve primary model ImageUnderstanding from config; pass into prompt builder,
  tool context, and API message stripping.
- Add ToolOptions / description_with_context for vision-aware tool schemas.
- Expand ComputerUse desktop host and tool docs; fix format string delimiter.
- Session snapshot and i18n tweaks for session config.
---
 src/apps/desktop/src/api/config_api.rs        |   7 +
 .../desktop/src/computer_use/desktop_host.rs  | 118 +++++++++--
 .../core/src/agentic/agents/agentic_mode.rs   |   1 +
 .../core/src/agentic/agents/cowork_mode.rs    |   1 +
 .../core/src/agentic/agents/debug_mode.rs     |   1 +
 .../core/src/agentic/agents/init_agent.rs     |   1 +
 .../core/src/agentic/agents/plan_mode.rs      |   1 +
 .../agents/prompt_builder/prompt_builder.rs   |  26 ++-
 .../core/src/agentic/agents/registry.rs       |  12 +-
 .../src/agentic/execution/execution_engine.rs | 119 +++++++----
 .../core/src/agentic/tools/framework.rs       |  21 ++
 .../tools/implementations/bash_tool.rs        |   4 +-
 .../tools/implementations/code_review_tool.rs |   8 +
 .../implementations/computer_use_locate.rs    |   5 -
 .../computer_use_mouse_click_tool.rs          |   7 +-
 .../computer_use_mouse_precise_tool.rs        |   7 +-
 .../computer_use_mouse_step_tool.rs           |   7 +-
 .../implementations/computer_use_tool.rs      | 197 ++++++++++++++++--
 src/crates/core/src/service/config/types.rs   |   2 +-
 .../core/src/service/snapshot/manager.rs      |   9 +
 .../en-US/settings/session-config.json        |   4 +-
 .../zh-CN/settings/session-config.json        |   4 +-
 22 files changed, 443 insertions(+), 119 deletions(-)

diff --git a/src/apps/desktop/src/api/config_api.rs b/src/apps/desktop/src/api/config_api.rs
index 1ab92667..23b5ca88 100644
--- a/src/apps/desktop/src/api/config_api.rs
+++ b/src/apps/desktop/src/api/config_api.rs
@@ -271,6 +271,13 @@ pub async fn get_mode_configs(state: State<'_, AppState>) -> Result<Value, Strin
             needs_save = true;
         } else if let Some(config) = mode_configs.get_mut(&mode_id) {
             config.default_tools = default_tools.clone();
+            // Migration: add ComputerUse to available_tools when the mode default includes it.
+            if default_tools.iter().any(|t| t == "ComputerUse")
+                && !config.available_tools.iter().any(|t| t == "ComputerUse")
+            {
+                config.available_tools.push("ComputerUse".to_string());
+                needs_save = true;
+            }
             // Migrate older Claw sessions that only allowlisted "ComputerUse" before split mouse tools existed;
             // All desktop automation is now consolidated into ComputerUse.
             // Remove any stale split tool names from available_tools.
diff --git a/src/apps/desktop/src/computer_use/desktop_host.rs b/src/apps/desktop/src/computer_use/desktop_host.rs
index accbbc72..7f092a20 100644
--- a/src/apps/desktop/src/computer_use/desktop_host.rs
+++ b/src/apps/desktop/src/computer_use/desktop_host.rs
@@ -1036,6 +1036,85 @@ end tell"#])
         Ok(())
     }
 
+    /// Ease 0..1 for pointer paths (smooth acceleration/deceleration).
+    fn smoothstep01(t: f64) -> f64 {
+        let t = t.clamp(0.0, 1.0);
+        t * t * (3.0 - 2.0 * t)
+    }
+
+    /// Move the pointer along a short visible path instead of warping in one event.
+    #[cfg(target_os = "macos")]
+    fn smooth_mouse_move_cg_global(x1: f64, y1: f64) -> BitFunResult<()> {
+        const MIN_DIST: f64 = 2.5;
+        const MIN_STEPS: usize = 8;
+        const MAX_STEPS: usize = 85;
+        const MAX_DURATION_MS: u64 = 400;
+
+        let (x0, y0) = macos::quartz_mouse_location().unwrap_or((x1, y1));
+        let dx = x1 - x0;
+        let dy = y1 - y0;
+        let dist = (dx * dx + dy * dy).sqrt();
+        if dist < MIN_DIST {
+            return Self::post_mouse_moved_cg_global(x1, y1);
+        }
+        let duration_ms = (70.0 + dist * 0.28).min(MAX_DURATION_MS as f64) as u64;
+        let steps = ((dist / 5.5).ceil() as usize).clamp(MIN_STEPS, MAX_STEPS);
+        let step_delay = Duration::from_millis((duration_ms / steps as u64).max(1));
+
+        for i in 1..=steps {
+            let t = i as f64 / steps as f64;
+            let te = Self::smoothstep01(t);
+            let x = x0 + dx * te;
+            let y = y0 + dy * te;
+            Self::post_mouse_moved_cg_global(x, y)?;
+            if i < steps {
+                std::thread::sleep(step_delay);
+            }
+        }
+        Ok(())
+    }
+
+    /// Windows/Linux: same smooth path using enigo absolute moves (single `Enigo` session).
+    #[cfg(not(target_os = "macos"))]
+    fn smooth_mouse_move_enigo_abs(x1: f64, y1: f64) -> BitFunResult<()> {
+        const MIN_DIST: f64 = 2.5;
+        const MIN_STEPS: usize = 8;
+        const MAX_STEPS: usize = 85;
+        const MAX_DURATION_MS: u64 = 400;
+
+        Self::run_enigo_job(|e| {
+            let (cx, cy) = e.location().map_err(|err| {
+                BitFunError::tool(format!("smooth_mouse_move: pointer location: {}", err))
+            })?;
+            let x0 = cx as f64;
+            let y0 = cy as f64;
+            let dx = x1 - x0;
+            let dy = y1 - y0;
+            let dist = (dx * dx + dy * dy).sqrt();
+            if dist < MIN_DIST {
+                return e
+                    .move_mouse(x1.round() as i32, y1.round() as i32, Coordinate::Abs)
+                    .map_err(|err| BitFunError::tool(format!("mouse_move: {}", err)));
+            }
+            let duration_ms = (70.0 + dist * 0.28).min(MAX_DURATION_MS as f64) as u64;
+            let steps = ((dist / 5.5).ceil() as usize).clamp(MIN_STEPS, MAX_STEPS);
+            let step_delay = Duration::from_millis((duration_ms / steps as u64).max(1));
+
+            for i in 1..=steps {
+                let t = i as f64 / steps as f64;
+                let te = Self::smoothstep01(t);
+                let x = x0 + dx * te;
+                let y = y0 + dy * te;
+                e.move_mouse(x.round() as i32, y.round() as i32, Coordinate::Abs)
+                    .map_err(|err| BitFunError::tool(format!("mouse_move: {}", err)))?;
+                if i < steps {
+                    std::thread::sleep(step_delay);
+                }
+            }
+            Ok(())
+        })
+    }
+
     fn map_button(s: &str) -> BitFunResult<Button> {
         match s.to_lowercase().as_str() {
             "left" => Ok(Button::Left),
@@ -2313,30 +2392,19 @@ impl ComputerUseHost for DesktopComputerUseHost {
     }
 
     async fn mouse_move_global_f64(&self, gx: f64, gy: f64) -> BitFunResult<()> {
-        #[cfg(target_os = "macos")]
-        {
-            tokio::task::spawn_blocking(move || {
-                Self::run_enigo_job(|_| Self::post_mouse_moved_cg_global(gx, gy))
-            })
-            .await
-            .map_err(|e| BitFunError::tool(e.to_string()))??;
-            self.clear_vision_pixel_nudge_block();
-            ComputerUseHost::computer_use_after_pointer_mutation(self);
-            return Ok(());
-        }
-        #[cfg(not(target_os = "macos"))]
-        {
-            self.mouse_move(gx.round() as i32, gy.round() as i32).await
-        }
-    }
-
-    async fn mouse_move(&self, x: i32, y: i32) -> BitFunResult<()> {
-        debug!("computer_use: mouse_move absolute ({}, {})", x, y);
+        debug!(
+            "computer_use: mouse_move_global_f64 smooth target ({:.2}, {:.2})",
+            gx, gy
+        );
         tokio::task::spawn_blocking(move || {
-            Self::run_enigo_job(|e| {
-                e.move_mouse(x, y, Coordinate::Abs)
-                    .map_err(|err| BitFunError::tool(format!("mouse_move: {}", err)))
-            })
+            #[cfg(target_os = "macos")]
+            {
+                Self::run_enigo_job(|_| Self::smooth_mouse_move_cg_global(gx, gy))
+            }
+            #[cfg(not(target_os = "macos"))]
+            {
+                Self::smooth_mouse_move_enigo_abs(gx, gy)
+            }
         })
         .await
         .map_err(|e| BitFunError::tool(e.to_string()))??;
@@ -2345,6 +2413,10 @@ impl ComputerUseHost for DesktopComputerUseHost {
         Ok(())
     }
 
+    async fn mouse_move(&self, x: i32, y: i32) -> BitFunResult<()> {
+        self.mouse_move_global_f64(x as f64, y as f64).await
+    }
+
     async fn pointer_move_relative(&self, dx: i32, dy: i32) -> BitFunResult<()> {
         if dx == 0 && dy == 0 {
             return Ok(());
diff --git a/src/crates/core/src/agentic/agents/agentic_mode.rs b/src/crates/core/src/agentic/agents/agentic_mode.rs
index 8efe9baf..25b6f0be 100644
--- a/src/crates/core/src/agentic/agents/agentic_mode.rs
+++ b/src/crates/core/src/agentic/agents/agentic_mode.rs
@@ -25,6 +25,7 @@ impl AgenticMode {
                 "AskUserQuestion".to_string(),
                 "Git".to_string(),
                 "TerminalControl".to_string(),
+                "ComputerUse".to_string(),
             ],
         }
     }
diff --git a/src/crates/core/src/agentic/agents/cowork_mode.rs b/src/crates/core/src/agentic/agents/cowork_mode.rs
index b0580b1b..2ccd3269 100644
--- a/src/crates/core/src/agentic/agents/cowork_mode.rs
+++ b/src/crates/core/src/agentic/agents/cowork_mode.rs
@@ -32,6 +32,7 @@ impl CoworkMode {
                 "Bash".to_string(),
                 "TerminalControl".to_string(),
                 "WebSearch".to_string(),
+                "ComputerUse".to_string(),
             ],
         }
     }
diff --git a/src/crates/core/src/agentic/agents/debug_mode.rs b/src/crates/core/src/agentic/agents/debug_mode.rs
index b9e9a1c9..9c460479 100644
--- a/src/crates/core/src/agentic/agents/debug_mode.rs
+++ b/src/crates/core/src/agentic/agents/debug_mode.rs
@@ -359,6 +359,7 @@ Below is a snapshot of the current workspace's file structure.
             "MermaidInteractive".to_string(),
             "Log".to_string(),
             "TerminalControl".to_string(),
+            "ComputerUse".to_string(),
         ]
     }
 
diff --git a/src/crates/core/src/agentic/agents/init_agent.rs b/src/crates/core/src/agentic/agents/init_agent.rs
index 3cb946c3..5e95c7a4 100644
--- a/src/crates/core/src/agentic/agents/init_agent.rs
+++ b/src/crates/core/src/agentic/agents/init_agent.rs
@@ -16,6 +16,7 @@ impl InitAgent {
                 "Write".to_string(),
                 "Edit".to_string(),
                 "Bash".to_string(),
+                "ComputerUse".to_string(),
             ],
         }
     }
diff --git a/src/crates/core/src/agentic/agents/plan_mode.rs b/src/crates/core/src/agentic/agents/plan_mode.rs
index 875ca244..dc77b8d2 100644
--- a/src/crates/core/src/agentic/agents/plan_mode.rs
+++ b/src/crates/core/src/agentic/agents/plan_mode.rs
@@ -19,6 +19,7 @@ impl PlanMode {
                 "Glob".to_string(),
                 "AskUserQuestion".to_string(),
                 "CreatePlan".to_string(),
+                "ComputerUse".to_string(),
             ],
         }
     }
diff --git a/src/crates/core/src/agentic/agents/prompt_builder/prompt_builder.rs b/src/crates/core/src/agentic/agents/prompt_builder/prompt_builder.rs
index 0e344297..1ea4bd49 100644
--- a/src/crates/core/src/agentic/agents/prompt_builder/prompt_builder.rs
+++ b/src/crates/core/src/agentic/agents/prompt_builder/prompt_builder.rs
@@ -42,6 +42,8 @@ pub struct PromptBuilderContext {
     pub remote_execution: Option<RemoteExecutionHints>,
     /// Pre-built tree text for `{PROJECT_LAYOUT}` when the workspace is not on the local disk.
     pub remote_project_layout: Option<String>,
+    /// When `Some(false)`, system prompt append Computer use text-only guidance (no screenshot tool output).
+    pub supports_image_understanding: Option<bool>,
 }
 
 impl PromptBuilderContext {
@@ -56,9 +58,15 @@ impl PromptBuilderContext {
             model_name,
             remote_execution: None,
             remote_project_layout: None,
+            supports_image_understanding: None,
         }
     }
 
+    pub fn with_supports_image_understanding(mut self, supports: bool) -> Self {
+        self.supports_image_understanding = Some(supports);
+        self
+    }
+
     pub fn with_remote_prompt_overlay(
         mut self,
         execution: RemoteExecutionHints,
@@ -93,10 +101,10 @@ impl PromptBuilder {
         let current_date = now.format("%Y-%m-%d").to_string();
 
         let computer_use_keys = match host_os {
-            "macos" => "Computer use / `key_chord`: the **local BitFun desktop** is **macOS** — use `command`, `option`, `control`, `shift` (not Win/Linux modifier names). **System clipboard (prefer over long type_text):** command+a (select all), command+c (copy), command+x (cut), command+v (paste). Spotlight: command+space; switch app: command+tab.",
-            "windows" => "Computer use / `key_chord`: the **local BitFun desktop** is **Windows** — use `meta`/`super` for the Windows key, `alt`, `control`, `shift`. **System clipboard:** control+a/c/x/v. Start menu: meta; Alt+Tab for window switch.",
-            "linux" => "Computer use / `key_chord`: the **local BitFun desktop** is **Linux** — typically `control`, `alt`, `shift`, and sometimes `meta`/`super` depending on the desktop; match the user's session. **System clipboard:** usually control+a/c/x/v (confirm in-app menus if unsure).",
-            _ => "Computer use / `key_chord`: match modifier names to the **local BitFun desktop** OS below. Prefer standard clipboard chords before retyping long text.",
+            "macos" => "Computer use / `key_chord`: the **local BitFun desktop** is **macOS** — use `command`, `option`, `control`, `shift` (not Win/Linux modifier names). **ACTION PRIORITY:** 1) Terminal/CLI/system commands (use Bash tool for `osascript`, AppleScript, shell scripts) 2) Keyboard shortcuts: command+a/c/x/v (clipboard), command+space (Spotlight), command+tab (switch app) 3) UI control (AX/OCR/mouse) only when above fail.",
+            "windows" => "Computer use / `key_chord`: the **local BitFun desktop** is **Windows** — use `meta`/`super` for Windows key, `alt`, `control`, `shift`. **ACTION PRIORITY:** 1) Terminal/CLI/system commands (use Bash tool for PowerShell, cmd, scripts) 2) Keyboard shortcuts: control+a/c/x/v (clipboard), meta (Start menu), Alt+Tab (switch) 3) UI control only when above fail.",
+            "linux" => "Computer use / `key_chord`: the **local BitFun desktop** is **Linux** — typically `control`, `alt`, `shift`, and sometimes `meta`/`super`. **ACTION PRIORITY:** 1) Terminal/CLI/system commands (use Bash tool for shell scripts, system commands) 2) Keyboard shortcuts: control+a/c/x/v (clipboard) 3) UI control (AX/OCR/mouse) only when above fail.",
+            _ => "Computer use / `key_chord`: match modifier names to the **local BitFun desktop** OS below. **ACTION PRIORITY:** 1) Terminal/CLI/system commands first 2) Keyboard shortcuts second 3) UI control (mouse/OCR) last resort.",
         };
 
         if let Some(remote) = &self.context.remote_execution {
@@ -466,6 +474,16 @@ Do not read from, modify, create, move, or delete files outside this workspace u
             result = result.replace(PLACEHOLDER_VISUAL_MODE, &visual_mode);
         }
 
+        if self.context.supports_image_understanding == Some(false) {
+            result.push_str(
+                "\n\n# Computer use (text-only primary model)\n\n\
+The configured **primary model does not accept image inputs**. When using **ComputerUse**:\n\
+- **Do not** use **`screenshot`** or **`click_label`**.\n\
+- **ACTION PRIORITY:** 1) Terminal/CLI/system commands (Bash tool) 2) Keyboard shortcuts (**`key_chord`**, **`type_text`**) 3) UI control: **`click_element`** (AX) → **`locate`** → **`move_to_text`** (use **`move_to_text_match_index`** when multiple OCR hits listed) → **`mouse_move`** (**`use_screen_coordinates`: true** with coordinates from tool JSON) → **`click`**.\n\
+- **Never guess coordinates** — always use precise methods (AX, OCR, system coordinates from tool results).\n",
+            );
+        }
+
         Ok(result.trim().to_string())
     }
 }
diff --git a/src/crates/core/src/agentic/agents/registry.rs b/src/crates/core/src/agentic/agents/registry.rs
index 536de37a..61c44cb2 100644
--- a/src/crates/core/src/agentic/agents/registry.rs
+++ b/src/crates/core/src/agentic/agents/registry.rs
@@ -385,10 +385,18 @@ impl AgentRegistry {
         match entry.category {
             AgentCategory::Mode => {
                 let mode_configs = get_mode_configs().await;
-                mode_configs
+                let mut tools = mode_configs
                     .get(agent_type)
                     .map(|config| config.available_tools.clone())
-                    .unwrap_or_else(|| entry.agent.default_tools())
+                    .unwrap_or_else(|| entry.agent.default_tools());
+                let defaults = entry.agent.default_tools();
+                const COMPUTER_USE: &str = "ComputerUse";
+                if defaults.iter().any(|t| t == COMPUTER_USE)
+                    && !tools.iter().any(|t| t == COMPUTER_USE)
+                {
+                    tools.push(COMPUTER_USE.to_string());
+                }
+                tools
             }
             AgentCategory::SubAgent | AgentCategory::Hidden => entry.agent.default_tools(),
         }
diff --git a/src/crates/core/src/agentic/execution/execution_engine.rs b/src/crates/core/src/agentic/execution/execution_engine.rs
index a3796947..9e4e48b4 100644
--- a/src/crates/core/src/agentic/execution/execution_engine.rs
+++ b/src/crates/core/src/agentic/execution/execution_engine.rs
@@ -12,6 +12,7 @@ use crate::agentic::image_analysis::{
     ImageLimits,
 };
 use crate::agentic::session::{CompressionTailPolicy, ContextCompressor, SessionManager};
+use crate::agentic::tools::framework::ToolOptions;
 use crate::agentic::tools::{get_all_registered_tools, SubagentParentInfo};
 use crate::agentic::util::build_remote_workspace_layout_preview;
 use crate::agentic::{WorkspaceBackend, WorkspaceBinding};
@@ -873,6 +874,50 @@ impl ExecutionEngine {
                     model_id, e
                 ))
             })?;
+
+        // Primary model vision capability (tools + system prompt appendix; also used below for API message stripping).
+        let (resolved_primary_model_id, primary_supports_image_understanding) = {
+            let config_service = get_global_config_service().await.ok();
+            if let Some(service) = config_service {
+                let ai_config: crate::service::config::types::AIConfig =
+                    service.get_config(Some("ai")).await.unwrap_or_default();
+
+                let resolved_id = Self::resolve_configured_model_id(&ai_config, &model_id);
+
+                let model_cfg = ai_config
+                    .models
+                    .iter()
+                    .find(|m| m.id == resolved_id)
+                    .or_else(|| ai_config.models.iter().find(|m| m.name == resolved_id))
+                    .or_else(|| {
+                        ai_config
+                            .models
+                            .iter()
+                            .find(|m| m.model_name == resolved_id)
+                    })
+                    .or_else(|| {
+                        ai_config.models.iter().find(|m| {
+                            m.model_name == ai_client.config.model
+                                && m.provider == ai_client.config.format
+                        })
+                    });
+
+                let supports = model_cfg.is_some_and(|m| {
+                    m.capabilities
+                        .iter()
+                        .any(|cap| matches!(cap, ModelCapability::ImageUnderstanding))
+                        || matches!(m.category, ModelCategory::Multimodal)
+                });
+
+                (resolved_id, supports)
+            } else {
+                warn!(
+                    "Config service unavailable, assuming primary model is text-only for image input gating"
+                );
+                (model_id.clone(), false)
+            }
+        };
+
         // Get configuration for whether to support preserving historical thinking content
         let enable_thinking = ai_client.config.enable_thinking_process;
         let support_preserved_thinking = ai_client.config.support_preserved_thinking;
@@ -894,7 +939,8 @@ impl ExecutionEngine {
                     workspace_path.clone(),
                     Some(context.session_id.clone()),
                     Some(ai_client.config.model.clone()),
-                );
+                )
+                .with_supports_image_understanding(primary_supports_image_understanding);
                 let overlayed = if let Some(ws) = context.workspace.as_ref() {
                     if ws.is_remote() {
                         if let Some(cid) = ws.connection_id() {
@@ -1035,6 +1081,7 @@ impl ExecutionEngine {
                 &allowed_tools,
                 context.workspace.as_ref(),
                 &agent_type,
+                primary_supports_image_understanding,
             )
             .await
         } else {
@@ -1043,49 +1090,6 @@ impl ExecutionEngine {
 
         let enable_context_compression = session.config.enable_context_compression;
         let compression_threshold = session.config.compression_threshold;
-        // Detect whether the primary model supports multimodal image inputs.
-        // When false, multimodal user messages are converted to text placeholders before the provider call.
-        let (resolved_primary_model_id, primary_supports_image_understanding) = {
-            let config_service = get_global_config_service().await.ok();
-            if let Some(service) = config_service {
-                let ai_config: crate::service::config::types::AIConfig =
-                    service.get_config(Some("ai")).await.unwrap_or_default();
-
-                let resolved_id = Self::resolve_configured_model_id(&ai_config, &model_id);
-
-                let model_cfg = ai_config
-                    .models
-                    .iter()
-                    .find(|m| m.id == resolved_id)
-                    .or_else(|| ai_config.models.iter().find(|m| m.name == resolved_id))
-                    .or_else(|| {
-                        ai_config
-                            .models
-                            .iter()
-                            .find(|m| m.model_name == resolved_id)
-                    })
-                    .or_else(|| {
-                        ai_config.models.iter().find(|m| {
-                            m.model_name == ai_client.config.model
-                                && m.provider == ai_client.config.format
-                        })
-                    });
-
-                let supports = model_cfg.is_some_and(|m| {
-                    m.capabilities
-                        .iter()
-                        .any(|cap| matches!(cap, ModelCapability::ImageUnderstanding))
-                        || matches!(m.category, ModelCategory::Multimodal)
-                });
-
-                (resolved_id, supports)
-            } else {
-                warn!(
-                    "Config service unavailable, assuming primary model is text-only for image input gating"
-                );
-                (model_id.clone(), false)
-            }
-        };
 
         let mut execution_context_vars = context.context.clone();
         execution_context_vars.insert(
@@ -1475,12 +1479,18 @@ impl ExecutionEngine {
         mode_allowed_tools: &[String],
         workspace: Option<&crate::agentic::WorkspaceBinding>,
         agent_type: &str,
+        primary_supports_image_understanding: bool,
     ) -> (Vec<String>, Option<Vec<ToolDefinition>>) {
         // Use get_all_registered_tools to get all tools including MCP tools
         let all_tools = get_all_registered_tools().await;
 
         // Filter tools: 1) Check if enabled 2) Check if mode allows
         let mut tool_definitions = Vec::new();
+        let mut tool_opts_custom = HashMap::new();
+        tool_opts_custom.insert(
+            "primary_model_supports_image_understanding".to_string(),
+            serde_json::Value::Bool(primary_supports_image_understanding),
+        );
         let description_context = crate::agentic::tools::framework::ToolUseContext {
             tool_call_id: None,
             message_id: None,
@@ -1492,7 +1502,20 @@ impl ExecutionEngine {
             safe_mode: None,
             abort_controller: None,
             read_file_timestamps: Default::default(),
-            options: None,
+            options: Some(ToolOptions {
+                commands: vec![],
+                tools: vec![],
+                verbose: None,
+                slow_and_capable_model: None,
+                safe_mode: None,
+                fork_number: None,
+                message_log_name: None,
+                max_thinking_tokens: None,
+                is_koding_request: None,
+                koding_context: None,
+                is_custom_command: None,
+                custom_data: Some(tool_opts_custom),
+            }),
             response_state: None,
             image_context_provider: None,
             computer_use_host: None,
@@ -1513,10 +1536,14 @@ impl ExecutionEngine {
                     .await
                     .unwrap_or_else(|_| format!("Tool: {}", tool.name()));
 
+                let parameters = tool
+                    .input_schema_for_model_with_context(Some(&description_context))
+                    .await;
+
                 tool_definitions.push(ToolDefinition {
                     name: tool.name().to_string(),
                     description,
-                    parameters: tool.input_schema_for_model().await,
+                    parameters,
                 });
             }
         }
diff --git a/src/crates/core/src/agentic/tools/framework.rs b/src/crates/core/src/agentic/tools/framework.rs
index 67e14721..1f20d7e0 100644
--- a/src/crates/core/src/agentic/tools/framework.rs
+++ b/src/crates/core/src/agentic/tools/framework.rs
@@ -68,6 +68,17 @@ impl ToolUseContext {
         self.workspace_services.as_ref().map(|s| s.shell.as_ref())
     }
 
+    /// Whether the session primary model accepts image inputs (from tool-definition / pipeline context).
+    /// Defaults to **true** when unset (e.g. API listings without model metadata).
+    pub fn primary_model_supports_image_understanding(&self) -> bool {
+        self.options
+            .as_ref()
+            .and_then(|o| o.custom_data.as_ref())
+            .and_then(|m| m.get("primary_model_supports_image_understanding"))
+            .and_then(|v| v.as_bool())
+            .unwrap_or(true)
+    }
+
     /// Resolve a user or model-supplied path for file/shell tools. Uses POSIX semantics when the
     /// workspace is remote SSH so Windows-hosted clients still resolve `/home/...` correctly.
     pub fn resolve_workspace_tool_path(&self, path: &str) -> BitFunResult<String> {
@@ -223,6 +234,16 @@ pub trait Tool: Send + Sync {
         self.input_schema()
     }
 
+    /// JSON Schema for the model when tool listing has a [`ToolUseContext`] (e.g. primary model vision capability).
+    /// Default: ignores context and delegates to [`input_schema_for_model`].
+    async fn input_schema_for_model_with_context(
+        &self,
+        context: Option<&ToolUseContext>,
+    ) -> Value {
+        let _ = context;
+        self.input_schema_for_model().await
+    }
+
     /// Input JSON Schema - optional extra schema
     fn input_json_schema(&self) -> Option<Value> {
         None
diff --git a/src/crates/core/src/agentic/tools/implementations/bash_tool.rs b/src/crates/core/src/agentic/tools/implementations/bash_tool.rs
index d2243690..8547dcfb 100644
--- a/src/crates/core/src/agentic/tools/implementations/bash_tool.rs
+++ b/src/crates/core/src/agentic/tools/implementations/bash_tool.rs
@@ -267,9 +267,9 @@ Usage notes:
                 base = base
             );
         }
-        if context.and_then(|c| c.agent_type.as_deref()) == Some("Claw") {
+        if !context.map(|c| c.is_remote()).unwrap_or(false) {
             base.push_str(
-                "\n\n**Claw (desktop automation):** Prefer this tool for anything achievable from the **workspace shell** (build, test, git, scripts, CLIs). On **macOS**, `open -a \"AppName\"` launches or foregrounds an app with fewer steps than GUI workflows. Use **`ComputerUse`** **`action: locate`** for **named** on-screen controls before guessing coordinates from **`action: screenshot`** alone.",
+                "\n\n**Desktop automation:** Prefer this tool for anything achievable from the **workspace shell** (build, test, git, scripts, CLIs). On **macOS**, `open -a \"AppName\"` launches or foregrounds an app with fewer steps than GUI workflows. When **Computer use** is enabled, use **`ComputerUse`** **`action: locate`** for **named** on-screen controls before guessing coordinates from **`action: screenshot`** alone.",
             );
         }
         Ok(base)
diff --git a/src/crates/core/src/agentic/tools/implementations/code_review_tool.rs b/src/crates/core/src/agentic/tools/implementations/code_review_tool.rs
index f3f9ffbb..fc395c2f 100644
--- a/src/crates/core/src/agentic/tools/implementations/code_review_tool.rs
+++ b/src/crates/core/src/agentic/tools/implementations/code_review_tool.rs
@@ -231,6 +231,14 @@ impl Tool for CodeReviewTool {
         Self::input_schema_value_for_language(lang.as_str())
     }
 
+    async fn input_schema_for_model_with_context(
+        &self,
+        context: Option<&crate::agentic::tools::framework::ToolUseContext>,
+    ) -> Value {
+        let _ = context;
+        self.input_schema_for_model().await
+    }
+
     fn is_readonly(&self) -> bool {
         true
     }
diff --git a/src/crates/core/src/agentic/tools/implementations/computer_use_locate.rs b/src/crates/core/src/agentic/tools/implementations/computer_use_locate.rs
index 9f54a6b8..f063eebe 100644
--- a/src/crates/core/src/agentic/tools/implementations/computer_use_locate.rs
+++ b/src/crates/core/src/agentic/tools/implementations/computer_use_locate.rs
@@ -15,11 +15,6 @@ pub(crate) async fn execute_computer_use_locate(
     input: &Value,
     context: &ToolUseContext,
 ) -> BitFunResult<Vec<ToolResult>> {
-    if context.agent_type.as_deref() != Some("Claw") {
-        return Err(BitFunError::tool(
-            "ComputerUse action locate is only available in Claw assistant mode.".to_string(),
-        ));
-    }
     if context.is_remote() {
         return Err(BitFunError::tool(
             "ComputerUse action locate cannot run while the session workspace is remote (SSH)."
diff --git a/src/crates/core/src/agentic/tools/implementations/computer_use_mouse_click_tool.rs b/src/crates/core/src/agentic/tools/implementations/computer_use_mouse_click_tool.rs
index ccf3f3ff..bd7fb0d7 100644
--- a/src/crates/core/src/agentic/tools/implementations/computer_use_mouse_click_tool.rs
+++ b/src/crates/core/src/agentic/tools/implementations/computer_use_mouse_click_tool.rs
@@ -1,4 +1,4 @@
-//! Mouse button click and wheel at the current pointer (Claw Computer use).
+//! Mouse button click and wheel at the current pointer (Computer use).
 
 use crate::agentic::tools::computer_use_capability::computer_use_desktop_available;
 use crate::agentic::tools::implementations::computer_use_tool::computer_use_execute_mouse_click_tool;
@@ -88,11 +88,6 @@ impl Tool for ComputerUseMouseClickTool {
     }
 
     async fn call_impl(&self, input: &Value, context: &ToolUseContext) -> BitFunResult<Vec<ToolResult>> {
-        if context.agent_type.as_deref() != Some("Claw") {
-            return Err(BitFunError::tool(
-                "ComputerUseMouseClick is only available in Claw assistant mode.".to_string(),
-            ));
-        }
         if context.is_remote() {
             return Err(BitFunError::tool(
                 "ComputerUseMouseClick cannot run while the session workspace is remote (SSH)."
diff --git a/src/crates/core/src/agentic/tools/implementations/computer_use_mouse_precise_tool.rs b/src/crates/core/src/agentic/tools/implementations/computer_use_mouse_precise_tool.rs
index 6b2f7e67..d93b72d7 100644
--- a/src/crates/core/src/agentic/tools/implementations/computer_use_mouse_precise_tool.rs
+++ b/src/crates/core/src/agentic/tools/implementations/computer_use_mouse_precise_tool.rs
@@ -1,4 +1,4 @@
-//! Absolute pointer positioning for Claw Computer use.
+//! Absolute pointer positioning for Computer use.
 
 use crate::agentic::tools::computer_use_capability::computer_use_desktop_available;
 use crate::agentic::tools::implementations::computer_use_tool::computer_use_execute_mouse_precise;
@@ -77,11 +77,6 @@ impl Tool for ComputerUseMousePreciseTool {
     }
 
     async fn call_impl(&self, input: &Value, context: &ToolUseContext) -> BitFunResult<Vec<ToolResult>> {
-        if context.agent_type.as_deref() != Some("Claw") {
-            return Err(BitFunError::tool(
-                "ComputerUseMousePrecise is only available in Claw assistant mode.".to_string(),
-            ));
-        }
         if context.is_remote() {
             return Err(BitFunError::tool(
                 "ComputerUseMousePrecise cannot run while the session workspace is remote (SSH)."
diff --git a/src/crates/core/src/agentic/tools/implementations/computer_use_mouse_step_tool.rs b/src/crates/core/src/agentic/tools/implementations/computer_use_mouse_step_tool.rs
index 086a9e3c..07145ecd 100644
--- a/src/crates/core/src/agentic/tools/implementations/computer_use_mouse_step_tool.rs
+++ b/src/crates/core/src/agentic/tools/implementations/computer_use_mouse_step_tool.rs
@@ -1,4 +1,4 @@
-//! Cardinal pointer step (up/down/left/right) for Claw Computer use.
+//! Cardinal pointer step (up/down/left/right) for Computer use.
 
 use crate::agentic::tools::computer_use_capability::computer_use_desktop_available;
 use crate::agentic::tools::implementations::computer_use_tool::computer_use_execute_mouse_step;
@@ -72,11 +72,6 @@ impl Tool for ComputerUseMouseStepTool {
     }
 
     async fn call_impl(&self, input: &Value, context: &ToolUseContext) -> BitFunResult<Vec<ToolResult>> {
-        if context.agent_type.as_deref() != Some("Claw") {
-            return Err(BitFunError::tool(
-                "ComputerUseMouseStep is only available in Claw assistant mode.".to_string(),
-            ));
-        }
         if context.is_remote() {
             return Err(BitFunError::tool(
                 "ComputerUseMouseStep cannot run while the session workspace is remote (SSH)."
diff --git a/src/crates/core/src/agentic/tools/implementations/computer_use_tool.rs b/src/crates/core/src/agentic/tools/implementations/computer_use_tool.rs
index 3dc9ecbe..61227040 100644
--- a/src/crates/core/src/agentic/tools/implementations/computer_use_tool.rs
+++ b/src/crates/core/src/agentic/tools/implementations/computer_use_tool.rs
@@ -1,4 +1,4 @@
-//! Desktop automation for Claw (Computer use).
+//! Desktop automation (Computer use).
 
 use super::computer_use_input::{
     coordinate_mode, ensure_pointer_move_uses_screen_coordinates_only, parse_screenshot_params,
@@ -93,6 +93,73 @@ impl ComputerUseTool {
         Self
     }
 
+    /// Tool description when the primary model is **text-only** (no `screenshot` / `click_label` / JPEG workflow).
+    fn description_text_only() -> String {
+        let os = Self::host_os_label();
+        let keys = Self::key_chord_os_hint();
+        format!(
+            "Desktop automation (host OS: {}). {} \
+The **primary model cannot consume images** in tool results — **do not** use **`screenshot`** or **`click_label`**.\n\
+**ACTION PRIORITY (CRITICAL):** Always think in this order:\n\
+1. **Terminal/CLI/System commands first** — Use Bash tool for terminal commands, system scripts (e.g., macOS `osascript`), shell automation. Most efficient.\n\
+2. **Keyboard shortcuts second** — Use **`key_chord`** / **`type_text`** for system/app shortcuts, navigation keys.\n\
+3. **Precise UI control last** — Only when above fail: **`click_element`** (AX) → **`move_to_text`** (OCR, use **`move_to_text_match_index`** from text `candidates` when multiple hits) → **`mouse_move`** (**`use_screen_coordinates`: true** with **`global_center_*`** / **`locate`** / **`pointer_global`**) → **`click`**.\n\
+**Rhythm:** one action at a time; use **`wait`** when UI animates. Observe **`interaction_state`** and **`computer_use_context`** in tool JSON.\n\
+**`click_element` / `locate`:** Accessibility (AX/UIA/AT-SPI). **`move_to_text`:** OCR match + move pointer only. **`click`:** at current pointer only — use **`mouse_move`** or **`move_to_text`** / **`click_element`** first.\n\
+**`mouse_move` / `drag`:** **`use_screen_coordinates`: true** with globals from tools. **`pointer_move_rel`:** relative nudge; host may block right after certain flows — follow tool errors.\n\
+**`key_chord` / `type_text` / `scroll` / `wait`:** standard desktop automation without any screenshot step.\n",
+            os, keys
+        )
+    }
+
+    /// JSON Schema without `screenshot`, `click_label`, or screenshot-only fields.
+    fn input_schema_text_only() -> Value {
+        json!({
+            "type": "object",
+            "properties": {
+                "action": {
+                    "type": "string",
+                    "enum": ["click_element", "move_to_text", "click", "mouse_move", "scroll", "drag", "locate", "key_chord", "type_text", "pointer_move_rel", "wait"],
+                    "description": "The action to perform. **Primary model is text-only — no `screenshot` or `click_label`.** **ACTION PRIORITY:** 1) Use Bash tool for CLI/terminal/system commands first. 2) Prefer `key_chord` for shortcuts/navigation. 3) Only when above fail: `click_element` (AX) → `move_to_text` (OCR, use `move_to_text_match_index` when multiple hits listed) → `mouse_move` (**`use_screen_coordinates`: true** with globals) + `click`. Never guess coordinates."
+                },
+                "x": { "type": "integer", "description": "For `mouse_move` and `drag`: X in **global display** units when **`use_screen_coordinates`: true** (required). **Not** for `click`." },
+                "y": { "type": "integer", "description": "For `mouse_move` and `drag`: Y in **global display** units when **`use_screen_coordinates`: true** (required). **Not** for `click`." },
+                "coordinate_mode": { "type": "string", "enum": ["image", "normalized"], "description": "Ignored for `mouse_move` / `drag` — host rejects image/normalized positioning; always set **`use_screen_coordinates`: true**." },
+                "use_screen_coordinates": { "type": "boolean", "description": "For `mouse_move`, `drag`: **must be true** — global display coordinates from `move_to_text`, `locate`, AX, or `pointer_global`. **Not** for `click`." },
+                "button": { "type": "string", "enum": ["left", "right", "middle"], "description": "For `click`, `click_element`, `drag`: mouse button (default left)." },
+                "num_clicks": { "type": "integer", "minimum": 1, "maximum": 3, "description": "For `click`, `click_element`: 1=single (default), 2=double, 3=triple click." },
+                "delta_x": { "type": "integer", "description": "For `pointer_move_rel`: horizontal delta (negative=left). For `scroll`: horizontal wheel delta." },
+                "delta_y": { "type": "integer", "description": "For `pointer_move_rel`: vertical delta (negative=up). For `scroll`: vertical wheel delta." },
+                "start_x": { "type": "integer", "description": "For `drag`: start X coordinate." },
+                "start_y": { "type": "integer", "description": "For `drag`: start Y coordinate." },
+                "end_x": { "type": "integer", "description": "For `drag`: end X coordinate." },
+                "end_y": { "type": "integer", "description": "For `drag`: end Y coordinate." },
+                "keys": { "type": "array", "items": { "type": "string" }, "description": "For `key_chord`: keys in order — modifiers first, then the main key. Desktop host waits after pressing modifiers so shortcuts register (important on macOS with IME)." },
+                "text": { "type": "string", "description": "For `type_text`: text to type. Prefer clipboard paste (key_chord) for long content." },
+                "ms": { "type": "integer", "description": "For `wait`: duration in milliseconds." },
+                "text_query": { "type": "string", "description": "For `move_to_text`: visible text to OCR-match on screen (case-insensitive substring)." },
+                "move_to_text_match_index": { "type": "integer", "minimum": 1, "description": "For `move_to_text`: **1-based** index from `candidates[].match_index` after disambiguation (multiple OCR hits). Omit on the first pass; set when choosing which hit to move to." },
+                "ocr_region_native": {
+                    "type": "object",
+                    "description": "For `move_to_text`: optional global native rectangle for OCR. If omitted, macOS uses the frontmost window bounds from Accessibility; other OSes use the primary display.",
+                    "properties": {
+                        "x0": { "type": "integer", "description": "Top-left X in global screen coordinates." },
+                        "y0": { "type": "integer", "description": "Top-left Y in global screen coordinates." },
+                        "width": { "type": "integer", "minimum": 1, "description": "Width in the same coordinate unit as x0/y0." },
+                        "height": { "type": "integer", "minimum": 1, "description": "Height in the same coordinate unit as x0/y0." }
+                    }
+                },
+                "title_contains": { "type": "string", "description": "For `locate`, `click_element`: case-insensitive substring match on accessible title (AXTitle)." },
+                "role_substring": { "type": "string", "description": "For `locate`, `click_element`: case-insensitive substring on AXRole." },
+                "identifier_contains": { "type": "string", "description": "For `locate`, `click_element`: case-insensitive substring on AXIdentifier." },
+                "max_depth": { "type": "integer", "minimum": 1, "maximum": 200, "description": "For `locate`, `click_element`: max BFS depth (default 48)." },
+                "filter_combine": { "type": "string", "enum": ["all", "any"], "description": "For `locate`, `click_element`: `all` (default, AND) or `any` (OR) for filter combination." }
+            },
+            "required": ["action"],
+            "additionalProperties": false
+        })
+    }
+
     /// Max OCR hits to attach as preview crops + AX (multimodal disambiguation).
     const MOVE_TO_TEXT_DISAMBIGUATION_MAX: usize = 8;
     /// Half-size in native screen pixels for each candidate preview (~400×400 logical crop).
@@ -177,6 +244,70 @@ impl ComputerUseTool {
         Ok(vec![ToolResult::ok_with_images(body, Some(hint), attachments)])
     }
 
+    /// Same as [`Self::move_to_text_disambiguation_response`] but **no image attachments** (primary model is text-only).
+    async fn move_to_text_disambiguation_text_only(
+        host_ref: &dyn crate::agentic::tools::computer_use_host::ComputerUseHost,
+        text_query: &str,
+        ocr_region_native: Option<OcrRegionNative>,
+        matches: &[ScreenOcrTextMatch],
+    ) -> BitFunResult<Vec<ToolResult>> {
+        let take = matches.len().min(Self::MOVE_TO_TEXT_DISAMBIGUATION_MAX);
+        let mut candidates: Vec<Value> = Vec::with_capacity(take);
+        for (i, m) in matches.iter().take(take).enumerate() {
+            let idx_1based = i + 1;
+            let ax = host_ref
+                .accessibility_hit_at_global_point(m.center_x, m.center_y)
+                .await?;
+            candidates.push(json!({
+                "match_index": idx_1based,
+                "ocr_text": m.text,
+                "confidence": m.confidence,
+                "global_center_x": m.center_x,
+                "global_center_y": m.center_y,
+                "bounds_left": m.bounds_left,
+                "bounds_top": m.bounds_top,
+                "bounds_width": m.bounds_width,
+                "bounds_height": m.bounds_height,
+                "accessibility": ax,
+            }));
+        }
+        let input_coords = json!({
+            "kind": "move_to_text",
+            "text_query": text_query,
+            "ocr_region_native": ocr_region_native,
+            "move_to_text_phase": "disambiguation",
+        });
+        let mut body = json!({
+            "success": true,
+            "action": "move_to_text",
+            "move_to_text_phase": "disambiguation",
+            "text_query": text_query,
+            "ocr_region_native": ocr_region_native,
+            "disambiguation_required": true,
+            "instruction": "Several OCR hits for this substring. The primary model **cannot** view screenshots — pick **`move_to_text_match_index`** using **`candidates`** (global_center_* + accessibility) only. Call **`move_to_text` again** with the same `text_query`, same `ocr_region_native`, and **`move_to_text_match_index`** = that index. Pointer was not moved.",
+            "candidates": candidates,
+            "total_ocr_matches": matches.len(),
+            "candidates_previewed": take,
+        });
+        if take < matches.len() {
+            if let Some(obj) = body.as_object_mut() {
+                obj.insert(
+                    "truncation_note".to_string(),
+                    json!(format!(
+                        "Only the first {} of {} OCR matches are listed; narrow `ocr_region_native` or `text_query` if needed.",
+                        take, matches.len()
+                    )),
+                );
+            }
+        }
+        let body = computer_use_augment_result_json(host_ref, body, Some(input_coords)).await;
+        let hint = format!(
+            "move_to_text: {} OCR matches — set move_to_text_match_index using text candidates (no image previews). Pointer not moved.",
+            matches.len(),
+        );
+        Ok(vec![ToolResult::ok(body, Some(hint))])
+    }
+
     fn primary_api_format(ctx: &ToolUseContext) -> String {
         ctx.options
             .as_ref()
@@ -190,6 +321,11 @@ impl ComputerUseTool {
     /// Screenshot tool results attach JPEGs via `tool_image_attachments`; only providers whose
     /// request converters emit multimodal tool output are supported (Anthropic + OpenAI-compatible).
     fn require_multimodal_tool_output_for_screenshot(ctx: &ToolUseContext) -> BitFunResult<()> {
+        if !ctx.primary_model_supports_image_understanding() {
+            return Err(BitFunError::tool(
+                "The primary model does not accept images; do not use ComputerUse action `screenshot` or other image-producing steps. Use `click_element`, `locate`, `move_to_text` (with `move_to_text_match_index` when listed), `mouse_move` with globals from tool JSON, `key_chord`, etc.".to_string(),
+            ));
+        }
         let f = Self::primary_api_format(ctx);
         if matches!(
             f.as_str(),
@@ -845,9 +981,12 @@ impl Tool for ComputerUseTool {
         let keys = Self::key_chord_os_hint();
         Ok(format!(
             "Desktop automation (host OS: {}). {} All actions in one tool. Send only parameters that apply to the chosen `action`. \
-**Cowork-style loop (default rhythm):** **`screenshot`** (observe current UI) → **one** input action (`key_chord`, `type_text`, `move_to_text` + `click`, `click_element`, `scroll`, …) → **`screenshot`** again to **verify** the outcome before the next decision. Use **`wait`** if the UI animates. When **`interaction_state.recommend_screenshot_to_verify_last_action`** is true, your **next** call should usually be **`screenshot`** (optionally **`screenshot_reset_navigation`**: true for full-screen context). This is separate from the host rule that requires a **fresh** capture **before** guarded **`click`** / Enter **`key_chord`** after pointer moves — follow both. \
-**Input priority:** Prefer **`key_chord`** / **`type_text`** over mouse when one key or typing completes the step (e.g. **Enter** to confirm default, **Escape** to cancel, **Tab** to move focus). Do not click “OK”/“Submit” when **Enter** is equivalent; use **`screenshot`** then Enter **`key_chord`** per host when required. \
-**Targeting priority (when pointing is required):** `click_element` → **`move_to_text`** (OCR + move pointer only) → `click_label` (when SoM exists) → **`screenshot`** (confirm / drill) + **`mouse_move`** (**`use_screen_coordinates`: true only**) + **`click`** last. **Screenshots are for confirmation — do not guess move targets from JPEG pixels.** \
+**ACTION PRIORITY (CRITICAL):** Always think in this order before choosing an action:\n\
+1. **Terminal/CLI/System commands first** — Use Bash tool for terminal commands, system scripts (e.g., macOS `osascript`, AppleScript), shell automation. This is the MOST EFFICIENT approach.\n\
+2. **Keyboard shortcuts second** — Use **`key_chord`** for system shortcuts, app shortcuts, navigation keys (Enter, Escape, Tab, Space, Arrow keys). Prefer over mouse when equivalent.\n\
+3. **Precise UI control last** — Only when above methods fail: use **`click_element`** (AX/accessibility) → **`move_to_text`** (OCR) → **`click_label`** (SoM) → **`mouse_move`** + **`click`** (coordinate-based, last resort).\n\
+**Screenshot usage:** **`screenshot`** is ONLY for observing/confirming UI state and extracting text/information — NEVER use screenshot coordinates to control mouse movement. Always use precise methods (AX, OCR, system coordinates) for targeting.\n\
+**Cowork-style loop:** **`screenshot`** (observe) → **one** action → **`screenshot`** (verify). Use **`wait`** if UI animates. When **`interaction_state.recommend_screenshot_to_verify_last_action`** is true, call **`screenshot`** next. \
 **`click_element`:** Accessibility tree (AX/UIA/AT-SPI) locate + click. Provide `title_contains` / `role_substring` / `identifier_contains`. On macOS, **`TextArea`** and **`TextField`** match both `AXTextArea` and `AXTextField` (many chat apps use TextField for compose). If several text fields match, the host deprioritizes known **search** controls (e.g. WeChat `_SC_SEARCH_FIELD`) and prefers **lower** on-screen fields (composer). Bypasses coordinate screenshot guard. \
 **`move_to_text`:** OCR-match visible text (`text_query`) and **move the pointer** to it (no click, no keys); **no prior `screenshot` required for targeting** (host captures **raw** pixels for Vision — no agent screenshot overlays; on macOS defaults to the **frontmost window** unless **`ocr_region_native`** overrides). Matching **strips whitespace** between CJK glyphs and allows **small edit distance** when Vision mis-reads one character. The host **trusts** the resulting globals — **next `click`** does **not** require an extra `screenshot` (same as AX). If **several** hits match, the host returns **preview JPEGs + accessibility** per candidate — pick **`move_to_text_match_index`** (1-based) and call **`move_to_text` again** with the same query/region, or narrow with **`ocr_region_native`**. When **`click_label`** lists a results table (`AXTable`), prefer **`click_label`** on that label over guessing OCR row text. Use **`click`** afterward if you need a mouse press. Prefer after `click_element` misses when text is visible. \
 **`click_label`:** After `screenshot` with `som_labels`, click by label number. Bypasses coordinate guard. \
@@ -863,9 +1002,16 @@ impl Tool for ComputerUseTool {
 
     async fn description_with_context(
         &self,
-        _context: Option<&ToolUseContext>,
+        context: Option<&ToolUseContext>,
     ) -> BitFunResult<String> {
-        self.description().await
+        let vision = context
+            .map(|c| c.primary_model_supports_image_understanding())
+            .unwrap_or(true);
+        if vision {
+            self.description().await
+        } else {
+            Ok(Self::description_text_only())
+        }
     }
 
     fn input_schema(&self) -> Value {
@@ -875,7 +1021,7 @@ impl Tool for ComputerUseTool {
                 "action": {
                     "type": "string",
                     "enum": ["screenshot", "click_element", "click_label", "move_to_text", "click", "mouse_move", "scroll", "drag", "locate", "key_chord", "type_text", "pointer_move_rel", "wait"],
-                    "description": "The action to perform. **Prefer `key_chord` over `click` when a key completes the same step (Enter, Escape, Tab, Space, app shortcuts).** `click_element` = find UI element by accessibility + click (for named controls when keyboard is not equivalent). `click_label` = click a numbered Set-of-Mark label from the latest screenshot. `move_to_text` = OCR visible text and move pointer **only** (no click). **After `click_element` fails, prefer `move_to_text` (visible substring, UI language) over vision guesses.** `click` = press at **current pointer only** — **do not** pass `x`, `y`, `coordinate_mode`, or `use_screen_coordinates` (use `mouse_move` first). `mouse_move` = absolute move with **`use_screen_coordinates`: true** (globals from tools — **no** JPEG pixel mode). `scroll` = mouse wheel. `drag` = drag between two globals (**`use_screen_coordinates`: true**). `screenshot` = confirmation JPEG (host may apply ~500×500 when required). `locate` = find UI element (no click). `key_chord` = keyboard shortcut. `type_text` = type string. `pointer_move_rel` = relative move — **host blocks right after `screenshot`** until a trusted absolute move (`move_to_text`, `mouse_move`, `click_element`, `click_label`). `wait` = pause."
+                    "description": "The action to perform. **ACTION PRIORITY:** 1) Use Bash tool for CLI/terminal/system commands (most efficient). 2) Prefer **`key_chord`** for shortcuts/navigation keys over mouse. 3) Only when above fail: `click_element` (AX) → `move_to_text` (OCR, move pointer only) → `click_label` (SoM) → `mouse_move` (globals only, **`use_screen_coordinates`: true**) + `click` (last resort). **`screenshot`** is for observation/confirmation ONLY — never derive mouse coordinates from screenshots. `click` = press at **current pointer only** (no x/y params). `scroll`, `type_text`, `drag`, `pointer_move_rel`, `wait`, `locate` = standard actions."
                 },
                 "x": { "type": "integer", "description": "For `mouse_move` and `drag`: X in **global display** units when **`use_screen_coordinates`: true** (required). **Not** for `click`." },
                 "y": { "type": "integer", "description": "For `mouse_move` and `drag`: Y in **global display** units when **`use_screen_coordinates`: true** (required). **Not** for `click`." },
@@ -922,6 +1068,20 @@ impl Tool for ComputerUseTool {
         })
     }
 
+    async fn input_schema_for_model_with_context(
+        &self,
+        context: Option<&ToolUseContext>,
+    ) -> Value {
+        let vision = context
+            .map(|c| c.primary_model_supports_image_understanding())
+            .unwrap_or(true);
+        if vision {
+            self.input_schema_for_model().await
+        } else {
+            Self::input_schema_text_only()
+        }
+    }
+
     fn is_readonly(&self) -> bool {
         false
     }
@@ -947,11 +1107,6 @@ impl Tool for ComputerUseTool {
     }
 
     async fn call_impl(&self, input: &Value, context: &ToolUseContext) -> BitFunResult<Vec<ToolResult>> {
-        if context.agent_type.as_deref() != Some("Claw") {
-            return Err(BitFunError::tool(
-                "ComputerUse is only available in Claw assistant mode.".to_string(),
-            ));
-        }
         if context.is_remote() {
             return Err(BitFunError::tool(
                 "ComputerUse cannot run while the session workspace is remote (SSH).".to_string(),
@@ -1050,6 +1205,11 @@ impl Tool for ComputerUseTool {
             }
 
             "click_label" => {
+                if !context.primary_model_supports_image_understanding() {
+                    return Err(BitFunError::tool(
+                        "click_label requires Set-of-Mark labels from a screenshot; the primary model is text-only. Use `click_element`, `move_to_text`, `locate`, or `mouse_move` with globals from tool JSON, then `click`.".to_string(),
+                    ));
+                }
                 let label = input
                     .get("label")
                     .and_then(|v| v.as_u64())
@@ -1141,9 +1301,18 @@ impl Tool for ComputerUseTool {
 
                     let n = matches.len();
                     if n > 1 && move_to_text_match_index.is_none() {
-                        return Self::move_to_text_disambiguation_response(
+                        if context.primary_model_supports_image_understanding() {
+                            return Self::move_to_text_disambiguation_response(
+                                host_ref,
+                                context,
+                                text_query,
+                                ocr_region_native.clone(),
+                                &matches,
+                            )
+                            .await;
+                        }
+                        return Self::move_to_text_disambiguation_text_only(
                             host_ref,
-                            context,
                             text_query,
                             ocr_region_native.clone(),
                             &matches,
diff --git a/src/crates/core/src/service/config/types.rs b/src/crates/core/src/service/config/types.rs
index 819bb55a..ea5aacb5 100644
--- a/src/crates/core/src/service/config/types.rs
+++ b/src/crates/core/src/service/config/types.rs
@@ -445,7 +445,7 @@ pub struct AIConfig {
     #[serde(default)]
     pub known_tools: Vec<String>,
 
-    /// Allow Claw Computer use (desktop automation) when the desktop host is available.
+    /// Allow Computer use (desktop automation) when the desktop host is available (all session modes).
     #[serde(default)]
     pub computer_use_enabled: bool,
 }
diff --git a/src/crates/core/src/service/snapshot/manager.rs b/src/crates/core/src/service/snapshot/manager.rs
index 5fe69396..45506467 100644
--- a/src/crates/core/src/service/snapshot/manager.rs
+++ b/src/crates/core/src/service/snapshot/manager.rs
@@ -352,6 +352,15 @@ impl Tool for WrappedTool {
         self.original_tool.input_schema_for_model().await
     }
 
+    async fn input_schema_for_model_with_context(
+        &self,
+        context: Option<&crate::agentic::tools::framework::ToolUseContext>,
+    ) -> Value {
+        self.original_tool
+            .input_schema_for_model_with_context(context)
+            .await
+    }
+
     fn input_json_schema(&self) -> Option<Value> {
         self.original_tool.input_json_schema()
     }
diff --git a/src/web-ui/src/locales/en-US/settings/session-config.json b/src/web-ui/src/locales/en-US/settings/session-config.json
index 0d0cf02c..27b12d39 100644
--- a/src/web-ui/src/locales/en-US/settings/session-config.json
+++ b/src/web-ui/src/locales/en-US/settings/session-config.json
@@ -13,10 +13,10 @@
     "sectionDescription": "Confirmation and timeout settings when tools run in a session."
   },
   "computerUse": {
-    "sectionTitle": "Computer use (Claw)",
+    "sectionTitle": "Computer use (desktop)",
     "sectionDescription": "In the BitFun desktop app, the assistant can capture the screen and control the mouse and keyboard; requires a multimodal model for vision.",
     "enable": "Enable Computer use",
-    "enableDesc": "When off, the ComputerUse tool stays disabled even in Claw mode.",
+    "enableDesc": "When off, the ComputerUse tool stays disabled in every session mode.",
     "accessibility": "Accessibility",
     "accessibilityDesc": "",
     "screenCapture": "Screen recording",
diff --git a/src/web-ui/src/locales/zh-CN/settings/session-config.json b/src/web-ui/src/locales/zh-CN/settings/session-config.json
index 9d611335..b5e32ad8 100644
--- a/src/web-ui/src/locales/zh-CN/settings/session-config.json
+++ b/src/web-ui/src/locales/zh-CN/settings/session-config.json
@@ -13,10 +13,10 @@
     "sectionDescription": "本会话中 AI 调用工具时的确认与超时策略。"
   },
   "computerUse": {
-    "sectionTitle": "Computer use（助理 Claw）",
+    "sectionTitle": "Computer use（桌面自动化）",
     "sectionDescription": "在 BitFun 桌面端允许助理截取屏幕并控制键鼠；需多模态模型理解画面。",
     "enable": "启用 Computer use",
-    "enableDesc": "关闭时，即使在 Claw 模式下也不会启用 ComputerUse 工具。",
+    "enableDesc": "关闭时，任何会话模式都不会启用 ComputerUse 工具。",
     "accessibility": "辅助功能",
     "accessibilityDesc": "",
     "screenCapture": "屏幕录制",