fix(cli): honor configured output token limits

USAGE.md

@@ -389,6 +389,27 @@ These are the models registered in the built-in alias table with known token lim
 
 Any model name that does not match an alias is passed through verbatim after provider routing is resolved. This is how you use OpenRouter model slugs (`openai/gpt-4.1-mini` with a custom `OPENAI_BASE_URL`), Ollama tags (`llama3.2` or `qwen2.5-coder:7b`), slash-containing local IDs (`local/Qwen/Qwen3.6-27B-FP8`), or full Anthropic model IDs (`claude-sonnet-4-20250514`).
 
+### Output token limit
+
+Set `plugins.maxOutputTokens` in a settings file, such as
+`.claw/settings.local.json`, to override the output token allowance sent by the
+`claw` CLI:
+
+```json
+{
+  "plugins": {
+    "maxOutputTokens": 1024
+  }
+}
+```
+
+The CLI loads this setting when constructing the conversation client and applies
+it to each model request, including requests after tool results. When omitted,
+the existing per-model defaults remain in effect. Choose a positive limit that
+fits the server's output allowance and leaves room for the input within its
+context window. This setting does not automatically query server limits and
+does not configure separate sub-agent clients or the `claw-analog` binary.
+
 ### User-defined aliases
 
 You can add custom aliases in any settings file (`~/.claw/settings.json`, `.claw/settings.json`, or `.claw/settings.local.json`):

rust/crates/rusty-claude-cli/src/main.rs

@@ -7391,6 +7391,7 @@ struct PromptHistoryEntry {
 
 struct RuntimePluginState {
     feature_config: runtime::RuntimeFeatureConfig,
+    max_output_tokens: Option<u32>,
     tool_registry: GlobalToolRegistry,
     plugin_registry: PluginRegistry,
     mcp_state: Option<Arc<Mutex<RuntimeMcpState>>>,
@@ -12467,6 +12468,7 @@ fn build_runtime_plugin_state_with_loader(
         .with_runtime_tools(runtime_tools)?;
     Ok(RuntimePluginState {
         feature_config,
+        max_output_tokens: runtime_config.plugins().max_output_tokens(),
         tool_registry,
         plugin_registry,
         mcp_state,
@@ -12896,6 +12898,7 @@ fn build_runtime_with_plugin_state(
     }
     let RuntimePluginState {
         feature_config,
+        max_output_tokens,
         tool_registry,
         plugin_registry,
         mcp_state,
@@ -12916,6 +12919,7 @@ fn build_runtime_with_plugin_state(
             tool_registry.clone(),
             progress_reporter,
             provider_config,
+            max_output_tokens,
         )?,
         CliToolExecutor::new(
             allowed_tools.clone(),
@@ -13189,6 +13193,7 @@ struct AnthropicRuntimeClient {
     client: ApiProviderClient,
     session_id: String,
     model: String,
+    max_output_tokens: Option<u32>,
     enable_tools: bool,
     emit_output: bool,
     terminal_tail: TerminalTail,
@@ -13212,6 +13217,7 @@ impl AnthropicRuntimeClient {
         tool_registry: GlobalToolRegistry,
         progress_reporter: Option<InternalPromptProgressReporter>,
         provider_config: Option<api::ProviderConfig>,
+        max_output_tokens: Option<u32>,
     ) -> Result<Self, Box<dyn std::error::Error>> {
         // Explicit persisted provider configuration takes precedence.
         // When absent, preserve the existing model-based provider detection.
@@ -13277,6 +13283,7 @@ impl AnthropicRuntimeClient {
             client,
             session_id: session_id.to_string(),
             model,
+            max_output_tokens,
             enable_tools,
             emit_output,
             terminal_tail,
@@ -13326,7 +13333,10 @@ impl ApiClient for AnthropicRuntimeClient {
         let enable_local_tools = self.enable_tools && model_supports_local_tools(&self.model);
         let message_request = MessageRequest {
             model: self.model.clone(),
-            max_tokens: max_tokens_for_model(&self.model),
+            max_tokens: api::max_tokens_for_model_with_override(
+                &self.model,
+                self.max_output_tokens,
+            ),
             messages: convert_messages(&request.messages),
             system: (!request.system_prompt.is_empty()).then(|| request.system_prompt.join("\n\n")),
             tools: enable_local_tools
@@ -18289,6 +18299,7 @@ mod tests {
 
     #[test]
     fn session_exists_resume_command_reports_json_contract() {
+        let _cwd_guard = cwd_guard();
         let session = Session::new();
         let path = PathBuf::from("missing-session.jsonl");
         let outcome = run_resume_command(
@@ -18811,6 +18822,8 @@ mod tests {
     #[test]
     fn boot_preflight_snapshot_reports_machine_readable_contract_fields() {
         let _guard = env_lock();
+        // Binary availability probes inherit cwd even though Git probes use workspace.
+        let _cwd_guard = cwd_guard();
         let workspace = temp_workspace("boot-preflight-json");
         fs::create_dir_all(&workspace).expect("workspace should create");
         git(&["init", "--quiet"], &workspace);
@@ -19084,6 +19097,8 @@ UU conflicted.rs",
 
     #[test]
     fn status_context_reads_real_workspace_metadata() {
+        // Reading cwd and launching helpers must serialize with cwd-changing tests.
+        let _cwd_guard = cwd_guard();
         let context = status_context(None).expect("status context should load");
         assert!(context.cwd.is_absolute());
         assert!(context.discovered_config_files >= context.loaded_config_files);
@@ -20231,6 +20246,7 @@ UU conflicted.rs",
             GlobalToolRegistry::builtin(),
             None,
             Some(config),
+            None,
         )
         .expect("runtime client should construct from explicit model provider hint");
 
@@ -20329,6 +20345,7 @@ UU conflicted.rs",
             GlobalToolRegistry::builtin(),
             None,
             Some(provider_config),
+            state.max_output_tokens,
         )
         .expect("runtime client should construct from persisted provider configuration");
 
@@ -20346,6 +20363,9 @@ UU conflicted.rs",
     #[test]
     #[allow(clippy::too_many_lines)]
     fn build_runtime_plugin_state_discovers_mcp_tools_and_surfaces_pending_servers() {
+        // MCP children inherit the process cwd. Keep it stable until shutdown so
+        // they cannot hold another test's temporary directory open on Windows.
+        let _cwd_guard = cwd_guard();
         let config_home = temp_dir();
         let workspace = temp_dir();
         fs::create_dir_all(&config_home).expect("config home");