@hatchway/cli 0.50.65 → 0.50.67

package/dist/index.js

@@ -1462,23 +1462,23 @@ var init_droid_strategy$1 = __esm$1({
 // src/lib/prompts.ts
 var CLAUDE_SYSTEM_PROMPT = `You are an elite coding assistant specialized in building visually stunning, production-ready JavaScript applications.
 
-## Platform Skills
+## Platform Skills (hatchway-platform plugin)
 
-You have access to platform skills that provide critical procedural knowledge. Load and follow these skills for EVERY task:
+You have platform skills from the \`hatchway-platform\` plugin. These are loaded via the skill system \u2014 invoke each one by name to read its full instructions.
 
-**Always load (required for every build):**
-- \`todo-workflow\` \u2014 You MUST use TodoWrite to track progress. Without it, users see no progress in the UI.
-- \`communication-style\` \u2014 Follow the Hatchway output formatting conventions.
-- \`build-verification\` \u2014 Use the fix-verify loop for all dependency and build errors.
-- \`context-awareness\` \u2014 Read existing code before modifying. Never write blind.
-- \`dependency-management\` \u2014 Install all dependencies upfront in a single operation.
+**BEFORE doing any work, load ALL 5 of these required skills:**
+1. \`hatchway-platform:todo-workflow\` \u2014 You MUST load this FIRST. It defines how to use TodoWrite for progress tracking. Without it, users see no progress in the UI.
+2. \`hatchway-platform:communication-style\` \u2014 Defines output formatting for the Hatchway platform.
+3. \`hatchway-platform:build-verification\` \u2014 Defines the fix-verify loop for dependency and build errors.
+4. \`hatchway-platform:context-awareness\` \u2014 Defines read-before-write discipline.
+5. \`hatchway-platform:dependency-management\` \u2014 Defines how to install all dependencies upfront.
 
-**Load when relevant:**
-- \`architectural-thinking\` \u2014 Load when starting a new feature or multi-file change.
-- \`design-excellence\` \u2014 Load when building or styling user-facing UI.
-- \`template-originality\` \u2014 Load only when building a new project from a template scaffold.
+**Also load these when the task involves them:**
+- \`hatchway-platform:architectural-thinking\` \u2014 Load for new features or multi-file changes.
+- \`hatchway-platform:design-excellence\` \u2014 Load when building or styling UI.
+- \`hatchway-platform:template-originality\` \u2014 Load when building from a template scaffold.
 
-Load each skill by reading its SKILL.md file, then follow its instructions throughout the task.
+Load each skill at the START of the task before writing any code. Follow the loaded skill instructions throughout the entire task.
 
 ## Plan Mode
 
@@ -4424,33 +4424,28 @@ function createNativeClaudeQuery(modelId = DEFAULT_CLAUDE_MODEL_ID, abortControl
         let messageCount = 0;
         let toolCallCount = 0;
         let textBlockCount = 0;
-        // Create the gen_ai.invoke_agent span using startSpanManual.
+        // Create the gen_ai.invoke_agent span as a child of the current active span.
         //
-        // Why startSpanManual and not startSpan?
-        // startSpan() takes a callback and ends the span when the callback returns.
-        // But this is an async generator — we can't yield from inside a callback.
-        // startSpanManual() makes the span active on the current scope AND gives us
-        // a handle to end it ourselves in the finally block.
+        // We use startInactiveSpan because this is an async generator — we can't use
+        // startSpan/startSpanManual (both require a callback, and yields can't cross
+        // callback boundaries). startInactiveSpan creates a span that inherits the
+        // parent from the current active span (build.runner, restored by engine.ts
+        // via Sentry.withActiveSpan).
         //
-        // Why this works now (it didn't before):
-        // engine.ts captures the parent build.runner span before creating the
-        // ReadableStream, then restores it via Sentry.withActiveSpan() inside the
-        // stream's start() callback. So when this generator runs, the build.runner
-        // span is the active parent, and our gen_ai.invoke_agent becomes its child.
-        // Tool spans created with startSpan() inside the loop become children of
-        // gen_ai.invoke_agent because it's the active span at that point.
-        const agentSpan = Sentry.startSpanManual({
+        // For tool spans, we use Sentry.withActiveSpan(agentSpan, ...) to temporarily
+        // make the agent span active so tool spans become its children.
+        const agentSpan = Sentry.startInactiveSpan({
             op: 'gen_ai.invoke_agent',
-            name: `Claude Agent (${modelId})`,
+            name: 'invoke_agent hatchway-builder',
             attributes: {
+                'gen_ai.operation.name': 'invoke_agent',
                 'gen_ai.agent.name': 'hatchway-builder',
                 'gen_ai.request.model': modelId,
-                'gen_ai.agent.input': finalPrompt.substring(0, 500),
-                'gen_ai.system_prompt.length': appendedSystemPrompt.length,
-                'gen_ai.agent.available_tools': JSON.stringify(['Bash', 'Read', 'Write', 'Edit', 'Glob', 'Grep', 'Task', 'TodoWrite', 'WebFetch']),
+                'gen_ai.request.messages': JSON.stringify([{ role: 'user', content: finalPrompt.substring(0, 1000) }]),
+                'gen_ai.request.available_tools': JSON.stringify(['Bash', 'Read', 'Write', 'Edit', 'Glob', 'Grep', 'Task', 'TodoWrite', 'WebFetch']
+                    .map(name => ({ name, type: 'function' }))),
             },
-        }, (span) => span // Return the span so we control its lifecycle
-        );
+        });
         try {
             // Stream messages directly from the SDK
             for await (const sdkMessage of query({ prompt: finalPrompt, options })) {
@@ -4465,20 +4460,21 @@ function createNativeClaudeQuery(modelId = DEFAULT_CLAUDE_MODEL_ID, abortControl
                                 toolCallCount++;
                                 debugLog$4(`[runner] [native-sdk] 🔧 Tool call: ${block.name}\n`);
                                 // Emit a gen_ai.execute_tool span as a child of gen_ai.invoke_agent.
-                                // Using startSpan (active) with an empty callback — the span is created,
-                                // becomes briefly active, records the tool invocation, and ends when
-                                // the callback returns. This gives Sentry the tool call event with
-                                // proper parent-child nesting.
-                                Sentry.startSpan({
-                                    op: 'gen_ai.execute_tool',
-                                    name: `Tool: ${block.name}`,
-                                    attributes: {
-                                        'gen_ai.tool.name': block.name,
-                                        'gen_ai.tool.call_id': block.id,
-                                        'gen_ai.tool.input': JSON.stringify(block.input).substring(0, 1000),
-                                    },
-                                }, () => {
-                                    // Span created and ended — marks the tool invocation point
+                                // withActiveSpan temporarily makes agentSpan the active span so
+                                // the startSpan inside creates a proper child.
+                                Sentry.withActiveSpan(agentSpan, () => {
+                                    Sentry.startSpan({
+                                        op: 'gen_ai.execute_tool',
+                                        name: `execute_tool ${block.name}`,
+                                        attributes: {
+                                            'gen_ai.tool.name': block.name,
+                                            'gen_ai.tool.call_id': block.id,
+                                            'gen_ai.tool.input': JSON.stringify(block.input).substring(0, 1000),
+                                            'gen_ai.request.model': modelId,
+                                        },
+                                    }, () => {
+                                        // Span created and ended — marks the tool invocation point
+                                    });
                                 });
                             }
                             else if (block.type === 'text') {
@@ -4499,6 +4495,11 @@ function createNativeClaudeQuery(modelId = DEFAULT_CLAUDE_MODEL_ID, abortControl
                         process.stderr.write(`[native-sdk] SDK init — plugins: ${JSON.stringify(loadedPlugins)}\n`);
                         process.stderr.write(`[native-sdk] SDK init — tools: ${toolCount} loaded\n`);
                     }
+                    // Set discovered skills on the agent span
+                    if (agentSpan) {
+                        agentSpan.setAttribute('gen_ai.agent.skills', discoveredSkills.join(', '));
+                        agentSpan.setAttribute('gen_ai.agent.skill_count', discoveredSkills.length);
+                    }
                     if (discoveredSkills.length > 0) {
                         Sentry.logger.info('SDK initialized with skills', {
                             skillCount: String(discoveredSkills.length),
@@ -4532,21 +4533,27 @@ function createNativeClaudeQuery(modelId = DEFAULT_CLAUDE_MODEL_ID, abortControl
                 if (sdkMessage.type === 'result') {
                     const resultMsg = sdkMessage;
                     if (agentSpan) {
+                        // Standard gen_ai token usage attributes (Sentry AI Agent Monitoring spec)
                         agentSpan.setAttribute('gen_ai.usage.input_tokens', resultMsg.usage?.input_tokens ?? 0);
                         agentSpan.setAttribute('gen_ai.usage.output_tokens', resultMsg.usage?.output_tokens ?? 0);
                         agentSpan.setAttribute('gen_ai.usage.total_tokens', (resultMsg.usage?.input_tokens ?? 0) + (resultMsg.usage?.output_tokens ?? 0));
-                        agentSpan.setAttribute('gen_ai.usage.cost_usd', resultMsg.total_cost_usd ?? 0);
-                        agentSpan.setAttribute('gen_ai.agent.num_turns', resultMsg.num_turns ?? 0);
-                        agentSpan.setAttribute('gen_ai.agent.num_tool_calls', toolCallCount);
-                        agentSpan.setAttribute('gen_ai.agent.result', resultMsg.subtype ?? 'unknown');
-                        agentSpan.setAttribute('gen_ai.agent.duration_ms', resultMsg.duration_ms ?? 0);
-                        agentSpan.setAttribute('gen_ai.agent.duration_api_ms', resultMsg.duration_api_ms ?? 0);
                         if (resultMsg.usage?.cache_read_input_tokens) {
-                            agentSpan.setAttribute('gen_ai.usage.cache_read_tokens', resultMsg.usage.cache_read_input_tokens);
+                            agentSpan.setAttribute('gen_ai.usage.input_tokens.cached', resultMsg.usage.cache_read_input_tokens);
                         }
                         if (resultMsg.usage?.cache_creation_input_tokens) {
-                            agentSpan.setAttribute('gen_ai.usage.cache_creation_tokens', resultMsg.usage.cache_creation_input_tokens);
+                            agentSpan.setAttribute('gen_ai.usage.input_tokens.cache_write', resultMsg.usage.cache_creation_input_tokens);
+                        }
+                        // Response text (truncated for span safety)
+                        if (resultMsg.result) {
+                            agentSpan.setAttribute('gen_ai.response.text', JSON.stringify(resultMsg.result.substring(0, 1000)));
                         }
+                        // Custom (non-spec) attributes for operational insight
+                        agentSpan.setAttribute('hatchway.cost_usd', resultMsg.total_cost_usd ?? 0);
+                        agentSpan.setAttribute('hatchway.num_turns', resultMsg.num_turns ?? 0);
+                        agentSpan.setAttribute('hatchway.num_tool_calls', toolCallCount);
+                        agentSpan.setAttribute('hatchway.result', resultMsg.subtype ?? 'unknown');
+                        agentSpan.setAttribute('hatchway.duration_ms', resultMsg.duration_ms ?? 0);
+                        agentSpan.setAttribute('hatchway.duration_api_ms', resultMsg.duration_api_ms ?? 0);
                     }
                     if (resultMsg.subtype === 'success') {
                         debugLog$4(`[runner] [native-sdk] ✅ Query complete - ${resultMsg.num_turns} turns, $${resultMsg.total_cost_usd?.toFixed(4)} USD\n`);