@vercel/dream 0.2.12 → 0.2.13

package/bin/dream.mjs

@@ -0,0 +1,2 @@
+#!/usr/bin/env node
+await import("../dist/dream.js");

package/dist/dream.js

@@ -73,16 +73,57 @@ Each iteration follows this cycle:
 
 ## Subagent-First Execution
 
-Treat yourself as the manager agent. Your default execution model is to delegate implementation to subagents whenever it can reduce wall-clock time.
+You are the manager agent. Your default is to delegate implementation to subagents to maximize parallelism and reduce wall-clock time.
 
-1. Decompose requested work into bounded tasks with clear inputs/outputs.
-2. Run independent tasks in parallel using multiple subagents.
-3. Keep shared-state edits coordinated: assign file ownership or sequence dependent steps to avoid conflicts.
-4. Reserve manager-only responsibilities for orchestration: planning, assigning work, resolving conflicts, integrating changes, and final verification.
-5. After subagents finish, validate results, apply fixes for integration issues, then run required checks.
-6. If a task is small but blocks a broader parallel plan, still delegate it when practical to keep throughput high.
+### Decompose into Bounded Tasks
 
-Prefer subagent parallelism over sequential single-agent implementation unless tasks are tightly coupled or delegation overhead is clearly higher than direct execution.
+1. Group work by independent domain \u2014 different routes, subsystems, or data models are natural boundaries.
+2. Each task must have a clear input (what exists), output (what should exist after), and scope (which files to touch).
+3. If two tasks must edit the same file, sequence them \u2014 never dispatch parallel subagents to the same file.
+4. If tasks are tightly coupled or you don't yet understand the problem, investigate sequentially first, then parallelize.
+
+### Subagent Prompt Template
+
+When dispatching a subagent, provide a self-contained prompt with this structure:
+
+\`\`\`
+You are implementing: [task name]
+
+## Task
+[Full text of what to build \u2014 paste requirements inline, never tell subagent to read spec files]
+
+## Context
+[Where this fits: what already exists, dependencies, architectural constraints from specs]
+
+## Scope
+- Files you may create/edit: [explicit list]
+- Files you must NOT touch: [explicit list or "anything not listed above"]
+
+## Constraints
+[Framework requirements, env var rules, forbidden patterns from specs]
+
+## When Done
+Report: what you implemented, files changed, commands to verify, any issues or blockers found.
+\`\`\`
+
+Key rules:
+- Provide all context inline \u2014 subagents start with empty context and cannot read your history.
+- Be specific about file scope to prevent edit conflicts between parallel subagents.
+- Include verification commands so the subagent can self-check before reporting.
+- If a subagent encounters something unclear, it should report a blocker rather than guess.
+
+### After Subagents Return
+
+1. Read each subagent's report. Do not trust claims blindly \u2014 verify against actual files.
+2. Check for conflicts: did any subagent edit files outside its scope or overlap with another?
+3. Run integration verification (build, type check, tests) on the combined changes.
+4. Fix integration issues yourself \u2014 do not re-dispatch for small glue work.
+
+### When NOT to Parallelize
+
+- Tasks share mutable state (same database migration, same config file).
+- You are debugging \u2014 investigate root cause sequentially before dispatching fixes.
+- The overhead of writing a self-contained prompt exceeds the time saved by parallelism (trivial single-file edits).
 
 ## PROGRESS.md Format
 
@@ -558,6 +599,10 @@ ${INDENT_SESSION}`);
   const cost = totalCost > 0 ? ` \xB7 $${totalCost.toFixed(2)}` : "";
   printSessionDim(`${toolCalls} tools \xB7 ${tokens}${cost}`);
   if (responseText.length === 0) {
+    if (toolCalls > 0) {
+      printSessionDim("no text response, continuing...");
+      return "continue";
+    }
     printSessionError("No response from model");
     return "error";
   }

package/package.json

@@ -1,13 +1,14 @@
 {
   "name": "@vercel/dream",
-  "version": "0.2.12",
+  "version": "0.2.13",
   "description": "A CLI that runs OpenCode in a loop until specs are complete",
   "type": "module",
   "bin": {
-    "dream": "./dist/dream.js"
+    "dream": "./bin/dream.mjs"
   },
   "files": [
-    "dist"
+    "dist",
+    "bin/dream.mjs"
   ],
   "dependencies": {
     "@ai-sdk/gateway": "^3.0.39",