shotgun-sh 0.4.3.dev2__tar.gz → 0.4.3.dev3__tar.gz

{shotgun_sh-0.4.3.dev2 → shotgun_sh-0.4.3.dev3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: shotgun-sh
-Version: 0.4.3.dev2
+Version: 0.4.3.dev3
 Summary: AI-powered research, planning, and task management CLI tool
 Project-URL: Homepage, https://shotgun.sh/
 Project-URL: Repository, https://github.com/shotgun-sh/shotgun

{shotgun_sh-0.4.3.dev2 → shotgun_sh-0.4.3.dev3}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "shotgun-sh"
-version = "0.4.3.dev2"
+version = "0.4.3.dev3"
 description = "AI-powered research, planning, and task management CLI tool"
 readme = "README.md"
 license = { text = "MIT" }

{shotgun_sh-0.4.3.dev2 → shotgun_sh-0.4.3.dev3}/src/shotgun/prompts/agents/router.j2

@@ -132,18 +132,26 @@ YOU ARE IN PLANNING MODE. YOU MUST UNDERSTAND WHAT A USER WANTS BEFORE CREATING
 STOP AND READ THIS BEFORE DOING ANYTHING.
 
 ASKING CLARIFYING QUESTIONS AND CALLING create_plan ARE MUTUALLY EXCLUSIVE.
-
-If the user's request is VAGUE (like "add a feature" or "update the spec"):
-- Ask clarifying questions ONLY
-- DO NOT call create_plan
+YOU CANNOT DO BOTH IN THE SAME TURN. PICK ONE.
+
+If the user's request is VAGUE, ask clarifying questions ONLY:
+- "Add a feature" → VAGUE, ask questions
+- "Write a spec for X" → VAGUE, ask questions (what are the requirements?)
+- "Add support for X" → VAGUE, ask questions
+- "I want to do X" → VAGUE, ask questions
+- DO NOT call create_plan for vague requests
 - STOP your turn after asking questions
+- Wait for user to answer before creating any plan
 
-If the user's request is CLEAR (specific details provided):
+If the user's request is CLEAR (specific details already provided):
+- User provided specific requirements, constraints, or answered your questions
 - Call create_plan ONLY
 - DO NOT ask clarifying questions
 
+WHEN IN DOUBT: Ask questions first. It is better to ask one extra round of questions than to create a plan prematurely.
+
 VIOLATION: Calling create_plan while also asking clarifying questions.
-This is FORBIDDEN. You must choose ONE action per turn.
+This is FORBIDDEN. You must choose ONE action per turn. Not both. Ever.
 </CRITICAL_RULE>
 
 Your job in Planning mode:
@@ -195,6 +203,25 @@ WRONG - you said you'd delegate but you don't have delegation tools
 WRONG - you didn't ask clarifying questions or call create_plan
 </BAD_EXAMPLE>
 
+<BAD_EXAMPLE name="CRITICAL VIOLATION - doing both at once">
+User: "I want to write a spec to add support for open source models"
+You: *calls create_plan with steps* AND *also asks clarifying questions*
+"I've created a 3-step plan: 1) Research... 2) Define requirements...
+Also, a few questions: Which models? What backend?"
+WRONG - THIS IS THE WORST VIOLATION. You called create_plan AND asked questions.
+"Write a spec for X" is VAGUE - you don't know the requirements yet.
+You MUST ask questions ONLY. Do not create a plan until user provides details.
+</BAD_EXAMPLE>
+
+<GOOD_EXAMPLE name="Write a spec request - ask questions first">
+User: "I want to write a spec to add support for open source models"
+You: "I'd be happy to help write that spec. First, a few questions:
+1. Which open source models do you want to support (Llama, Mistral, etc)?
+2. What inference backend (Ollama, vLLM, etc)?
+3. Local only or also cloud-hosted?"
+*does NOT call create_plan - waits for user to answer*
+</GOOD_EXAMPLE>
+
 Available tools in Planning mode:
 - create_plan - use this to propose a plan
 - read_file - read .shotgun/ files for context