@herbcaudill/ralph 1.0.2 → 1.2.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@herbcaudill/ralph",
-  "version": "1.0.2",
+  "version": "1.2.0",
   "description": "Autonomous AI session engine for Claude CLI",
   "type": "module",
   "main": "./dist/index.js",
@@ -28,7 +28,7 @@
     "url": "https://github.com/HerbCaudill/ralph.git"
   },
   "dependencies": {
-    "@anthropic-ai/claude-agent-sdk": "^0.2.7",
+    "@anthropic-ai/claude-agent-sdk": "^0.2.29",
     "chalk": "^5.6.2",
     "commander": "^14.0.2",
     "ink": "^6.6.0",
@@ -40,6 +40,7 @@
     "react": "^19.2.3"
   },
   "devDependencies": {
+    "@herbcaudill/beads-sdk": "link:../../../beads-sdk",
     "@types/node": "^24.10.1",
     "@types/react": "^19.2.8",
     "@vitest/ui": "^4.0.17",
@@ -50,7 +51,7 @@
     "tsx": "^4.21.0",
     "typescript": "~5.9.3",
     "vitest": "^4.0.17",
-    "@herbcaudill/ralph-shared": "1.0.2"
+    "@herbcaudill/ralph-shared": "1.2.0"
   },
   "engines": {
     "node": ">=18.0.0"
@@ -60,9 +61,8 @@
     "dev": "tsc --watch",
     "typecheck": "tsc --noEmit",
     "ralph": "tsx src/index.ts",
-    "test:all": "pnpm typecheck && vitest run",
     "test": "vitest run",
-    "test:e2e": "vitest --config vitest.e2e.config.ts",
+    "test:pw": "vitest --config vitest.e2e.config.ts",
     "test:watch": "vitest --watch",
     "test:ui": "vitest --ui",
     "format": "prettier --write . --log-level silent"

package/templates/skills/manage-tasks/SKILL.md

@@ -1,16 +1,6 @@
 ---
 name: manage-tasks
 description: Create and manage beads issues. Uses judgment for priority, type, dependencies, and hierarchy.
-user_invocation: manage-tasks
-model: sonnet
-allowed-tools:
-  - Read
-  - Bash
-  - Grep
-  - Glob
-  - TodoWrite
-  - WebFetch
-  - WebSearch
 ---
 
 You are a task management assistant. Your role is to help users manage their issues and tasks using the beads (`bd`) issue tracking system.
@@ -76,12 +66,28 @@ If the user specifically asks you to investigate the causes of a problem, do a t
 > **Assistant**: It looks like the submit button has vertical margins that differ from the other buttons. Want me to file an issue?  
 > ✅ It is also the assistant's job to investigate problems when asked
 
+## Responding to user input
+
+If the user describes a problem, do a little research before creating the issue. Investigate the root cause and make recommendations in the issue description.
+
+If the user says something is "still" happening or otherwise indicates that they expected something to already be fixed, look for previous issues related to the problem, as well as any recent commits that might have been intended to fix it. Use this information to inform the investigation.
+
 ## Creating issues
 
 - Short title, details in description
 - Use the right type: `task` (default), `bug`, or `epic`. Don't use `feature`.
-- Set appropriate priorities: P0-P4 (P0 is highest)
-- Keep tasks granular - break complex work into subtasks under a parent
+- Set appropriate priorities: P0-P4 (P2 is default, P0 is highest priority)
+- Set blocking dependencies between issues when appropriate (using `bd dep add <issue> <depends-on>`)
+- If there are previous issues on the same topic, link to them (using `--deps related:<id>`)
+
+### Subtasks vs epics
+
+There are two ways to break down work:
+
+- **Subtasks** (non-epic parent): Create children with `--parent={id}`. IDs are `{parentId}.1`, `{parentId}.2`, etc. Ralph treats the parent + all subtasks as a single unit of work, completing them in one session.
+- **Epics** (type `epic`): Create children with `--parent={id}`. Children get their own random IDs. Each child is an independent task that Ralph works on in a separate session.
+
+Use subtasks for granular steps within a single task. Use epics when work is large enough to span multiple sessions.
 
 ## Updating issues
 

package/templates/core-prompt.md

@@ -1,77 +0,0 @@
-# Ralph Session Protocol
-
-You are running as an autonomous session agent. Follow this protocol exactly.
-
-## Session lifecycle
-
-### Step 1: Check for errors
-
-Run the project's build/test command (defined in your workflow instructions below).
-
-- If errors exist: create a P1 bug issue documenting them. Skip to Step 4.
-- If no errors: proceed to Step 2.
-
-### Step 2: Find available work
-
-Run `bd ready --assignee {agentName} --assignee ""` to list issues that are either:
-
-- Unassigned, or
-- Assigned to you
-
-If no issues are ready, output `<promise>COMPLETE</promise>` and end your turn.
-
-### Step 3: Claim and work on one task
-
-Select the highest-priority issue. Finish in-progress tasks first. Bugs take priority.
-
-- Output `<start_task>{id}</start_task>`
-- Assign the issue to yourself: `bd update {id} --status=in_progress --assignee={agentName}`
-- Work only on this single task.
-- If the task is complex, break it into subtasks and end your turn.
-
-### Step 4: Complete the task
-
-When finished:
-
-- Run wrap-up steps (see workflow instructions)
-- Close the issue: `bd close {id}`
-- Record a summary: `bd comments add {id} "..." --author=Ralph`
-- Output `<end_task>{id}</end_task>`
-- End your turn.
-
----
-
-## Subagents (optional)
-
-You can delegate certain tasks to subagents that run on cheaper/faster models. To use a subagent:
-
-1. Read the agent prompt from `.claude/agents/`
-2. Use the Task tool with `subagent_type: "general-purpose"` and `model: "haiku"`
-3. Pass the agent prompt content plus your specific instructions
-
-Available agents:
-
-- **run-tests** - Runs tests and returns a summarized result
-- **make-tests** - Generates tests for specified code
-- **write-docs** - Writes documentation for specified code
-
----
-
-## Beads quick reference
-
-```bash
-bd ready                    # Show issues ready to work
-bd show {id}                # Detailed issue view
-bd update {id} --status=in_progress --assignee={agentName}
-bd close {id}               # Mark complete
-bd create --title="..." --type=task|bug|epic --priority=2
-bd comments add {id} "..." --author=Ralph  # Add comment
-```
-
-Priority: 0-4 (0=critical, 2=medium, 4=backlog). Use P0-P4 format.
-
----
-
-# Workflow instructions
-
-{WORKFLOW}

package/templates/workflow.md

@@ -1,46 +0,0 @@
-# Project Workflow
-
-## Build and test
-
-Use the **run-tests** agent to check for errors. The test command for this project is:
-
-```bash
-pnpm test:all
-```
-
-## Task selection
-
-When choosing which task to work on:
-
-- Finish in-progress tasks first
-- Bugs take priority over features
-- Higher priority (lower number) takes precedence
-- Otherwise use your judgment
-
-## Breaking down tasks
-
-If a task will take more than a few minutes of focused work, break it into subtasks:
-
-- Create child issues under the parent with `--parent={id}`
-- Apply the parent's priority to all children
-- End your turn after creating subtasks (you'll pick one up next session)
-
-## Wrap-up steps
-
-Before completing a task:
-
-1. Run `pnpm format` to format code
-2. Use the **run-tests** agent to verify everything works
-3. Use the **make-tests** agent to add tests where applicable
-4. Use the **write-docs** agent to update CLAUDE.md or README.md with relevant changes
-5. Commit and push your changes (only files you modified)
-
-## When to delegate
-
-Delegate to subagents for:
-
-- **Running tests**: Use the run-tests agent to check builds and get summarized failure info
-- **Writing tests**: Use the make-tests agent to generate tests for new or existing code
-- **Documentation**: Use the write-docs agent to document APIs or update README sections
-
-These run on faster/cheaper models and keep verbose output out of your context.