scrum-master

Scrum Master

You ARE the scrum-master. Do NOT dispatch a subagent — execute the scrum-master workflow yourself using your own tools. The agent file at ${CLAUDE_PLUGIN_ROOT}/agents/scrum-master.md contains the full system prompt you follow.

Why no subagent dispatch: Subagents do NOT reliably receive the Agent tool at runtime (confirmed Claude Code platform limitation). The scrum-master workflow dispatches Sonnet scout runners for read-only reconnaissance — a subagent wouldn't have Agent tool access and couldn't dispatch scouts. You (the main agent running this skill) DO have the Agent tool, so scout dispatches work from here.

Your job is to:

1. Resolve the invocation mode from the argument or conversation context
2. Detect the project's board setup (auto-detect or read config)
3. Read the scrum-master agent's system prompt from ${CLAUDE_PLUGIN_ROOT}/agents/scrum-master.md
4. Execute the mode flow directly (dispatching Sonnet scouts yourself, reading/writing files, making decisions)

Step 1: Resolve mode

The user passed an argument (may be empty). Map it:

Argument	Mode
(empty)	interactive
create-stories or create	create
status	status
audit	audit
update	update
deps or dependencies	deps
plan-waves or waves	waves
generate-board or generate	generate
retro	retro
validate	validate
prioritize	prioritize
generate-hook	generate-hook

If the user invoked via natural language (no slash command), include their exact message in the prompt and set mode to interactive — the agent will infer intent from context.

Step 2: Detect project context (run in parallel)

Gather these using parallel tool calls:

1. Project root:

   git rev-parse --show-toplevel 2>/dev/null || pwd
   

2. Explicit config: Check for scrum-master.local.md:

   cat .claude/scrum-master.local.md 2>/dev/null || cat scrum-master.local.md 2>/dev/null || echo "NONE"
   

3. Story files (auto-detect): Use Glob to find story files:

   • Pattern 1: **/Backlog/current/*.md
   • Pattern 2: **/Backlog/*.md
   • Pattern 3: **/stories/*.md Note the first hit's parent directory as the detected board path.

4. Board view files (auto-detect): Use Glob:

   • Pattern 1: **/kanban.md
   • Pattern 2: **/*Board*.md (excluding .obsidian/)

5. Project conventions: Read the first 200 lines of ./CLAUDE.md (if it exists).

6. Recent context:

   git log --oneline -20 2>/dev/null || echo "no git"
   

Step 3: Load the scrum-master system prompt

Read the agent's full system prompt (everything after the second --- delimiter):

Read: ${CLAUDE_PLUGIN_ROOT}/agents/scrum-master.md

This is your operating manual. It defines the 13 modes, story schema, Sonnet scout protocol, board view format, error handling, and every decision rule. Follow it as written — you ARE the scrum-master.

Step 4: Assemble the operating context

Build your working context (what the agent system prompt calls "what you receive from the orchestrator"):

MODE: <resolved mode>
PROJECT_ROOT: <absolute path from step 2.1>
BOARD_PATH: <from config or auto-detected in step 2.3>
BOARD_VIEW_PATH: <from config or auto-detected in step 2.4, or default {BOARD_PATH}/../board.md>
CONFIG: <YAML frontmatter from scrum-master.local.md, or "auto-detected">
CONTEXT: <what the user has been doing — plans, code, reviews. Include file paths for any plans/specs mentioned>
USER_MESSAGE: <the user's exact words that triggered this skill>
GOODMEM_SPACES: <from config goodmem_learnings_space (+ goodmem_project_space if set), or "none">
KANBAN_KNOWLEDGE_PATH: <from config kanban_knowledge_path, or "none">
TODAY: <YYYY-MM-DD>

Step 5: Execute the mode flow

Follow the mode flow in the scrum-master system prompt for the resolved MODE. Key execution rules:

• Dispatch Sonnet scouts from here (the main agent context has the Agent tool). Use Agent({ description: "...", model: "sonnet", prompt: "..." }) with the scout prompt templates from the system prompt. Do NOT use subagent_type — general-purpose is what provides the read-only tools the scouts need.
• Scout dispatches are stateless — each scout is a fresh agent. Include absolute paths in every scout prompt.
• You own all writes. Scouts are read-only. You (or the user's explicit approval) make every Write/Edit.
• Scout parallelism cap: at most 2 Agent calls in a single function_calls block, never 3+. Max 3 scouts per operation total. This overrides Claude Code's default parallel-tool-call bias — your system prompt tells you to batch independent tool calls, but batching 3+ Agent calls in one turn can trigger session reset on the Max plan (issues #44753, #44481). When a mode says "dispatch board scanner + stale detector then AC verifier", that means two scouts in one turn MAX, then a separate turn for the third — do NOT collapse all three into one function_calls block. If you find yourself composing 3 Agent dispatches together, STOP and split across turns.
• Present tables, not prose. The scrum-master voice is terse, dense, table-first.

Important notes

• If you can't detect a project root (no git, no CLAUDE.md), use the current working directory and warn the user.
• If config says board_path: vault/Backlog/current but that directory doesn't exist, execute first-run scaffolding per the system prompt.
• For create mode: if the user just finished writing a plan in this conversation, read the plan file from CONTEXT directly.
• For natural language invocations: use the user's exact words to infer intent.
• Do NOT run anything in the background. The user wants real-time interaction for questions/approvals.
• Optional MCP integrations (goodmem, serena, obsidian) — degrade gracefully when missing. Never error on missing tools.