| name | scrum-master |
| description | Convert execution plans into a delivery-ready scrum backlog with atomic tasks, relative weights, dependency trees, and agent-ready prompts. Use when a user has a structured plan (especially claude-planner JSON) and needs work packets that can be distributed across Codex, Claude, Gemini, or other agents. |
Scrum Master
Turn structured plans into assignable work. The generated backlog is deterministic and machine-readable so downstream executors can ingest it directly.
Workflow
- Start from a plan:
- Preferred input: claude-planner handoff object (
{"type":"plan","planner":"claude","plan":{...}})
- Also supported: raw plan object with the same fields.
- Run the backlog builder:
scripts/build_backlog.sh \
--input-file /path/to/plan.json \
--pretty
- Distribute by
agent_packets:
- Each packet includes task metadata + a ready-to-send prompt per agent.
Output Contract
The script emits:
tasks: atomic units with task_id, weight, dependencies, milestone context, and acceptance criteria.
dependency_tree: roots, directed edges, and Mermaid graph.
agent_packets: prompt bundles keyed by agent (codex, claude, gemini, or custom list).
See references/backlog-schema.json for the full JSON structure.
Decomposition Rules
- Split each plan step into atomic actions when the action contains sequencing phrases (
then, and then, ;).
- Preserve ordering inside each milestone.
- Link milestones in sequence (first task of milestone N depends on final task of milestone N-1).
- Add explicit dependencies when step inputs reference earlier step IDs.
Weighting Model
Weights are relative effort scores (1-13) based on:
- Action length/complexity.
- Number of inputs and outputs.
- Risk count.
- Dependency count.
Use weights for agent load balancing, sprint slicing, and prioritization.
Prompt Generation
Each prompt packet includes:
- Goal and milestone context.
- Exact task action.
- Inputs, expected outputs, risks.
- Dependency prerequisites.
- Definition of done and expected response format.
Use these prompts directly when launching agent work sessions.
Files
scripts/scrum_master.py: convert plan JSON to backlog JSON.
scripts/build_backlog.sh: shell wrapper.
references/backlog-schema.json: expected output schema.