granary-orchestrate

Orchestrating Implementation with Granary

This skill is for implementation-time orchestration. Your job is to coordinate sub-agents to implement planned work. You do NOT plan or design—you delegate and track.

Critical: Trust the Scheduler

NEVER determine task parallelism by:

• Checking if blocked_by is null/empty on individual tasks
• Assuming tasks are independent because they don't have explicit blockers
• Launching more tasks than granary next --all returns
• Inspecting task metadata to make your own parallelism decisions

granary next --all is the ONLY source of truth for what can run in parallel. It considers:

• Explicit task dependencies (task_dependencies table)
• Project dependencies (project_dependencies table)
• Task status (draft tasks are never returned)
• Claims and leases
• Implicit ordering requirements

If granary next --all returns 2 tasks, spawn exactly 2 agents. Not 6. Not 4. Exactly 2.

Output Format

Always use --format=prompt instead of --json for granary commands. The --format=prompt output is specifically designed for LLM consumption—it's token-efficient and easy to parse without wasting context on JSON syntax.

Only use --json when you need to extract specific fields via jq, and always filter the output:

# Only if you need specific fields
granary task <id> --json | jq '.id, .status'

Primary Use-Case

You are prompted to implement a project. Follow these steps:

Step 1: Get Granary Overview

granary summary

This shows all projects, their status, and task counts. Understand what exists before proceeding.

Step 2: Find Your Target Project

If given a project name, find it:

granary projects
granary project <project-id>

Finding Projects and Tasks

Search for projects and tasks by title:

granary search "query"                # Find matching projects and tasks
granary search "api" --format=prompt  # LLM-friendly output

This is useful when you know part of the project or task name but not the exact ID.

Step 3: Assess Project Readiness

Examine the project structure:

granary project <project-id> tasks

Decision tree:

Situation	Action
Project has no tasks	Use /granary:plan-work skill to plan first
Project has tasks but they lack detail	Use /granary:plan-work skill to refine
Single simple project with clear description	Implement directly as one task
Project has tasks with dependencies	Happy path - proceed to Step 4

Step 4: Start Orchestration Session

granary session start "implementing-<project-name>" --mode execute
granary session add <project-id>

Step 5: The Orchestration Loop

Your sole focus: hand off tasks to sub-agents. You do NOT implement tasks yourself.

Key principle: Maximize parallelism. Always look for opportunities to run multiple tasks concurrently. Tasks without dependencies on each other can and should run in parallel.

Get ONLY the Tasks Granary Says Are Ready

# Get ONLY tasks the scheduler has determined are ready
granary next --all --format=prompt

This returns only the tasks that are ready to execute. The scheduler has already evaluated dependencies, project ordering, claims, and status. Spawn agents for exactly the tasks returned—no more, no fewer.

If no tasks are returned, either all tasks are complete or remaining tasks are blocked.

DO NOT bypass this by:

• Listing all project tasks and filtering by blocked_by == null
• Assuming tasks without explicit blockers can run in parallel
• Making your own judgment calls about task independence

The scheduler knows things you don't. Trust it.

Spawning Agents in Parallel

When multiple tasks are unblocked, spawn them all at once using parallel Task tool calls:

# If granary next --all returns task-1, task-2, task-3:

Use Task tool (call all three in a SINGLE message):
  1. prompt: "Execute granary task task-1. Use /granary:execute-task skill."
     subagent_type: "general-purpose"
     run_in_background: true

  2. prompt: "Execute granary task task-2. Use /granary:execute-task skill."
     subagent_type: "general-purpose"
     run_in_background: true

  3. prompt: "Execute granary task task-3. Use /granary:execute-task skill."
     subagent_type: "general-purpose"
     run_in_background: true

Important: Use run_in_background: true for parallel execution. This allows multiple agents to work simultaneously.

The Loop

1. Ask the scheduler what's ready → granary next --all --format=prompt
2. Spawn agents for exactly those tasks → One Task tool call per task, all in the same message
3. Monitor progress → Check on background agents, wait for completions
4. Repeat → When agents complete, ask the scheduler again for newly unblocked tasks

Critical: Never spawn more agents than tasks returned by granary next --all. If it returns 2 tasks, spawn 2 agents. Period.

Single Task Fallback

If only one task is available or tasks must be sequential:

granary next --format=prompt

Then spawn a single agent and wait for completion before checking for the next task.

What Sub-Agents Do

Each sub-agent is responsible for:

• Building context (granary handoff --tasks <id>)
• Starting the task (granary task <id> start)
• Doing the actual implementation
• Marking the task done (granary task <id> done) or blocked

Note: You do NOT need to run granary handoff yourself. Sub-agents handle their own context building. Your job is to identify unblocked tasks and spawn agents.

Step 6: Handle Completion

When all tasks are done:

granary session close --summary "Completed implementation of <project-name>"

Handling Edge Cases

Sub-Agent Gets Blocked

If a sub-agent cannot complete:

# Sub-agent should have blocked the task:
granary task <task-id> block --reason "..."

# You see it when checking next task
granary next --format=prompt  # Will skip blocked tasks

Conflict Prevention

Sub-agents should claim tasks with leases:

granary task <task-id> claim --owner "Worker-1" --lease 30
# Exit code 4 = conflict (task claimed by another)

Checkpointing

Before risky operations:

granary checkpoint create "before-major-change"

# If things go wrong
granary checkpoint restore before-major-change

Common Mistakes

Inspecting tasks to determine parallelism

# WRONG - Don't do this
granary project <id> tasks --json | jq '.[] | select(.blocked_by == null)'

This bypasses the scheduler and leads to race conditions. Just because a task has no explicit blocked_by value doesn't mean it's ready to run. The scheduler considers implicit dependencies, project ordering, and other factors you can't see in task metadata.

Assuming task independence

# WRONG - Don't assume tasks can run together
granary show task-1  # blocked_by: null
granary show task-2  # blocked_by: null
granary show task-3  # blocked_by: null
# "All three have no blockers, I'll run them all!"

Tasks may have implicit logical dependencies. Task-3 might implement a service that uses schemas from task-1. Running them in parallel causes compiler errors.

Overriding scheduler decisions

# WRONG - Don't second-guess the scheduler
granary next --all  # Returns 2 tasks
# "But I checked and 6 tasks have no blockers, I'll launch 6 agents"

If granary next --all returns 2 tasks, there's a reason. Trust it.

The right approach

# RIGHT - Trust the scheduler completely
granary next --all --format=prompt
# Returns task-1, task-2
# Spawn exactly 2 agents, one for each task

Only spawn agents for tasks returned by granary next --all. When they complete, run the command again to discover what's newly unblocked.

Summary

1. Get overview → granary summary
2. Find project → granary projects
3. Assess readiness → Are tasks planned? If not, use plan-work skill
4. Start session → granary session start
5. Loop: delegate tasks → Use granary next --all, spawn agents for exactly the tasks returned
6. Close session → granary session close

Your job is coordination, not implementation. Sub-agents use granary handoff to build their own context and do the actual work—you just spawn agents for the tasks that granary next --all returns.

Remember: granary next --all is the scheduler. Trust it completely. If it returns 2 tasks, spawn 2 agents. Don't inspect task metadata to make your own parallelism decisions—that path leads to race conditions and compiler errors.