Build Task

Implement a task autonomously through the full development lifecycle.

Input

$ARGUMENTS

Formats:

• (empty) → Auto-detect most recent docs/plans/*.md
• docs/plans/*.md → Blueprint file from Skill("blueprint")
• #123, PROJ-123, URL → Task reference (GitHub, Beads, Linear)
• --no-ship → Flag to skip shipping (used by run-blueprint for batch shipping)

Step 0: Detect Input Type

# If empty, find most recent plan
ls -t docs/plans/*.md 2>/dev/null | head -1

Parse flags:

• --no-ship: Set skip_ship=true, remove flag
• --ac "<criteria>": Set ac_items to provided criteria, remove flag

Input	Source	Skip Steps
Empty + plan found	plan	1, 2, 5
*.md file path	plan	1, 2, 5
Task reference	task	(none)
Empty + no plan	Ask user	—

For plans: Extract TASK_ID from filename slug, TITLE from first # heading.

---

Workflow (Context Gathering)

1. Fetch Task (task source only)

Task (majestic-engineer:workflow:task-fetcher):
  prompt: Task: <reference>

2. Claim Task (task source only)

Task (majestic-engineer:workflow:task-status-updater):
  prompt: Action: claim | Task: <ID>

3. Set Terminal Title

Run /rename <task-title> to set the terminal title for visibility.

4. Setup Workspace

Read config values:

WORKFLOW = config_read("workflow", "branches")
BRANCH_NAMING = config_read("branch_naming", "issue-desc")
POST_CREATE = config_read("workspace_setup.post_create", "")

DEFAULT_BRANCH=$(git remote show origin | grep 'HEAD branch' | awk '{print $NF}')

Then setup workspace:

Task (majestic-engineer:workflow:workspace-setup):
  prompt: |
    Task ID: <ID>
    Title: <title>
    Type: <type>
    Workflow: <workflow>
    Branch Naming: <branch_naming>
    Default Branch: <default_branch>
    Post-Create Hook: <post_create>

5. Verify Branch (MANDATORY)

After workspace setup, verify we are NOT on a protected branch:

CURRENT_BRANCH=$(git branch --show-current)

Current Branch	Action
main	STOP - workspace setup failed
master	STOP - workspace setup failed
<default_branch>	STOP - workspace setup failed
Feature branch	Continue

If on protected branch: STOP and report error. Do not proceed.

6. Resolve Toolbox

Task (majestic-engineer:workflow:toolbox-resolver):
  prompt: |
    Stage: build-task
    Task Title: <title>
    Task Description: <description>

Stores: methodology, build_agent, fix_agent, coding_styles, design_system_path, research_hooks, pre_ship_hooks, quality_gate.reviewers

7. Auto Research (if triggers match)

For each mode: auto hook where triggers match task text:

Task (majestic-engineer:workflow:context-proxy):
  prompt: agent: <hook.agent> | budget: 2000 | prompt: Research for: <title> | Context: <description>

8. Context Check (Post-Research)

If research agents returned outputs and combined output > 4000 chars:

• Summarize research findings before planning (KEEP task title, description, key patterns; DISCARD verbose tool outputs)
• Preserve task title, description, and key patterns

9. Plan (task source only)

Task (majestic-engineer:workflow:context-proxy):
  prompt: agent: architect | budget: 3000 | prompt: Task: <title> | Description: <description> | Research: <findings>

Note: Skip if source is plan - use plan file content instead.

---

Delegate to Build Workflow Manager

Pass all gathered context to the build-task-workflow-manager agent:

agent build-task-workflow-manager "
Task ID: <ID or 'plan'>
Title: <title>
Branch: <branch>
Plan: <plan content>
Acceptance Criteria:
  <ac_items>
Methodology: <methodology>
Build Agent: <build_agent>
Fix Agent: <fix_agent>
Coding Styles: <styles>
Design System Path: <path>
Pre-Ship Hooks: <hooks>
Quality Gate Reviewers: <reviewers>
Source: <task or plan>
Skip Ship: <skip_ship>
"

AC source:

Input	AC Source
--ac flag provided	Use provided AC directly
Plan file	Extract from **Acceptance Criteria:** section
GitHub Issue	Extract from issue body (look for AC section)
Linear/Beads	Extract from task description

The agent handles:

1. Loading design system (if configured)
2. Activating coding style skills
3. Building the implementation
4. Slop removal (MANDATORY)
5. AC Verification (MANDATORY)
6. Quality gate (MANDATORY)
7. Fix loop (if needed, max 3 attempts)
8. Capture learnings
9. Pre-ship hooks
10. Shipping (PR creation)
11. Task completion (if task source)

Agent returns:

• AC verification results (which criteria passed/failed)
• Learnings discovered
• Status (PASS or FAIL)

Caller responsibility: If caller needs to persist results (e.g., update blueprint checkboxes), it handles that based on the returned results.

---

Output

Task source:

## Build Complete: #<ID> - <title>
- Backend: <github|beads|linear>
- Branch: <branch>
- PR: #<number>
- Quality: Passed (attempt <n>)
- Next: PR awaits review, task closes on merge

Plan source:

## Build Complete: <title>
- Plan: <file-path>
- Branch: <branch>
- PR: #<number>
- Quality: Passed (attempt <n>)
- Next: PR awaits review

---

Examples

Skill("build-task")                           # Auto-detect recent plan
Skill("build-task") docs/plans/add-auth.md    # Explicit plan
Skill("build-task") #42                       # GitHub issue
Skill("build-task") PROJ-123                  # Beads/Linear

---

Notes

• This command gathers context (steps 1-9), then delegates execution
• The workflow manager ensures no build/verify/ship steps are skipped
• Branch safety check prevents accidental commits to main/master
• Research agents run conditionally based on toolbox triggers