/ultrawork

Ultrawork — Autonomous Execution

Execute the following task autonomously from start to finish. Do not stop until the task is complete and verified.

Task: {{arguments}}

Workflow

Step 1: Intent Gate

Classify this task:

• Bug fix: Focus on diagnosis first, then minimal targeted fix
• Feature: Implement with full verification
• Refactor: Ensure behavioral equivalence
• Migration: Plan rollback strategy before starting

Step 2: Explore (PARALLEL, BACKGROUND)

Launch 2-3 explorer subagents with run_in_background: true to map the relevant codebase while you plan:

• Explorer 1: Find all files related to the task subject
• Explorer 2: Find tests, configs, and dependencies
• Explorer 3: Find similar patterns/implementations elsewhere in the codebase

Continue to Step 3 while explorers run. Collect their results before Step 4.

Step 3: Research (BACKGROUND)

If external knowledge is needed (libraries, APIs, frameworks):

• Launch librarian agent with run_in_background: true for comprehensive documentation research
• Use context7 MCP for library docs, grep_app for GitHub code examples
• Continue planning while research runs

Step 4: Plan

Create a comprehensive todo list with TodoWrite:

• Break the task into atomic, verifiable steps
• Order by dependency
• Include verification steps (run tests, check types, etc.)

Step 5: Implement

For each todo item:

1. Mark it in_progress
2. Execute the implementation
3. Read every modified file to verify changes
4. Run relevant tests
5. Mark it completed

Delegation: For complex subtasks (touching 3+ files), delegate to a sisyphus-junior subagent using the 6-section format.

Step 6: Final Verification

• Run diagnostics: bash .claude/plugins/omo/scripts/diagnostics.sh
• Run the full test suite if available
• Read all modified files one last time
• Self-review: look for edge cases, regressions, inconsistencies

Step 7: Report

Summarize what was accomplished, files changed, and any follow-up items.

Failure Recovery

If you fail 3 times on the same step:

1. STOP the current approach
2. Revert unsuccessful changes
3. Launch the oracle agent for advice
4. If Oracle can't help, ask the user

Rules

• Do NOT stop until all todos are complete
• Do NOT skip verification steps
• Do NOT make changes beyond the scope of the task
• PREFER delegation for complex subtasks
• PARALLELIZE independent explorations