Implement Yolo – Autonomous Specification Implementation

Role: Implementation Loop Orchestrator

Orchestrate the complete implementation of a Speckit specification by executing /speckit.implement in automated cycles, committing at each checkpoint, and looping until all tasks are complete.

Capabilities

• Execute implementation in autonomous loop (yolo mode)
• Detect checkpoints in tasks.md by **Checkpoint** marker
• Auto-commit at each checkpoint with descriptive message
• Retry lint/type-check/test errors up to 3 times
• Progress tracking with iteration and task counts
• Clean context isolation between iterations

Requirements

• Plugin @openspoon/subtask2 installed for loop orchestration
• tasks.md must exist with checkpoints defined
• Active feature branch (numbered branch from Speckit)
• /speckit.implement command available

Workflow Overview

/implement-yolo
     │
     ├──► [INIT] Validate prerequisites
     │
     ├──► [SCAN] Parse tasks.md
     │         - Count pending tasks (- [ ])
     │         - Identify next checkpoint
     │         - Determine current phase
     │
     └──► [LOOP] while has_pending_tasks:
               │
               ├──► SUBTASK: /speckit.implement
               │    (implement until next checkpoint)
               │
               ├──► [VERIFY] Check result
               │    - lint/type/test error → retry (max 3x)
               │    - other error → STOP
               │    - success → continue
               │
               ├──► [COMMIT] Auto-commit checkpoint
               │
               └──► [NEXT] Re-scan and continue

Execution Steps

Step 0: Prerequisites Check

Run the prerequisites script to get FEATURE_DIR:

.specify/scripts/bash/check-prerequisites.sh --json --require-tasks --include-tasks

Parse the JSON output for FEATURE_DIR and verify:

• tasks.md exists
• At least one task is pending (- [ ])
• Currently on a feature branch

On failure: Stop and report missing prerequisites.

Step 1: Parse tasks.md

Read {FEATURE_DIR}/tasks.md and extract:

1. Total tasks: Count all lines matching - [ ] or - [x] or - [X]
2. Pending tasks: Count lines matching - [ ]
3. Completed tasks: Count lines matching - [x] or - [X]
4. Next checkpoint: Find first **Checkpoint** line after any pending task
5. Current phase: Extract phase name from nearest ## Phase header

Checkpoint format in tasks.md:

**Checkpoint**: Description of what should be achieved ✅

Step 2: Execute Implementation Loop

For each iteration while pending tasks exist:

2.1 Launch Implementation Subtask

Use the Task tool to launch /speckit.implement with specific instructions:

Task(
  description="[Iteration N] speckit.implement until checkpoint",
  prompt="/speckit.implement Implement all tasks until the next checkpoint: '{checkpoint_description}'. Mark each completed task as [x] in tasks.md. STOP when you reach the checkpoint. Do NOT proceed past the checkpoint.",
  subagent_type="general"
)

CRITICAL: Each iteration runs as an independent subtask for clean context isolation.

2.2 Verify Result

After subtask completes, check for errors:

Recoverable errors (retry up to 3 times):

• Lint errors: output contains “ruff” or “lint” failures
• Type errors: output contains “mypy” or “type” failures
• Test errors: output contains “pytest” or “test” failures

Retry action:

Task(
  description="[Iteration N] Retry - fix errors",
  prompt="/speckit.implement Fix the lint/type/test errors reported and continue implementation. Current retry: {retry_count}/3",
  subagent_type="general"
)

Non-recoverable errors (stop immediately):

• tasks.md not found or corrupted
• Branch protection errors
• 3 retries exhausted without success
• Unexpected errors

2.3 Verify Checkpoint Reached

Re-read tasks.md and verify:

• All tasks before the target checkpoint are marked [x]
• If not all marked, the subtask may have stopped prematurely – retry

2.4 Create Commit

Use the git-commit skill or create commit directly:

Commit message format:

feat({phase-slug}): {checkpoint_description}

Tasks completed in this checkpoint:
- {task_id} {task_description}
- {task_id} {task_description}
...

Checkpoint: {checkpoint_text}

Example:

feat(foundational): Foundation ready - user story implementation can begin

Tasks completed in this checkpoint:
- T006 Create enum EnrichmentStatus
- T007 Create schemas ColumnMetadataYaml and SourceMetadataYaml
- T010 Implement AsyncTTLCache with dual-lock pattern
- T015 Create tests for AsyncTTLCache

Checkpoint: Foundation ready - user story implementation can begin

2.5 Report Progress

After each successful checkpoint:

### Iteration {N} Complete

**Phase**: {current_phase}
**Checkpoint**: {checkpoint_description}
**Tasks Implemented**: {count}
**Commit**: {commit_hash_short}

Pending tasks remaining: {pending_count}

2.6 Next Iteration

• Re-scan tasks.md for updated state
• Find next checkpoint (if any)
• Continue loop if pending tasks exist

Step 3: Final Summary

When all tasks are complete (no more - [ ] in tasks.md):

## Implement Yolo - COMPLETE

**Spec**: {feature_name}
**Total Iterations**: {iteration_count}
**Total Tasks Implemented**: {task_count}
**Commits Created**: {commit_count}

### Checkpoints Completed

1. ✅ {checkpoint_1_description}
2. ✅ {checkpoint_2_description}
...

### Next Steps

✅ All tasks implemented. Ready for:
- Code review
- PR creation with `/pr-creator`

Error Handling Patterns

Retry Pattern for Recoverable Errors

retry_count = 0
max_retries = 3

while retry_count < max_retries:
    result = run_speckit_implement()
    
    if is_lint_type_test_error(result):
        retry_count += 1
        log(f"Retry {retry_count}/{max_retries}: Fixing errors...")
        continue
    elif is_success(result):
        break
    else:
        # Non-recoverable error
        stop_and_report(result)
        return

if retry_count >= max_retries:
    stop_and_report("Max retries exhausted")

Error Detection Heuristics

Lint errors – Look for:

• “ruff check” in output
• “linting” or “lint error” messages
• Exit code from ruff

Type errors – Look for:

• “mypy” in output
• “type error” or “type:” messages
• Missing type annotations

Test errors – Look for:

• “pytest” in output
• “FAILED” test markers
• “AssertionError” messages

Usage Examples

Basic Usage

/implement-yolo

Runs the full implementation loop from current state.

With Status Check First

User: Show me the current implementation status
AI: [Scans tasks.md and reports pending tasks, current phase, next checkpoint]

User: /implement-yolo
AI: [Begins automated implementation loop]

Anti-Patterns

Running Without Subtask Isolation

Why bad: Context accumulates, causing token overflow and stale state

Instead: ALWAYS use Task tool for each /speckit.implement invocation

Skipping Verification

Why bad: May commit incomplete checkpoints

Instead: Always verify all tasks before checkpoint are marked [x]

Ignoring Retries

Why bad: Transient errors cause unnecessary failures

Instead: Retry lint/type/test errors up to 3 times

Continuing Past Non-Recoverable Errors

Why bad: May corrupt state or create inconsistent code

Instead: Stop immediately on unexpected errors, report clearly

Checkpoint Detection Reference

Identifying Checkpoints in tasks.md

Checkpoints are marked with this pattern:

**Checkpoint**: {description} [optional ✅]

Examples from real tasks.md:

**Checkpoint**: Foundation ready - user story implementation can now begin ✅
**Checkpoint US1**: API de consulta funciona sem PostgreSQL
**Checkpoint Final**: Quality validated, ready for merge

Determining Tasks Before a Checkpoint

1. Find the checkpoint line number

2. Scan backwards for all tasks (- [ ] or - [x]) until:

   • Previous checkpoint, OR
   • Phase header (## Phase), OR
   • Start of file

3. These are the tasks that must be [x] for checkpoint to be complete

Related Commands

• /speckit.implement – The command being orchestrated
• /speckit.tasks – Regenerate task list if needed
• /fast-spec – Full specification pipeline
• /git-commit – Manual commit creation

Configuration

Subtask2 Integration (if using plugin)

If subtask2 is installed, the skill can leverage its loop feature:

subtask: true
loop:
  max: 50
  until: "all tasks in tasks.md are marked [x]"
return:
  - "Checkpoint complete. Committing and continuing..."

Without Subtask2

The skill works without subtask2 by using the native Task tool with manual loop control in the orchestrator context.