18 KiB
description
| description |
|---|
| Execute implementation plan with automatic sequential/parallel orchestration using git-spice and worktrees |
You are executing an implementation plan.
Architecture
The execute command uses an orchestrator-with-embedded-instructions architecture for cognitive load reduction:
Orchestrator Skills (executing-sequential-phase, executing-parallel-phase)
- Responsibilities: Setup, worktree management, dispatch coordination, code review orchestration
- Size: ~464 lines (sequential), ~850 lines (parallel)
- Dispatches: Task tool with embedded execution instructions (~150 lines per task)
- Focus: Manages workflow lifecycle, embeds focused task instructions for subagents
Verification Skill (phase-task-verification)
- Responsibilities: Shared git operations (add, branch create, HEAD verify, detach)
- Size: ~92 lines
- Used by: Task subagents via Skill tool (both sequential and parallel)
- Focus: Eliminates duplication of branch creation/verification logic
Cognitive Load Reduction:
- Original: 750-line monolithic skills (orchestration + task + verification mixed)
- Current: Subagents receive ~150 lines of focused task execution instructions via Task tool
- Benefit: 80% reduction in content subagents must parse (only see task instructions, not orchestration)
- Verification: Shared 92-line skill eliminates duplication
Maintainability benefit: Orchestrator features (resume support, enhanced error recovery, verification improvements) can be added without affecting task execution instructions. Subagents receive focused, consistent instructions while orchestrators handle workflow complexity.
Testing framework: This plugin uses Test-Driven Development for workflow documentation. See tests/README.md and CLAUDE.md for the complete testing system. All changes to commands and skills must pass the test suite before committing.
Available Skills
Skills are referenced on-demand when you encounter the relevant step. Do not pre-read all skills upfront.
Phase Execution (read when you encounter the phase type):
executing-parallel-phase- Mandatory workflow for parallel phases (Step 2)executing-sequential-phase- Mandatory workflow for sequential phases (Step 2)
Support Skills (read as needed when referenced):
understanding-cross-phase-stacking- Reference before starting any phase (explains base branch inheritance)validating-setup-commands- Reference if CLAUDE.md setup validation needed (Step 1.5)using-git-spice- Reference for git-spice command syntax (as needed)requesting-code-review- Reference after each phase completes (Step 2)verification-before-completion- Reference before claiming completion (Step 3)finishing-a-development-branch- Reference after all phases complete (Step 4)troubleshooting-execute- Reference if execution fails (error recovery)using-git-worktrees- Reference if worktree issues occur (diagnostics)
Input
User will provide: /spectacular:execute {plan-path}
Example: /spectacular:execute @specs/a1b2c3-magic-link-auth/plan.md
Where a1b2c3 is the runId and magic-link-auth is the feature slug.
Workflow
Step 0a: Extract Run ID from Plan
User provided plan path: The user gave you a plan path like .worktrees/3a00a7-main/specs/3a00a7-agent-standardization-refactor/plan.md
Extract RUN_ID from the path:
The RUN_ID is the first segment of the spec directory name (before the first dash).
For example:
- Path:
.worktrees/3a00a7-main/specs/3a00a7-agent-standardization-refactor/plan.md - Directory:
3a00a7-agent-standardization-refactor - RUN_ID:
3a00a7
# Extract RUN_ID and FEATURE_SLUG from plan path (replace {the-plan-path-user-provided} with actual path)
PLAN_PATH="{the-plan-path-user-provided}"
DIR_NAME=$(echo "$PLAN_PATH" | sed 's|^.*specs/||; s|/plan.md$||')
RUN_ID=$(echo "$DIR_NAME" | cut -d'-' -f1)
FEATURE_SLUG=$(echo "$DIR_NAME" | cut -d'-' -f2-)
echo "Extracted RUN_ID: $RUN_ID"
echo "Extracted FEATURE_SLUG: $FEATURE_SLUG"
# Verify RUN_ID and FEATURE_SLUG are not empty
if [ -z "$RUN_ID" ]; then
echo "❌ Error: Could not extract RUN_ID from plan path: $PLAN_PATH"
exit 1
fi
if [ -z "$FEATURE_SLUG" ]; then
echo "❌ Error: Could not extract FEATURE_SLUG from plan path: $PLAN_PATH"
exit 1
fi
CRITICAL: Execute this entire block as a single multi-line Bash tool call. The comment on the first line is REQUIRED - without it, command substitution $(...) causes parse errors.
Store RUN_ID and FEATURE_SLUG for use in:
- Branch naming:
{run-id}-task-X-Y-name - Filtering:
git branch | grep "^ {run-id}-" - Spec path:
specs/{run-id}-{feature-slug}/spec.md - Cleanup: Identify which branches/specs belong to this run
Announce: "Executing with RUN_ID: {run-id}, FEATURE_SLUG: {feature-slug}"
Step 0b: Verify Worktree Exists
After extracting RUN_ID, verify the worktree exists:
# Get absolute repo root (stay in main repo, don't cd into worktree)
REPO_ROOT=$(git rev-parse --show-toplevel)
# Verify worktree exists
if [ ! -d "$REPO_ROOT/.worktrees/${RUN_ID}-main" ]; then
echo "❌ Error: Worktree not found at .worktrees/${RUN_ID}-main"
echo "Run /spectacular:spec first to create the workspace."
exit 1
fi
# Verify it's a valid worktree
git worktree list | grep "${RUN_ID}-main"
IMPORTANT: Orchestrator stays in main repo. All worktree operations use git -C .worktrees/{run-id}-main or absolute paths.
This ensures task worktrees are created at the same level as {run-id}-main, not nested inside it.
Announce: "Verified worktree exists: .worktrees/{run-id}-main/"
Step 0c: Check for Existing Work
Check if implementation work already exists:
# Get repo root
REPO_ROOT=$(git rev-parse --show-toplevel)
# Check current branch in main worktree
CURRENT_BRANCH=$(git -C "$REPO_ROOT/.worktrees/${RUN_ID}-main" branch --show-current 2>/dev/null || echo "")
# Count existing task branches for this RUN_ID
EXISTING_TASKS=$(git branch 2>/dev/null | grep -c "^ ${RUN_ID}-task-" || echo "0")
# Report status
if [ "$EXISTING_TASKS" -gt 0 ]; then
echo "📋 Found $EXISTING_TASKS existing task branch(es) for RUN_ID: $RUN_ID"
echo " Current branch: $CURRENT_BRANCH"
echo ""
echo "Resuming from current state. The execution will:"
echo "- Sequential phases: Continue from current branch"
echo "- Parallel phases: Skip completed tasks, run remaining"
echo ""
else
echo "✅ No existing work found - starting fresh execution"
echo ""
fi
CRITICAL: Execute this entire block as a single multi-line Bash tool call. The comment on the first line is REQUIRED - without it, command substitution $(...) causes parse errors.
Resume behavior:
- If
EXISTING_TASKS > 0: Execution continues from current state in main worktree - If
EXISTING_TASKS = 0: Execution starts from Phase 1, Task 1 - Main worktree current branch indicates progress (latest completed work)
Note: Orchestrator proceeds immediately to Step 1. Phase execution logic in Step 2 handles resume by checking current branch and existing task branches.
Step 1: Read and Parse Plan
Read the plan file and extract:
- Feature name
- All phases (with strategy: sequential or parallel)
- All tasks within each phase
For each task, extract and format:
- Task ID and name
- Files to modify (explicit paths)
- Acceptance criteria (bullet points)
- Dependencies (which tasks must complete first)
Store extracted task info for subagent prompts (saves ~1000 tokens per subagent):
Task 4.2:
Name: Integrate prompts module into generator
Files: - src/generator.ts - src/types.ts
Acceptance Criteria: - Import PromptService from prompts module - Replace manual prompt construction with PromptService.getCommitPrompt() - Update tests to mock PromptService - All tests pass
Dependencies: Task 4.1 (fallback logic removed)
Verify plan structure:
- ✅ Has phases with clear strategies
- ✅ All tasks have files specified
- ✅ All tasks have acceptance criteria
- ✅ Dependencies make sense
Step 1.5: Validate Setup Commands (REQUIRED)
Use the validating-setup-commands skill:
This skill validates that CLAUDE.md defines required setup commands BEFORE creating worktrees. It provides clear error messages with examples if missing.
The skill will extract and return:
INSTALL_CMD- Required dependency installation commandPOSTINSTALL_CMD- Optional post-install command (codegen, etc.)
Store these commands for use in dependency installation steps throughout execution.
Step 1.6: Detect Project Commands (Optional)
Optionally detect project-specific quality check commands for subagents to use.
This is optional - most projects define commands in CLAUDE.md that subagents can discover.
If you want to provide hints, check for common patterns:
- TypeScript/JavaScript:
package.jsonscripts - Python:
pytest,ruff,black - Go:
go test,golangci-lint - Rust:
cargo test,cargo clippy
If detected, mention in subagent prompts:
TEST_CMD- Command to run testsLINT_CMD- Command to run lintingFORMAT_CMD- Command to format codeBUILD_CMD- Command to build project
If not detected, subagents will check CLAUDE.md or skip quality checks with warning.
IMPORTANT: Do NOT read constitution files here. Let subagents read them as needed to reduce orchestrator token usage.
Step 1.7: Configure Code Review Frequency
Determine when to run code reviews:
# Check if REVIEW_FREQUENCY env var is set
REVIEW_FREQUENCY=${REVIEW_FREQUENCY:-}
If not set, prompt user for preference:
If REVIEW_FREQUENCY is empty, use AskUserQuestion tool to prompt:
Question: "How frequently should code reviews run during execution?"
Header: "Review Frequency"
Options:
1. "After each phase"
Description: "Run code review after every phase completes (safest - catches errors early, prevents compounding issues)"
2. "Optimize automatically"
Description: "Let Claude decide when to review based on phase risk/complexity (RECOMMENDED - balances speed and quality)"
3. "Only at end"
Description: "Skip per-phase reviews, run one review after all phases complete (faster, but errors may compound)"
4. "Skip reviews"
Description: "No automated code reviews (fastest, but requires manual review before merging)"
Store user choice:
- Option 1 →
REVIEW_FREQUENCY="per-phase" - Option 2 →
REVIEW_FREQUENCY="optimize" - Option 3 →
REVIEW_FREQUENCY="end-only" - Option 4 →
REVIEW_FREQUENCY="skip"
Announce decision:
Code review frequency: {REVIEW_FREQUENCY}
Note: This setting applies to all phases in this execution. Phase skills check REVIEW_FREQUENCY to determine whether to run code review step.
Step 2: Execute Phases
If resuming: Start from the incomplete phase/task identified in Step 0.
For each phase in the plan, execute based on strategy:
Code Review Gates: Phase execution skills check REVIEW_FREQUENCY (set in Step 1.7) to determine when to run code reviews:
per-phase: Review after each phase before proceedingoptimize: LLM decides whether to review based on phase risk/complexityend-only: Skip per-phase reviews, review once after all phasesskip: No automated reviews (manual review required)
Cross-Phase Stacking
Use the understanding-cross-phase-stacking skill:
This skill explains how sequential and parallel phases automatically chain together through base branch inheritance. Key concepts:
- Main worktree tracks progress (current branch = latest completed work)
- Parallel phases inherit from current branch (not original base)
- Natural chaining creates linear stack across all phases
Read this skill before starting any new phase to understand how phases build on each other.
Sequential Phase Strategy
Use the executing-sequential-phase skill:
This skill provides the complete workflow for executing sequential phases. Key concepts:
- Execute tasks one-by-one in existing
{runid}-mainworktree - Trust natural git-spice stacking (no manual
gs upstack onto) - Tasks build on each other cumulatively
- Stay on task branches for automatic stacking
The skill includes detailed subagent prompts, quality check sequences, and code review integration.
CRITICAL: When dispatching subagents, substitute {run-id} and {feature-slug} with the values extracted in Step 0a. Subagents need these to read the spec at specs/{run-id}-{feature-slug}/spec.md.
Parallel Phase Strategy
Use the executing-parallel-phase skill:
This skill provides the complete mandatory workflow for executing parallel phases. Key requirements:
- Create isolated worktrees for EACH task (even N=1)
- Install dependencies per worktree
- Spawn parallel subagents in single message
- Verify completion before stacking
- Stack branches linearly, then cleanup worktrees
- Code review after stacking
The skill includes the 8-step mandatory sequence, verification checks, N=1 edge case handling, and stacking algorithm.
CRITICAL: When dispatching subagents, substitute {run-id} and {feature-slug} with the values extracted in Step 0a. Subagents need these to read the spec at specs/{run-id}-{feature-slug}/spec.md.
Step 3: Verify Completion
After all phases execute successfully:
Use the verification-before-completion skill:
This skill enforces verification BEFORE claiming work is done.
Required verifications (if commands detected):
# Run full test suite
if [ -n "$TEST_CMD" ]; then
$TEST_CMD || { echo "❌ Tests failed"; exit 1; }
fi
# Run linting
if [ -n "$LINT_CMD" ]; then
$LINT_CMD || { echo "❌ Linting failed"; exit 1; }
fi
# Run production build
if [ -n "$BUILD_CMD" ]; then
$BUILD_CMD || { echo "❌ Build failed"; exit 1; }
fi
# Verify all detected checks passed
echo "✅ All quality checks passed - ready to complete"
If no commands detected:
⚠️ No test/lint/build commands found in project.
Add to CLAUDE.md or constitution/testing.md for automated quality gates.
Proceeding without verification - manual review recommended.
Critical: Evidence before assertions. Never claim "tests pass" without running them.
Step 4: Finish Stack
After verification passes:
Use the finishing-a-development-branch skill to:
- Review all changes
- Choose next action:
- Submit stack as PRs:
gs stack submit(per using-git-spice skill) - Continue with dependent feature:
gs branch create - Mark complete and sync:
gs repo sync
- Submit stack as PRs:
Step 5: Final Report
✅ Feature Implementation Complete
**RUN_ID**: {run-id}
**Feature**: {feature-name}
**Worktree**: .worktrees/{run-id}-main/
**Stack**: {count} task branches (all stacked on {run-id}-main)
## Execution Summary
**Phases Completed**: {count}
- Sequential: {count} phases
- Parallel: {count} phases
**Tasks Completed**: {count}
**Commits**: {count}
**Isolation**: All work completed in worktree. Main repo unchanged.
## Parallelization Results
{For each parallel phase:}
**Phase {id}**: {task-count} tasks in parallel
- Estimated sequential time: {hours}h
- Actual parallel time: {hours}h
- Time saved: {hours}h
**Total Time Saved**: {hours}h ({percent}%)
## Quality Checks
Quality checks are project-specific (detected from CLAUDE.md, constitution, or common patterns):
✅ Tests passing (if `TEST_CMD` detected)
✅ Linting clean (if `LINT_CMD` detected)
✅ Formatting applied (if `FORMAT_CMD` detected)
✅ Build successful (if `BUILD_CMD` detected)
If no commands detected, quality gates are skipped with warning to user.
✅ {total-commits} commits across {branch-count} task branches
## Next Steps
### Review Changes (from main repo)
```bash
# All these commands work from main repo root
gs log short # View all branches and commits in stack
gs log long # Detailed view with commit messages
git branch | grep "^ {run-id}-" # List all branches for this run
# To see changes in worktree:
cd .worktrees/{run-id}-main
git diff main..HEAD # See all changes in current stack
cd ../.. # Return to main repo
```
Submit for Review (from main repo)
# git-spice commands work from main repo
gs stack submit # Submits entire stack as PRs (per using-git-spice skill)
Or Continue with Dependent Feature (from worktree)
cd .worktrees/{run-id}-main # Navigate to worktree
gs branch create # Creates new branch stacked on current
cd ../.. # Return to main repo when done
Cleanup Worktree (after PRs merged)
# From main repo root:
git worktree remove .worktrees/{run-id}-main
# Optional: Delete the {run-id}-main branch
git branch -d {run-id}-main
Important: Main repo remains unchanged. All work is in the worktree and task branches.
## Error Handling
**Use the `troubleshooting-execute` skill:**
This skill provides comprehensive diagnostic and recovery strategies for execute command failures:
- Phase execution failures (sequential and parallel)
- Parallel agent failures with 4 recovery options
- Merge conflicts during stacking
- Worktree not found errors
- Parallel task worktree creation failures
The skill includes diagnostic commands, recovery strategies, and prevention guidelines.
Consult this skill when execution fails or produces unexpected results.
## Important Notes
- **Orchestrator delegates, never executes** - The orchestrator NEVER runs git commands directly. All git operations are delegated to subagents (setup, implementation, cleanup)
- **Subagents own their git operations** - Implementation subagents create branches and commit their own work using `gs branch create -m`
- **Skill-driven execution** - Uses using-git-spice, using-git-worktrees, and other superpowers skills
- **Automatic orchestration** - Reads plan strategies, executes accordingly
- **Git-spice stacking** - Sequential tasks stack linearly; parallel tasks branch from same base (per using-git-spice skill)
- **No feature branch** - The stack of task branches IS the feature; never create empty branch upfront
- **Worktree isolation** - Parallel tasks run in separate worktrees (per using-git-worktrees skill)
- **Critical: HEAD detachment** - Parallel task subagents MUST detach HEAD after creating branches to make them accessible in parent repo
- **Context management** - Each task runs in isolated subagent to avoid token bloat
- **Constitution adherence** - All agents follow project constitution (@docs/constitutions/current/)
- **Quality gates** - Tests and linting after every task, code review after every phase
- **Continuous commits** - Small, focused commits with [Task X.Y] markers throughout
Now execute the plan from: {plan-path}