Agent Skills: Implementation Orchestration

|

UncategorizedID: malhashemi/dotfiles/implementation-orchestration

Install this agent skill to your local

pnpm dlx add-skill https://github.com/malhashemi/dotfiles/tree/HEAD/dot_config/opencode/skill/implementation-orchestration

Skill Files

Browse the full folder contents for implementation-orchestration.

Download Skill

Loading file tree…

dot_config/opencode/skill/implementation-orchestration/SKILL.md

Skill Metadata

Name
implementation-orchestration
Description
|

Implementation Orchestration

CRITICAL: You Are an Orchestrator, Not an Implementer

NEVER implement phases yourself. You do not write code, tests, or make commits.

ALWAYS spawn implement subagent for each phase via the Task tool.

Your job is to:

  • Read and analyze plans
  • Create worktrees and branches
  • Spawn Implement agents (they do the actual work)
  • Spawn Validator after each phase
  • Track progress and update checkboxes
  • Handle decomposition when needed

If you find yourself writing code, editing source files, or running test commands to verify your own changes — STOP. You are doing the Implement agent's job. Spawn an Implement agent instead.


Overview

After creating a plan, the Planner becomes the orchestrator for that plan's implementation. This skill provides the protocol for managing phases, worktrees, branches, validation, and decomposition triggers.

When to Use

  • After plan is approved by Validator (via plan-review)
  • When RPIV hands off a plan for implementation
  • When managing multiple phases that may be sequential or parallel

Prerequisites

  1. Plan exists and is approved
  2. Feature branch exists (e.g., polish)
  3. Git working tree is clean

Scripts

Execute via justfile or directly. The base directory is provided when skill is loaded.

| Recipe | Command | Purpose | |--------|---------|---------| | worktree-create | just -f {base_dir}/justfile worktree-create {plan-slug} {base-branch} | Create plan worktree | | worktree-cleanup | just -f {base_dir}/justfile worktree-cleanup {plan-slug} | Remove worktree after merge |

Orchestration Protocol

Step 1: Setup Worktree

just -f {base_dir}/justfile worktree-create {plan-slug} {base-branch}

See references/worktree-patterns.md for naming conventions and git commands.

Step 2: Analyze Phase Dependencies

Categorize phases:

  • Sequential: One after another → commit to plan branch
  • Parallel: Can run simultaneously → separate branches + worktrees
  • Blocked: Waiting on dependency → defer until unblocked

Step 3: Execute Phases

For each sequential phase, spawn Implement:

task({
  subagent_type: "implement",
  description: "Execute Phase N",
  prompt: "Execute Phase N from plan at [plan-path].
           Worktree: .trees/{plan-slug}
           Branch: implement/{plan-slug}
           Load skill(name='phase-execution') for protocol.
           Return PHASE_COMPLETE or NEEDS_DECOMPOSITION."
})

For parallel phases, issue multiple Task calls simultaneously with separate worktrees.

CRITICAL: Parallel Work Isolation

Parallel phases MUST have separate worktrees. Never run parallel sub-phases in the same worktree—they will conflict on file edits and git operations. Each parallel track needs:

  • Its own worktree: .trees/plan-N-phase-X
  • Its own branch: implement/plan-N-phase-X

Step 4: Validation Gates

After each phase, validate. See references/validation-gates.md for:

  • Validator spawning pattern
  • Handling PROCEED / PROCEED_WITH_NOTES / BLOCKED
  • RPIV trigger for decomposition

Step 5: Progress Tracking

After successful phase completion and validation, update the plan document to reflect progress. See the Progress Tracking section below for detailed guidance.

Step 6: Plan Completion & PR Opening

  1. Final validation with Validator
  2. Create PR: gh pr create --base {feature-branch} --head implement/{plan-slug}
  3. Return to RPIV with structured completion report (see below)

CRITICAL: Do NOT handle PR review yourself. Return to RPIV immediately after opening the PR. RPIV orchestrates the PR review phase separately.

Completion Report Format:

IMPLEMENTATION_COMPLETE

Plan: {plan-path}
PR: #{pr_number} - {pr_title}
URL: {pr_url}
Worktree: {worktree_path}
Branch: {branch}

Phases completed: {n}/{total}
Commits: {commit_count}
Tests: passing
Build: passing

Ready for PR review phase.

After PR is merged (RPIV will tell you when):

  • just -f {base_dir}/justfile worktree-cleanup {plan-slug}

Progress Tracking

CRITICAL: Only YOU (the orchestrator) update plan checkboxes. Implement agents do NOT touch the plan file—they return PHASE_COMPLETE and you record the progress.

After receiving PHASE_COMPLETE from Implement and PROCEED from Validator, update the plan to reflect what was accomplished. Use judgment to determine completion status.

Checkbox Types in Plans

| Type | Format | When to Update | |------|--------|----------------| | Phase header | ## Phase 1: Title## Phase 1: Title ✓ | All phase success criteria pass | | Automated checks | - [ ] Tests pass | After running and confirming pass | | Success criteria | - [ ] File exists at path | After verifying condition is met | | Manual verification | - [ ] Verify UI works | After checking with available tools |

Update Protocol

After receiving PHASE_COMPLETE from Implement and PROCEED from Validator:

  1. Mark phase header complete: Add ✓ to phase header

    ## Phase 2: Core Implementation
    

    becomes:

    ## Phase 2: Core Implementation ✓
    
  2. Update success criteria checkboxes: Change [ ] to [x] for verified items

    - [ ] All tests pass
    - [ ] TypeScript compiles without errors
    

    becomes:

    - [x] All tests pass
    - [x] TypeScript compiles without errors
    
  3. Handle manual verification items: Attempt ALL verification, not just "automated" ones

    • Use available tools: browser automation, CLI tools, API calls, file checks
    • If verified successfully → mark [x]
    • If genuinely cannot verify → leave [ ] and note reason in commit or report

Verification Approach

Attempt everything. The distinction is not "automated vs manual" but "can verify vs genuinely cannot."

Available verification methods:

  • CLI tools: Run commands, check output, verify behavior
  • Browser automation: Playwright, browser MCP for UI verification
  • API calls: Test endpoints, verify responses
  • File system: Check files exist, content matches expected
  • Process checks: Verify services running, ports listening

Only defer verification when genuinely impossible (e.g., "requires production access", "needs stakeholder approval").

Edit Tool Usage

Use the Edit tool to update plan checkboxes directly:

edit({
  filePath: "thoughts/shared/plans/my-plan.md",
  oldString: "## Phase 2: Core Implementation",
  newString: "## Phase 2: Core Implementation ✓"
})

For checkbox updates:

edit({
  filePath: "thoughts/shared/plans/my-plan.md",
  oldString: "- [ ] All tests pass",
  newString: "- [x] All tests pass"
})

Partial Completion

If a phase is partially complete (some criteria pass, some fail):

  • Do NOT mark phase header complete
  • Mark individual passing criteria with [x]
  • Leave failing criteria as [ ]
  • Document what remains in validation report

Anti-Patterns

  • NEVER implement phases yourself - Always spawn Implement agents via Task tool
  • NEVER write code, tests, or edit source files - That's the Implement agent's job
  • NEVER run implementation commands - Only run commands to verify Implement's work
  • Don't skip validation gates - Every phase must be validated
  • Don't commit to feature branch directly - Always use plan branches
  • Don't leave worktrees orphaned - Always cleanup after PR merge
  • Don't continue after BLOCKED verdict - Wait for resolution
  • Don't assume "manual" means "skip" - Attempt all verification with available tools
  • Don't mark phase complete if any criteria failed - Only mark when ALL pass