Ibrahim-3d/executing-plans

Executing Plans

Overview

Load plan, review critically, execute all tasks autonomously to completion.

Core principle: Execution behavior depends on conductor/config.json → "mode":

• "agentic": Autonomous execution — never stop for human feedback. Resolve all issues by consulting lead agents or best-judgment.
• "human-in-the-loop": Stop at blockers and report to user. Wait for feedback between batches of 3 tasks.

Announce at start: "I'm using the executing-plans skill to implement this plan."

The Process

Step 1: Load and Review Plan

1. read_file plan file
2. Review critically - identify any questions or concerns about the plan
3. If concerns:
   • "agentic" mode: Resolve autonomously — consult lead agents (Architecture Lead, Tech Lead) via subagent dispatch. NEVER ask the user.
   • "human-in-the-loop" mode: Raise concerns with the user before starting.
4. Create TodoWrite and proceed

Step 2: Execute Tasks (Mode-Dependent)

"agentic" mode — Execute ALL tasks sequentially without stopping:

For each task:

1. Mark as in_progress
2. Follow each step exactly (plan has bite-sized steps)
3. Run verifications as specified
4. If verification fails: attempt fix autonomously (max 3 retries per task), log the issue, continue to next task
5. Mark as completed

"human-in-the-loop" mode — Execute in batches of 3 tasks, then pause:

For each batch (3 tasks at a time):

1. Mark tasks as in_progress
2. Follow each step exactly
3. Run verifications as specified
4. If verification fails: STOP and ask the user for help before continuing
5. After each batch completes: report progress, say "Ready for feedback.", and wait before proceeding to next batch

Step 3: Report Progress

After tasks complete:

• Show what was implemented
• Show verification output
• "agentic" mode: Show any autonomous decisions made; do NOT wait for feedback — proceed to completion
• "human-in-the-loop" mode: Say "Ready for feedback." and wait

Step 5: Complete Development

After all tasks complete and verified:

• Announce: "I'm using the finishing-a-development-branch skill to complete this work."
• REQUIRED SUB-SKILL: Use orchestrator-supaconductor:finishing-a-development-branch
• Follow that skill to verify tests, present options, execute choice

Blocker Resolution (Mode-Dependent)

If mode = "agentic": Resolve all blockers autonomously. NEVER stop. If mode = "human-in-the-loop": STOP and ask the user for help on any blocker.

Agentic Mode Resolutions:

• Missing dependency → Install it if safe (<50KB), or skip the task and log the blocker
• Test fails → Attempt fix (max 3 retries), then log failure and continue with remaining tasks
• Instruction unclear → Spawn a Plan subagent to interpret based on codebase context, or consult Product Lead
• Plan has critical gaps → Consult Architecture Lead via subagent to fill gaps autonomously
• Verification fails repeatedly → Log the issue with details, mark task as completed-with-warnings, continue

When to Revisit Earlier Steps

Return to Review (Step 1) when:

• Fundamental approach is failing (>50% of tasks failing) — re-plan autonomously
• Architecture Lead subagent recommends a different approach

Remember

• Review plan critically first
• Follow plan steps exactly
• Don't skip verifications
• Reference skills when plan says to
• "agentic" mode: Execute ALL tasks without stopping — resolve blockers autonomously, log all decisions
• "human-in-the-loop" mode: Execute in batches of 3 — stop at blockers and ask the user, wait for feedback between batches
• Never start implementation on main/master branch — use feature branches

Conductor Integration (Autonomous Mode)

When invoked with --plan, --track-dir, and --metadata parameters (from Conductor orchestrator):

• read_file plan from --plan path
• "agentic" mode: Execute ALL tasks without stopping — run autonomously
• "human-in-the-loop" mode: Execute in batches of 3 — stop at blockers, wait for user feedback between batches
• After each task: use replace tool to mark [x] in plan.md with commit SHA
• After all tasks: update --metadata checkpoint to EXECUTE: PASSED
• Return concise verdict: {"verdict": "PASS", "tasks_completed": N}
• If --resume-from is provided, skip tasks before that task ID

When these parameters are absent, fall back to the standalone mode-aware workflow above.

Integration

Required workflow skills:

• orchestrator-supaconductor:using-git-worktrees - REQUIRED: Set up isolated workspace before starting
• orchestrator-supaconductor:writing-plans - Creates the plan this skill executes
• orchestrator-supaconductor:finishing-a-development-branch - Complete development after all tasks