Use when executing an implementation plan produced by Radberi-PlanWrite, or when resuming plan execution after a break. Trigger when the user says "execute the plan", "start implementing", approves a plan, or references a previously written plan to begin work on.
After context compaction: Re-read this skill AND ./spec-reviewer-prompt.md before continuing.
Execute plan tasks sequentially, with Direction Checks verifying each non-trivial task and Production Readiness + Code Simplification reviews at the end.
Direction Checks exist because without independent verification, implementers tend to silently miss requirements or over-engineer. The subagent reads actual code and compares it to the spec — catching drift early before it compounds across tasks.
1. Setup
└─ Locate plan file (from plan mode context or conversation)
└─ Read plan → Extract tasks with complexity tiers
└─ Create TodoWrite with all tasks
└─ Identify first non-trivial task (for Direction Check agent creation)
2. Per Task
└─ Implement task (invoke required skills)
└─ Trivial tasks: Skip Direction Check
└─ Standard/Complex tasks:
└─ First time: Launch Direction Check agent (see ./spec-reviewer-prompt.md)
└─ Subsequent: Resume stored agent ID
└─ If issues: fix → re-check. Summarise issues and fixes.
└─ If passes: next task.
3. Completion (after all tasks)
└─ Launch Production Readiness agent with full files-changed list
└─ If issues: fix → re-review. Summarise issues and fixes.
└─ If passes: Resume same agent for Code Simplification pass
└─ Present suggestions → User confirms → Apply
└─ Invoke Radberi-FinishingDevelopmentBranch
Pre-warming agents at setup caused timeouts on longer plans. Instead, create each agent when first needed and reuse it via resume for the rest of the session.
./spec-reviewer-prompt.md — Pre-Warm + Resume templatesresume with stored agent ID for every subsequent non-trivial task. This preserves the agent's understanding of the plan and codebase across tasks.Radberi-ProductionReadiness)Complexity tiers are defined in PlanWrite. The depth table and check criteria live in ./spec-reviewer-prompt.md — the orchestrator does not need to duplicate them here.
Trivial Batching: When multiple contiguous Trivial tasks appear, execute them all as a batch. Report completion as a group.
Commit after each task passes its Direction Check (or after a trivial batch completes). This creates clean rollback points — if a later task goes wrong, earlier work is safely preserved.
After each task (or trivial batch), briefly report to the user what was completed and what's next. Keep it to 1-2 sentences — enough to track progress without interrupting flow.
Task failure: Retry once with a targeted fix → if still fails, escalate to user.
Review failure (Direction Check or Production Readiness):
Radberi-ReceivingCodeReview to evaluate the feedback — verify each suggestion against codebase reality, apply YAGNI checks, push back on technically incorrect suggestionsRequired skills:
Radberi-PlanWrite — Creates the planRadberi-ProductionReadiness — Production Readiness + Code Simplification (two-pass agent)Radberi-ReceivingCodeReview — Evaluate review feedback before acting on itRadberi-FinishingDevelopmentBranch — Completion