By name: trycycle-executing description: Internal trycycle subskill — do not invoke directly.
Executing Plans
Overview
Load the plan, create TodoWrite entries for all tasks, then execute every task sequentially without pausing.
Core principle: Execute all tasks continuously. The plan has already been reviewed. Your job is to implement it precisely.
Step 1: Load Plan
- Read the implementation plan file and the test plan file
- Create a TodoWrite entry for every task in the plan
Step 2: Execute All Tasks
For each task, in order:
- Mark as
in_progress - Follow each step exactly — the plan has bite-sized steps
- Work through red/green/refactor cycles until the task is genuinely complete:
- Run the failing check (red)
- Make it pass (green)
- Refactor and tighten the implementation and tests, then re-run all relevant automated checks
- Continue until all required automated checks named or implied by the plan are passing — including any broader regression suite or full project suite the repo requires before completion
- Mark as
completed
Repeat until all tasks are done, then commit your changes.
Execution is not complete merely because one targeted check turned green. Keep working until all required automated checks pass for legitimate reasons.
Test quality is non-negotiable. Do not weaken, delete, or dilute a valid automated test to obtain a passing result. If a test is wrong, obsolete, or lower-fidelity than a better replacement, you may correct or replace it — but do not manufacture green by making good tests less demanding.
Blockers
There are two states: execute or stop for a blocker.
A blocker is something where the agent cannot use its best judgment because there is no path forward, or because being wrong could cause harm — a missing dependency that cannot be worked around, a test environment that is down, a suddenly dirty file in the repo that, if handled incorrectly, could cause data loss.
"I have concerns about the approach" is not a blocker. The plan is already reviewed. Red tests and review findings are inputs to keep iterating — not blockers — unless the environment itself is broken.
If you hit a blocker: stop and report your findings. Do not guess your way through something that could produce silently broken output or cause harm.
Everything else: proceed.
Remember
- Follow plan steps exactly
- Don't skip verifications
- Work red/green/refactor cycles — don't stop at the first green
- All required automated checks must pass legitimately before execution is complete
- Never weaken or delete a valid test to achieve green
- Reference skills when the plan says to
- Stop only for genuine blockers — not concerns, not uncertainty, not preference
- Never implement on main/master branch without explicit user consent