Durable Execution Plans for Agents: using-plan Skill

Where this fits

Your agent is mid-way through a multi-step refactor: it has already changed three files, but the terminal session dies and the context window resets. Without a durable plan, the agent starts over, re-reads everything, and risks diverging from the original intent. The using-plan skill gives the agent a persistent, coarse-grained execution plan stored in LWC (Local Workspace Context), so it can resume exactly where it left off — no matter how many interruptions occur.

Why agents benefit

• Resume cleanly — After any interruption, the agent runs lwc plan brief PLAN_ID to recover the current step, next step, and progress without hidden reasoning or stale assumptions.
• Avoid plan drift — Each mutation checks the current revision with --if-revision, preventing concurrent edits from overwriting newer decisions.
• Separate concerns — The plan stays independent from Todo lists; it tracks the coarse execution path, not granular chores, so the agent never conflates "what to do" with "how to do it."
• Explicit done criteria — Every plan defines one objective, constraints, and ordered steps, making it unambiguous when the work is actually complete.
• Opt-in safety — The skill respects the user's plan.setting; if disabled, it explains how to enable rather than silently forcing a new workflow.

Practical scenarios

Long-running code migration

An agent is migrating a legacy service to a new framework. The plan holds the migration order (e.g., models → routes → tests). When the user switches tasks and returns, the agent calls lwc plan brief and picks up at the exact step, without re-scanning the codebase.

Multi-session debugging

A tricky bug spans several files and requires hypothesis testing. The agent records each hypothesis as a step and marks it blocked or revised when new evidence appears. On the next session, the plan's plan.tracking shows which theory was last active and what to try next.

Collaborative code review

Two agents work on the same repository. One owns the plan, the other suggests changes. Because every mutation uses revision guards, the plan never silently overwrites the other agent's progress — it blocks or asks for a merge decision.

Add it to your agent workflow

Install the skill directly from the CLI:

npx skills add JanYork/llm-wiki-cli --skill using-plan

Then, in your agent's system prompt or skill instructions, tell it to check LWC config first and only use Plan when enabled. A minimal usage example:

Agent: I need to refactor the payment module.
Skill: Run `lwc config show` → if plan.setting is enabled, create a plan with one objective, done criteria, and ordered steps. Use `lwc plan brief` before any mutation.

For a deeper look at this skill and its ecosystem, see the using-plan skill page on OpenAgentSkill.

Compare before adopting

Before wiring this into your agent, evaluate it against alternatives:

• Quality signals — Check the GitHub repo (JanYork/llm-wiki-cli) for recent commits, open issues, and whether the LWC spec is stable.
• Maintenance freshness — Is the skill actively updated to match LWC version changes? A stale skill can break silently.
• Alternatives — Compare with Todo-based approaches or external plan files; using-plan is durable but only works if LWC is enabled.
• Workflow fit — Does your agent already handle long-running tasks? If not, the extra state management may be overkill.

Why it is worth tracking

With 44 stars and a focused scope, using-plan is a small but growing utility that solves a real pain point: agent continuity. It's worth tracking because it demonstrates a pattern — durable, revision-safe plans — that many coding agents will adopt. Evaluate it when you need your agent to survive context resets or work across sessions, and watch for updates that align with LWC's evolution.