executing-plans

version	2
name	executing-plans
description	当已有书面实现计划且需要按检查点执行、验证和收尾时使用；适用于 MoviePilot Agent 可用工具链，不依赖外部任务清单工具。
allowed-tools	read_file list_directory execute_command edit_file write_file ask_user_choice

Executing Plans

Purpose

Execute an existing written plan without turning it into endless discussion. This skill is for implementation discipline, not for designing the plan itself.

Use it after a plan already exists and the next step is to carry it out safely inside the MoviePilot Agent environment.

Entry Check

Before executing:

1. Read the plan or the referenced task context.
2. Identify the target files, commands, or MoviePilot operations.
3. Classify risk:
   • Low-risk /config/agent or /config local plugin asset change → execute directly when the user asked to continue.
   • High-impact action such as deletion, download start, credential change, plugin install/uninstall, restart, scheduler/workflow execution, or broad cleanup → ask for confirmation, preferably with buttons.
   • Ambiguous target or missing required input → ask one focused clarification.
4. If the plan is unsafe or internally inconsistent, stop before mutation and report the concrete issue.

Execution Loop

For each plan item, preserve the Vibe-style governed execution checkpoints adapted to MoviePilot Agent:

1. Requirement freeze — restate the item as target, scope, risk, and acceptance evidence before mutation.
2. Phase execute — perform the smallest correct action with the narrowest tool path.
3. No-regression check — run fresh verification before moving on; for runtime-affecting changes, include a representative dry-run, router check, command discovery, safe execution, or authoritative state query.
4. Failure narrowing — if verification fails, diagnose once with a narrower check; do not claim completion.
5. Phase cleanup — update the local plan/status if needed, preserve only reusable context, and continue to the next independent item only after evidence exists.

Do not wait for the user after every tiny step when the plan is clear and authorized. Do not bypass requirement freeze, phase_cleanup, or no-regression discipline for XL or multi-file work.

MoviePilot Agent Adaptation

• Use available MCP tools and shell diagnostics directly; do not reference unavailable external task-list tools.
• Use subagents only for independent read-only investigation or planning; main agent handles writes and confirmations.
• For /config/agent capability assets, prefer editing the relevant skill or memory file, then verify by re-reading or running structural checks.
• For MoviePilot media operations, follow the domain router: sites → identity → resources → download/subscription/transfer → verification.
• For repository synchronization, hand off to moviepilot-agent-git-maintenance after local verification.

Spec Execution Discipline

When executing a spec-style plan:

• preserve the frozen problem/scope/non-goals while editing;
• do not silently expand from one owner file to a new subsystem;
• if a phase reveals a new risk gate, stop that phase and ask with buttons;
• if a candidate idea belongs to another owner, record it in runtime rather than mixing it into the current patch;
• phase cleanup must update only the active runtime plan/admission note, not long-term memory.

File-Aware Planning

For plans that involve several files or capability assets:

1. Map each intended change to an owner file before editing.
2. Keep a per-file acceptance condition: inserted rule, removed duplicate, updated status, or verified behavior.
3. Execute in dependency order: routing/memory before dependent skill references only when needed.
4. After each file, run a focused assertion instead of waiting for a broad final sweep.
5. During phase_cleanup, update runtime plan state and admission notes, but do not write one-off process history into memory.

If the file owner is uncertain, pause mutation and use agent-capability-map or skill-architecture-governance first.

Verification Contract

A plan item is not complete until fresh evidence exists:

• File/config change → re-read changed lines or run syntax/structure checks.
• Script/plugin change → compile/check and, if safe, run the minimal command or reload after confirmation when required.
• Media task → query the relevant MoviePilot state after action.
• Job task → verify frontmatter status, schedule, last_run behavior, and command result.

Failure Handling

If execution fails:

1. Report the exact command/tool error, not a vague “工具不可用”.
2. Try one smaller diagnostic path if safe.
3. If still blocked, state the blocker and the smallest user decision needed.

Output Contract

Final response should include:

• what was executed;
• what verification passed;
• what remains blocked, if anything;
• the next safe handoff such as repository sync when relevant.