agent-growth-cycle

name	agent-growth-cycle
version	2
description	Use this skill when the Agent needs to turn experience into a repeatable growth loop: observe signals, decide whether learning is needed, distill reusable lessons, place them into the right asset layer, verify the change, and keep future behavior lighter rather than heavier. Trigger when the user asks to 主动成长、持续进化、经验闭环、复盘后改进, or when repeated failures show a missing habit rather than a one-off bug.
allowed-tools	read_file list_directory write_file edit_file execute_command

Agent Growth Cycle

Purpose

Act as the Agent's growth metabolism loop. This is not a place to store every lesson; it decides whether an experience deserves durable change and makes the smallest useful update.

Trigger

Use this skill when:

• The user asks the Agent to 主动成长、持续进化、长期能力化、复盘后改进.
• A behavior failure repeats across turns, such as over-confirming, not executing, missing verification, or misrouting skills.
• A new reusable pattern emerges that should improve future behavior.

Do not use it for a one-off task with no reusable lesson.

Growth Loop

1. Observe

   • Identify the concrete signal: user correction, repeated failure, successful pattern, or missing capability.
   • Separate fact from emotion and transient context.
   • Check context pressure: is the proposed lesson worth being loaded, routed, or retained?

2. Classify

   • Memory: only high-frequency global boundary or preference.
   • Skill: reusable workflow, checklist, routing discipline, or domain method.
   • Script: deterministic repeated action.
   • Job: scheduled or recurring task.
   • Runtime: non-sensitive local anchor, active workstream state, candidate inventory, or current-state mapping.
   • Docs/archive: historical explanation, long reference, external corpus notes, or non-runtime record.

3. Distill

   • Keep the rule operational and short.
   • Remove blame, noise, stale evidence, and one-time details.
   • Prefer improving an existing asset over creating a duplicate.
   • If evidence is ambiguous or low-value, keep the item in shadow/runtime/docs rather than promoting it to memory or routing.

4. Apply

   • Make the smallest file change needed.
   • Avoid bloating memory with low-frequency process details.
   • Preserve single control plane: growth notes may advise, but cannot become a second router, second truth source, or automatic promotion engine.

5. Verify

   • Re-read changed files or run minimal assertions.
   • Check that the new rule is discoverable by the routing layer.
   • Check owner-boundary preservation and fallback: if quality evidence is stale, duplicate, or unsafe, the safe state is no write / shadow / archive.
   • Do not claim growth unless there is evidence.

Promotion Thresholds

Promote a lesson only when at least one is true:

• it prevents a repeated failure the user has actually experienced, such as over-confirming, not executing, wrong owner, or weak verification;
• it strengthens a high-frequency MoviePilot/Agent workflow: media chain, local plugin work, repository sync, UI verification, or safety boundary;
• it reduces future context load by replacing repeated investigation with a small owner rule;
• it gives a clear proof floor for a claim the Agent often needs to make.

Reject or shadow candidates that add general cleverness but no MoviePilot owner, no trigger, or no verification path.

Anti-Patterns

Avoid:

• turning every correction into memory
• copying the user's wording as a rule without distillation
• creating a new skill when an existing one should be edited
• adding broad slogans that do not change execution
• claiming evolution without a file change or verification

Handoff

If the update changes /config/agent capability assets, recommend repository sync through moviepilot-agent-git-maintenance after verification.