| name | agent-evolution-governor |
| version | 1 |
| description | Use this skill when the user asks to 自检自身、重新蒸馏、进化版 Agent、增强回去、避免死循环处理、能力增益评估,或在瘦身与增强之间取舍。它先盘点已处理对象、能力风险和真正增益点,再决定增强、修复、拆分、瘦身或停止,不把行数减少当成默认目标。 |
| allowed-tools | read_file list_directory write_file edit_file execute_command task subagent_task |
Agent Evolution Governor
Purpose
Govern Agent self-improvement so it produces a stronger Agent, not a smaller but weaker one.
Use this skill before broad self-audit or capability refactor work. It decides whether to enhance, restore, split, slim, or leave an asset untouched.
Trigger Examples
Use when the user says:
- “自检自身”
- “重新蒸馏”
- “进化版 Agent”
- “增强回去”
- “可以瘦身的接着瘦身”
- “不要死循环处理”
- “不是精简版的残废”
- “检查哪些能力被削弱了”
Evolution Decision Model
Classify each candidate asset into one of five outcomes:
- Enhance
- Trigger, workflow, tool routing, validation, fallback, or safety boundary is missing.
- Recent failures show the Agent could not act, repeated work, or chose the wrong route.
- Restore
- A prior slimming pass removed operationally necessary details, examples, or guardrails.
- The skill became too thin to execute its real job.
- Split
- Essential route is mixed with large rare references.
- Keep
SKILL.md compact and move rare material to supporting files.
- Slim
- The file is bloated with examples, duplicate slogans, stale platform assumptions, or copied reference text.
- Slimming must preserve trigger, workflow, tools, safety, and verification.
- No-op
- Already verified and no new problem or capability gain exists.
- Do not reprocess just because the user said “下一步”.
Required Self-Audit Inputs
Before writing, gather only relevant evidence. For a full-Agent distillation request, run the self-check as an asset-wide program, not as a single local patch:
- recent user correction and activity context;
- inventory of
memory/, skills/, runtime/, jobs/, and scripts/;
- skill frontmatter quality and line counts;
- stale references such as unavailable platform/tool names;
- missing
allowed-tools where the workflow uses tools;
- missing verification or fallback steps;
- business-critical skills that may be too thin after slimming;
- duplicate responsibilities across memory and skills;
- assets created by the Agent that may be the wrong layer, especially scripts made for a one-off governance check.
Use subagents for independent read-only inspection when helpful, but the main Agent owns all writes and final decisions. A failed subagent does not stop the program; continue with direct local inspection.
Campaign State
For broad self-refactor campaigns, keep a short non-secret runtime note or local processed set with:
- current phase and goal;
- assets already audited, changed, verified, or intentionally skipped;
- open capability gaps and their owner skill;
- latest verification command or evidence.
Use runtime/ for current campaign anchors, not memory. Never treat activity history alone as the campaign state when deciding what to revisit.
Anti-Loop Rules
- Maintain a processed set during the current improvement campaign.
- Do not revisit an asset that was already changed and verified unless one of these is true:
- new evidence of a defect appears;
- the user explicitly asks to revisit it;
- a dependent architecture change requires a targeted update.
- Do not rank by line count alone. Rank by capability impact first, then maintenance cost.
- Stop slimming a file once it is below the practical threshold and still complete.
- If the next step is unclear, first read the campaign state and choose the highest-impact unprocessed owner; do not restart global scans.
- If the user explicitly asks to蒸馏整个 Agent or start the self-check program, broaden from the current owner to the asset-wide inventory, then return to bounded edits.
- Do not create new scripts as the default expression of self-improvement; first improve routing, skills, memory boundaries, runtime state, and verification behavior. Keep or add scripts only when the check is deterministic, repeatable, and owned by an existing workflow.
Safe Change Standard
Every change must preserve or improve:
- trigger clarity;
- operational steps;
- tool declarations;
- confirmation and safety boundaries;
- fallback behavior;
- verification evidence;
- MoviePilot Agent domain routing.
If any of these would be lost, choose restore/enhance/split instead of slimming.
Verification
After changes:
- re-read or assert the changed sections;
- check skill frontmatter, name, version, description, and tool declarations;
- scan for stale platform/tool terms;
- run any available self-audit or targeted script;
- summarize enhanced, restored, slimmed, and intentionally untouched assets separately.
Output Contract
Report:
- what was audited;
- what was enhanced/restored/slimmed/no-op;
- what waste was discarded;
- verification evidence;
- next highest-impact capability target, not merely the next longest file.