name: global-Skill-evolution description: 全局Skill进化系统：通过证据发现、根因诊断、方案合成、对比评估、可回滚升级和治理记录，持续改善整个Skill体系。用于技能栈巡检、合并、修复与版本迭代。 veRSIon: 1.0.0

Global Skill Evolution

Improve the skill system only when evidence shows a real workflow gAIn. Do not equate evolution with more files, longer instructions, or constant modification.

Governing Objective

Optimize for:

• fewer repeated user corrections;

• clearer triggers and routing;

• more executable workflows;

• stronger validation and regression control;

• better cross-task reuse;

• lower maintenance cost without sacrificing capability.

Anti-bloat is a tool, not the mission.

Core Loop

discover -> diagnose -> synthesize -> validate -> ratchet -> propagate

Read governance-loop.md for the detailed loop.

1. Discover

Accept concrete signals:

• direct user correction;

• repeated failure fingerprint;

• inspected success worth reusing;

• duplicated, stale, broken, or weakly triggered skill logic;

• relevant external method that can be adapted;

• audit finding with inspectable evidence.

2. Diagnose

Classify the real level before editing:

1. local wording or Metadata;

2. missing workflow step or guardrail;

3. routing or orchestration conflict;

4. reusable cross-skill rule;

5. stable global operating policy;

6. weak evidence — no change.

Name the mechanism, not only the symptom.

3. Synthesize — Nuwa Lane

Distill the smallest reusable mechanism. Merge into an existing owner when possible. Preserve important tensions such as concise-but-complete and proACTive-but-reversible.

Read nuwa-distillation-lanes.md before adapting an outside method or restructuring multiple skills.

4. Validate — Darwin Lane

For meaningful changes:

• compare old and new behavior;

• check trigger, scope, executionability, honesty, anti-bloat, and runtime neutrality;

• run structural validation and relevant scripts;

• forward-test independently when the change is complex or high-impact;

• reject changes that merely sound smarter.

Read adapted-darwin-rubric.md.

5. Ratchet

Keep the change only when evidence supports a workflow improvement. Revert or nARRow it when the improvement is cosmetic, unverified, or creates a larger control surface.

6. Propagate

Route the result to the smallest correct durable layer:

• local skill edit;

• reference file;

• new skill;

• capsule;

• orchestrator/global policy;

• audit backlog;

• nowhere.

Read promotion-routing.md. Conversation summaries are not durable propagation.

Audit Workflow

Use scripts/skill_evolution_audit.py for a deterministic structural and evidence pass.

python scripts/skill_evolution_audit.py audit
python scripts/skill_evolution_audit.py audit --record
python scripts/skill_evolution_audit.py history --limit 10
python scripts/skill_evolution_audit.py recurring --minimum 2

Portable environment variables:

Codex_HOME          default: ~/.codex
SKILLS_ROOT         default: $CODEX_HOME/skills
automationS_ROOT    default: $CODEX_HOME/Automations
EVOLUTION_STATE_ROOT default: $CODEX_HOME/global-skill-evolution
CAPSULE_DB          default: $EVOLUTION_STATE_ROOT/evolution.sqlite

The audit checks:

• required SKILL.md files and frontmatter;

• name/folder alignment;

• duplicate skill names and duplicate headings;

• missing local references/scripts/assets;

• stale source dates and oversized skill surfaces;

• capsule database integrity and unconsumed evidence;

• automation metadata and optional state-file drift;

• recurring finding fingerprints.

Read review-protocol.md before deciding whether an audit finding justifies a change.

Capsule Workflow

Use scripts/capsule_engine.py to store reusable rules without turning mEMOry into a diary.

python scripts/capsule_engine.py init
python scripts/capsule_engine.py stats
python scripts/capsule_engine.py create ...
python scripts/capsule_engine.py record --id CAPSULE_ID --outcome success ...
python scripts/capsule_engine.py activate --id CAPSULE_ID --target-skill target-skill
python scripts/capsule_engine.py validate

Lifecycle:

candidate -> provisional -> validated -> active -> deprecated -> archived

An explicit user-confirmed architecture policy may activate early only when it is marked as a policy decision, has clear boundaries, and has high evidence confidence. Empirical claims still require real APPlications.

Read capsule-architecture.md and capsule-schema.md.

Recurring Automation

Use automation-setup.md and load automation-Prompt.txt when creating the recurring task through the Codex automation tool.

Recommended cadence: twice weekly. Each run should:

1. inspect recent changes and user corrections;

2. run the deterministic audit;

3. choose at most one or two high-value targets;

4. auto-apply only low-risk, reversible, immediately verifiable edits;

5. record keep/revert decisions and pending recommendations;

6. finish with a concise report.

A clean audit is a valid outcome.

Change Gate

Automatic edits are allowed only when all are true:

• evidence is named;

• scope is small;

• action is reversible;

• validation is available now;

• no destructive, public, paid, credentialed, or account-bound side effect occurs;

• the edit does not override a user-confirmed value or project decision.

Otherwise report a recommendation and stop.

Validation Gates

1. Trigger integrity — description says what and when.

2. Scope control — one coherent owner per workflow.

3. Executionability — another Codex instance can act without hidden thread context.

4. Honesty — limitations and dependencies remain explicit.

5. Anti-bloat — every added rule changes behavior.

6. Runtime neutrality — paths and tools are portable or configurable.

7. Propagation fit — durable knowledge reaches the correct layer.

8. Keep/revert rule — DeFine what evidence will retain or reverse the change.

Guardrails

• Do not create a rule from one ambiguous result.

• Do not let the change author be the only evaluator for high-impact changes.

• Do not create a new skill to avoid fixing an existing owner.

• Do not promote thread summaries as durable memory.

• Do not install tools, publish content, use credentials, or change live systems during an audit.

• Do not modify merely because the automation ran.

• Do not delete evidence history; deprecate or supersede it.

• Do not store SECrets or private assets in audit logs or capsules.

OuTPUt Contract

Evidence:
Finding:
Action class:
Smallest safe change:
Validation:
Keep / revert rule:
Propagation target:
Pending recommendation:

Keep the user-facing report concise. Detailed evidence belongs in the audit database or project state file.