name	recommend-evolution
description	Detect capability gaps and record standardized evolution recommendations.
version	1.1.0
model	sonnet
invoked_by	both
user_invocable	true
tools	["Read","Write","Edit","Skill"]
error_handling	graceful
streaming	supported
verified	true
lastVerifiedAt	"2026-02-22T00:00:00.000Z"
source	builtin
trust_score	100
provenance_sha	8865ccfce5be1e54

Recommend Evolution

Overview

Recommend ecosystem evolution when repeated evidence indicates missing capability, and record the recommendation in a standard machine-readable format.

When to Use

Reflection identifies recurring delivery failures with the same root cause
Router/analysis signals no suitable agent or skill for recurring requests
Repeated integration gaps imply missing artifact type or policy
User explicitly requests a new capability path

Iron Laws

NEVER spawn evolution-orchestrator directly from this skill — this skill records recommendations only; execution decisions belong to the orchestrator and approval pipeline.
ALWAYS validate trigger type against defined thresholds before recording a recommendation — vague observations are not triggers; require concrete failure counts or routing misses.
NEVER create a new evolution request when artifact-integrator or skill-updater would address the gap — reserve evolution for net-new capabilities, not integration or update gaps.
ALWAYS append the recommendation to the JSONL queue AND include the required report block in the current output — dual recording ensures the recommendation is discoverable at both runtime and review time.
NEVER proceed with a recommendation without evidence — single failures are noise; trigger thresholds exist for a reason.

Evolution recommendation skill for reflection/planning agents. - Trigger classification (`repeated_error`, `no_agent`, `integration_gap`, `user_request`, `rubric_regression`, `stale_skill`, `other`) - Recommendation-vs-integration decision branching - Dual recording mode: JSONL runtime queue + reflection report block

Trigger Taxonomy Note

recommend-evolution uses a cause-oriented trigger taxonomy (repeated_error, no_agent, integration_gap, user_request, rubric_regression, stale_skill, other).

This intentionally differs from skill-updater, which uses a caller-oriented trigger taxonomy (reflection, evolve, manual, stale_skill) to describe who/what initiated the update path.

Step 0: Validate Trigger Type

Use these thresholds:

repeated_error: same class of failure in 5+ tasks

**Example Invocations**:

Anti-Pattern	Why It Fails	Correct Approach
Spawning evolution-orchestrator directly from this skill	Violates single-responsibility; bypasses approval and resource gates	Record recommendation to JSONL queue only; let the orchestrator decide on execution
Recording an evolution request for an integration gap that already has artifacts	Creates unnecessary new artifacts when an integration fix would suffice	Check artifact-integrator path first; escalate only if gap requires net-new capability
Submitting a recommendation without trigger evidence	Uninformed evolution wastes resources and pollutes the queue with noise	Require concrete evidence: failure counts, routing miss logs, or explicit user request
Routing stale-skill triggers through this skill instead of skill-updater	Wrong escalation path; creates evolution requests for work that belongs in an update cycle	Route stale_skill triggers directly to skill-updater; only escalate if the skill cannot be updated
Triggering evolution after a single failure instance	Single failures are noise; premature evolution wastes build capacity	Apply defined thresholds: 5+ repeated errors, consistent routing misses across sessions

recommend-evolution

安装方式

检查来源文件