Self-learning & skill evolution
What this is for. Two operator surfaces sit behind this chapter. Self-Learning Insights (/self-learning) reports how runs are spending tokens and what patterns keep showing up. Skill Evolution (/skill-evolution) is the human gate for AI-suggested new skills. Neither writes behaviour until you approve. Enable the related autonomy loops only if you want background proposals — they make paid model calls.
Routes: /self-learning (sidebar Self-learning), /skill-evolution.
When to use it
Section titled “When to use it”- You want a weekly efficiency picture: tokens, cost, sessions, success rate.
- Repeated work looks like it should become a skill, and you want the suggestion — not a silent auto-write.
- Skill Evolution has pending candidates and you need to Approve or Reject them.
- You already use Forge for identity/soul changes and want the skill-side equivalent.
Typical workflow
Section titled “Typical workflow”- Open Self-learning (
/self-learning). Read the four summary cards, then Execution Insights, Activity Patterns, and Skill Suggestions. - Press Run Analysis when you want a fresh pass (
POST /self-learning/analyze). - Open Skill Evolution (
/skill-evolution) for candidates with Pending / Approved / Rejected. - Expand Show details, read reasoning and suggested content, then Approve (adopts into the catalogue, still subject to the auto-adoption gate) or Reject.
- Confirm the new skill under Skills → Inventory.
Features
Section titled “Features”| Area | Meaning |
|---|---|
| Insights | Patterns learned from usage |
| Skill suggestions | Names + reasons from the insights page — not yet candidates |
| Skill evolution | Proposed skill files with content, confidence, and session count |
| Review / apply | Human gate before behaviour changes |
Prefer reviewing Forge/skill proposals before apply. Generated skills still pass the Skills auto-adoption gate.
Which model writes them. Self-learning suggestions, Forge description proposals and skill authoring for Skill Evolution run on EYAS’s background model, one isolated call each: the Heartbeat routing tier (primary, then fallback), then the install default, then API providers, then CLIs that can run isolated calls. They never go to a provider the gateway picks on its own. When no model qualifies (for example a Grok-only or Kimi-only install before their isolation is verified) or the budget is at stop, no model call is made: Self-learning shows its generic suggestion sentences, Forge keeps the concatenated description proposal, and Skill Evolution writes the template SKILL.md. The Autonomy feature flags still gate these calls. See Routing & budget — The background model.
Fields and controls
Section titled “Fields and controls”Self-Learning Insights (`/self-learning`)
Subtitle: Efficiency reports, activity patterns, and optimization suggestions.
| Control | Meaning |
|---|---|
| Run Analysis | Trigger a fresh analysis pass |
| Total Tokens | Tokens in the weekly report |
| Total Cost | USD in the weekly report |
| Sessions | Session count |
| Success Rate | Share of successful sessions |
Execution Insights — type badges Optimization / Cost / Quality / Speed, plus current vs suggested metric, confidence, and reasoning. Empty: No insights yet. Run an analysis to generate them.
Activity Patterns — name, category, seen N×, last seen. Empty: No activity patterns detected yet.
Skill Suggestions — name, description, reason, confidence. Empty: No skill suggestions yet. These are suggestions; they become reviewable candidates on Skill Evolution.
Skill Evolution (`/skill-evolution`)
Subtitle: Review AI-suggested skill candidates based on usage patterns.
| Control | Meaning |
|---|---|
| Stats | Pending / Approved / Rejected counts, Avg Confidence |
| Search | Search candidates… |
| Filter All / Pending / Approved / Rejected | Status filter |
| N sessions | How many sessions the suggestion is based on |
| Show details / Hide details | Reasoning + suggested skill content |
| Approve / Reject | Human decision; pending only |
Empty: No skill evolution candidates yet. The system will suggest new skills based on your usage patterns.