Problem

There are two independent skill auto-creation systems that confuse users and produce poor results separately:

1. auto_evolve / background_skill_review — LLM reviewer runs after each turn. Generates good semantic content BUT:

   • 30s timeout too short for local models (Ollama/qwen) — most reviews time out
   • Review prompt is too conservative — almost always returns skip
   • Creates skills directly without user approval

2. skill_workshop (PR feat(skill-workshop): closes #3328 — passive after-turn capture pipeline #4741 / Skill workshop: capture and refine reusable skills from successful agent interactions #3328) — Heuristic scanner detects repeated tool patterns, user corrections, explicit instructions. Has approval flow (pending drafts) BUT:

   • RepeatedToolPattern scanner generates useless drafts like tool_sequence_file_read_x5 with just tool names listed
   • No semantic understanding of what the user was doing

From a user perspective, these should be one system: the agent learns from interactions and proposes skills for human review.

Proposed fix

1. Increase background_skill_review timeout from 30s to 120s (local models need more time)
2. Lower the review threshold — add criteria like "3+ different tools in sequence" and "multi-step procedure that could be reused"
3. Route create actions through skill_workshop::storage::save_candidate instead of evolution::create_skill — so LLM-proposed skills land in the Pending tab for human approval instead of being created directly
4. Keep update/patch actions as direct operations (modifying skills the user already approved)

Result

• One unified flow: agent turn → LLM review → pending draft with semantic content → human approval
• No more useless tool_sequence_* drafts — the LLM generates meaningful names, descriptions, and procedures
• Users see proposed skills in the Pending tab and decide what to keep

Refs: #3328