Learning Loop

How Org in a Box automatically learns from agent behaviour and proposes reusable skills.

Overview

After every agent session that reaches idle status, the reflection plugin:

Loads the full session transcript
Extracts a tool trajectory (ordered sequence of tool calls with input/output summaries)
Hashes the trajectory using SHA-256
Generates a reflection via LLM: "What worked? What failed? Any reusable patterns?"
Stores the trajectory and reflection in the trajectories table with a pgvector embedding
Checks for similar past trajectories (cosine similarity > 0.8, tool overlap ≥ 70%)
When 3+ similar trajectories exist, auto-proposes a skill

Trajectory Hashing

Two trajectories are "similar" if:

Their tool sequences overlap by ≥ 70% (Jaccard-style: |intersection| / |max(A,B)|)
Their embeddings have cosine similarity > 0.8

This catches the same task approached slightly differently (different tool order, minor phrasing variations) without treating every unique task as a new pattern.

Skill Lifecycle

proposed → promoted → deprecated

Status	Meaning
`proposed`	Auto-created by reflection plugin; not yet active
`promoted`	Admin approved; available to all agents in the tenant
`deprecated`	Retired; no longer shown to agents

Transition rules: proposed → promoted, proposed → deprecated, promoted → deprecated. No backwards transitions.

Managing Skills

Via admin API

GET /v1/skills?status=proposed
PATCH /v1/skills/:id/promote
PATCH /v1/skills/:id/deprecate

Via the TUI

/skills proposed     # list proposed skills awaiting review
/skills promoted     # list active skills

In the web dashboard

The admin panel Skills page shows proposed skills with their source trajectories. Click Promote to make a skill available to agents, or Deprecate to retire it.

What a Skill Contains

{
  "name": "competitive-analysis",
  "description": "Research 3+ competitors and produce a structured comparison",
  "body": "# Skill: Competitive Analysis\n\n## When to use\n...\n\n## Steps\n...",
  "triggers": ["compare competitors", "competitive analysis", "market research"],
  "status": "promoted",
  "version": 1
}

Promoted skills are written to SKILL.md files in the sandbox so opencode can reference them during planning.

Storage

Table	Description
`trajectories`	One row per session; `tools` JSONB, `embedding vector(1536)`, `reflection` text
`skills`	Versioned skill definitions; `source_trajectories` array of trajectory IDs

Disabling

Set OIAB_REFLECTION_DISABLED=true to skip the reflection hook for a session, or remove @orginabox/plugin-reflection from opencode.jsonc.