Method, tools and authority
A skill describes how to work. A Tool provides an action; a connection configures access. A model supplies inference and a harness runs a Task. Instructions alone cannot grant permission to use a service.
Inspect and administer procedures
A skill package contains SKILL.md and optional scripts, templates or references. Its files and source are visible. Assign relevant procedures to agents and review the effective authorisations before changing a category or global rule.
Dream skill learning is disabled by default. Its mode and evidence thresholds can be configured in administration.
Skills for understanding and evaluating Galaris
Galaris Knowledge guides search and reading in the bundled official documentation. Galaris Lab guides evaluation preparation and analysis. Both skills are assigned individually when the Galaris assistant is created; they remain globally disabled for other agents.
Their tools have separate permissions. In particular, the Lab connection remains inactive until you enable it: having a procedure alone does not grant permission to run an experiment.
Use a skill for repeatable work
A document review procedure can specify source checks, output structure and acceptance criteria. It should help the agent perform a recurring job, while task-specific facts remain in the mission context. Learned procedures carry separate evidence and scores.
Administer a reusable library
Authorised skills are synchronised before a fresh execution on managed Codex, Claude Agent, DeepSeek Harness and Hermes runtimes. Edits and revocations apply at the next start without depending on a UI refresh. An ongoing remote continuation keeps its environment until a fresh execution.
Import a package, browse files, read its Markdown and download its archive. Categories organise the library; scanning finds valid directories and flags those needing correction. System skills remain protected.
Permissions combine global activation, category rules and per-agent overrides. The matrix shows effective state and affected agents. Compatible harnesses receive authorised procedures.
Dream-learned skills form a separate collection with evidence and scores. A procedure can be suspended or reactivated. Understand learning
Read permissions and inheritance
The matrix distinguishes Global (all), Category and This agent, with Active or Blocked choices. The skill’s agent rule takes precedence, then its category rule, then the global setting. An unselected column may indicate an inherited rule or one masked by an override.
Skill and file changes are applied before the next new task on a compatible managed harness. A running execution keeps its context; failed projection prevents starting with a stale copy. Choose models and harnesses.


