Documentation
¶
Overview ¶
Package learning provides autonomous skill distillation from completed agent sessions. After a task is completed successfully, the Learner reviews what happened and, if the task is novel enough, synthesises a reusable skill following the agentskills.io standard.
Learned skills are persisted in two places:
- MutableRepository — filesystem-backed, used by the skills framework.
- Vector store — used for semantic search so the orchestrator can discover relevant skills via memory_search before starting new tasks.
Design influences:
- SkillRL (2026): hierarchical skill bank, failure lessons as first-class data
- ExpeL: cross-task experiential learning with insight extraction
- Self-RAG: relevance-scored retrieval with reflection
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
This section is empty.
Types ¶
type Config ¶
type Config struct {
// MinimumNoveltyScore is the LLM-assigned novelty score (1-10) above
// which a completed task is considered novel enough to distill into a
// reusable skill. Lower values create more skills; higher values are
// more selective. Default is 7.
MinimumNoveltyScore int `yaml:"minimum_novelty_score,omitempty" toml:"minimum_novelty_score,omitempty"`
}
Config tunes the post-session skill distillation pipeline.
func DefaultConfig ¶
func DefaultConfig() Config
DefaultConfig returns a Config with sensible defaults.
type ILearner ¶
type ILearner interface {
Learn(ctx context.Context, req LearnRequest) error
}
ILearner is the interface for the skill distillation pipeline. The orchestrator depends on this interface so background learning can be faked in unit tests.
type LearnRequest ¶
type LearnRequest struct {
// Goal is the original user question / task.
Goal string
// Output is the final answer / result from the agent.
Output string
// ToolsUsed lists the names of tools that were called during execution.
ToolsUsed []string
// ToolTrace is a compressed summary of tool calls and their outcomes.
// Unlike ToolsUsed (which only has names), this captures what each tool
// returned, enabling the distillation LLM to accurately describe what
// worked and what didn't. Inspired by SkillRL's failure lesson extraction.
ToolTrace string
}
LearnRequest contains the context needed for a learning attempt.
type Learner ¶
type Learner struct {
// contains filtered or unexported fields
}
Learner reviews completed sessions and distills reusable skills.
func NewLearner ¶
func NewLearner( exp expert.Expert, skillRepo skills.ISkillWriter, vectorStore vector.IStore, auditor audit.Auditor, cfg Config, ) *Learner
NewLearner creates a Learner backed by the given expert, skill repository, vector store (for discoverability), and auditor. cfg.MinimumNoveltyScore controls novelty filtering; zero-value fields fall back to DefaultConfig.
Source Files
Directories