learning-engine
AdvisoryAudited by Static analysis on Apr 30, 2026.
Overview
No suspicious patterns detected.
Findings (0)
Artifact-based informational review of SKILL.md, metadata, install specs, static scan signals, and capability signals. ClawScan does not execute the skill or run runtime probes.
A generated or mistaken rule could silently change how other skills behave in future tasks.
This instructs the agent to modify installed skill instruction files rather than only report suggestions, with no described approval, diff review, backup, skill allowlist, or rollback.
Auto-add learned rules to relevant skill SKILL.md ... Location: `skills/{skill-name}/SKILL.md`Require explicit user approval and a visible diff before editing any skill file; limit edits to user-selected skills and keep backups for rollback.
Private operational history or bad lessons could be stored and reused as future instructions.
The skill turns persistent logs, evaluations, and performance data into reusable rules. Those sources may contain sensitive details or untrusted/incorrect content, and no validation or retention controls are described.
Learning Sources ... `memory/errors/` ... self-eval Results ... performance Data ... Convert learned patterns to rules ... `memory/learned-rules/`
Keep learned rules separate from executable skill instructions until reviewed; add source filtering, retention limits, confidence scoring, conflict checks, and user approval.
One incorrect lesson could affect several workflows and keep influencing the agent after the original task is over.
The pipeline propagates one learned pattern into memory, skill files, events, and reports, so a bad inference can spread across future sessions and multiple skills.
Extract patterns + Create rules → Save to memory/learned-rules/ → Auto-update relevant skill SKILL.md → Publish event
Add containment: stage changes as proposals, apply them one skill at a time, validate rules before propagation, and provide easy rollback.
The agent may continue generating reports or updating learning state outside a direct user request if connected to a hook engine.
The skill describes autonomous hook-triggered and scheduled activity, but the artifacts do not define explicit opt-in, disable, scoping, or review controls for those recurring actions.
hook-engine Integration ... on-error hook ... post-hook ... scheduled hook: Every Monday → Generate weekly learning report
Only enable hooks after explicit user consent, document how to disable them, and require review before any hook-triggered skill edits.