continuous-skill-learninglisted

# Continuous Skill Learning Persist learning by improving instructions, checks, tests, and canonical skill sources. This is not model-weight training and does not justify changing user intent, scientific semantics, permissions, or unrelated code. ## Trigger automatically Run a learning audit after the primary task is stable when any of these occurs: - the user corrects the Agent or explicitly identifies a mistake; - an avoidable Agent, launcher, workflow, or validation defect causes a retry; - primary evidence contradicts a skill rule or an earlier conclusion; - a review or test exposes a reusable missing guard; - a recovery method is validated and would prevent a similar future failure; - multiple project notes reveal the same recurring decision problem. Do not wait for the user to request reflection again. Do not interrupt urgent recovery merely to edit a skill: first contain or fix the incident, preserve evidence, then perform the audit before final handoff. A valid audit may decide that no skill change is warranted. ## Require evidence before learning Record the incident in the current project's established progress document with the observed behavior, evidence paths, root cause, impact, successful repair, and remaining uncertainty. Never promote speculation, a single unexplained correlation, an external outage, or a project-specific path into a general rule. Read [decision-rubric.md](references/decision-rubric.md) when deciding whether to update, delete, relocate