The flagged behavior is conditional on the user starting watch-sandbox, whose stated purpose is to inject memory-based course corrections. The token is sent to the Ellipsis API for that session action, while transcript context goes to the package’s own memory API; neither flow requires a block under the stated rules.
Static reason
High-risk behavior combination matched malicious policy.
Trigger
Explicit invocation of watch-sandbox or --watch-sandbox; the Ellipsis send additionally requires a session ID, token, and novel intervention rule.
Impact
Enables transcript-based memory interventions and course corrections in the user's session; no unrelated credential recipient was identified.
Mechanism
The watcher submits transcript context to the SlashML memory API. For a novel intervention, it conditionally posts the intervention message to the Ellipsis session API using an Ellipsis token.
Network endpoints2
api.slashml.aiapi.ellipsis.dev (overridable with ELLIPSIS_API_URL)