Post by Chris Nighswonger

Claude Code v2.1.169 shipped today (June 9). Among the changes: a previously-dormant code path is now armed. A section of the agent's system prompt that returned null in v2.1.168 now returns a compiled-in string starting with "You are operating autonomously. The user is not watching in real time..." — gated by a GrowthBook flag (tengu_amber_sextant, compiled default true) and a server-controlled model-identifier substring selector (tengu_velvet_falcon_model, compiled default empty). On a fresh install the path is dormant. The moment Anthropic sets a non-empty selector server-side, it activates for every operator running a matching model — no CC release, no changelog entry, no notification. This is the third widening in eight days of the binary-level surface through which Anthropic's servers can influence the local agent's prompt. The June 1 post (Three Claude Code releases, one surface) traced v2.1.150 and v2.1.158/159. v2.1.169 widens it again with a different shape: server-supplied targeting for compiled-in content. The string is in the binary; the flag channel decides whether and for which models. A few things worth flagging: — The operator-side override exists but is undocumented. CLAUDE_CODE_VELVET_FALCON="" (literal empty string) disables it. Don't set it to "0" or "false" — those are non-empty strings, and JS truthiness will force the predicate true instead. — No log line, no statusline indicator, no /config surface. From inside the session, there's no observable signal that the autonomy-append is engaged. — The content surface is narrower than the v2.1.150 and v2.1.158 channels (which could carry arbitrary server-supplied text). v2.1.169's path carries only the string Anthropic compiled in. Same trust frame; narrower content surface. Worth saying out loud, not eliding. We're not filing this as a CC issue — the HackerOne disposition on our v2.1.150 disclosure (closed Informative on May 27) established Anthropic's position on this class of surface. The post itself is the record. Reproducible from the public npm tarballs at @anthropic-ai/[email protected] and @2.1.169. Writeup: https://lnkd.in/e-h_EU4b #ClaudeCode #Anthropic #AI #DevTools #Security #BinaryAnalysis #Transparency #LLMOps #PromptEngineering #VSITS