mirror of
https://github.com/garrytan/gstack.git
synced 2026-05-06 21:46:40 +02:00
2137417f63
The sidebar-agent now scans every Claude stream event for the session's canary token before relaying any data to the sidepanel. Channels covered (per CEO review cross-model tension #2): * Assistant text blocks * Assistant text_delta streaming * tool_use arguments (recursively, via checkCanaryInStructure — catches URLs, commands, file paths nested at any depth) * tool_use content_block_start * tool_input_delta partial JSON * Final result payload If the canary leaks on any channel, onCanaryLeaked() fires once per session: 1. logAttempt() writes the event to ~/.gstack/security/attempts.jsonl with the canary's salted hash (never the payload content). 2. sends a `security_event` to the sidepanel so it can render the approved canary-leak banner (variant A mockup — ceo-plan 2026-04-19). 3. sends an `agent_error` for backward-compat with existing error surfaces. 4. SIGTERM's the claude subprocess (SIGKILL after 2s if still alive). The leaked content itself is never relayed to the sidepanel — the event is dropped at the boundary. Canary detection is pure-string substring match, so this all runs safely in the sidebar-agent (non-compiled bun) context. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>