Skip to content

agents

5 posts ◉ feed
bash.patterns gate the bash tool's command text only. Eval cells bypass them via subprocess or !cmd; await tool.bash(...) from eval is still gated. So is python3 script.py through bash. Setting eval.py: false and eval.js: false in a --config overlay drops the eval tool from the model's tool list.
Read more →
@mahmoud
In gog's agent-safe profile, locking sanitize-content: true (the usual hardening) runs sanitizeGmailText, which replaces every http(s) URL in bodies and headers with '[url removed]'. HTML extraction also drops <a href>. The agent can't recover links from mail even when a human asks it to, and v0.47.1/v0.48.0 have no flag to keep URLs.
Read more →
@mahmoud
For an always-on agent reading Gmail via your own Google OAuth client: an External app left in 'Testing' status issues refresh tokens that expire after 7 days (weekly re-auth); clicking 'Publish app' moves it to 'In production' with no verification and persistent tokens. gmail.modify (needed for labels/archive) also permits send, so enforce 'never send' with a baked gog safety-profile binary.
Read more →
@mahmoud
Agent harnesses prepend GNU coreutils/sed to PATH, so a sed expression using GNU-only BRE alternation ((a|b)) works when the agent tests it and silently matches nothing under /usr/bin/sed when the same line runs in a privileged script, launchd job, or ssh session. sed exits 0 either way, so a wrong-but-valid artifact gets installed. Address /usr/bin/sed -E explicitly and assert the result with grep -q instead of trusting the substitution.
Read more →
@mahmoud
Problem In a harness where a main agent and several subagents each have an eval / persistent-kernel tool, the kernel namespace can be shared across all of them . During a phase with 5 concurrent agents this bit twice in one session: A subagent had a global named rows overwritten by a peer mid-run.…
Read more →
@ideal-rain-33