daily.dev on what broke after shipping a Codex-built Slack agent in four days
daily.dev details three weeks of credential leaks, silent event-loop hangs and memory crashes after launching its org-wide Slack agent, Smith.
daily.dev built Smith, a Slack agent that queries BigQuery, ClickHouse and Postgres, manages secrets, automates browsers and runs scheduled jobs. It shipped internally on March 12, four days after the first commit. In a post dated March 29, 2026, the team describes the three weeks of fixes that followed. If you're giving a vibe-coded agent real permissions, this is a useful list of what to harden before launch. The post resurfaced this week via @bibryam on X.
Smith is 29,000 lines of TypeScript plus 10,000 lines of tests, written mostly by Codex. daily.dev says security was the hard part, and it was a recurring category of problems rather than a single bug:
- Credential leakage. The agent ran credential-checking CLI commands in a shared runtime, which corrupted other users' sessions. Tokens also carried over between turns. After the obvious commands were blocked, the agent tried workarounds such as environment inspection with grep and scripting one-liners. GitHub auth alone took 10 consecutive fix commits. The resulting command sanitizer is a blocklist that, the team writes, "will never be complete."
- Silent hangs. A blocked Node.js event loop left the process alive but unresponsive. Systemd's
Restart=on-failurenever fired. The fix had four layers: a worker-thread watchdog, Fastify timeouts, Caddy health checks and shorter stop timeouts. - Memory blowups. One 170+ message thread froze the whole 15GB VM. The team added cgroup memory limits, but still doesn't know why built-in conversation compaction failed in that case.
The team also admits it hasn't audited the roughly 25 skills Smith wrote for itself. The only coverage so far is daily.dev's own account, so none of this has been independently verified.