What shipped: more agents, and agents that ask before risky steps
Gemini CLI and Codex now have native Pushary hooks, and setup writes a proactive-ask instruction so they volunteer questions over push.
Two things landed in agent-hooks. Gemini CLI and Codex are now first-class, and setup teaches both of them to ask you before a risky step instead of waiting to be stopped at a gate.
More agents, wired natively
Gemini CLI connects through MCP plus native hooks. Codex gets a native hooks adapter with auto-trust and a file-based key, so it works without you pasting credentials into a config by hand. Both join Claude Code, Cursor, and Hermes on the same CLI hook path, which means the same permission policy and the same audit trail apply no matter which agent is running. See Codex notifications for the Codex setup, or the quickstart to connect any of them.
Why "ask before" needed a fix
A hook is a gate. It fires when an agent reaches for a tool, and Pushary decides allow, ask, or deny. That covers shell commands and file writes well. It does not cover the cases where the agent should pause and ask a judgment question on its own, like "which of these two refactors do you want" or "this migration will drop a column, proceed?" Those only happen if the agent decides to ask.
Codex and Gemini were not volunteering those questions, because nothing told them they could. The hook gates them, but it never drives the asking.
What changed
Setup now writes a short use-Pushary instruction into the file each CLI already loads on startup: AGENTS.md for Codex, GEMINI.md for Gemini CLI. The instruction tells the agent to send a push and wait for your answer before anything risky or irreversible, and to notify you when a long task finishes. So now you get both layers: the hook still enforces your permission policy, and the agent also raises its own questions to your phone.
The read-only safe floor still applies underneath. Proven read-only commands like ls, cat, and git status auto-approve so you only hear about decisions that matter. That floor was set from 1,721 real production questions, not a guess.
Turning it on
Run setup or doctor and it writes the instruction file for you. Existing installs pick it up on the next setup run. If you manage AGENTS.md or GEMINI.md yourself, the block is additive and you can edit the wording.
Every question and answer lands in the audit log with its answer source, so a teammate or an auditor can see what was asked and how it was decided. More on the agent surface at Pushary for AI agents.
Frequently asked questions
Do Gemini CLI and Codex support Pushary notifications?
Yes. Gemini CLI connects through MCP plus native hooks, and Codex gets a native hooks adapter with auto-trust and a file-based key. Both join Claude Code, Cursor, and Hermes on the same CLI hook path, so the same permission policy and audit trail apply.
How do I make Codex or Gemini CLI ask before a risky step?
Pushary setup writes a short use-Pushary instruction into the file each CLI loads on startup, AGENTS.md for Codex and GEMINI.md for Gemini CLI, telling the agent to send a push and wait for your answer before anything risky or irreversible. Run setup or doctor and it writes the file for you; existing installs pick it up on the next setup run.
What is the difference between a hook gate and the proactive-ask instruction?
A hook is a gate that fires when an agent reaches for a tool and enforces your permission policy with allow, ask, or deny. The proactive-ask instruction is separate: it tells the agent to volunteer judgment questions on its own, like which refactor to use, so you get both enforced gating and agent-raised questions on your phone.