Gemini CLI¶

Google's Gemini CLI. Hooks and MCP entries land in ~/.gemini/settings.json.

What gets installed¶

Both hooks and the MCP server live in the same settings.json. Repowire appends to existing arrays without clobbering user-defined entries.

Hooks¶

Gemini CLI uses different event names than Claude Code or Codex:

Gemini event	Mapped to	What it does
`SessionStart`	session hook	Registers the peer with the daemon
`BeforeAgent`	prompt hook (`UserPromptSubmit` equivalent)	Marks peer `busy`
`AfterAgent`	stop hook (`Stop` equivalent)	Captures response, fetches pending asks, marks peer `online`

The native normalizer in daemon-go/hooks/common.go maps these to canonical names so the handler code does not branch on runtime.

Hook output¶

Gemini hooks require an explicit decision in their JSON output. Repowire emits {"decision": "allow"} from AfterAgent so the agent continues normally. Claude Code and Codex Stop hooks emit a decision only when blocking to resurface an open ask.

MCP server¶

The Repowire MCP identity shim is added under mcpServers.repowire in the same settings.json, runs as repowire mcp over stdio, sets REPOWIRE_BACKEND=gemini, and proxies tools to the daemon's /mcp endpoint.

Response field¶

Where Claude Code dumps a transcript JSONL that the stop hook parses, and Codex emits last_assistant_message, Gemini emits prompt_response. The adapter pulls the right field per runtime — you don't need to think about this unless you're patching hooks.

Verifying¶

repowire status

Open a Gemini session, send one message, watch repowire peer list. Gemini fires hooks promptly on first prompt.

Troubleshooting¶

Hook errors visible in Gemini output → confirm repowire is on PATH for the shell Gemini was launched from. Gemini surfaces hook stderr; Claude Code and Codex tend to swallow it.
Peer stuck busy → the AfterAgent hook didn't fire or didn't return {"decision": "allow"}. Lazy repair can reset stale busy/working state after daemon.stale_busy_timeout_seconds; see Ghost peers and stuck busy state.