Exercise --steer against a running Gemini CLI

[Chirag6722]

--steer on a Gemini CLI session is derived from the source (google-gemini/gemini-cli: docs/hooks/reference.md, packages/core/src/core/client.ts, hooks/types.ts) and validated against a fixture. Nobody has run it against a real Gemini CLI yet, and the README says so. This issue is that run — no code unless something disagrees.

What agit claims (src/steer.ts header, README "--steer")

• AfterAgent: a blocking decision whose reason is sent to the agent as the next prompt, history kept — the "agent just finished" case.
• BeforeAgent: hookSpecificOutput.additionalContext appended to the prompt for that turn — the idle case.
• The hook's session_id is the recording's own session id, which is what the adapter reports and what the steer inbox is keyed by.

Steps

1. npm install -g agitsh; in one terminal agit relay.
2. In a project, agit hook --config gemini-cli and add the fragment to Gemini CLI's settings (agit must be on PATH).
3. Start a Gemini CLI session; find its recording under ~/.gemini/tmp/<project>/chats/session-*.jsonl.
4. agit share <that file> --steer — copy the link and the steer key.
5. From another terminal, agit steer <link> "<message>" --steer-key <key> (or type it on the share page with the key), once while the agent is idle and once while it is mid-turn.

Report

• Gemini CLI version.
• Idle case: did the message ride along with your next prompt? Finished case: did the agent continue with the message as its next prompt, with history kept?
• What the agent said it did (it is asked to say), and whether the delivered text matched what agit hook prints for the same event.
• Anything surprising: an escaped character, a message delivered twice, a hook timeout.

If reality disagrees, the fix goes in src/steer.ts (and the README paragraph), with a test in test/steer.test.ts. Sanitize anything you paste — a recording holds whatever the agent saw — and never attach a raw one.

[Chirag6722]

Partial groundwork from this machine, for whoever takes it. Gemini CLI 0.55.1 is installed here, and its bundle confirms the channel agit relies on: AfterAgent / BeforeAgent hook events, hookSpecificOutput.additionalContext read through getAdditionalContext(), and the hook's session_id set from config.getSessionId() — the value the recorder writes as the recording's sessionId. One detail worth knowing before you set it up: hooks declared in a project .gemini/settings.json are skipped unless the folder is trusted (hookConfig.source === Project && !isTrustedFolder()), so either trust the folder or put the fragment in ~/.gemini/settings.json.

I could not run it to the end: this machine has no cached Gemini login and gemini -p stops at the browser sign-in prompt, which is a step for a person. A one-file stub that logs the hook's stdin and answers BeforeAgent with a marker in additionalContext (and AfterAgent with a one-time decision: block) is a quick way to check both claims before wiring the real agit hook; asking the model to repeat the marker shows whether the context reached it.