Set up TRW for Gemini CLI

Prerequisites

• Gemini CLI installed and authenticated.
• Python 3.10+ (3.11+ recommended) and pip 22+.
• A git repository you can write to — TRW writes repo-local files only.

Install

curl -fsSL https://trwframework.com/install.sh | bash

When the installer prompts for target clients, select Gemini CLI. The bootstrap installs trw-mcp and writes the Gemini-native surface into the repo.

What the installer writes

• GEMINI.md — repo-wide instructions with the TRW ceremony protocol. Gemini reads it on startup, so the tool lifecycle is in context from the first turn.
• .gemini/settings.json — registers trw-mcp under mcpServers via a JSON deep-merge (only the trw entry is touched).
• .gemini/agents/ — trw-*.md role-based subagent definitions translated from the TRW roster.
• .trw/config.yaml — repo-local TRW config listing Gemini in target_platforms.

Run your first session

From the project root, start Gemini the way you normally would. It loads GEMINI.md and the MCP server from .gemini/settings.json, which tell the model to begin with trw_session_start().

1. trw_session_start() — loads prior learnings and any active run state.
2. Do the work. Checkpoint after each milestone so you can resume after compaction.
3. trw_learn(summary, detail) whenever you discover a non-obvious gotcha.
4. trw_build_check() then trw_deliver() to persist state for the next session.

Differences from Claude Code

• MCP lives in .gemini/settings.json. Gemini registers the server under mcpServers in its settings file — there is no .mcp.json at the repo root.
• One instruction file. GEMINI.md carries the ceremony protocol, similar in spirit to CLAUDE.md.
• Agent teams are a subset. Not all 12 Claude Code specialists translate cleanly to .gemini/agents/; expect the core roster plus any specialists Gemini can host.
• Same memory store. Learnings captured here are visible to every other TRW-wired client in the same repo.

Common problems