feat: add support for Gemini CLI

[ryzizub]

Description

Wires up MCP, hooks, and the reviewer agent for Gemini CLI. Skills were already portable after #128. Closes #125.

Gemini CLI reads the Agent Skills standard directly, scanning .agents/skills alongside its own .gemini/skills, so gemini skills install <repo-url> --path skills lands all 15 with no Gemini-specific file. The enforcement layer does not carry over as cleanly — different hook event names, different tool names, a stricter agent schema — so gemini/ holds those three, ported.

What changes:

• gemini/settings.json — registers the same dart and very-good-cli servers as .mcp.json and wires the same hook scripts under Gemini's event names. Generated with gemini hooks migrate --from-claude, then corrected for the three things that tool leaves behind, all of which fail silently: ${CLAUDE_PLUGIN_ROOT} is not rewritten and has no Gemini equivalent for settings-level hooks (hence ${VGV_PLUGIN_ROOT}); timeout is copied verbatim, but Gemini reads milliseconds where Claude Code reads seconds, so a 10 becomes a 10 ms budget; and MCP matchers keep mcp__<server>__<tool> when Gemini names them mcp_<server>_<tool>.
• hooks/scripts/vgv-cli-common.sh — deny() and allow() read hook_event_name and emit whichever response shape the firing harness reads (hookSpecificOutput.permissionDecision for Claude Code, top-level decision/reason for Gemini), so one script serves both. Claude Code behavior is unchanged: a payload without that field still gets the Claude shape, and check-vgv-cli.sh still auto-approves there. Gemini has no auto-approve for a BeforeTool hook, so that path enforces the version gate only and "trust": true on the MCP server covers approval.
• gemini/agents/flutter-reviewer.md — Gemini cannot bundle a subagent in an install, so users copy this file. It cannot be the same file as agents/flutter-reviewer.md: Gemini validates agent frontmatter against a strict allowlist and drops any agent carrying a key outside it, so skills: and hooks: are both rejected. Read-only is enforced by granting no shell tool at all rather than by the agent-scoped hook Claude Code uses — same guarantee, different mechanism, worth a look.
• Docs across README.md, CONTRIBUTING.md, CLAUDE.md, AGENTS.md.

Everything else is shared verbatim: both harnesses put the shell command in tool_input.command and the edited path in tool_input.file_path, so block-cli-workarounds.sh, analyze.sh and format.sh need no payload changes.

Marketplace/extension packaging is deliberately out of scope. Gemini's own gemini extensions install would need this repo rooted as an extension, and Gemini then reads hooks/hooks.json and agents/ with its own schemas — printing Invalid hook event name: "PreToolUse" ... Skipping. twice per session and erroring on the Claude agent file. Not worth degrading either path for.

Verification

Tested against a real Gemini CLI 0.59.0 (current latest on npm) in a throwaway HOME, installed the way a user would.

• gemini skills install https://github.com/VeryGoodOpenSource/vgv-ai-flutter-plugin --path skills — 15/15 installed from the real URL. Also 15/15 discovered from .agents/skills.
• Merged gemini/settings.json into ~/.gemini/settings.json and copied the agent to ~/.gemini/agents/: agent loads clean, zero Invalid hook event name, zero hook failures, the SessionStart hook actually ran from ${VGV_PLUGIN_ROOT} and returned its warning, and gemini mcp list shows both dart and very-good-cli Connected. Verified for both the user-scope and project-scope install described in the README.
• The Claude Code agent file, run through Gemini's real loader for contrast, fails with Unrecognized key(s) in object: 'skills', 'hooks' — which is why there are two files.
• Both hook suites pass (19, 26), including the new Gemini response-shape assertions; every deny/allow branch exercised by hand with both payload shapes.
• cspell, markdownlint and claude plugin validate . clean.

Cross-checked every claim in the docs against the 0.59.0 source and the official docs, not just the version that happened to be installed. Byte-identical between 0.52.0 and 0.59.0: hook event names, the milliseconds timeout, the decision/reason block shape, the absence of any auto-approve for BeforeTool, gemini hooks migrate --from-claude's event and tool-name mappings, the strict local-agent schema, the .agents/skills discovery path, the agent directories, the hook env vars (there is still no plugin-root variable for settings-level hooks, hence ${VGV_PLUGIN_ROOT}), and the settings env-var interpolation regex.

One behavior did change in 0.59.0 — workspace trust is now fail-closed, so a headless run in an untrusted directory refuses to start rather than loading with workspace features disabled. That only affects the by-hand verification steps, which now say to set GEMINI_CLI_TRUST_WORKSPACE=true; interactive users just get the normal trust prompt.

Not verified: a live BeforeTool block or AfterTool analyze/format inside a real Gemini session — that needs credentials and a model call. Those paths are covered at the script level with real Gemini payloads, and the wiring that reaches them (event names, matchers, script resolution) is verified above.

Type of Change

• New feature (feat)
• Bug fix (fix)
• Code refactor (refactor)
• Documentation (docs)
• CI change (ci)
• Chore (chore)

🤖 Generated with Claude Code