Maestro in Gemini CLI

Use this skill whenever you run Maestro inside Gemini Code. Keep it native to Gemini and the Gemini Skills standard.

• Commands: /maestro, /maestro:setup, /maestro:newTrack, /maestro:implement, /maestro:orchestrate, /maestro:status, /maestro:revert, /maestro:leindex, /maestro:tui.
• Skill install path (user scope): ~/.config/gemini/skills/maestro/ (installed by the Maestro wizard).
• MCP: gemini.mcpServers.leindex → { "command": "maestro", "args": ["mcp", "tool-search"], "type": "stdio" (if required) } so Gemini reaches the Maestro MCP pool through the dynamic broker. Do not point to any TLDR/legacy endpoints.
• No cross-tool bleed: Avoid references to ~/.claude, ~/.amp, or OpenCode paths.

Quick flow

1. /maestro setup — initialize/refresh product, tech stack, workflow, track registry.
2. /maestro newTrack "<goal>" — generate spec + plan with clarifying questions.
3. /maestro implement <track> — execute plan with LeIndex 5-phase analysis for scoped file access.
4. /maestro orchestrate — cockpit/orchestrator panel (Rust TUI pane).
5. /maestro status — report progress and blockers.
6. /maestro revert [track|phase|task] — controlled rollback.
7. /maestro leindex — analysis via LeIndex; no TLDR imports.
8. /maestro tui — launch the Rust cockpit (primary TUI; Go/Python TUIs are deprecated).

TrackLens Review Protocol

You MUST call tracklens_review in these situations:

• After drafting or substantially revising any spec.md or plan.md
• After generating any markdown document the user will need to approve or act on
• When the user asks you to "review", "check", or "look over" a document you produced
• After maestro:setup generates product/tech-stack/workflow documents

You MUST call tracklens_walkthrough after completing all tasks in a track or when the user wants a review of the completed implementation as a whole.

You MUST NOT call tracklens_review:

• For trivial edits (typo fixes, formatting, single-line changes)
• When the user explicitly says to skip review
• For intermediate drafts the user hasn't asked to see yet
• For files you are only reading, not producing

If the user asks for review but there is no clear reviewable artifact yet:

• Identify the exact file, markdown artifact, or track output first
• Do not guess which document to open in TrackLens
• If the request is for implementation output rather than a document, prefer tracklens_walkthrough

When a review is denied:

1. Read every annotation — pay attention to severity (ERROR > WARNING > INFO)
2. Address each annotation in severity order
3. If edited_content is returned, use the user's edited version as your new baseline
4. Re-call tracklens_review with the updated content
5. Do NOT mark work as complete until review is approved
6. After 3 consecutive denials on the same document, ask the user what they want changed instead of guessing

When a walkthrough is denied:

1. Convert the denial into remediation work
2. Complete the remediation work before re-running tracklens_walkthrough
3. Do NOT mark the track complete until the walkthrough is approved

When a review is approved:

• If edited_content is present, write that version to disk (the user refined your draft)
• If no edited_content, your draft was accepted as-is
• Proceed to the next workflow step

Guardrails

• Treat maestro/archive/tldr/* as reference-only; do not import or call maestro.tldr.
• Keep token usage lean; ask LeIndex for scoped file lists.
• If LeIndex MCP is missing, request rerun of /maestro:configure / installer.

Validation checklist

• Skill exists at ~/.config/gemini/skills/maestro/.
• Gemini config (~/.config/gemini/settings.json) has gemini.mcpServers.leindex pointing to maestro mcp tool-search.
• Commands are available through Gemini’s command surface (per its command discovery).