Skills & Commands

The hub skill that orchestrates everything. Takes a task description and walks through four phases: Research, Plan, Implement, Review. A hard gate leaves each of the first three and needs approval; nothing leaves Review, where a balrog agent attacks the implementation, /warden checks conventions, the work is verified, and the PR is opened. Research provisions the isolated worktree, loads task context, and studies the reference files for conventions; /lembas compacts context between every phase. Supports plan-driven mode: provide a pre-existing plan file and the quest skips Research and Plan, starting at Implement with one gate. Supports promoted mode: a quest promoted from a scout validates and supplements the scout findings instead of researching from scratch. Includes a notes board for cross-quest knowledge sharing: quests scan it at the start of Research and post discoveries during Research and Implement, shared across all worktrees via the main repo root.

For multiple independent tasks, Gandalf (the coordinator) spawns quest and scout teammates. Quests run in isolated worktrees and produce PRs. Scouts research questions and deliver findings. Say 'status' during a fellowship for a progress table. Gates surface to you for approval by default. Supports plan-driven quests: provide a plan file and Gandalf spawns quests that skip to Implement. For large plans, Gandalf can fan out into multiple parallel quests after confirming the split with you. Supports scout-to-quest promotion: say 'promote scout-X to a quest' and Gandalf spawns a quest pre-loaded with the scout's findings. The notes board enables cross-quest knowledge sharing — cleared automatically on disband.

Autonomous research agent that investigates questions with configurable depth. For complex questions, spawns a fresh adversarial validator subagent to verify findings. Produces a structured report with confidence levels. Writes findings to a namespaced file (.fellowship/scout-findings-{name}.md) that can be promoted into a quest. Use alongside /quest in a fellowship for research questions that don’t need code changes.

Loads task-relevant files, conventions, and architecture, scoped to the relevant package in monorepos. Invoke it yourself at the start of a session; quest inlines the same orientation as its Research step 2, so it does not call this. It does not look for checkpoints — quest and /rekindle handle resuming.

Analyzes your codebase to extract patterns before writing code. Examines existing implementations to understand naming conventions, file organization, testing patterns, and architectural decisions. Prevents ‘wrong approach’ rework by learning from what’s already there. Quest inlines the same extraction as its Research step 4, so this is for when you want it on its own.

Writes a structured checkpoint at a phase transition and continues from that summary instead of the full history, keeping the context window in the reasoning sweet spot. The checkpoint at .fellowship/checkpoint.md is also what a crashed quest resumes from. Invoked at every phase transition during a quest, where the hooks verify it before a gate can be submitted.

Compares your changes against reference files and documented patterns in CLAUDE.md. Catches convention violations before they reach PR review. Checks naming, file organization, testing patterns, and architectural consistency.

Analyzes a completed fellowship’s gate history, palantir alerts, and quest metrics to surface patterns. Identifies which gates added value, which phases caused delays, and interactively recommends configuration changes like auto-approving gates with zero rejection rates.

Pulls structured context from GitHub issues via gh CLI — title, body, labels, and recent comments. Returns a package with issue context, a suggested branch name incorporating the issue number, and PR closing keywords (Closes #N). Gandalf invokes it automatically when issue references (#N) are detected in quest descriptions. Also usable standalone: /missive 42 to preview what context a quest would receive.

Invoked at the start of each quest phase when your spawn prompt includes a TEMPLATE: assignment. Resolves the template from the project (.claude/fellowship-templates/), user (~/.claude/fellowship-templates/), or built-in directory, reads the section matching the current phase — one each for Research, Plan, Implement, and Review — and applies it as advisory context that never waives a gate. Fellowship ships one example template to copy; /scribe writes real ones.

Walks through your project to extract conventions into CLAUDE.md. Run once when setting up Fellowship in a new codebase. Examines your code structure, testing patterns, naming conventions, and documents them for future quests.

Starts a local HTTP server (default http://localhost:3000) in the background and prints its URL. Shows every quest's phase and pending gates, scouts, groups (with one-click "Approve All"), the recent event stream, the notes board, and health — with gate approve/reject directly from the page.

After a PR review, extracts conventions from reviewer comments and adds them to CLAUDE.md. Closes the convention learning loop — reviewer feedback becomes documented patterns that future quests will follow.

Interactive setup for all configuration options in ~/.claude/fellowship.json. View current settings, edit individual values, or reset to defaults.

Analyzes your project to create quest templates that encode project-specific knowledge — conventions Claude wouldn’t know, domain rules that aren’t in code, team workflows that matter. A template carries one guidance section per quest phase (Research, Plan, Implement, Review), loaded by /lorebook as each phase starts. Stored in .claude/fellowship-templates/ (project) or ~/.claude/fellowship-templates/ (personal).

Interactive walkthrough that teaches fellowship by running a real task on your codebase. Walks you through research, planning, and implementation with natural-language checkpoints at each stage. Produces a real PR. Afterward, introduces /quest and /fellowship for continued use. Designed for people new to agent orchestration.

Scans worktrees and state files to reconstruct fellowship state from on-disk artifacts. Presents a recovery dashboard showing which quests are resumable (have checkpoints), stale (no checkpoint), or already complete (branch merged). On confirmation, re-spawns Gandalf and quest runners with recovered context.