Genie
Wishes in, PRs out: it interviews you into a plan, dispatches parallel agents, and reviews the code before you see it.
Least privilege is well documented: workflows pin actions to exact SHAs with minimal permissions; NATS loads only on the omni serve path with a white-box test; deletions require proven ownership. Deducted because these are mostly prose claims — the core src/ files proving actual permission boundaries were not provided, and the WhatsApp bridge plus ed25519 key registration touch host-sensitive operations with limited documentation depth.
Internally consistent across README, SECURITY.md, package. and tests; frozen lockfile and pinned Bun version support dependency availability; stable diagnostic strings are asserted in tests. Deducted because the 2026-dated versions and incident timeline cannot be validated statically, and the retired npm distribution creates confusing historical version references.
Clear developer audience; platform matrix covers linux/darwin and explicitly excludes darwin-x64; skills are runtime-neutral across agent homes. Deducted because capability boundaries and trigger precision rest mainly on prose — no skill-body trigger constraint files were supplied, and Orca compatibility is runtime-probed rather than statically guaranteed.
Excellent information architecture: command and skill tables, clear lifecycle, thorough install notes with verification scripts and channel manifests, full MIT license. Deducted for the absence of a CHANGELOG file (history is described only via channel manifests), few end-to-end examples, no FAQ, and maintenance responsibility resting on an unverified publisher identity.
Output usability is designed for (versioned JSON payloads, doctor diagnostic lines, board view); marginal value over ad-hoc orchestration is real (git documents + one SQLite file, zero daemons). Deducted because actual output quality (mergeable PRs) needs execution evidence not available here, and the learning cost is nontrivial (modes, channels, legacy migrations).
Claim traceability is comparatively strong: verification script, pinned identity triple, multi-channel cross-checks with concrete paths and exit codes; tests corroborate README claims (MCP retirement diagnostics, action-pin matcher). Deducted because key claims (signed releases, the supply-chain incident and remediation) cannot be independently confirmed in a static review, marketing and fact are occasionally blended in the README, and the 2026 timeline is unverifiable.
- README and SECURITY.md contain 2026 dates, versions, and a supply-chain compromise narrative; static review cannot confirm authenticity — verify independently via git history and out-of-band channels before use.
- The install path is curl | bash; although the docs claim the installer verifies cosign/SLSA signatures, read install.sh itself and verify the pinned identity offline before executing.
- Skills are written into multiple agent-global skill homes (~/.claude/skills etc.) and may back up and overwrite directories owned by other tools; check backups and digest records before upgrading.
- The Omni WhatsApp bridge involves key registration and a remote-approval channel; it is off by default, but assess messaging risk before enabling and confirm the store-don't-answer behavior for unmapped chats.
- npm distribution is retired and historical 4.x versions are declared compromised; never install from npm — use only repository-manifest-designated channels.
- This is a static review (low confidence) with no tests executed; no runtime verification was performed.
What does this agent do, and when should you use it?
Genie is a planning-and-execution layer for AI coding agents. You describe what you want in one sentence; it interviews you into DESIGN.md and WISH.md plan documents, dispatches native subagents to build the work in parallel, and an independent reviewer delivers a SHIP, FIX-FIRST, or BLOCKED verdict against the acceptance criteria. The architecture is deliberately lightweight: no daemons and no Postgres — operational state lives in a single per-repo SQLite file (.genie/genie.db), while plan documents are plain markdown in git. The CLI ships 15 commands and 4 runtime dependencies in a ~0.9 MB single-file bundle built on Bun. Skills are distributed through the skills.sh channel into the skill homes of Claude Code, Codex, Goose, Windsurf and other agents, with an optional Orca plugin providing a managed lifecycle mode. Every release is cosign-signed (keyless OIDC) with SLSA provenance, and the installer verifies the binary before running it.
The lifecycle runs: /brainstorm refines a vague idea into DESIGN.md with a mandatory design review → /wish converts an accepted design into a scoped WISH.md with execution groups → /review performs plan review (APPROVED or concrete gaps) → /work dispatches native role subagents wave by wave → /review independently judges the implementation as SHIP, FIX-FIRST, or BLOCKED. genie context --wish <slug> emits one line of versioned JSON (branch + resolved base SHA + ready tasks) consumed at spawn time; genie board derives a live kanban snapshot; genie task drives task state in the single SQLite file. genie.db is resolved from the git common directory, so all linked worktrees share one state. genie omni bridges a running agent to WhatsApp via an Omni hub over NATS, enabling remote approvals (reply y/n or emoji reactions) and inbound one-shot tasks. genie setup --orchestration-mode orca transfers lifecycle authority to Orca (requires Orca 1.4.192+); switching back to standalone is equally explicit.
- A developer working in Claude Code or Codex who wants to turn a one-line vague request into a reviewed, parallel-executed, mergeable change
- Teams maintaining multi-worktree repositories who need all worktrees to see the same task state with no manual sync
- Engineering leads who want code judged by a reviewer agent independent of the author, against explicit wish criteria
- Users who want to approve agent actions or issue short tasks from their phone via the genie omni WhatsApp bridge
- Teams already on the Orca orchestration platform who want Orca to own the task/board/roadmap lifecycle
What are this agent's strengths and limitations?
- Lightweight footprint: no daemons, no Postgres — one SQLite file per repo, each command opens it, runs one transaction, and exits; ~0.9 MB single-file bundle with only 4 runtime dependencies
- Reviewer independence is a hard design rule: the review subagent is separate from the one that wrote the code (reviewer ≠ engineer), and verdicts are measured against the wish criteria
- Strong supply-chain security: every release is cosign-signed (keyless OIDC) with SLSA provenance and verified before install; uninstall removes only ownership-proven files and preserves modified ones with a report
- Multi-agent skill distribution: one runtime-neutral skill set lands simultaneously in Claude Code, Codex, Goose, and Windsurf skill homes, all sharing the same lifecycle
- Native multi-worktree state sharing: genie.db resolves from the git common directory, so tasks created in one worktree are instantly visible in another
- Depends on a host agent runtime: skills require Claude Code, Codex, Goose, or Windsurf skill homes; Genie itself provides no model-calling layer
- v4 to v5 is a breaking cutover: the Postgres backend, pane-based orchestration, executor registries, full-screen console, and desktop app are all gone
- Legacy MCP and UI bridge are retired: genie mcp and genie ui-bridge only return non-zero retirement diagnostics, so workflows relying on them must migrate or roll back to a pre-A7 release
- Orca mode is a one-way authority switch: Genie stops reading or writing genie.db and keeps no fallback database — if Orca is unavailable, operations fail rather than silently degrading
- Skills channel can diverge from releases: npx skills add serves the repository's default branch, which may be ahead of or behind any signed release
How do you install or deploy this agent?
One-line install: curl -fsSL https://raw.githubusercontent.com/automagik-dev/genie/main/install.sh | bash. The installer verifies the cosign-signed binary via gh attestation verify (falling back to cosign verify-blob) before running. Then run genie install to converge the skills channel (--integrations auto|codex|claude|all|none scopes which agent skill homes are written); alternatively use npx skills add automagik-dev/genie (add -g --all for every agent home). Inside a trusted initialized repo, run genie init to scaffold state, then genie doctor to confirm the install. Requires Bun and git.
How do you use this agent?
Invoke the skills in order inside a repo: /brainstorm (or "brainstorm this") → /wish → /review → /work → /review. In Claude Code use slash commands; in Codex and other agents invoke by name or plain language. Run genie board anytime for a kanban snapshot, genie task to inspect or drive task state, and genie context --wish <slug> to get spawn context. For the WhatsApp bridge, connect a WhatsApp instance to an Omni hub, run genie omni handshake once per host to register an ed25519 keypair, then keep genie omni serve running. Run genie doctor after install or update; use genie update to upgrade and genie update --rollback to roll back.
How does this agent compare with similar options?
The README itself contrasts v4 (preserved on the v4 branch): v4 used a Postgres backend, pane-based process orchestration, and a desktop app, while v5 deliberately cuts down to skills + git documents + one SQLite file. The other named alternative surface is Orca: standalone and Orca modes are mutually exclusive, and in Orca mode Orca becomes the sole lifecycle authority.