nWave
A human-gated agent workflow that turns product ideas into tested, working code.
Per-dimension scores and reasoning
The evidence shows an empty-permissions PR workflow that neither checks out nor executes PR content and uses structured JSON encoding. Repository hooks are opt-in by default, the seven-wave process requires human approval, and installation, update checks, housekeeping, and uninstall effects are described. Authors, contributors, and license attribution are clear. Deductions apply because the recommended curl-to-shell command executes content from the moving main branch, installation modifies user-level Claude/Codex configuration, and the data boundaries of update checks and project-reading behavior are incomplete. The privacy claim lacks the referenced policy or implementation support, dependencies mostly have lower bounds rather than locks or hashes, and credential rotation, log redaction, and artifact-level rollback are not demonstrated.
The README, package version, Commitizen version, and release notes are broadly consistent. Python requirements, dependencies, test groups, offline-install routes, DES failure messages, and doctor guidance are documented. Deductions apply because the primary raw-GitHub installation URL contains a suspicious extra `}`, potentially breaking the critical setup path; the offline guide and runtime implementation are absent, dependency ranges are loose, and several legacy tests are described as skipped or exempted from undefined-name checks during an architecture migration.
The material gives distinct routes for greenfield work, brownfield features, bugs, refactoring, teams, CI, offline environments, and selectable rigor levels. It also identifies the degraded plugin-marketplace boundary and claims support for Claude Code, OpenCode, and Codex. Deductions apply because DES step-ID matching has documented false positives requiring manual exemptions, cross-platform details are deferred to unavailable guides, Windows is limited to WSL guidance, and environment probing or compatibility code is not shown.
The README is well organized around setup, waves, commands, troubleshooting, updates, uninstall, privacy, community, and breaking changes. Naming migration, worked commands, known limitations, the complete MIT license, version history, and maintenance releases are explicit. Deductions apply because the preferred install example appears malformed and many important details exist only behind links not supplied here. Maintainer names and an issue channel are present, but part of the release workflow is managed from an unavailable private repository and publisher identity is unverified, so responsibility and the update path are not fully auditable.
The sources describe reviewable wave artifacts, a single-file feature-delta format, an outcomes registry, validation, and adjustable documentation density, all oriented toward usable delivery outputs. Rigor profiles clearly compare models, review, TDD, mutation testing, and relative cost. Deductions apply because agent prompts, representative outputs, and core implementation are absent, preventing static confirmation of output quality or incremental value over ordinary Claude/Codex workflows. Cost is qualitative only, without token, time, or monetary measurements.
Version, license, and authorship align across README, LICENSE, and pyproject. Release claims point to issue numbers, ADRs, guides, and commands, while experimental v4 scope, an incomplete fix, and degraded installation are explicitly distinguished. Deductions apply because the cited issues, ADRs, tests, privacy policy, changelog, and implementation are not included, so many feature, security, and testing claims cannot be independently checked. Statements such as no data collection and enforcement of proven practices remain largely self-asserted, with no systematic claim-versus-inference labeling.
- Do not run the current curl-to-shell command as written: its URL contains a suspicious extra `}`, and it executes an unpinned, unhashed script from the moving main branch. Inspect it and pin it to the reviewed revision first.
- Installation changes user-level agent directories and hook configuration and enables update checks and background log cleanup. Back up configuration, enumerate affected files, and test disable/uninstall behavior in a disposable repository.
- The no-data-collection statement is supported only by the README here. Before allowing the buddy to inspect a sensitive codebase, review PRIVACY.md, network-request code, logged fields, and retention behavior.
- Dependencies use open lower bounds and no lockfile or artifact hashes are shown. Produce a controlled lock, scan resolved packages, and verify offline dependency provenance before deployment.
What does this agent do, and when should you use it?
nWave is a structured software-delivery agent workflow built primarily for Claude Code, with documented support for Codex CLI and OpenCode. It divides feature delivery into seven waves: DISCOVER, DIVERGE, DISCUSS, DESIGN, DEVOPS, DISTILL, and DELIVER. Specialized product, architecture, platform, acceptance-testing, and software-crafting agents create filesystem artifacts at each wave, and the user reviews and approves them before work proceeds. Delivery follows TDD, using RED→GREEN→COMMIT as the current default, while DES hooks validate phases, file actions, and audit state. The distribution includes the `nwave-ai` CLI, slash commands, 40 agents, configurable documentation density, and adjustable rigor profiles. It suits teams that want an explicit engineering process, but adoption requires Python 3.10+, a supported coding host, and installation of agents, skills, commands, configuration, and hooks.
After installation, /nw-buddy reads the current project, locates existing artifacts, recommends an entry wave, and gives a concrete next action. Greenfield work can begin with /nw-discover or /nw-diverge; /nw-discuss develops requirements, while /nw-design routes work to system-designer, ddd-architect, and solution-architect for architecture and domain modeling. /nw-devops covers infrastructure and deployment, /nw-distill creates Given-When-Then acceptance tests, and /nw-deliver assigns TDD implementation to software-crafter. Bug fixes can jump directly to DISTILL and DELIVER, while tested refactoring can start at DELIVER. During delivery, DES inspects Bash and file actions, TDD phases, step markers, audit logs, and session state. /nw-rigor selects lean, standard, thorough, exhaustive, or custom quality settings; nwave-ai project enable|disable, mode, and status control per-project activation.
- A greenfield product team that wants to explore the problem space before moving through requirements, architecture, acceptance tests, and TDD implementation.
- An engineer adding a feature to an established codebase who needs reviewable requirements and design artifacts without restarting the full discovery process.
- A developer fixing a defect who wants to establish a failing acceptance test with
/nw-distillbefore implementing the correction through/nw-deliver. - A team refactoring code already protected by passing tests and intending to enter directly at the DELIVER wave.
- A team that wants lean handling for documentation or configuration work but stronger reviews and mutation testing for production-critical features.
- An engineering organization using Claude Code or Codex CLI that wants enforcement around TDD phases, file operations, and delivery audit records.
What are this agent's strengths and limitations?
- The seven-wave model connects discovery, requirements, architecture, platform work, acceptance testing, and implementation while allowing task-specific entry points.
- Human approval is required between waves, preserving user control at product and engineering decision gates.
- DES checks delivery-phase Bash and file operations, TDD progress, and audit state rather than relying only on prompt instructions.
- Forty specialized agents and peer-review roles divide product and engineering responsibilities more explicitly than a single general coding agent.
- Configurable rigor and documentation density let teams trade token cost against review depth, TDD structure, and mutation testing.
- It requires Python 3.10+ and a supported host such as Claude Code, Codex CLI, or OpenCode; it is not documented as a standalone service.
- The complete setup installs global hooks, user-level configuration, and project activation markers, creating more adoption overhead than a prompt-only workflow.
- Claude Code marketplace installation cannot run DES, so phase enforcement, TDD validation, rigor profiles, and audit logging are absent in that mode.
- Each supported host has its own installation integration; multi-host support does not provide a universal, no-change agent protocol.
- Version 3.22.1 retains a known commit-metadata limitation in which some generated metadata may not be recognized as Git trailers.
How do you install or deploy this agent?
Python 3.10+ and Claude Code are listed as the primary requirements; separate paths are documented for Codex CLI, OpenCode, and offline environments. The README publishes this one-step installer:
sh -c "$(curl -fsSL https://raw.githubusercontent.com/nWave-ai/nWave/main}/scripts/install/install.sh)"It installs the nwave-ai CLI and connects nWave to Claude Code, preferring uv while also supporting pipx. The main} segment is reproduced exactly from the supplied source. Restart Claude Code after installation. No API key or other credential is specified in the source; an offline or air-gapped installation path is also available.
How do you use this agent?
After restarting the host, begin in Claude Code with:
/nw-buddy What should I do next?
The buddy reads the project and recommends the appropriate entry point. A representative feature sequence is:
/nw-diverge "user authentication approaches"
/nw-discuss "user login with email and password"
/nw-design --architecture=hexagonal
/nw-distill "user-login"
/nw-deliverReview and approve the artifacts after each wave. Use /nw-rigor lean or /nw-rigor standard to change the quality-cost balance. To update, run uv tool upgrade nwave-ai or pipx upgrade nwave-ai, followed by nwave-ai install.
How does this agent compare with similar options?
Compared with the Claude Code plugin marketplace route, the recommended CLI installation enables the complete DES hook layer, including phase enforcement, TDD validation, rigor profiles, and audit logging. The marketplace route is blocked by an upstream ${CLAUDE_PLUGIN_ROOT} limitation and supplies only agents, commands, and skills as a degraded preview. Claude Code is the primary host, while Codex CLI and OpenCode have separate supported installation paths that require host-specific integration.
FAQ
Will it autonomously run from an idea all the way to committed code?
Is Claude Code mandatory?
Does installation require internet access or API credentials?
What should I do if DES blocks a write or reports an incomplete TDD phase?
/nw-deliver session or wait until that session is finalized. An agent that returns before completing required TDD phases should be dispatched again to finish the missing phases.Will uninstalling nWave remove my project files?
nwave-ai uninstall removes installed agents, commands, configuration, and DES hooks; uv tool uninstall nwave-ai or pipx uninstall nwave-ai removes the CLI. Project files remain untouched.