- Home
- Skills
- Agents & Orchestration
- Claude Wave Plugin: Evidence-First Engineering Waves
Works with the AI Agents you already use
Claude Wave Plugin: Evidence-First Engineering Waves
Run Claude Code like an engineering org: a dedicated sub-agent per phase, strict TDD, visual verification gates and hook-enforced phase order.
Free
Claude Wave Plugin: Evidence-First Engineering Waves
Example session with this skill installed
/wave-start --demo "make the save button show a loading spinner"
- Read your context and instructions
- Compiled the claude wave plugin
Summary
The orchestrator runs the 5-phase demo wave (AC, DR, RED, GREEN, TEET). Each phase is a separate sub-agent with lean context, and every hand-off is gated on evidence files under .wave/.
What happens, phase by phase
| # | Phase | Sub-agent | Evidence produced |
|---|---|---|---|
| 1 | Acceptance Criteria | ac-writer | .wave/ac.md: brutal, falsifiable ACs (spinner visible while saving, button disabled, spinner gone on success/error) |
| 2 | Design Review | design-reviewer | .wave/dr.md: audit of your design system's installed dist files, per-route mockup, computed-style visual ACs. Any design-system gap becomes an OPEN item that pauses the wave |
| 3 | RED | red-tests | Failing tests written against the ACs, with the failure output printed as proof |
| 4 | GREEN | green-impl | Minimum implementation, live stack started, full test suite run, Playwright screenshot taken |
| 5 | Visual approval | orchestrator | The wave stops, shows you the screenshot and asks you to confirm before going on |
| 6 | TEET | teet-verify | Cross-system computed-style and boundingBox assertions, plus the final verdict |
Example of a visual AC it enforces
// Not toBeVisible(): that passes for a 0x0 element
const box = await spinner.boundingBox();
expect(box.width).toBeGreaterThan(0);
expect(await spinner.evaluate(e => getComputedStyle(e).animationName)).not.toBe('none');
The rule to remember
If the agent says "done" without showing you a screenshot, refuse the hand-off. That is exactly the failure mode the plugin exists to prevent.
Next steps
- After two or three successful
--demowaves, drop the flag:/wave-start "implement the new payment flow"runs the full 17-phase pipeline with [DR] and optional [CR]. - If context fills mid-wave, run
/wave-checkpointand resume in a fresh session.
Connects securely to your tools. The creator never sees your data.
About this skill
Stop vibe-coding. claude-wave-plugin turns Claude Code into an orchestration layer instead of an autocomplete. Every phase of shipping a feature gets its own dedicated sub-agent with lean context, and the orchestrator only accepts a hand-off when it sees real evidence: live-stack screenshots, computed-style assertions and true end-to-end runs. "Build clean" or "tsc passes" is never enough.
The defaults of vibe-coding optimise for time-to-first-demo. This plugin optimises for time-to-shippable.
The problem it solves
One-shot prompting produces fake-done features in production codebases: tests that pass while the page renders blank, configs that validate but are never read at runtime, a data-theme="dark" attribute that is set while the screenshot stays light. Wave execution makes those failures impossible to hand off.
What you get
- Full Wave Execution Framework v2: 17 base phases plus two conditional gates, [DR] Design Review (mandatory for UI waves) and [CR] Code Review (opt-in). Brutal ACs, Silent Error Analysis, Dependency Audit, Brutalised E2E, version bump, commit, cleanup, checkpoint.
- Model routing: Opus advises (plan, critique, decide), Sonnet executes (code, tests, scans), Haiku handles mechanical steps.
- Hook enforcement (v0.2.x): phase order, valid phase/role/model, hand-off artifacts and markers, UI screenshot approval and bug-fix strategy approval are gated by executable hooks, not by trust.
- Checkpoint resume:
/wave-checkpoint(and an automatic PreCompact checkpoint) writes resume state to.wave/checkpoints/, so auto-compact or a dead battery never kills a run. - Token ledger and scorecard: per-agent model and token accounting, budgets and rework tracking.
Skills included
| Skill | Phase | Owns |
|---|---|---|
| wave-orchestrator | Router | Full/demo/solo dispatch and hand-off gates |
| ac-writer | 1. Acceptance Criteria | Brutal, falsifiable ACs |
| design-reviewer | 1.5 Design Review | Component API audit against installed dist files, per-route mockup, visual ACs |
| red-tests | RED | Failing tests first, proven to fail |
| green-impl | GREEN | Minimum code, then live stack + full tests + screenshot |
| teet-verify | TEET | True end-to-end with computed-style and boundingBox assertions |
| wave-execution-framework-v2 | Canonical spec | The 17-phase pipeline, 10 invariants and 12 reference docs |
Commands
/wave-start "<feature>": full v2 pipeline/wave-start --demo "<feature>": trimmed 5-phase entry subset (AC, DR, RED, GREEN, TEET). Start here./wave-start --solo "<feature>": direct task with explicit models, commit guard, checkpoint and ledger/wave-checkpoint: save wave state to disk
The 5 hard-won rules
- Verify before scan: never scan un-verified GREEN.
- Visual effect over DOM presence:
toBeVisible()is true for a 0x0 element, so assert computed styles and bounding boxes. - Visual vocabulary grounding: never invent a colour or token the app does not already render.
- Never implement without wiring: trace config to runner to handler end-to-end.
- Permissions once, upfront.
Best for
Production codebases where bugs cost money and "the agent said it's done" is not acceptable evidence. For quick prototypes or one-line fixes, plain prompting is faster.
Open source (MIT): https://github.com/Harshvardhan86/claude-wave-plugin
Requires local test infrastructure and live stack for verification. Not for single-file edits or research-only tasks. Effective only for projects with 3+ vertical slices of value.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
1 use
Times used to date
Free forever
No account required to browse
Trust & safety
Security scan
Security scan passed 1 month ago
- Free with an Agensi account
Requires
Permissions
File Scopes
Shell: hooks are Bash scripts that gate phase order, and phases run your build/test commands. Read/Write: sub-agents write code and tests in your repo, and wave state, approvals, checkpoints and the token ledger live under .wave/ (excluded from git). Browser: Playwright drives the live stack for screenshots and computed-style assertions in UI waves. Network: only for your own project's dependency installs and local dev server. The plugin itself sends no data anywhere.