Warpweave

Spec-driven minimalism with compressed feedback loops. One organism, four systems, zero waste.

An AI assistant left to its own defaults will happily skip planning, ship code it never tested, over-build the implementation, and flood your context with raw command output — four failure modes, four separate opinions needed to stop them. Warpweave fuses four independently-proven systems into a single enforced pipeline, so no layer has to be remembered or configured by hand:

| Layer | Role | Source | |-------|------|--------| | Spec | WHAT to build | OpenSpec | | Process | HOW to build | Superpowers | | Quality | HOW MUCH to write | Ponytail | | Feedback | HOW FAST to iterate | RTK |

Philosophy

Warpweave doesn't add one opinion on top of your AI assistant — it runs every change through four, one per phase, and never skips a layer:

USER INTENT
  → SPEC GATE   (Spec)     — proposal, specs, design, tasks; you approve
  → PLAN        (Process)  — tasks decomposed to file-level granularity
  → EXECUTE     (Quality)  — RED → GREEN → REFACTOR, climbing the Ladder
  → FEEDBACK    (Feedback) — every shell command compressed through RTK
  → REVIEW      (Process)  — two-stage review, archive, spec sync

Four hard gates back this up, and none of them are optional:

1. No code without spec — implementation is blocked until specs are approved.
2. No commit without test — the RED → GREEN TDD cycle has to pass first.
3. No merge without review — a two-stage review (spec compliance, then code quality) runs before archive.
4. No raw output without RTK — every shell command is compressed before it reaches context, so signal survives and noise doesn't.

The Ladder governs how much gets written, not just whether it works. Before a line of code exists, it climbs seven rungs and stops at the first one that holds: YAGNI → reuse → stdlib → native → dependency → one-liner → minimum. Validation, error handling, security, and accessibility are never sacrificed to climb higher — deliberate shortcuts get logged as // ponytail: <reason> instead of disappearing silently into the diff.

What running everything through four layers gets you, instead of one:

• A plan you both read before code exists, instead of finding out what the AI assumed after it already wrote it.
• Every line justified against the ladder, instead of a codebase that grows because nothing stopped it.
• Every task built test-first, instead of code that "looks right" until it doesn't.
• Every shell command's output compressed before it hits context, instead of noise crowding out the one line that mattered.

Quick Start

npm install -g warpweave@latest
cd your-project
warpweave init

Then, in your AI assistant's chat:

/ww:explore                   →  (optional) think it through with the AI first
/ww:propose add-dark-mode     →  AI drafts proposal, specs, design, tasks
        (you read and adjust the plan)
/ww:apply                     →  AI builds it, checking off tasks
/ww:archive                   →  specs updated, change archived

Full command list: warpweave --help. Setup runs in your terminal; the four commands above are slash commands typed in your AI assistant's chat — if that split is unfamiliar, see How Commands Work.

The core profile installs twelve skills automatically — the six core-loop skills below plus the three quality skills and three check skills that run without configuration. Three of them work in the background without ever being invoked:

• warpweave-translator — works like a secretary at the intake. When a request is underspecified, it asks 1–5 short multiple-choice clarifying questions (reply e.g. 1b 2a or defaults) before implementation starts, so the AI never guesses wrong or over-asks.
• warpweave-superpowers-tdd — RED-GREEN-REFACTOR, subagent-driven development, and a two-stage review (spec compliance, then code quality) on every implementation task.
• warpweave-ponytail-minimal-output — the YAGNI ladder (skip → reuse → stdlib → native → dependency → one-liner → minimum) plus // ponytail: debt markers, so every line earns its place.

Every skill — and what you get

Warpweave ships 27 skills. Twelve install automatically with the default core profile; the rest are opt-in, invoked as slash commands (/ww:*) or CLI subcommands when you need them. Any tool that supports skills.sh can install all of them at once: npx skills add dadosh1984/warpweave.

Core loop — installed by default

| Skill | Command | What you get | |-------|---------|---------------| | warpweave-explore | /ww:explore | A no-stakes thinking partner. Reads your code, lays out options, turns a fuzzy idea into a concrete plan — before any change file exists. | | warpweave-propose | /ww:propose | One command drafts the whole plan — proposal, specs, design, tasks — including a Ladder Decision table showing what options were weighed and why. | | warpweave-apply-change | /ww:apply | Implements tasks.md task by task, checking off progress, with TDD and the minimal-output ladder running underneath automatically. | | warpweave-update-change | /ww:update | Revise the plan mid-change — fold in a new decision, fix a scope mistake — without touching code and without artifacts drifting out of sync with each other. | | warpweave-sync-specs | /ww:sync | Fold a change's delta specs into your main specs whenever you want, without archiving the change yet. | | warpweave-archive-change | /ww:archive | Closes the loop: merges delta specs into the source of truth, moves the change to changes/archive/, reports token savings. |

Ships by default — implementation quality

| Skill | Runs | What you get | |-------|------|---------------| | warpweave-translator | Automatically, on any underspecified request | Never lets the AI silently guess your intent — asks a handful of quick multiple-choice questions instead. | | warpweave-superpowers-tdd | Automatically, during /ww:apply | Test-first implementation with RED-GREEN-REFACTOR and a two-stage review, instead of code that "looks right." | | warpweave-ponytail-minimal-output | Automatically, during /ww:apply | Guards against over-engineering on every task; deliberate shortcuts get logged as // ponytail: markers instead of silently accumulating. |

Ships by default — checks

| Skill | Command | What you get | |-------|---------|---------------| | warpweave-security-scan | /ww:security-scan (also runs automatically after each task during /ww:apply) | A native security scan over changed code — hardcoded secrets, injection surfaces, unsafe patterns — without semgrep or Docker. | | warpweave-drift-detection | /ww:drift-check (also runs automatically after each task during /ww:apply) | Catches spec/code mismatches while they're still a one-line fix, not a surprise at archive time. You choose: fix the code, fix the spec, or note it and continue. | | warpweave-guardrails | /ww:guardrails (also runs automatically before each commit) | One check before you commit: all four pipeline gates (SPEC, TDD, LADDER, RTK) at once, instead of trusting that everything's fine. |

Expanded workflow — opt-in

| Skill | Command | What you get | |-------|---------|---------------| | warpweave-ff-change | /ww:ff | The same artifacts as propose, minus the Ladder Decision analysis — the fastest path from idea to code when you don't need the extra thinking. | | warpweave-new-change | /ww:new | Start a change scaffold now, fill in each artifact as you're ready, instead of generating everything up front. | | warpweave-continue-change | /ww:continue | Picks the next ready artifact and creates exactly that one — for step-by-step rather than all-at-once workflows. | | warpweave-verify-change | /ww:verify (also runs automatically when all tasks are complete in /ww:apply) | Checks the finished implementation against specs, tasks, and design before you archive, so drift doesn't reach production unnoticed. | | warpweave-bulk-archive-change | /ww:bulk-archive | Archive several parallel changes at once, with conflict detection and per-delta spec sync — no one-by-one babysitting. | | warpweave-onboard | /ww:onboard | A guided, narrated walkthrough of the whole workflow that does one real cycle of work in your own codebase, not a toy example. |

Quality & analysis — opt-in

| Skill | Command | What you get | |-------|---------|---------------| | warpweave-ladder-audit | /ww:ladder-audit | Finds deletable code, over-engineering, and unjustified dependencies in the current diff before someone else has to. | | warpweave-debt-ledger | /ww:debt-ledger | Every // ponytail: marker you've left behind, collected into a structured backlog you can actually triage instead of grepping for later. | | warpweave-token-budget | /ww:token-budget | Set a token ceiling per change and track consumption across pipeline phases — cost becomes visible instead of a surprise at the end. | | warpweave-learn | /ww:learn | Mines your own archived changes for trends — estimation accuracy, discipline patterns — that no single change would show you on its own. | | warpweave-parallel-execute | /ww:parallel-execute | Splits tasks.md into independent layers and runs them concurrently via subagents, instead of one task at a time. | | warpweave-release-compare | /ww:release-compare | Compares the project before vs after a release, scores the improvement on defined criteria, and warns when a change falls short of the minimum improvement threshold. |

Setup

| Skill | Command | What you get | |-------|---------|---------------| | warpweave-init-unified | /ww:init-unified | Bootstraps the full environment in one pass — Warpweave, RTK, Superpowers, and Ponytail together — instead of four separate installs. |

Fast Lane

Small changes skip the ceremony. When the AI detects a change touches ≤3 files and <30 lines (or is a pure style/config/typo fix), it sets complexity: minimal in the change metadata. This auto-completes design.md and tasks.md — the change goes straight from proposal → specs → apply. No manual flags, no extra config. For normal changes, the full artifact chain runs as always.

| Signal | Complexity | Artifacts | |--------|-----------|-----------| | ≤3 files, <30 lines, or style/config fix | minimal | proposal → specs → apply | | 4+ files, new component, new API | normal | proposal → specs → design → tasks → apply |

Tessl Registry Integration

Warpweave can pull verified library skills from the Tessl Registry — a catalog of 10,000+ evaluated, versioned instructions for popular libraries. This reduces AI hallucination by giving agents accurate, up-to-date API knowledge instead of relying on stale training data.

Enable it during warpweave init (you'll be prompted) or manually:

warpweave config registry --enable

Once enabled, Warpweave automatically scans your package.json dependencies, queries the Tessl Registry, caches matching skills locally, and injects them into agent instructions as a ## Registry Skills section. No manual configuration needed per library.

| Command | Description | |---------|-------------| | warpweave config registry --status | Show current configuration | | warpweave config registry --enable | Enable registry integration | | warpweave config registry --disable | Disable and clear cache | | warpweave config registry --endpoint <url> | Set custom API endpoint | | warpweave config registry --auto-detect | Enable auto-scanning of dependencies | | warpweave config registry --no-auto-detect | Disable auto-scanning | | warpweave config registry --clear-cache | Clear cached skills |

Integration is opt-in and disabled by default. No new runtime dependencies — uses built-in node:fetch.

Documentation

• Core Concepts at a Glance — the whole mental model on one screen
• Getting Started — first steps
• Cheat Sheet — situation-first map of every command
• How Commands Work — using slash commands
• Concepts — the long version of the philosophy above
• All Docs — complete documentation index

Contributing

• Small fixes — submit PRs directly
• Large changes — create a change proposal first

AI-generated code is welcome if tested and verified.

Development

pnpm install
pnpm build
pnpm test

License

MIT License — see LICENSE.

Warpweave is a fork of OpenSpec, Copyright (c) 2024 OpenSpec Contributors. Modifications Copyright (c) 2026 dadosh1984.