@coldpress/core
Advanced tools
AI-native development framework: lite lane (Spec → Build → Verify → Ship) + 11-phase full lane, 8 subagents, plugin-distributed skills, sacred-doc governance, enforcement hooks, forward-carry quartet.
An AI-native development framework that drives the full lifecycle of AI-assisted software projects — from bootstrap to deployment to evolution.
coldpress-os is Agent Skills-compatible and implements the Spec → Plan → Implement → Review spine — a lite lane (Spec → Build → Verify → Ship, the scaffolded default) and a full lane of 11 phases with sacred-doc governance, hook-enforced gates, multi-agent orchestration, and the forward-carry quartet.
The 11 full-lane phases: Bootstrap · Discovery · Tech Stack · Planning · Design · Architecture · Breakdown · Implementation · Deployment · Operate · Evolve.
One framework. One format. Two runtimes. Pluggable stack packs.
npm install -g @coldpress/core
Requires Node.js >=22 and Claude Code (the CLI or the Agent SDK) on your machine. No Python is needed for the core framework. (Optional document-ingest and brownfield code-indexing capabilities install their own external tools on demand — see the brownfield capability pack — but the core lifecycle has no Python dependency.)
coldpress init my-project
cd my-project
claude # or: use the Agent SDK
coldpress init prompts you for a project name, slug, and user name, then scaffolds:
coldpress.yaml — project configuration (Phase-1 fields only; later phases write back as you progress)CLAUDE.md — framework routing for Butler (your main Claude Code session).claude/agents/ — 8 subagent definitions (analyst, architect, pm, ux-designer, developer, verifier, devops, reviewer).claude/settings.json — auto-enables the coldpress skill plugin (a local directory marketplace); the plugin ships the full skill library, so there are no per-skill wrappers to generate_context/ — produced artefacts (planning, design, implementation, testing, tracking, handoffs, audit, sacred docs)_input/ — raw inputs, legacy refs, vendor drops, assetssecure/ — credential manifest + pre-commit secret-scan hookcoldpress-os/ — framework files (read-only; upgraded via coldpress update)AGENTS.md + .cursor/rules/ + .roomodes + .openhands/microagents/ + .clinerules/ — interop outputs so the project feels native in any agent toolNew to coldpress-os? Read the First 10 Minutes or see the TaskPulse example for a full lifecycle demo.
coldpress-os's .claude/ tree loads unchanged under both runtimes:
claude in the project directory.@anthropic-ai/claude-agent-sdk (programmatic / CI pipelines) — import the SDK, point it at the project; all subagents and skills resolve.See test/agent-sdk-compat.test.ts for the compatibility smoke test.
# Everyday
coldpress init [project-name] Scaffold a new project
coldpress doctor Verify the local environment (Node, git, Claude Code CLI)
coldpress update Regenerate interop outputs (AGENTS.md, Cursor, Roo, OpenHands, Cline)
coldpress dashboard Start the localhost project dashboard
# Lifecycle
coldpress gate check|enter Evaluate a phase's exit gate / stamp phase entry
coldpress waves Compute implementation waves from the story graph
coldpress trace <verb> [id] Traceability queries over schema'd artefacts
coldpress tokens build|contrast Design-token build + WCAG 2.1 AA contrast validation
coldpress outcomes check Validate outcome contracts (P4 gate)
coldpress verdict record Record a clean-room verifier verdict
coldpress evals Run headless golden-task evals
coldpress evolve Aggregate run-logs into failure/cost leaderboards
coldpress security aggregate Merge scanner results into the security gate
coldpress run list|inspect Inspect recorded orchestration runs
coldpress wiring check Verify the cross-phase artefact manifest
# Misc
coldpress import bmad <dir> One-way BMAD module import
coldpress lane-upgrade Upgrade a lite-lane project to the full lane
coldpress feedback Open GitHub Issues in your browser
coldpress --version | --help Version / full command reference
11 phases, structured as Spec → Plan → Implement → Review expanded with the forward-carry quartet and a silent-divergence guard at the Design → Architecture boundary.
| Phase | Name | Owner | Purpose |
|---|---|---|---|
| 1 | Bootstrap | butler | Project init, intake, graph-prime, working-mode detection |
| 2 | Discovery | @analyst | Research, personas, context, idea validation, product brief |
| 3 | Tech Stack | @architect | Stack discovery, evaluation rubric, locking, env provisioning |
| 4 | Planning | @pm | PRD authoring + section-scoped re-validation (PRD-only post-split) |
| 5 | Design (new) | @ux-designer | UX spec, brand guidelines, prototype, narrative, design-deltas |
| 6 | Architecture (new) | @architect | Sacred architecture.md + ADRs incl. REQUIRED ADRs for flagged design-deltas (silent-divergence guard) |
| 7 | Breakdown | @pm | Epics, story-slice (contracts keyed to architecture), story-graph → computed waves (coldpress waves) |
| 8 | Implementation | @developer + @verifier | Dev-story (plan mode), clean-room verification vs spec + tokens, wave orchestration |
| 9 | Deployment | @devops | Readiness hard-checklist (SBOM, headers, budgets, license), deploy packs (staging → human-gated prod) |
| 10 | Operate | @devops | Sprint status, correct-course, incident-response, ops-deltas |
| 11 | Evolve (final) | @reviewer | Retrospective, product evolution, innovation strategy → next-iteration Phase 1 |
Forward-carry quartet — design-deltas (P5 exit) · architecture-deltas (P7 entry) · implementation-deltas (P11 retrospective) · ops-deltas (P11 retrospective). Each delta resolves via four reconciliation options: accept_into_prd, reject, flag_for_architecture_ADR, park_for_phase_11.
8 subagents + Butler. Butler is your main Claude Code session (not a file); it dispatches the 8 specialists, each with its own context window, tool allowlist, and model.
Model routing is deliberate: the highest-leverage reasoning roles — planning, architecture, clean-room verification, and the evidence-linked retrospective — default to the most capable model (Opus), while high-throughput execution roles default to the fast frontier model (Sonnet) so a full lifecycle stays affordable. The verifier deliberately outranks the developer it checks. Every default is a one-line frontmatter override per project — see Subagent customization.
| Slug | Default model | Primary phases | Role |
|---|---|---|---|
analyst | sonnet | 2 | Research, personas, idea validation (against explicit kill criteria), product brief with outcome metrics |
architect | opus | 3, 6 | Stack + deploy lock, walking skeleton (P3); sacred architecture + ADRs, three-way keyed, silent-divergence guard (P6) |
pm | opus | 4, 7 | Slice-able PRD (P4); story-graph breakdown — owns/produces/consumes + estimates → coldpress waves (P7) |
ux-designer | sonnet | 5 | tokens.json, styleguide + live /styleguide route, ux-spec keyed to requirements, perf/a11y budgets |
developer | sonnet | 8 | Implementation one story at a time in plan mode, red stubs → green within the packet boundary. Does not self-verify |
verifier | opus | 8 (Butler-dispatched only) | Clean-room verification vs spec + tokens — dispatched only by Butler with spec + acceptance + diff, never the developer's reasoning. Read-only. Replaces the old @qa |
devops | sonnet | 9, 10 | Readiness (SBOM/headers/budgets), staging → human-gated prod via the deploy pack (P9); steady-state ops digests (P10) |
reviewer | opus | 11 | Evidence-linked retrospective — every claim cites a run-log event ID (final phase) |
Each subagent's definition lives at .claude/agents/<slug>.md in your scaffolded project.
v0.4 roster surgery:
@qa→@verifier(structurally independent, Butler-only dispatch);@scrum-master(wave planning →@pm+coldpress waves),@communicator(→ forkable creative skills), and@valet(→ the framework-internal loop) were removed.
vibe-coder-fullstack (Convex + Next.js + Clerk), cli-npm-publishable (TypeScript + tsup + Vitest), browser-extension (WXT + Manifest V3), and static-single-page / static-multipage-blog (Astro variants). Activated via stack_pack: in coldpress.yaml after Phase-3 stack-locking. (seo-pack is a capability pack — cross-archetype audit/content/local/schema/technical — activated independently of the stack.)_context/sacred/{context,tech-stack,prd,architecture}.md — are protected by governance change workflows and the sacred-guard hook. (The full lane uses this five-doc-minus-one set; the lite lane — the default — uses a single spec.md.)_context/ vs _input/ — produced artefacts vs material fed into the project. Inputs are not written by any skill.@coldpress/core/
├── src/ # CLI + generators + enforcement hooks + trace/waves/evals (TypeScript)
├── template/ # Scaffolded into consumer projects (8-subagent set + plugin auto-enable)
├── lifecycle/ # 11-phase full lane + lite-lane skill organisation
├── skills/ # atomic reusable skills + 6 stack packs
├── agents/ # Subagent schema + registry
├── schemas/ # JSON / Zod schemas (sacred docs, handoffs, audit, design, deploy, evals)
├── governance/ # Sacred-doc change workflows + 4 reconciliation options
├── data/ # Portable knowledge assets (CSV/YAML method playbook, profiles, deploy packs, taxonomy)
├── templates/ # Document / design / infrastructure / prompt-snippet templates
├── plugin/ # Claude Code plugin marketplace tree (141 spec-compliant SKILL.md)
└── docs/ # Framework documentation
/plugin marketplace add coldpress-labs/coldpress-os
/plugin install @coldpress/core
Installs the full skill library as an Agent Skills–compliant plugin. Source lives under plugin/skills/, regenerated by npm run build:skills from the rich internal SKILL.md corpus.
npm update -g @coldpress/core # pull the latest release
Inside an existing project:
coldpress update # regenerate AGENTS.md / Cursor / Roo / OpenHands / Cline outputs
plugin/skills/). Installable as a Claude Code plugin via /plugin marketplace add coldpress-labs/coldpress-os. We wrap Anthropic's first-party skills where they overlap with ours (docx / pdf / pptx / xlsx under the forkable creative/export skills, webapp-testing under @verifier, mcp-builder under @architect). See docs/agent-skills-compatibility.md.NOTICE.md.docs/spec-plan-implement-review-mapping.md.coldpress trace, and AST code-indexing is an optional brownfield-pack capability that can call Graphify (or an equivalent) as an on-demand external install — never re-vendored. See NOTICE.md §4.| Guide | Description |
|---|---|
| First 10 Minutes | Hands-on setup walkthrough — zero to a running project |
| Example Walkthrough | Full lifecycle demo with a sample project (TaskPulse) |
| Troubleshooting & FAQ | Common issues and solutions |
| Architecture | Internal technical reference |
| Spec → Plan → Implement → Review mapping | 11-phase full-lane lifecycle grouped under the industry-standard spine |
| Agent Skills compatibility | How coldpress-os fits the Anthropic Agent Skills ecosystem |
| Template Registry | Every template — by category, phase, consuming skill |
| Skill Discovery Index | Every skill grouped by phase + cross-cutting utilities |
| Subagent × Phase Matrix | 8-subagent × 11-phase reference — which subagents do what in which phases |
| Phase → Subfolder Mapping | Canonical _context/* destinations per phase |
coldpress.yaml schema | Per-field phase ownership + write-back contract |
| Interop generator | AGENTS.md, Cursor, Roo, OpenHands, Cline — spec + tool translation |
| SKILL.md generator spec | Agent Skills spec compliance + field mapping |
| Anthropic skill wrapping | Where coldpress-os delegates to Anthropic's first-party skills |
| Secure pattern | secure/manifest.yaml + credential loader pattern |
| Decision trees | How Butler routes your intent to skills |
| Flow map | Visual mapping of phases, skills, and subagents |
| Glossary | Defined terms |
| Stack pack guide | Authoring a new technology stack pack |
| Subagent customization | Modes, overrides, custom agents |
| Step-file spec | Format specification for workflow step files |
Feedback flows from projects to coldpress-os via GitHub Issues and PRs. Use coldpress feedback to open the issue form in your browser. See CONTRIBUTING.md for the full guide and CODE_OF_CONDUCT.md for community standards.
Security: please do not open public issues for vulnerabilities. See SECURITY.md for private disclosure.
coldpress-os stands on the shoulders of four open-source projects, each of which contributed substantial ideas, code, and craft. All are MIT-licensed, and all are credited in full in NOTICE.md and docs/attribution-audit.md.
ux-designer subagent.coldpress trace, and tree-sitter AST code-indexing is an optional brownfield-pack capability that can call Graphify (or an equivalent) as an on-demand external install — never re-vendored. See NOTICE.md §4."BMad", "BMad Method", "BMad Core", "Whiteport", and "Whiteport Design System" are trademarks of their respective owners. coldpress-os is an independent project and is not affiliated with or endorsed by any of the above.
MIT — see LICENSE.
Built by ColdPress Labs. Built with intention. Scaled with purpose.
FAQs
AI-native development framework: lite lane (Spec → Build → Verify → Ship) + 11-phase full lane, 8 subagents, plugin-distributed skills, sacred-doc governance, enforcement hooks, forward-carry quartet.
The npm package @coldpress/core receives a total of 4 weekly downloads. As such, @coldpress/core popularity was classified as not popular.
We found that @coldpress/core demonstrated a healthy version release cadence and project activity because the last version was released less than a year ago. It has 1 open source maintainer collaborating on the project.
Security News
Agent Baseline’s new reference architecture shows how to secure software factories with four boundaries enforced outside the AI agent.
Security News
arXiv now limits authors to two submissions a month as AI slop overwhelms moderators, delays good papers, and sparks debate over applying the limit to everyone.
Research
/Security News
A new GhostAction wave hits hundreds of GitHub repos, expanding CI/CD secret theft to cloud and AI credentials in source code and git history.