SpecWeave 3.0 发布：用 handoff 和 pickup 在 Claude、Codex、Grok 之间交接任务
anton-abyzov/specweave
SpecWeave 发布 3.0 版本，把开发工作拆成 spec.md 规格，通过 handoff 和 pickup 两条命令在 Claude、Codex、Grok 等工具或账号之间交接任务，无需复制路径。
Change agents. Keep the thread.
Plan work as small specs, hand it between Claude, Codex, Grok and any other tool in two words, and close it with evidence.
Install
npm install -g specweave # Node.js 20.12.0+ cd your-project specweave init .
init writes AGENTS.md (read by Codex, Grok, Cursor, Gemini and Copilot), a two-line CLAUDE.md that imports it, and the skills for Claude Code and Codex. Nothing else runs in the background.
Out of tokens? Say "hand off"
you (in Claude, account 1): hand off
you (in Codex, or account 2): pick up
That is the whole handoff. "Hand off" runs specweave handoff: it releases your task claims, records why you stopped, and pushes your branch plus a snapshot of your uncommitted edits. "Pick up" runs specweave pickup in any other tool, account, machine or cloud session (a Claude Code Projects thread, a Codex cloud task): it brings that work into the checkout and prints the next task with its acceptance criteria. Nothing to copy, no paths to paste.
Running out mid-task? specweave auto-handoff on, once per machine, makes it automatic: when a session reaches 90% of your plan's 5-hour or weekly limit, it hands off by itself and tells you to say "pick up" elsewhere. Claude Code reads the limit through its status line (your own status line keeps working) and Codex through a Stop hook. Under the threshold it costs no tokens. --at 85 changes the threshold and off undoes every change.
specweave report writes an HTML timeline of who did what on an increment (tools, sessions, handoffs, pickups, test evidence), straight from the ledger.
The loop
| # | Say or run | What happens |
|---|---|---|
| 1 | "pick up" · specweave pickup |
The open increment, the next task with its acceptance criteria, claims held by others, branch state, notes and project memory, in one read. |
| 2 | /sw:increment · specweave create-increment "<title>" |
One spec.md: Problem, Scope, Acceptance Criteria, Approach and the Tasks. |
| 3 | /sw:do · specweave task claim T-01 → task done T-01 --run "<test>" |
Work a task. done refuses a failing test and stores the evidence in the ledger. |
| 4 | specweave verify · /sw:review |
Runs your test, lint and build; a fresh-context review cites path:line. |
| 5 | /sw:done · specweave complete <id> |
Closes on a green verify. Acceptance criteria are met when their tasks are done; nobody ticks boxes. |
| 6 | "hand off" · specweave handoff |
Stop anywhere; the next tool picks up. |
An increment is one folder, .specweave/increments/NNNN-slug/, with one file you read (spec.md) and one the CLI appends to (ledger.jsonl). Increments from 2.x with a tasks.md keep working unchanged.
Works the way Claude Code Projects work
| Claude Code Projects | SpecWeave |
|---|---|
| A thread: one session, one branch, one PR | One increment |
| The thread's checklist | The ## Tasks of that increment's spec.md, state in ledger.jsonl |
Project memory (MEMORY.md + one file per fact) |
.specweave/memory/, same format, committed, so every tool and account sees it |
| Threads passing notes | specweave note "<text>" on another increment |
Project memory in claude.ai stays with one account. .specweave/memory/ travels with the code, so Codex, Grok and a second Claude subscription start from the same decisions.
The eleven skills
The CLI is the product and runs in any tool or in CI. The skills expose it to coding agents: /sw:<name> in Claude Code, sw-<name> in .claude/skills/ (for Projects threads, where plugins do not load) and .agents/skills/ (Codex, Grok).
| Skill | Use it for |
|---|---|
increment |
Plan the work as one spec.md. |
do |
Claim a task, implement it, close it with evidence. |
auto |
The same loop, unattended, until the tasks run out. |
team |
A worktree per agent, claims arbitrated by the ledger. |
review |
Fresh-context adversarial review; findings cite path:line. |
done |
Verify, review check, specweave complete. |
handoff |
"Hand off" and "pick up". |
sync |
GitHub, Jira and Azure DevOps, only when you run it. |
project |
Shared goals, artifacts and briefs across tools. |
brainstorm |
Framed alternatives, ending in a pick. |
jev |
Closed-set decisions in about 250 ms via Jev. |
Upgrade to 3.0
npm i -g specweave@3 specweave update
specweave update rewrites AGENTS.md into the lean form, turns CLAUDE.md into an import of it, keeps your own sections, and backs up the old files under .specweave/backups/. Existing increments need no migration. The 3.0.0 changelog lists what changed and what was removed. To have sessions hand off by themselves near the usage limit, run specweave auto-handoff on once.
Built With SpecWeave
Examples from the maintainer's portfolio. These are usage examples, not controlled productivity measurements.
| App | Platform | What It Does |
|---|---|---|
| EasyChamp | Web (GCP) | Enterprise sports league management. 20+ microservices, ML video analytics. 4 years in production. |
| SketchMate | App Store | AI drawing game — multi-model evaluation judges player art semantically. |
| Lulla | App Store | Baby sleep app with Apple Watch. ML cry classification (tired/hungry/pain). |
| Football 2026 | App Store + Web | World Cup 2026 companion. AI travel planner, live tickets, team stats. |
| SkillUp Football | App Store | Coaches monetize training via Stripe. Instagram-like feed, scheduling. |
| BizZone | App Store | Student & business events with AI-powered news generation. |
| EduFeed | Web | NotebookLM meets Zoom. Upload videos, get quizzes, flashcards, live rooms. |
| JobWeave | Web | AI-powered job search. Smart matching, resume optimization. |
| SpecWeave | npm | The framework itself. 600+ increments, 538+ releases. |
| SpecWeave Umbrella | GitHub | Multi-repo orchestration workspace for all repositories. |
| vskill | npm | Package manager for AI skills. Security scanning, 49 platforms. |
| verified-skill.com | Web | Skill marketplace & studio. 105K+ verified skills, eval system. |
Browse increments on GitHub — full transparency.
How It Compares
| Capability | Cursor Rules | Copilot Instructions | Windsurf | Cline | Vibe Coding | SpecWeave |
|---|---|---|---|---|---|---|
| Structured specs (Problem, ACs, Approach) | — | — | — | — | — | Yes |
One closure gate you can actually see (verify.json) |
— | — | — | — | — | Yes |
| Autonomous execution (hours, unattended) | — | — | — | — | — | Yes |
| Multi-agent teams (parallel, contract-first) | — | — | — | — | — | Yes |
| External sync (GitHub / JIRA / ADO) | — | — | — | — | — | Yes |
| Append-only ledger (claims, evidence, no lost work) | — | — | — | — | — | Yes |
| LSP code intelligence (198x faster) | — | — | — | — | — | Yes |
| Cross-tool handoff (any vendor, any subscription) | — | — | — | — | — | Yes |
Cursor tells AI "use Tailwind." SpecWeave tells AI "build a checkout flow against these five acceptance criteria, prove the tests pass, review the diff, then close."
Key Features
Spec-First Planning — Every feature starts as one spec.md: Problem, Scope, ACs, Approach and Tasks.
Evidence, not vibes — specweave task done --run "<test>" refuses a failing command and stores the exit code and output tail in the ledger.
Multi-agent, any vendor — A worktree per agent, claims through ledger.jsonl, one closure. Coordination happens only through committed files.
┌──────────────────┬──────────────────┬──────────────────┐
│ Agent 1 (auth) │ Agent 2 (payments)│ Agent 3 (catalog)│
│ T-01..T-04 │ T-05..T-08 │ T-09..T-12 │
│ ████████░░ 80% │ ██████░░░░ 60% │ ████░░░░░░ 40% │
└──────────────────┴──────────────────┴──────────────────┘
LSP Code Intelligence — 198x faster than grep, 0 false positives. Semantic references, definitions, and types.
11 skills, one source — the same skills for Claude Code, Codex and Grok; see The eleven skills.
External Sync — specweave sync push|pull|status|setup. GitHub is first-class; Jira and Azure DevOps are opt-in. Nothing calls a tracker unless you run sync.
Enterprise Ready — Compliance audit trails. Brownfield analysis. Multi-repo workspaces.
Dashboard — specweave dashboard shows intents, increments and evidence from local files, with no model calls.
Skills Ecosystem
SpecWeave skills are published and verified at verified-skill.com. The vskill package manager provides:
- Security scanning — 52 attack patterns, SHA-256 pinning, blocklist API
- 49 agent platforms — one install deploys to Claude Code, Cursor, Copilot, Windsurf, and 45 more
- Skill evals — unit tests, A/B comparisons, cross-model testing. Skills tested like programs.
- Visual Skill Studio —
vskill eval servefor benchmarks, comparisons, and history
npx vskill install remotion-best-practices # Install from registry npx vskill eval run my-skill # Run eval suite
Documentation
spec-weave.com — SpecWeave 3.0 · handoff · commands · skills · configuration
Contributing
Inside this repo dependency install scripts are disabled (.npmrc): run npm ci, then npm run setup (rebuilds the allowlisted native deps), and npm run security:scan before pushing — see SECURITY.md.
Community
Discord · YouTube · GitHub Issues
License
来源：GitHub · github.com