anton-abyzov/specweave/do
Claude Code and agent skill by anton-abyzov/specweave
- community
Checking that the SKILL.md file is still in the repository.
Work a SpecWeave increment task by task - claim, implement inside its Files, test, commit, done with evidence - alone, unattended or in lanes, then close. Use for "implement", "next task".
Install
npx vskill@latest install anton-abyzov/specweave/doSemi-extensibleMentions customization, but needs skill-specific knowledge. Read the extensible skills standard
SKILL.md
sw-do: work an increment to done
Tasks are defined in spec.md under ## Tasks; their state lives only in the increment's ledger.jsonl, which the CLI appends to. Never write state into spec.md.
Start
specweave pickup applies any handoff pushed from another tool or account, then prints the open increment, the next task with its acceptance criteria, claims other agents hold, and notes left for you. No increment yet? Run sw-increment first.
The loop
specweave task nextprints the next claimable task with its AC text.specweave task claim T-03 --line "Writing the ledger fold". Exit 3: someone holds it (your own earlier claim from
a crashed session: say so, then --force); 4: its Files overlap a live claim; 6: its dependencies are not done. Otherwise pick another. Never edit before claiming.
- Implement inside that task's
Filesonly. Need another file? Claim the task that
owns it, or add a task to spec.md ## Tasks. Use the words in GLOSSARY.md, if the project has one, for every new name.
- Run the task's
Test. Red means fix the code; never skip, weaken or delete a test.
Test-first order is optional: follow it only when the user, testing.mode: TDD or the project's conventions ask. Don't add unit tests the task did not ask for; check the AC end to end, the way a user would hit it.
- Commit with the increment id first:
git commit -m "0042: fold the ledger". specweave task done T-03 --run "npm run test:e2e -- double-claim" --line "Wrote the ledger fold" --link <PR URL>.
Exit 5: the command failed and the task stays open. No Test? --evidence "<sha> + what you checked". --line is the card text (one line, past tense when done, at most 100 characters); --link is the PR, report or artifact behind it.
- Stuck:
specweave task block T-03 --reason "<what is missing>". Not needed:
specweave task skip T-03 --reason "<why>". Then back to 1.
Commands take the increment id last (specweave task next 0042) when more than one increment is active. specweave task list shows the whole board. For a status, print specweave cards --md: one ✓ / ✱ / ○ text (link) line per task. Work with no increment: specweave card add "<text>", then card start|done|wait C-01.
Unattended (auto mode)
"Auto", "run until done", "ship this while I sleep": the same loop without stopping to ask. The bar does not drop.
- Start only when the spec has nothing under
## Decisions>### Openand every
task has Files and a Test; otherwise fix the spec first.
specweave auto 0042records the session (--dry-runpreviews,--resetclears
stale state). A tool that runs SpecWeave's Stop hook hands you the remaining work after every turn; in any other tool, keep looping yourself.
- Blocked (missing secret, ambiguous spec, outside dependency):
task blockand move
on. When only blocked tasks remain, specweave cancel-auto and tell the user what each one needs, instead of looping to the turn limit.
specweave auto-statusshows progress; one auto session per project.
Parallel lanes
Three or more disjoint lanes, or 15+ tasks: you are the lead. Split, start, watch, merge and close; do not implement.
- Group tasks into 2 to 5 lanes whose
Filesdo not overlap. Shared contracts
(types, schema) go in an early task the others list under **Dependencies**: T-01.
- Lane agents share the existing checkout and branch, each with a unique
SPECWEAVE_AGENT(such asapiorweb);
stage only your lane's files and commit with an explicit path list, for example git add -- src/api.ts && git commit -m "0042: build API" -- src/api.ts. Use the same unique identity for every task command in the lane. Give a lane its own worktree only when the user or the project's settings ask for worktrees (your tool's or app's worktree feature, or git worktree add .worktrees/0042-api -b inc/0042-api, never ../). Brief each one: its tasks, this loop, "edit only their Files", and on finish specweave task release --all-mine plus specweave note "api: done" 0042.
- Watch with
specweave task list 0042;blockedrows are your queue. A claim older
than 2 hours with no progress may be taken with --force.
- If lanes had their own worktrees, merge their branches in dependency order (keep every
ledger.jsonl line from both sides) and remove the worktrees. Then close as below. Lane agents never run specweave handoff. specweave team opens agent panes for the same setup in tools that have them.
Finish without asking
When task next has nothing left, run sw-done: verify, review, complete, in the same session. "All tasks are done, should I close?" is the anti-pattern.
Plan for it from the start: a progress checklist or status you keep for the user ends with "Verify and close 0042", not with "merge" or "release". Work that shipped through a PR still gets specweave task done T-03 --evidence "<commit sha>" first, or the ledger never shows it finished. A session never ends with every task done and the increment still open; if a close gate stops you (verify fails, a manual check only the user can do), say which one in your report. A lane agent leaves the close to its lead.
Stopping early
Out of tokens, switching tool or account: specweave handoff --reason "<why>" (sw-handoff).
Manual path (no CLI)
Read the tasks from spec.md and the state from ledger.jsonl: the next task is the lowest T-id nobody holds whose Files no live claim holds. Append one JSON object per line, keys in this order, UTF-8, LF:
PowerShell never uses >> here: it writes UTF-16 and every other tool fails to read it. Events: claim, done (needs evidence), release, block and skip (both need note); any of them may carry "line" (card text) and "link", for example {"t":"T-02","e":"skip","by":"codex@mbp","at":"2026-09-02T12:00:00Z","note":"covered by T-01"}. by is SPECWEAVE_AGENT if set, else <tool>@<host>. Unattended or in lanes, follow the same steps by hand: there is no hook and no session file.
Rules
- A claim older than 2 hours with no
doneis stale and may be taken over. - Append only: never edit, reorder or delete a ledger line. On a merge conflict in
ledger.jsonl keep every line from both sides.
- Never widen a task's
Filesor weaken a test to keep a loop moving. - Never edit
.specweave/state/auto-mode.jsonby hand;specweave auto --resetclears it.
Security
- Security scanPassed, 96 of 100Pattern checks on the skill's files.
- Model reviewPassed
Quality evaluation
- Task evaluationNot runNo evaluation yet. Run one in Skill Studio.
Source
- SKILL.md in the repositoryCheckingplugins/specweave/skills/do/SKILL.md
Popularity
- Installs
- 6
- Unique users
- 1
- Repository stars
- 85
- Repository forks
- 9
- Repository npm downloads, 7 days
- 6,183
- 7-day trend score
- 48
Installs come from the vskill CLI. Stars and forks are for the whole GitHub repository, not this skill alone. Downloads are for the repository's npm package. Refreshed 16 Mar 2026, 7 months ago.
Works with
Any agent vskill installs to: 56 targets in all. These 39 have a page on this site.
- Claude Code
- Cursor
- GitHub Copilot
- Windsurf
- Codex
- Gemini CLI
- Cline
- Amp
- Roo Code
- Goose
- Continue
- Kilo Code
- Junie
- Trae
- OpenHands
- Qwen Code
- AI Agent
- Tabnine Chat
- Sourcegraph Cody
- Amazon Q
- Firebase Genkit
- Sweep
- Devin
- Phind
- Mentat
- Claude Desktop
- Aider
- GPT Pilot
- Plandex
- bolt.new
- v0
- Replit Agent
- Anse
- TypingMind
- BetterChatGPT
- LibreChat
- Warp AI
- Pieces
- Fig AI
Embed badge
[](https://verified-skill.com/skills/anton-abyzov/specweave/do)