Lattice

English | 简体中文

Ticket-Driven Development for coding agents.

Lattice is an open-source toolkit of Agent Skills and optional Claude Code hooks that gives coding agents a disciplined path from product intent to a merged pull request. It runs locally with your existing Git and GitHub credentials, and is packaged for Claude Code and Codex. It is discipline-first, not heavy — it constrains the path (every shippable branch binds to a tkt-N/spc-N, every PR carries its Why/How/Spec line) but leaves the model free to think and code inside each step; the guardrails are a floor, not a ceiling.

Quick start

Install once per machine, then open any repo and start a ticket.

# 1) Portable skills (all supported agents: Claude Code, Codex, Cursor, …)
npx skills add percena/lattice -g -y

# 2) Claude Code plugin (same skills + optional hooks)
/plugin marketplace add percena/lattice
/plugin install lattice@percena

# 3) In your repo
/start-work

Advanced install (org roll-out, private forks, local dev, refresh) → docs/getting-started.md.

The loop

/start-work is the universal entry — it classifies scope and routes into this loop (for a new feature it delegates to /create-spec; to resume an existing ticket, /start-work tkt-N).

  /create-spec  |  /create-review
       ↓
  /create-tickets
       ↓
  /start-work
       ↓
  implement
       ↓
  /create-pr
       ↓
  /finish-work

# 1) Plan — lock scope, split into tickets
/create-spec           # e.g. "Add a payment feature"
/create-tickets

# 2) Execute the ticket
/start-work tkt-N

# 3) Ship
/create-pr
/finish-work pr N

Philosophy

Constrain the path, not the model. Lattice scripts the backbone — Spec → ticket → worktree → PR → merge — so the chain never skips a step and never loses its lineage. It does not script how the model reasons, what order to think, or what to output inside each step. Strong models stay creative; the framework keeps the chain resumable. Guardrails (worktree isolation, alignment checks, lineage) are a floor, not a ceiling.

Local-first retrieval, transparent memory. Spec, ticket, and review are written as templated local files under .lattice/ (ADRs under docs/adr/) — grep-able, reviewable, and committed with your repo. Most context lookups (acceptance criteria, last review outcome, which tickets a spec split into) resolve with a local cat/grep in milliseconds; GitHub is queried only for facts that can only come from the remote (issue/PR live state, comments, CI). Memory is not a framework black box — it is an explicit engineering artifact you and your team own.

Lineage. Each delivery leaves a traceable chain — a Spec (spc-N) splits into tickets (tkt-N), each ticket lands in a PR (pr-N), every review is a rev-…. These IDs are the binder file names, so the whole chain is recoverable with one grep -r spc-N .lattice/, across sessions, without a network hop.

Skills

Skill	Purpose	Slash
start-work	Classify S/M/C, bind ticket + worktree, resume by id	/start-work
create-spec	Persist a Lattice Spec (spc-n) with acceptance criteria	/create-spec
create-review	Persist a Lattice Review (rev-YYYYMMDD-HHMMSSZ) with an explicit outcome	/create-review
create-tickets	Split locked scope into GitHub issues + binders	/create-tickets
create-pr	Open/update a well-formed GitHub PR	/create-pr
finish-work	Update base, alignment-check, merge, cleanup	/finish-work
_lattice-lib	Shared scripts backing the above (co-install, not a slash entry)	—

Not part of the delivery loop — three tiers, none create lineage nodes:

Tier	Skill(s)	Notes
PR-scoped quality side-paths	review-code · review-production	Optional, before/after /create-pr; no _lattice-lib
standalone doc tool	generate-wiki	wiki/ + llms.txt; anytime; no _lattice-lib
out-of-band companion (create-* family)	create-adr	writes docs/adr/NNN; co-installs _lattice-lib; not a lineage node — invoked alongside /create-spec//create-review (same worktree, promotes cross-feature decisions); never a loop entry or a Spec substitute

Documentation

Doc	Topic
getting-started	Install, auto-ensure, profiles, daily path, advanced install
github-surface	Kind + priority labels, optional Project auto-add
CONTRIBUTING	Changing skills/plugins in this monorepo
SECURITY	Vulnerability reporting
CODE OF CONDUCT	Community standards
CHANGELOG	Plugin SemVer

Requirements

git, gh, jq, python3 (≥ 3.8), curl — plus an agent that runs Agent Skills or Claude Code plugins. Hook tests need bats.

License

MIT — see LICENSE.

Acknowledgements

Thanks to the linuxdo community for discussion, sharing, and feedback.