lattice

name	lattice
description	Lattice agent coordination guide — mental model, CLI commands, lifecycle discipline, and reference for file-based task tracking across agents and sessions.
homepage	https://github.com/Stage-11-Agentics/lattice
metadata	{"openclaw":{"emoji":"clipboard","requires":{"bins":["lattice"]},"install":[{"id":"pip","kind":"command","command":"pip install lattice-tracker","bins":["lattice"],"label":"Install Lattice (pip)"},{"id":"pipx","kind":"command","command":"pipx install lattice-tracker","bins":["lattice"],"label":"Install Lattice (pipx)"},{"id":"uv","kind":"command","command":"uv tool install lattice-tracker","bins":["lattice"],"label":"Install Lattice (uv)"}]}}

Lattice — Agent-Native Task Tracker

Lattice is a file-based, event-sourced task tracker. Everything lives in .lattice/ in the project root. Every change is an immutable event. No accounts, no API keys, no network required.

The Lifecycle — Opening and Closing Rituals

Every unit of work follows this arc: claim → understand → work → complete. The opening and closing rituals are non-negotiable.

Opening Ritual

# 1. Claim the next task (or create one)
lattice next --actor agent:claude-cli --claim --json

# 2. If the task needs planning (empty plan file), plan first
lattice status <task_id> in_planning --actor agent:claude-cli
# ... write the plan to .lattice/plans/<task_id>.md ...
lattice status <task_id> planned --actor agent:claude-cli
lattice status <task_id> in_progress --actor agent:claude-cli

lattice next --claim atomically assigns the highest-priority ready task to you and moves it to in_progress. If you already have a task in progress, it returns that one (resume-first logic).

If there's no existing task, create one:

lattice create "Fix the login bug" --actor agent:claude-cli --priority high

Closing Ritual

lattice complete is THE way to finish work. It performs the full completion ceremony in one command: posts a review comment, attaches a review artifact, transitions through review to done.

lattice complete <task_id> --review "What was done. Key decisions. Test results. What remains." --actor agent:claude-cli

The --review text is your breadcrumb for every future agent and human who reads this task. Be specific: files changed, approach taken, edge cases considered, anything left undone.

Do not use raw lattice status ... done to finish work. The complete command exists because completion requires evidence — a review comment and artifact. Skipping this ceremony leaves the task without an audit trail.

Outcome	Action
Done	lattice complete <task_id> --review "..." --actor agent:claude-cli
Need human input	lattice needs-human <task_id> "<what you need>" --actor agent:claude-cli (the task keeps its status — the flag is orthogonal)
Blocked on dependency	lattice status <task_id> blocked --actor agent:claude-cli + comment explaining the blocker

The Work In Between

Between opening and closing:

1. Read before writing. lattice show <task_id> --json. Check .lattice/plans/<task_id>.md and .lattice/notes/<task_id>.md for context from previous minds.
2. Check previous work. If the task has prior events, investigate what happened. git log --oneline --grep="<short_id>" for prior commits.
3. Baseline tests. Run the test suite before changing anything. You own new failures, not pre-existing ones.
4. Commit as you go. Each meaningful unit of progress gets a commit.
5. Push when done. Each task is durable on the remote immediately.

Core Commands

# Create
lattice create "Title" --actor agent:claude-cli --priority high --type bug

# List
lattice list                           # All active tasks
lattice list --status in_progress      # Filter by status
lattice list --assigned agent:claude-cli

# Show
lattice show PROJ-1                    # Summary
lattice show PROJ-1 --events           # Full event history

# Status transitions
lattice status PROJ-1 in_progress --actor agent:claude-cli

# Complete (the closing ritual)
lattice complete PROJ-1 --review "Review text" --actor agent:claude-cli

# Assign
lattice assign PROJ-1 agent:claude-cli --actor agent:claude-cli

# Comment
lattice comment PROJ-1 "Found root cause" --actor agent:claude-cli

# Link
lattice link PROJ-1 blocks PROJ-2 --actor agent:claude-cli

# Flag for human attention (orthogonal to status — task keeps its current status)
lattice needs-human PROJ-1 "Which OAuth provider?" --actor agent:claude-cli
lattice needs-human PROJ-1 --clear --note "Decided: use Google" --actor agent:claude-cli
lattice list --needs-human          # scannable queue of flagged tasks, any status

# Next task
lattice next --actor agent:claude-cli --claim --json

# File-decision links
lattice file-link PROJ-1 src/auth/jwt.ts --reason "JWT validation" --actor agent:claude-cli
lattice file-unlink PROJ-1 src/auth/jwt.ts --actor agent:claude-cli
lattice explain src/auth/jwt.ts              # what decisions shaped this file?
lattice explain src/auth/                    # directory prefix
lattice explain "src/auth/*.ts"              # glob

# Health
lattice weather    # Daily digest
lattice stats      # Statistics
lattice doctor     # Data integrity check

# Archive
lattice archive PROJ-1 --actor agent:claude-cli

Options for create: --priority (critical/high/medium/low/none), --type (task/bug/chore), --description "...", --assign agent:claude-cli

No epic or spike types — just items of work with a dependency graph. Lattice intentionally rejects umbrella/exploratory ticket types. Express multi-phase or umbrella work as a plain task with subtask_of links from its children. Express exploratory/investigation work as a plain task whose deliverable is a concrete artifact (plan doc, prototype, decision). The subtask + dependency graph (subtask_of, blocks, depends_on) gives you epic-shape and spike-shape without dedicated types. Every ticket is a chunk of work with a real output, not a bucket or a question.

Task description depth: Match description detail to task ambiguity. Bug fixes and chores can be one-liners ("Add regex validation to frequency names"). Features and integration tasks should include: (1) what it does, (2) acceptance criteria, (3) architectural context, (4) what the user/operator experiences when done.

Relationship types for link: blocks, blocked_by, subtask_of, parent_of, depends_on, depended_on_by, related_to

Status Workflow

backlog → in_planning → planned → in_progress → review → in_validation → pr_open → done
                                       ↕
                                    blocked

Transitions are enforced. Use --force --reason "..." to override when genuinely needed.

needs-human is a flag, not a status. It rides orthogonally on top of whatever status a task is in — a task can be in_progress and flagged, blocked and flagged, even done and flagged. Set it with lattice needs-human <task> "<reason>" (reason required) and clear it with lattice needs-human <task> --clear. The flag never moves the task. blocked stays a status for generic external dependencies; needs-human means "waiting on a human specifically," and the two can coexist.

in_validation is the e2e gate. After local review passes, prove the change works against a running system — browser automation for web, simulator MCP for mobile, curl flows for APIs. Exercise the actual flow the ticket touched, then record evidence with lattice attach <task> --role validation (or lattice comment <task> --role validation). Transitioning to pr_open is blocked until validation evidence is recorded; if e2e genuinely doesn't apply, record a one-line N/A justification instead — explicit, never silent. Validation failure routes back to in_progress (impl-level) or in_planning (plan-level), and counts toward the 3-cycle rework valve. The bar: "I saw it work," not "I think it should work."

Actor IDs

Every command requires --actor. Format: prefix:identifier

• agent:claude-cli — default for your own actions
• agent:worker-1, agent:worker-2 — multi-agent setups
• human:username — when acting on behalf of a human

Task IDs

• Short ID: PROJ-1, PROJ-42 (use these in conversation)
• Full ULID: task_01HQ... (internal, always accepted)

All commands support --json for structured output: {"ok": true, "data": ...} or {"ok": false, "error": {"code": "...", "message": "..."}}.

Heartbeat Mode

Check if enabled: look for "heartbeat": {"enabled": true} in .lattice/config.json.

When enabled, keep advancing after each task: complete the current task → lattice next --claim → work the next one → repeat. Stop after max_advances (default 10), when the backlog is empty, or when a task is flagged needs-human or hits blocked.

Rules

1. Open before you work. Every unit of work starts with a Lattice task. Not after. Not during. Before.
2. Close with lattice complete. Never raw lattice status ... done.
3. One task at a time. Finish or transition before claiming the next.
4. Don't force transitions. If a transition fails, investigate why.
5. Don't cancel human tasks. Flag them with lattice needs-human instead — let the human decide.
6. Comment liberally. The next agent has no hallway to find you in.

References

Detailed guides live in {baseDir}/references/:

• Multi-Agent Guide — orchestrator/worker patterns, actor ID conventions, concurrency safety
• Worktree Guide — git worktree protocol for parallel development with shared Lattice state