Your First Task

This walkthrough shows what happens when you ask Weave to handle a task complex enough to trigger the full Plan → Review → Execute → Audit cycle. You'll see each stage in action — Loom orchestrating, Pattern planning, Weft reviewing, Tapestry executing, and Weft + Warp automatically auditing when done.

Prerequisites

Complete the Getting Started guide first to ensure Weave is installed and verified.

1. Start OpenCode with Weave

Open your project in a terminal and start OpenCode:

opencode

You should see Weave's confirmation:

✓ Weave loaded. Ready.

Loom is now your default agent. It's the orchestrator — it assesses every request and decides whether to handle it directly or delegate to specialists.

2. Ask Loom for Something Complex

Type a prompt that involves multiple distinct pieces of work:

Add a search feature to this project with an index service, an API endpoint, and a UI component.

Loom reads the request, assesses its complexity, and determines this is multi-step work that benefits from a structured plan. It delegates to Pattern to research the codebase and produce a plan.

When does Loom plan vs. act directly?

Simple requests ("rename this function", "fix this bug") are handled by Loom directly or delegated to Shuttle. Only tasks with enough complexity — multiple files, multiple steps, cross-cutting concerns — trigger the full plan cycle.

3. Pattern Creates a Plan

Pattern explores your codebase (via Thread), consults any relevant docs (via Spindle), and produces a structured plan file at .weave/plans/add-search.md. The plan contains checkboxed tasks — these are what Tapestry will execute later.

You'll see output like:

Pattern: Researching codebase structure...
Pattern: Creating implementation plan...
✓ Plan saved to .weave/plans/add-search.md

The plan file looks something like:

# Add Search Feature

## Tasks

- [ ] Create search index service (`src/services/search.ts`)
- [ ] Build search API endpoint (`src/routes/search.ts`)
- [ ] Add search UI component (`src/components/SearchBar.tsx`)
- [ ] Wire up endpoint in the router
- [ ] Add index population on data mutations

Each task is specific and actionable, with file references so Tapestry knows exactly where to work.

4. Weft Reviews the Plan

Before execution begins, Loom delegates to Weft to review the plan for quality:

Weft: Reviewing plan...
Weft: Checking file references...
Weft: Verifying task completeness...

[APPROVE]
Plan is solid. Tasks are well-scoped and file references
resolve correctly. No blocking issues found.

Weft is approval-biased — it approves unless there's a genuine blocking problem (unresolvable references, contradictory tasks, critical missing steps). If Weft rejects, Pattern revises and resubmits automatically.

5. Execute with /start-work

Once the plan is approved, run:

/start-work

Tapestry takes over. It reads the plan and executes each task sequentially — writing code, running commands, verifying output. As each task completes, Tapestry marks its checkbox:

Tapestry: Starting task 1/5: Create search index service
Tapestry: ✓ Completed task 1/5

Tapestry: Starting task 2/5: Build search API endpoint
Tapestry: ✓ Completed task 2/5

Tapestry: Starting task 3/5: Add search UI component
...

Interrupted? Close the terminal, reboot, whatever — when you come back and run /start-work again, Tapestry reads the plan, sees which boxes are checked, and resumes from the first unchecked task.

How continuation works during execution

Preview

Automatic continuation behavior is currently in preview. Manual resume via /start-work remains the most stable and predictable resume path.

There are two resume paths:

• Manual resume — run /start-work again
• Automatic recovery — if the session is compacted and Weave restores state, Tapestry can receive a recovery prompt and continue automatically

By default, Weave keeps post-compaction recovery on and generic idle nudges off. That means a quiet session does not automatically get a “keep going” prompt unless you opt into idle continuation settings.

Todo cleanup is separate from execution continuation:

• if Weave can update todos directly, it may silently mark leftover in_progress items as completed
• if direct todo writing is unavailable, a fallback todo-finalization prompt exists but is off by default

See Execution & Continuation for the full behavior model and configuration details.

6. Automatic Review & Audit

When all tasks are checked off, Tapestry automatically triggers a mandatory review gate — no user action required:

Weft: Reviewing code quality...
Warp: Auditing security...
Weft: [APPROVE] No quality issues found
Warp: [APPROVE] No security vulnerabilities detected
✓ All reviews passed

Weft reviews code quality — checking that the implementation is complete, consistent, and well-structured. Warp audits for security vulnerabilities — looking for injection risks, auth weaknesses, unsafe token handling, and similar issues. Both run in parallel. If either flags a blocking issue, it's surfaced to you before the workflow completes.

This gate is automatic and mandatory. You don't need to request it. If both agents are enabled (the default), it always runs after execution.

7. See the Results

When all tasks are complete:

✓ 5/5 tasks complete

Your plan file now shows all boxes checked:

- [x] Create search index service (`src/services/search.ts`)
- [x] Build search API endpoint (`src/routes/search.ts`)
- [x] Add search UI component (`src/components/SearchBar.tsx`)
- [x] Wire up endpoint in the router
- [x] Add index population on data mutations

All the files have been created and modified according to the plan.

What Just Happened?

Each agent played a specialized role:

Agent	What it did
Loom	Received your request, assessed complexity, orchestrated the workflow
Thread	Explored your codebase so Pattern could understand the project structure
Pattern	Produced a structured, checkboxed implementation plan
Weft	Validated the plan before execution; then reviewed code quality afterward
Tapestry	Executed every task, verified each one, marked it complete
Warp	Audited the completed code for security vulnerabilities

You typed one prompt and ran one command. The rest was handled automatically.

Simple Tasks

Not every task needs a plan. For simple requests — "rename this variable", "add a comment here", "fix this typo" — Loom handles them directly or delegates straight to Shuttle (the coding specialist). The planning cycle only kicks in when Loom determines the work is complex enough to benefit from it.

Next Steps

Now that you've seen the full workflow, explore how to customize it:

• Customization Overview — The full map of what you can configure
• Agent Configuration — Override models, temperatures, and tools per agent
• Skills — Inject domain knowledge into agent prompts
• Workflow Customization — Recipes for reshaping the delegation workflow