Write Vibe ADR

Vibe ADRs are concise, agent-facing architecture rules in docs/adr/. Writing an ADR is a two-step process: write the document, then register it in the Architecture Decisions table in AGENTS.md.

When To Create An ADR

• A design discussion creates a new architecture rule, constraint, or direction.
• An existing convention is undocumented and caused a bug, review comment, or repeated confusion.
• A pattern applies across a Vibe area or boundary, not just one file.
• Future work should follow a direction that may differ from the current architecture.

Do not create ADRs for one-off implementation choices, formatting rules, or behavior already enforced by linting.

Steps

1. Confirm the scope

Use one ADR when the decision is cohesive. Split ADRs when separate agents would need to read them for unrelated change areas.

Existing ADR topics:

• 0001 - architecture principles
• 0002 - core engine and delivery surfaces
• 0003 - event-driven agent loop
• 0004 - typed permissioned tools
• 0005 - layered configuration
• 0006 - local sessions
• 0007 - extension mechanisms
• 0008 - feature instrumentation

2. Determine the next number

ls docs/adr/

Use the next sequential number, zero-padded to 4 digits: 0009-my-decision.md.

3. Write the ADR

Use this format:

# 0009 Decision Title

## Decision

What we decided. Include concrete rules and current-vs-aspirational direction when relevant.

## Rationale

Why this decision exists. Name the ambiguity, pressure, or tradeoff.

## Agent Guidance

- Concrete instructions an agent should follow while changing code.
- Keep guidance task-oriented and easy to scan.

## Flag To User When

- Situations where an agent must stop and ask because current code or user request conflicts with the ADR direction.

Keep ADRs concise. Match the existing 20-50 line style. Do not add a status field.

4. Register in AGENTS.md

Add or update a row in the Architecture Decisions table:

| <task trigger> | [0009 Decision Title](docs/adr/0009-my-decision.md) |

The trigger text must describe what the agent is changing, not vague architecture language.

Good triggers:

• "Adding a new delivery surface, protocol bridge, or UI-owned behavior"
• "Changing tool args/results, permissions, output limits, or adapters"
• "Changing session transcript shape, metadata, resume, rewind, or migrations"

Bad triggers:

• "Working on architecture"
• "When relevant"
• "Making decisions"

If the ADR is not registered in AGENTS.md, agents will not reliably discover it.

Common Mistakes