Relay Guide

Not Yet Active

The relay infrastructure is implemented but not yet wired into the CLI. See Context Relay & Compaction for the full design document.

Relay is Wave's context compaction mechanism. When an agent approaches its token limit, relay summarizes the conversation and resumes with fresh context.

Why Relay?

LLMs have finite context windows. As agents work on complex tasks, conversations grow and quality degrades. Relay solves this by:

1. Detecting when context usage exceeds a threshold
2. Using a summarizer persona to create a checkpoint
3. Resuming with fresh context and the checkpoint

How It Works

Agent working (context growing)
         |
Token threshold reached (80%)
         |
Summarizer creates checkpoint
         |
Fresh agent starts with checkpoint
         |
Work continues

Configuration

Global Settings

In wave.yaml:

runtime:
  relay:
    token_threshold_percent: 80
    strategy: summarize_to_checkpoint

Per-Step Settings

In pipeline steps:

steps:
  - id: implement
    persona: craftsman
    handover:
      compaction:
        trigger: "token_limit_80%"
        persona: summarizer

Field	Default	Description
token_threshold_percent	80	Context usage trigger (50-95)
trigger	token_limit_80%	Step-level condition
persona	summarizer	Checkpoint persona

The Summarizer Persona

A specialized persona for creating checkpoints:

personas:
  summarizer:
    adapter: claude
    system_prompt_file: .agents/personas/summarizer.md
    temperature: 0.0
    permissions:
      allowed_tools: ["Read"]
      deny: ["Write(*)", "Bash(*)"]

Summarizer System Prompt

# Summarizer Persona

Create structured checkpoints for context relay.

## Output Sections

### Completed Work
- List actions with file paths
- Note decisions made

### Current State
- Modified files
- Test/build status

### Remaining Work
- Tasks not started
- Partial tasks
- Blockers

### Resume Instructions
- Specific next step
- Important context

Checkpoint Structure

# Checkpoint: Feature Implementation

## Completed Work
- Created user model at src/models/user.go
- Implemented CreateUser endpoint
- Tests: 12/12 passing

## Current State
- Modified: src/models/user.go, src/api/users.go
- Build: successful

## Remaining Work
- [ ] Implement GetUser endpoint
- [ ] Add integration tests

## Resume Instructions
Next: Implement GetUser. Follow pattern in CreateUser.

Enabling Relay

For long-running steps:

- id: implement
  persona: craftsman
  handover:
    compaction:
      trigger: "token_limit_80%"
      persona: summarizer
    contract:
      type: test_suite
      command: "go test ./..."

Relay Limits

Prevent infinite loops:

runtime:
  relay:
    token_threshold_percent: 80
  meta_pipeline:
    max_total_tokens: 500000

If the summarizer hits its token limit, the pipeline halts.

Monitoring

When activated, relay events will appear in the NDJSON event stream indicating compaction triggers, checkpoint creation, and resumption.

Best Practices

When to Enable

Enable for:

• Complex implementation work
• Multiple iterations expected
• Large codebase processing

Threshold Tuning

Scenario	Threshold
Short tasks	85-90%
Complex tasks	70-75%
Large codebases	75-80%

Checkpoint Quality

Good checkpoints are:

• Complete - All relevant information
• Structured - Clear sections
• Actionable - Explicit next steps
• Concise - No repetition

Troubleshooting

Relay loops: Check if task is completable, review checkpoint quality, lower threshold.

Lost context: Improve summarizer prompt, ensure "Completed Work" is detailed.

Summarizer failures: Check permissions (needs Read), verify temperature is 0.0.

Related Topics

• Manifest Schema Reference - Relay configuration
• Pipeline Schema Reference - Compaction settings
• Personas Guide - Summarizer persona setup