Context Relay & Compaction

Not Yet Active

The relay infrastructure is implemented but not yet wired into the CLI. The configuration fields exist in the manifest schema and the relay logic is in internal/relay/, but wave run does not currently instantiate the relay monitor. This page documents the intended design for when relay is activated.

Long-running agent tasks degrade silently as context fills up. The relay mechanism solves this by compacting context at a configurable threshold, handing off to a fresh instance with a structured checkpoint.

The Problem

LLM agents have finite context windows. As a step executes, the context accumulates:

• System prompt
• Tool call history
• File contents read
• Intermediate reasoning

At ~80% utilization, output quality degrades — the agent loses track of earlier decisions, repeats work, or contradicts itself.

How Relay Works

1. Monitor — Wave tracks token usage during step execution.
2. Trigger — At the configured threshold (default 80%), relay fires.
3. Summarize — A separate summarizer persona receives the full session history and produces a structured checkpoint.
4. Resume — A fresh instance of the original persona starts with the checkpoint injected as initial context.

Configuration

Manifest Settings

runtime:
  relay:
    token_threshold_percent: 80          # Trigger at 80% context usage
    strategy: summarize_to_checkpoint    # Only strategy currently

Per-Step Override

steps:
  - id: implement
    handover:
      compaction:
        trigger: "token_limit_80%"      # Per-step threshold
        persona: summarizer             # Which persona summarizes

Summarizer Persona

The summarizer needs minimal permissions — it only reads the session and writes a checkpoint:

personas:
  summarizer:
    adapter: claude
    system_prompt_file: .agents/personas/summarizer.md
    temperature: 0.0    # Deterministic — no creativity needed
    permissions:
      allowed_tools: ["Read"]
      deny: ["Write(*)", "Bash(*)"]

Checkpoint Structure

The summarizer produces a structured document:

# Relay Checkpoint

## Completed Work
- Created user model in src/models/user.go
- Added migration in db/migrations/003_users.sql
- Implemented GET /users endpoint

## Current State
- Working on POST /users endpoint
- User validation logic partially complete
- Test file created but not populated

## Modified Files
- src/models/user.go (new)
- src/handlers/users.go (new, incomplete)
- db/migrations/003_users.sql (new)
- src/routes/api.go (modified — added /users routes)

## Remaining Work
1. Complete POST /users handler with validation
2. Add authentication middleware integration
3. Write unit tests for user model
4. Write integration tests for /users endpoints
5. Update API documentation

## Resume Instructions
Continue from step 1 of remaining work. The POST handler
skeleton exists in src/handlers/users.go — complete the
validation logic and error handling.

Relay Events (Planned)

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

Edge Cases

Summarizer Hits Its Own Limit

The summarizer has a hard token cap. If the session is too large to summarize within that cap, the pipeline halts rather than entering infinite compaction. This is a safety mechanism — it's better to fail clearly than to loop.

Multiple Relays Per Step

A single step can relay multiple times. Each relay increments the relay count. The fresh instance receives only the most recent checkpoint, not the full chain.

Relay During Matrix Execution

Each matrix worker has its own independent relay tracking. Worker A relaying does not affect worker B.

Best Practices

1. Set thresholds conservatively — 80% is a good default. Going higher risks degradation before relay fires.
2. Keep summarizer prompts focused — the summarizer should produce structured, actionable checkpoints, not narrative summaries.
3. Use relay for implementation steps — navigation and review steps rarely need relay because they're shorter.
4. Monitor relay frequency — frequent relays may indicate steps that should be broken into smaller sub-steps.

Further Reading

• Manifest Schema — RelayConfig — field reference
• Pipeline Schema — CompactionConfig — per-step config
• Architecture — how relay fits into the system