Optimizations

Cost-consciousness and performance are core design principles. This page covers optimization across agent execution, context management, and build pipeline.

Agent Execution Optimization

Model routing recommendations are on the Agents page →. Below covers output hygiene and researcher budgets.

Output Hygiene

Agents are configured to minimize token waste:

• Native tool flags: Agents use --quiet, --oneline, maxResults to truncate output at source
• Pipeline truncation: head/tail/grep used only when native flags are insufficient
• Context pruning: The context envelope is filtered per-agent: each agent receives only the sections relevant to its task

Researcher Budget Controls

The gem-researcher agent has configurable budgets to prevent runaway exploration:

orchestrator:
  researcher:
    max_searches: 10 # Max search calls per task
    max_files_to_read: 25 # Max files read per task
    max_depth: 3 # Max exploration depth (follow-up searches)

Context & Cache Optimization

Context envelope structure and memory tiers are covered in Core Concepts →. This section covers eviction and efficiency settings.

The context envelope is filtered per-agent (each receives only relevant sections). Facts are persisted only at confidence ≥ 0.85, promoted to durable cache at ≥ 0.90 after ≥ 3 uses. Auto-eviction at 90 days of no access.

Next

Configuration → · Core Concepts → · Workflow →