harness

To prevent context drift, you can provision a disposable dev container and use a Planner/Generator/Evaluator split to manage the agent lifecycle. This ensures context health by applying compaction and reset rules when prompts approach large-context thresholds.

A Planner/Generator/Evaluator split improves output reliability by ensuring the agent never grades its own work. A dedicated Evaluator path enforces critique-only validation and triggers retries across phases, preventing self-validated outputs.

The best way to offload large tool-call outputs is to keep large artifacts like logs and build files in the filesystem. The orchestration layer loads only summaries back into the context, preventing token overflow and maintaining session stability.

A dev container is recommended for managing context health because it provisions a disposable, per-task environment. This ensures work never inherits hidden state, providing a clean baseline for multi-step coding and architecture tasks.

To initialize a multi-step implementation session, run the harness to provision a fresh environment, load CONTEXT.md then AGENTS.md, and configure the Planner/Generator/Evaluator execution split before beginning your task.

You should use context compaction and reset protocols when prompts approach large-context thresholds. These rehydration rules manage token zones and enforce evaluator-only critique, maintaining context health during complex multi-step tasks.