Zero-Read Orchestration Protocol: Preventing Context Rot in Complex Claude Code Workflows

[neoncapy]

The Problem

Context rot is real in Claude Code. As your context window fills up, output quality degrades. Read a few large files, run some analyses, and your agent starts making mistakes it would not have made at the start.

The Zero-Read Orchestration Protocol

claude-code-orchestration-protocol is a /orchestrate slash command that turns your main Claude Code session into a pure router.

Core Rules

• 100-line hard limit: The main session never reads anything over 100 lines
• Delegate thinking: When someone asks "how should we do X?" — launch a reasoning agent, do not think in the main context
• Agent output protocol: Every agent writes a ## Summary header above a --- line. The orchestrator reads only above the line

QC Loops

No arbitrary iteration caps. QC runs until genuinely zero issues at ALL severity levels (CRITICAL, MAJOR, MINOR). Safety valves: same-issue escape after 3 attempts, token-budget escape at YELLOW threshold.

Anti-Polling

Never poll agent output while agents run. Each poll dumps 50-150K tokens of agent transcript into the orchestrator context. Use zero-poll waiting instead.

Token Budget Management

5-tier system: GREEN (0-40%), GREEN+ (40-50%, distillation trigger), YELLOW (50-55%, no new waves), ORANGE (55-65%, stop), RED (65%+, emergency handover).

Progressive Documentation

Every task gets a plan file before work begins. Handover protocol means you can /clear and resume exactly where you left off.

Results

27 cloners in 2 weeks with zero promotion. The protocol was developed over months of real production use processing hundreds of documents, writing thesis content, and building complex pipelines.

MIT licensed. What patterns have you found for managing context in Claude Code?

https://github.com/orangefineblue/claude-code-orchestration-protocol

[hesreallyhim]

27 cloners in 2 weeks with zero promotion

Maybe you should keep it that way then.