comment-discipline

Stops Claude Code from flooding the context window with command output. Blocks streaming/foreground long-running commands before they run and coaches toward background tasks and redirect-then-grep.

Routes work through a six-role agent hierarchy — Orchestrator, Ultra-Advisor, Architect, Reviewer, Implementor, Task-Runner — each on the model priced for its job, with Ultra-Advisor escalation gated on your approval and zero-token usage tracking. See README for the full mechanics.

Claude Code plugin marketplace — context-window discipline tooling

Dispatches tool-heavy, high-output, and information-gathering work to a cheaper Haiku subagent that carries out explicit orders and returns a compact report, so the main high-reasoning agent spends its expensive tokens on judgment instead of on tool output. Session hooks never reach subagents, so a PreToolUse hook rewrites each Agent dispatch in flight to stamp the directive onto the spawned subagent's prompt. The same hook intercepts destructive and outward-facing commands from the runner — it cannot judge whether destroying something is right — and asks YOU to approve each one, since no model can accept that risk on your behalf; it denies instead when the permission mode means no prompt reaches a person. Toggle with /task-gopher.