dsh-ctm

In long multi-turn agent sessions, the context the model actually sees remains a black box: which tokens were cache hits, which segments have gone stale, what the current system prompt looks like—developers often have no way to inspect any of it until something goes wrong. dsh-ctm breaks the full context into a turn-to-segment flow, color-coded by role, token count, and cache status, with provider-measured MECE token accounting (uncached input, cache hit, output) and a derived cache hit rate that makes per-turn context cost legible at a glance.

Beyond visibility, dsh-ctm ships an effectiveness engine that auto-classifies every segment as effective, redundant, stale, or injected with manual override support, and lets users replace, delete, roll back, or restore individual segments. When Apply for Real is enabled, edits are written into the session log as surface replace events, guaranteeing that replay, fork, and token accounting always reflect exactly what the model saw. The initial system prompt is editable in place, and changes take effect at the next prompt assembly without restarting the session.

The plugin communicates with the host over a single plain HTTP POST /ctm endpoint, carries zero runtime dependencies, and inlines Zod into its build artifacts so it works out of the box. It is aimed at prompt engineers, agent developers, and researchers who need to inspect, prune, or audit long context windows in production DeepSeek Harness deployments.