DeepSeek Harness 0.1.5-alpha.1 Preserves the KV Cache for Dynamic Prompts but Reworks Session and Plugin Interfaces

DeepSeek released [Harness 0.1.5-alpha.1](https://github.com/deepseek-ai/deepseek-harness/releases/tag/dsh-v0.1.5-alpha.1) on September 8. The change with the most direct impact on inference costs allows models that explicitly declare support for the capability to preserve the KV cache after dynamic modifications to the system prompt. Previously, when an agent rebuilt its prompt in response to changes in tools, skills, or policies, the server often had to reprocess the entire prefix. The new path can avoid this invalidation, but whether it takes effect still depends on the model provider’s capability flag; users should not assume that every OpenAI-compatible endpoint will benefit.

The greater upgrade risk lies in the state layer. Session V3 converts old sessions into the new log format while retaining the original files. It also writes the system prompt into the message history and migrates legacy PTC events and the `code` preset. The official notes explicitly state that downgrade reads are unsupported. Backup, indexing, and audit tools that parse logs directly must also be adapted to the new format. On the plugin side, `ctx.agent` has been removed, and callers must now pass the Agent explicitly. `Inbox` retains only its type interface; pending messages are now accessed through `agent.inbox`, while the previously public `hasPending` and `claim` methods are no longer available.

The interface layer adds a right-hand Sidebar that supports pagination, split views, and full-screen display. Generated files and in-chat file links can be opened directly in the Sidebar. Fixes also ensure that a paused goal cannot be resumed autonomously by the model, that project-root discovery does not fall back to a parent directory after encountering a permissions error, and that the native build dependency introduced by `fs-ext` on macOS and Linux is removed.

This remains a Developer Preview. An independent roundup indicates that npm’s `latest`, `next`, and `alpha` channels do not track the same release line. Before deployment, users should pin an exact version, copy session data, and test log migration, plugin startup, and rollback in an isolated profile instead of treating the alpha tag as a routine patch upgrade. The [third-party release roundup](https://freedom.tech/posts/2026-09-08-deepseek-harness-0-1-5-alpha-1/) also provides no latency or cache-hit-rate benchmarks, so the actual benefits of the KV cache change must still be measured against each deployment’s prompt-change frequency and provider-side billing model.