dsh-trilogy

Working with an AI coding assistant, the most frustrating part is rarely writing the code itself—it is starting every new session as if meeting the project for the first time: the architecture is forgotten, the rationale behind past decisions is lost, the pitfalls hit last week have vanished. dsh-trilogy maintains three Markdown memory files per workspace (current state, decisions, session log), auto-injects them at session start, and nudges for missing records at the end, so cross-session project memory is no longer at the mercy of the model.

Rather than hoping a prompt reminds the model to read files, the plugin guarantees the behaviour at the host level: it scaffolds missing files, injects context at session start (skipping unchanged content to stay KV-cache friendly), and queues a gentle reminder on your next message if real work was done without being logged. The first time you open a project it even drives a bootstrap pass—reading documentation, running tests, filling in the state—after which you simply get on with writing code.

Built for engineers who use DeepSeek Harness in daily development: sessions that open and close frequently, no desire to re-explain context every time, and a need for AI decisions and project state to be traceable. The plugin performs no vector search, makes no external model calls, and all memory lives in plain Markdown files you can open in any editor.