dsh-context-maid

The most frustrating moment in a long session is rarely the conversation itself — it is the tool output and dead logs that bloat the context. A failed build spits three hundred lines of stderr, the same file gets read five times, an intermediate experiment result is already superseded. Worse, the requirement you stated thirty minutes ago, the goal you are halfway through, the decision you just confirmed — they are buried in that noise, and the moment compression triggers they risk being flattened into the same vague summary as the stderr they sit beside. dsh-context-maid exists to keep garbage and critical context apart so that neither drags the other down.

Its approach is a five-tier curation model (PIN / KEEP / SLIM / SWEEP / FOLD) applied to every message node. Tool output is trimmed in a content-aware way — errors keep their tail, JSON keeps its skeleton, logs keep head and tail — rather than blindly chopping a fixed number of characters from both ends. User-stated requirements, corrections, and the current goal are pinned in a no-loss-compression zone and re-injected into the summarisation prompt every cycle. Before anything is removed it is archived into the ACP ledger, and the compression summary itself carries a supersedes chain so you can always trace back. You can also point the summarisation step at a cheaper or local OpenAI-compatible model, keeping the main conversation on your preferred engine.

If your daily dsh workflow involves multi-turn debugging, batch experiments, or any pipeline heavy on tool calls, think of this plugin as a context maid: it tosses the trash, files what matters into a drawer, wipes the table — and that drawer is open to you whenever you need it back.