dsh-workspace-change-awareness

In a DeepSeek Harness session, an agent edits files while a human developer, an IDE, or an external script may be modifying the same repository concurrently. Without an external-change awareness layer, the agent keeps acting on stale cached content and produces results built on incorrect premises. dsh-workspace-change-awareness closes that blind spot: before every model request it invalidates affected paths through a shared Chokidar watcher, then re-resolves and stats each path via ctx.fs so the model sees the true workspace state.

The plugin enforces three layers of protection. First, a pre-step notice: when a path changes outside the session's acknowledged baseline, a user-role notice is injected at agent/pre-step, asking the model to re-read the affected files and inspect the current diff; the notice is persisted before the model can see it. Second, a turn-stopping final check: if a new filesystem state arrives or the watcher boundary degrades while the model is still reasoning, agent/turn-stopping performs one additional step to feed the latest state into the model. Third, graceful degradation and recovery: a watcher error first emits a model-visible coverage-degradation warning before the registry attempts a fresh watcher, and previously acknowledged paths continue to be revalidated through ctx.fs.stat() on every subsequent step, so a brief outage never silently drops a change.

This mechanism fits real-world engineering environments where multiple actors touch the same files simultaneously: a developer's terminal, IDE auto-save, CI hooks, and the harness agent all operate on the same codebase. A brand-new agent starts from the current workspace state and does not report initial files as changes; a resumed or compacted session receives a conservative re-read reminder because process-local observations cannot cover edits made while DSH was stopped. The bundle works out of the box with no configuration, is released under the MIT license, and suits teams and individuals who want an external change-awareness layer in their harness workflow.