dsh-approval-comment

The default approval flow in DSH Web has a frustrating blind spot. Clicking Reject does not stop the model-it speculates about why you said no and pivots to a different plan on its own. Force-stopping the session instead means the rejection never enters the context, so the next turn starts as if nothing happened. Either way, you are losing a tug-of-war with an agent that cannot take a clear instruction at the decision point.\n\ndsh-approval-comment keeps the familiar approval banner and adds a single annotation field, expanding the outcome into four meaningful combinations. Allow with a note lets the tool execute while the model reads the note on its very next step and self-corrects. Reject with a note tells the model exactly what to change, and it adjusts and continues seamlessly, still inside the same approval window. A plain Reject terminates the current turn instantly and waits for your next instruction. A plain Allow executes as usual. The note is injected via the steer mechanism so it lands in the live context, and both the rejection and the note remain in the transcript, so a manual stop does not erase them. Under the hood the plugin is pure JavaScript with zero third-party dependencies and zero build steps, and it does not patch any core package, meaning upstream upgrades will never overwrite it.\n\nIt is built for DSH Web users who run agent workflows with frequent tool-approval prompts. You do not need to rewrite the approval layer. Install the plugin, type a one-line note when you reject, and the model actually hears you, acts on it, and moves forward without guessing.