dsh-prompt-injector

Rules parked in the system prompt or persona layer sound explicit on paper, yet measured in practice they silently stop being followed once a session passes twenty turns. Refresh the code graph before coding, check the wiki first for factual questions — clear instructions that quietly die round after round. dsh-prompt-injector takes the straightforward fix: if a rule matters every turn, surface it in the model context every turn.

On the agent/pre-step hook, each enabled prompt from the settings page is appended to the current round context as a single collapsed Context injection notice line — the same channel dsh-mem0-plugins uses for memory recall, generalized from one plugin per discipline into one reusable prompt list. Each row picks its own trigger mode: everyTurn for constant reminders, or postCompaction to fire exactly once on the first substantive round after a context compaction. The default skipTrivial filter quietly skips bare acknowledgements such as ok, thanks, or 好的, so reminders never turn into nagging, while inputs carrying real content still trigger injection. All prompts are managed through a visual card on the settings page — add, edit, toggle — taking effect immediately with no restart.

The design boundary is deliberately strict: the plugin reminds, it never executes. No tool calls, no network access, no filesystem writes; the judgment call is always the model's. It ships as a standard DSH bundle plugin with zero source modifications, relying only on three peer packages the host provides at build time. It is aimed at DSH Web-profile users who maintain a growing set of behavioral rules and want them enforced per round without writing a separate plugin for each one.