dsh-evolve-modes

By default, an agent works in a fairly rigid way: it receives a task and just executes it, occasionally pauses to plan, and rarely looks back to question its own answer. But real-world tasks vary enormously — some are routine chores that should be done quickly, some are high-impact decisions that demand careful thought before action, and some are deliverables that need strict verification. dsh-evolve-modes was built to solve exactly this problem: it takes "making the agent work smarter" out of hardcoded behavior and turns it into a set of workflow decisions you can adjust at any moment.

The plugin adds a compact control to the DeepSeek Harness input area, letting you freely combine four dimensions per task: work state (normal vs. plan), thinking strategy (standard vs. first-principles), quality gate (off, adversarial review, or acceptance review), and self-evolving (off or on). It doesn't fork Harness, doesn't replicate the agent loop, and doesn't touch core code. Once installed, the currently active combination is always visible next to the input box, and one click opens the controls to recompose it.

The most interesting part is the self-evolving mode. The plugin analyzes multiple completed sessions to identify stable user identity, preferences, and work requirements. After every 3 parent-agent replies it runs a learning pass and produces rule proposals backed by verbatim user evidence. But the learning process is deliberately isolated: it doesn't inherit parent session history, doesn't spawn sub-agents, doesn't carry tools, and doesn't read AGENTS.md from project folders. Proposals are just proposals by default — you have to manually approve or ignore them in the settings page before they take effect as global rules. Quality gates are equally restrained: the review agent only reports evidence, gaps, and follow-up actions, never silently rewrites, retries, or patches your reply.

If you're using DeepSeek Harness and want your workflows to be composable, auditable, and continuously improvable — without being locked into a single "persona mode" — this plugin is for you. Whether you need to finish routine tasks fast, make high-impact decisions, verify implementations with confidence, or challenge high-risk answers for hidden flaws, you can recompose a suitable workflow in seconds, and every learning or review action stays traceable, controllable, and recoverable.