dsh-rivermind: A Texas Hold'em AI Training Table in DeepSeek Harness

What Does It Do?

RiverMind is a DSH plugin for DeepSeek Harness that turns the harness into a two-player no-limit Texas Hold'em training table where you play against an AI opponent named Iris. It solves the problem of AI opponents that only produce one-off answers with no independent strategy or memory, giving Iris its own decision budget, conditional profile and long-term memory stored per player. The project is maintained by jiangzeyuan under the MIT license, written in TypeScript, currently at v0.3.2 and last updated in October 2026.

Core Features

• Two-player no-limit Hold'em table: play heads-up against Iris with BB and pot-ratio quick bets, dual BB/chip unit display and decimal BB input.
• Independent strategy with bounded tools: each Iris action defaults to a 60-second budget and at most 10 tool calls, adjustable and saved in the table, with timeouts or failed submissions clearly marked as safe fallbacks.
• Traceable decisions and hand review: after a hand ends you review the log and memory inside RiverMind, including the model and reasoning level actually requested, plus a compact Trace copy entry.
• Long-term memory per player: memory, budgets and finished hands persist in the data directory across DSH restarts.
• Rule-based sparring mode: a local preview that calls no model, used to verify the UI and poker rules and switchable between hands.
• Model selection follows DSH: Iris reads the current default provider, model and reasoning level on its first action, and the plugin never stores a separate API key.

How to Use This Plugin?

After enabling it in DSH, open "RiverMind 德扑训练场" from the left sidebar as a standalone main page and click the button to start the first hand. Configure a model in DSH before using AI mode; Iris reads the current default model on its first action. To switch models, pick one in the DSH session model selector and wait for the selection to finish, then fully quit and reopen DSH so a newly created Iris picks it up — clicking restart training alone does not reselect the model. During play you can adjust and save Iris's decision budget in the table, and after a hand you review the log, memory reference status and history on the same page. The rule-based sparring mode lets you validate the UI and rules without spending model quota.

How to Troubleshoot This Plugin?

• An old Iris instance shows up under DSH's ungrouped list and cannot be opened: update the build and restart DSH to release the stale instance.
• A GitHub source install is blocked during build: follow the DSH and pnpm prompts and merge the build authorization into the workspace configuration of the profile you use, then retry.
• Review records are missing model information: model details and traces absent from old records are not backfilled and are explicitly marked as unrecorded.
• The web page does not change after installing: restart the running web service process and open the table again.