DeepSeek Harness In multi-turn conversations, progress often relies on the Agent’s subjective descriptions and lacks objective evidence. The dsh-rich-tracking plugin introduces a kanban docked at the bottom of the conversation, converting progress into artifact-fact-based percentages, and provides a complete checkpoint and instruction injection mechanism.

Plugin Positioning

This is a client-side plugin maintained by svgop. It provides percentage progress scoreboards, evidence-based row records, host-captured Git checkpoints, and operational instructions such as pursue/align/dismiss.

Core Capabilities

  • Percentage scoreboard: Displays real-time progress for each row without relying on subjective descriptions.
  • Evidence binding: Percentages must be based on artifact facts; “hallucinations” are strictly prohibited. Rows with percentage >= 1 must include evidence. 100% progress requires all items to be checked and have receipts.
  • Git checkpoints: Through the tracking_checkpoint tool, the host can automatically capture the current branch, HEAD, and dirty state, and freeze the kanban.
  • Instruction injection: Operations such as Pursue/Align/Dismiss are directly converted into the Agent’s next-step instructions, taking over the execution flow.
  • Row record dialog: Click ? to view full details, evidence sources, and Markdown-format records.
  • Research and automatic execution: Scout supports launching subagents for each row’s task to perform competitive research; Play mode can automatically execute high-value todos.
  • Forced synchronization: The /track command can force the current ledger state to be injected into the Agent.

Installation and Enablement

dsh plugin --profile web add dsh-rich-tracking

After installation, you must restart the dsh web process before the tracking_write and tracking_checkpoint tools take effect for all presets, and the kanban will be rendered in the conversation dock (order 5, below the todo pill).

Typical Usage

  1. View progress: After restarting, the kanban appears below the todo pill. Row count is limited to 1-12 (recommended 3-7).
  2. View details: Click the ? button on the left side of a row to open a record dialog containing status, evidence, details, and sources.
  3. Competitive research: Click the Scout button. The system launches subagents for each row’s task, researches 3-6 competitors or implementation approaches, and writes the results back to detail and sources.
  4. Forced synchronization: In the conversation, enter /track (or /track <message>) to inject the full ledger and correction instructions into the Agent.
  5. Automatic execution: Enable Play mode. The system automatically prioritizes the highest-value todo row at the end of each turn.
  6. Manual writing: After completing a task/step, call tracking_write to update the status.

Notes

  • System requirements: DeepSeek Harness version must be >=0.1.2-alpha.1.
  • Evidence rules: The kanban is not reset on turn/start until all rows reach 100% or are manually cleared.
  • Scout dependencies: The Scout feature requires a persistent research digest file located in .docs/digest/ or .docs/research/.
  • Registration mechanism: The host must register tracking/write, tracking/checkpoint, and tracking/decision in ctx.sessionEventTypes.

dsh-rich-tracking resolves the problems of opaque progress and lack of verifiability by enforcing artifact-fact-based percentage calculation. Combined with Git checkpoints and operational instructions, it makes multi-turn task management trackable and auditable.

Plugin Directory | GitHub Repository