@deepseek-ai/dsh-ops-todo-tree

An investigation tree tool for DeepSeek Harness ops mode — replaces todo_write with a diverge-converge tree of steps, milestones, dead ends, and a resolved terminal.

What it does

Agent-driven investigation tracking: the agent maintains a tree of investigation steps via the todo_tree model tool. Each call appends an incremental event to the session log; a session projection folds these into the current tree state; the client renders a git-graph-style flat list with colored lanes, expandable rows, and status glyphs.

• 8 actions: create_tree, add_step, add_milestone, start, complete, abandon, resolve, note
• 6 statuses: goal, pending, in_progress, done, dead_end, resolved
• Dead ends are not deleted — they stay on the tree as part of the exploration record
• Branch with branch=true to explore side paths in parallel lanes
• Every call returns the full tree + a status summary (advisor, not gatekeeper)

Installation

Add to dsh-web-app dependencies and reference in the ops preset's agent.cordis.yml:

- id: tool-ops-todo-tree
  name: '@deepseek-ai/dsh-ops-todo-tree'

Model Experience

todo_tree tool

What the model sees

A tool description explaining the 8 actions and when to use each. A system prompt section with usage guidance.

Token effect

Tool schema + description (~200 tokens). System prompt section (~300 tokens).

KV Cache effect

Stable across turns — tool description and prompt section are static.

Known Limitations and Deferred Work

• No cross-session continuity (v1: one tree per session)
• No human editing (pure agent-driven)
• Lane/depth computed client-side (layout is derived, not stored)
• Subagents cannot directly write to the tree — main agent must relay their results
• No replay/timeline scrub UI (events are in the session log, but no dedicated timeline view)