dsh-todo-guard

dsh-todo-guard ✅

English | 简体中文

A reliable todo panel plugin for DeepSeek Harness (dsh) web: the todo strip survives restarts (works around the official panel disappearing after a restart), and completion is verified against evidence paths.

Unofficial project: independently developed and maintained by a community member, not an official DeepSeek product.

Screenshot

Requirements

• DSH web >= 0.1.0-rc.7 (run with npx @deepseek-ai/dsh web)
• rc.6 users: install the frozen rc6-compat tag instead: dsh plugin add github:a903067276-rgb/dsh-todo-guard#rc6-compat (no maintenance; upgrade to rc.7+ recommended)

Features

• Restart-proof panel — replaces the official todo strip (conversation.input.dock cell, shadowed at priority: -1); after a dsh restart, reopening the session shows the todo list again
• Completion verification (three states) — when the agent marks a todo completed, evidence is checked automatically:
  • （证据：路径） and the path exists → ✅ green check (verified)
  • （证据：路径） but the path does not exist → 🚫 blocked, the agent gets a clear error and must fix it
  • no evidence marker → ⚠️ allowed, the item shows a yellow "unverified" badge (no false blocks on "did it but forgot to write evidence"; visible at a glance)

Evidence syntax

Write （证据：路径） inside a todo item (multiple allowed). Relative paths resolve against the session working directory; absolute paths work as-is:

Fix the button（证据：lib/index.js）
Run tests（证据：test/run.log）（证据：docs/result.md）

Install

dsh plugin --profile web add "github:a903067276-rgb/dsh-todo-guard#main"
# restart dsh web to activate

How it works (why it survives restarts)

• Data: todos live in the session event stream (official todo/write, last-wins whole-list), persisted on disk — the official panel just fails to re-render them after a restart
• Panel: official useProjection('todos') projection + same-id slot shadowing (priority: -1, lowest renders) — official interfaces only
• Verification: official tools/pre-execute waterfall intercepts todo_write before it commits; failed evidence denies the write with a readable reason

Notes

• Pure local logic: zero model calls, zero token cost
• Fake files are the verification boundary: an existing path does not prove the work was done, but the unverified badge keeps the loop honest
• Strict mode (block when evidence is missing) is planned for v2

License

MIT