dsh-todo-freshness-guard
在 DeepSeek Harness 中完成 todo_write 对账前，提醒并阻止过时的工具工作
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:lamost423/dsh-todo-freshness-guard#6ea5d64d751d5174b4bc3d0f4e0a9810fbd5024eTodo Freshness Guard 以 dsh-todo-freshness-guard 发布，当前版本为 0.1.2。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
English | 简体中文
Problem it solves: During long agent turns, the model may keep calling tools without updating its unfinished todo_write list. The visible plan then stops matching the work actually being done.
Most painful failure: The agent is still active, but the user is looking at stale status—completed work still appears pending, new blockers are missing, and the next step is unclear. This removes progress visibility exactly when a long task needs it most.
How it fixes it: This out-of-tree DeepSeek Harness Guard reminds the model to reconcile the complete Todo list, then blocks ordinary tools if the list remains stale. todo_write and the outer run_code recovery path always remain available.
**解决的问题：**Agent 在长任务中可能持续调用工具，却不更新未完成的 todo_write 列表，导致用户看到的计划与真实执行状态脱节。
**最痛的问题：**Agent 明明还在运行，用户看到的却是过期状态——已完成事项仍显示未完成，新阻塞没有记录，下一步也无法判断。任务越长，进度越不可信。
DSH Todo Freshness Guard keeps long-running task progress visible
After a successful todo_write contains unfinished items, the Guard counts non-bookkeeping tool attempts for that Session:
reminderAfterCalls, it injects one model-visible reminder to reconcile the complete todo list.blockAfterCalls, it denies ordinary tools until another complete todo_write replaces the list.todo_write always remains reachable.run_code transport remains reachable so Code Mode can call todo_write.This plugin fixes stale todo_write state. It does not replace or patch the filesystem Write tool.
0.1.0-rc.6 through 0.1.7-rc.2 (Session formats 0–4)^22.19.0 || >=24.0.0Install the compatible DSH CLI, then add the release archive to the profile that should enforce the Guard:
npm install --global @deepseek-ai/dsh@0.1.7-rc.2
dsh plugin --profile web add https://github.com/lamost423/dsh-todo-freshness-guard/releases/download/v0.1.2/dsh-todo-freshness-guard-0.1.2.tgz
dsh web
To install from a checkout instead:
git clone https://github.com/lamost423/dsh-todo-freshness-guard.git
cd dsh-todo-freshness-guard
corepack enable
pnpm install --frozen-lockfile
pnpm build
dsh plugin --profile web add .
dsh web
The package contributes this default patch layer:
- insert:
- id: todo-freshness-guard
name: dsh-todo-freshness-guard
config:
reminderAfterCalls: 5
blockAfterCalls: 8
Profile and command-line patch layers applied after this bundle may replace the row configuration. blockAfterCalls must be an integer greater than reminderAfterCalls; both values must be positive integers.
Remove it with:
dsh plugin --profile web remove dsh-todo-freshness-guard
On DSH 0.1.7, v0.1.1 fails the turn with format v4 message requires a producer-owned source kind as soon as its reminder fires, and every later turn of that Session fails the same way. Add the v0.1.2 archive with the install command above, which replaces v0.1.1, then restart dsh web: a running Host keeps the old code until it restarts. The affected Session opens again, with the failed turn marked as interrupted; the part of that turn that was never saved, and anything sent in that Session between the failure and the restart, cannot be recovered.
pnpm check
pnpm pack --pack-destination /tmp
The test suite covers the native and Code Mode policies, concurrent resets, todo recovery through each Host's Session log reader, Loader composition, reminder attribution under Session format 4, the packed Bundle contract, and an actual Web startup through official DSH 0.1.0-rc.6.
MIT. Derived portions retain the DeepSeek Harness license; see NOTICE.