dsh-todo-guard
可靠的 DSH 待办面板：审查停靠面板可在重启后保留，并验证通过其自身工具提交的完成证据（官方待办面板保持整洁）
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:a903067276-rgb/dsh-todo-guard#4b7c24a48bbc9939b1bb6db9b16af8ab6e7cfc2fTodo Guard 以 dsh-todo-guard 发布，当前版本为 2.1.2。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。
A todo review plugin for DeepSeek Harness (dsh) web: the official todo panel stays untouched (item text carries nothing but the task itself), and a review dock hangs right below it — unfinished tasks persist across turns/restarts (checkpoint resume), completed items get evidence submitted through the plugin's own todo_evidence tool and verified automatically (verified / evidence-missing / text-only / unsubmitted), and missing evidence never blocks the check: it is only annotated and left for your one-click confirm (works while the dock is collapsed).
Unofficial: a community project, not a DeepSeek product.
npx @deepseek-ai/dsh web)main (default).tools.register, the mirror-projection contract (stateSchema+wire), the tools/post-execute hook and reading tool/call events from the session log all work. Contract points: mirror projection stateSchema+wire (0.1.1-rc.1+), multi-id dock coexistence (conversation.input.dock), settings card dual field key+id (satisfies rc.6 & rc.7+), tool registration via defineTool from @deepseek-ai/dsh-tools (peer dependency, provided by the host), tools/post-execute nudges (degrade silently on hosts without this hook — panel annotation only)v0.1.6 (dsh plugin add github:a903067276-rgb/dsh-todo-guard#v0.1.6)rc6-compat (no longer maintained)turn/start (by design); the plugin's mirror projection todo-guard/todos is not: the full list persists across turns, session switches and dsh restarts in the review dock (also the fallback when the official panel fails to render after a restart)content/status, and the plugin never stuffs anything into that text (the pre-2.1 habit of writing （证据：…） inside an item is retired — it showed up verbatim in the official panel); the review dock is an additional conversation.input.dock entry right below ittodo_evidence tool; the host verifies path existence (stat only, never reads content; relative paths resolve against the session cwd). Four annotations:
⚠ N items without evidence (call todo_evidence) / N evidence paths missing / too many checks at once)Confirm N button, so you never have to expand the list to confirm; all checked + nothing pending → auto-hides (confirmation is persisted per session in the browser)（证据：…） written inside item text before the upgrade still displays and verifies in the review dock (old sessions keep working), but is no longer recommendedThe agent writes clean task names into the official todo_write and submits evidence afterwards with todo_evidence:
todo_write { todos: [ { content: "Fix the button", status: "completed" }, … ] }
todo_evidence { items: [ { content: "Fix the button", evidence: ["lib/index.js"] } ] }
content must match the todo_write item text (that is how evidence is attached; a mismatch is reported back so it can be resubmitted)evidence accepts several entries: file / artifact / log paths, absolute paths and ~/ as-is; plain text credentials (e.g. "pnpm test all green") are accepted too, just not existence-checkedtool/call event — it survives restarts, session switches and session restoredsh plugin --profile web add "github:a903067276-rgb/dsh-todo-guard#main"
# restart dsh web
todo/write events (last-wins whole list), is never cleared at turn/start, and persists with the session projection cache — recovery after restart yields the "last valid list"conversation.input.dock id (order 10, below official todo / above queue) — pure official interfaces, zero model calls, zero tokenstodo_evidence (host-side tools.register) writes evidence into the session log; the dock joins it by item text; nothing extra is persisted on disk — a restart rebuilds it from the log/api/dsh-todo-guard/review (host aggregates evidence + stat existence only, never reads content); relative paths resolve against the session cwd; the dock refreshes immediately on list changes and polls every 4s to cover a "check first, submit evidence after" orderingtools/post-execute waterfall appends a non-blocking text notice after a successful write (not an Error); on older hosts without the hook it degrades to panel annotation onlyMIT