dsh-undo

Context undo/redo plugin for DeepSeek Harness (dsh), inspired by the undo/redo experience in coding agents like opencode.

The plugin adds two model-facing tools to every root agent:

• undo — rolls the model context back to the end of the last completed step. Every assistant message and tool result of that step is shadowed out of the visible context; the durable session log keeps every rolled-back event, so nothing is lost.
• redo — restores the messages the most recent undo removed, re-adding them to the context exactly as they were.

How it works

The session log is append-only, so undo never deletes events. Instead it appends one empty-content assistant/message whose surface replace op shadows the target step's model-visible nodes. Empty assistant content derives to no message, so the context the model sees is exactly the pre-step context, while the log remains a complete, replayable, reversible history. redo re-appends fresh copies of the shadowed events as ordinary surface appends (with re-minted message identities; tool results keep their call correlation, so transcript tool-call → tool-result adjacency is preserved).

Design notes:

• Undo unit is one completed step (the newest step/end whose step still has visible messages). Repeated undo calls walk back one step at a time.
• User messages are never rolled back — a user prompt stays in context; only the assistant work done in response can be undone.
• Redo is invalidated by new user input — any new user-role message clears the redo stack (standard undo/redo semantics). The undo/redo tools' own calls, results, and redo copies do not invalidate it.
• Durability — every operation flushes through the shared session durability barrier (ctx.sessions.flush) before reporting success; if persistence cannot be proven, the tool returns persistence_uncertain rather than lying.
• Undo history is process-local — like a browser's undo stack, the redo stack is in memory. Restarting the process keeps the log consistent (markers stay shadowed, copies stay restored) but clears the redo stack.
• Scope — the tools are installed only for root agents published after the plugin loads (subagents and pre-existing sessions are not retrofitted).

Install

The package is a dsh bundle: package.json declares dsh.bundle pointing at cordis.patch.yml, which activates the plugin row.

From npm:

dsh plugin --profile demo add dsh-undo

From git (source checkout — the prepare script builds lib/ on install; authorize the build in the profile's pnpm-workspace.yaml first):

dsh plugin --profile demo add github:LingLambda/dsh-undo#<sha>

For local development against a source checkout of deepseek-harness, load the plugin directly as an overlay:

pnpm dsh web --patch ./cordis.patch.yml --patch /absolute/path/to/dsh-undo/overlay.yml

where overlay.yml references the source entry:

- insert:
    - id: undo
      name: /absolute/path/to/dsh-undo/src/index.ts

Usage

The model calls undo when you ask it to undo, rewind, or roll back its latest action, and redo to restore what was undone:

> That edit was wrong. Undo it and try a different approach.

Each call's result reports what happened (rolled_back with the shadowed count and marker seq, restored, nothing_to_undo, nothing_to_redo, or an error).

Develop

yarn install
yarn typecheck
yarn test
yarn build

License

MIT