dsh-message-editor

DeepSeek Harness stores every conversation as an append-only event log, so there is
no built-in "undo". dsh-message-editor brings back the three moves every chat deserves —
撤回 (recall), 编辑重发 (edit-and-resend) and 重新生成 (regenerate).

Recall / edit remove the target messages from the conversation view and the model
context — that is exactly the effect you see. What stays untouched is the underlying
durable transcript: it remains append-only, old events are never rewritten or deleted,
and the plugin merely appends one valid replacement event (the same replace primitive
the built-in compaction uses) to rewind the surface — so the log keeps a full audit trail
of every rewind.

✨ Features

Why it's different

• 🎯 Whole-round recall — one click removes the input and its output (including tool rows), not just a single bubble.
• 🖥️ Web + Desktop — the same plugin covers both surfaces of DeepSeek Harness.
• 🔒 Removed from view & context, not from the log — recalled/edited messages disappear from the conversation view and the model context, while the durable transcript is never rewritten or deleted; the plugin only appends valid, typed session events (the same replace primitive the built-in compaction uses), so the log keeps a full audit trail.
• 🧠 View ⇄ context in sync — the conversation view always reflects exactly what the agent sees.
• ⚡ Try in 30 seconds — the dynamic form installs in your current session with no rebuild.

🚀 Quick start

Requires DeepSeek Harness with the dsh CLI. Installs the plugin as a profile
bundle and automatically rebuilds the Web client:

# DSH Desktop (desktop profile)
dsh plugin --profile desktop add dsh-message-editor

# standalone Web (`dsh web` / web profile)
dsh plugin --profile web add dsh-message-editor

⚠️ Restart after install. A running app keeps the previously loaded bundle
in memory, so quit and reopen DSH Desktop (or restart the dsh process for
a standalone Web deployment) before the plugin activates.

That's it — after the restart, hover any assistant reply, or any user message,
and use ↩ / ✎ / ↻.

📦 Installation

1. Profile bundle (recommended)

The package declares a dsh.bundle manifest, so it installs through the official
plugin path into any profile:

dsh plugin --profile <name> add dsh-message-editor

⚠️ Restart required. The install writes the new files and re-renders the
profile composition, but a running app does not hot-reload bundles — quit
and reopen DSH Desktop (or restart the dsh process for a standalone Web
deployment) to load the plugin. To uninstall:
dsh plugin --profile <name> remove dsh-message-editor (then restart again).

It also shows up in dsh-market for
one-click install from inside Settings (same restart applies).

2. Manual install (no dsh CLI)

The same result with plain file edits and pnpm — exactly the steps
dsh plugin add performs for you:

1. Open the profile manifest (defaults: ~/.dsh/profiles/desktop on DSH
   Desktop, ~/.dsh/profiles/web for standalone Web) and add both the
   dependency and the bundle-layer entry:

   {
    "dependencies": {
      "dsh-message-editor": "^0.2.0"
    },
    "dsh": {
      "profile": {
        "bundles": [
          "@deepseek-ai/dsh-base",
          "@deepseek-ai/dsh-web-app",
          "dsh-message-editor"
        ]
      }
    }
   }

   (Keep whatever entries your profile already has; only add the two
   dsh-message-editor lines.)

2. Install inside the profile directory:

   cd ~/.dsh/profiles/<name> && pnpm install

3. Restart DSH Desktop / the dsh process (see above).

For local development, point the dependency at a checkout instead of the
registry: "dsh-message-editor": "file:/path/to/dsh-message-editor" — or let
dsh do it: dsh plugin --profile <name> add /path/to/dsh-message-editor.

3. npm package + composition (classic)

npm i dsh-message-editor

Add the package to the harness composition (cordis.yml of the app/deployment you use):

- name: 'dsh-message-editor'

The client half is picked up automatically from the package's dsh.client metadata and
bundled into the Web client (a client-module rebuild happens automatically when the
composition changes). The Host half registers the same-origin HTTP route
/api/plugins/message-editor/* for the browser UI.

4. Dynamic plugin (current session — no install, no rebuild)

Use the dynamic entries shipped in the package. In the session where you want the
feature:

1. Open the plugin editor and define a new plugin from lib/dynamic-host.js
   (Host half) and lib/dynamic-client.js (Client half).
2. Approve and run the Client half.
3. Done — hover any assistant reply, or any user message, and use ↩ / ✎ / ↻.

The dynamic host registers the same operations behind the package-private
harness.handle RPC (messageEditor.recall / messageEditor.editAndResend /
messageEditor.regenerate).

⚙️ Settings → General

🧠 How it works

 durable transcript (append-only)          model context & view
 ┌────────────────────────────────┐        ┌──────────────────┐
 │ ... target message             │        │  … target message │
 │     ↓ shadow span              │        │       ↓ rewind    │
 │ [target … last surface node]   │ ─────▶ │  (empty replace   │
 │     ↳ one replacement          │        │   = context cut)  │
 │       assistant/message (empty)│        └──────────────────┘
 │     ↳ optional original-input  │        agent.followup(new prompt)
 └────────────────────────────────┘        → next turn rebuilds request

1. Host core (lib/host-core.js, zero runtime imports) locates the target
   message in the session's live surface, computes the shadow span
   [message … last surface node], and appends one replacement
   assistant/message with an empty body — a valid surface node that
   derives to no model message, so the LLM context simply rewinds.
2. Edit / regenerate additionally call agent.followup(...) with the
   (new) prompt text; the agent's next turn builds its request from the
   rewound session.deriveMessages().
3. Client (lib/client.js) registers:
   • a user-actions conversation node under every user message
     (编辑 / 撤回 row with an inline editor); recall echoes the text into the
     composer,
   • the recall-marker node renderer: a notice row that injects CSS hiding
     every shadowed message row from the flow (view and model context stay in
     sync), plus the optional original-input comparison block,
   • the message-editor entry in the conversation.chat.assistant-actions
     strip (撤回 / 重新生成),
   • two preference toggles under Settings → General.

Two different layers are at play: the durable transcript (append-only; old
events are never rewritten or deleted) and the model-visible surface (rewound
by an appended replacement event). So the old events stay in the log as an audit
trail — but they are synchronized out of both the model context and the visible
conversation, and the view always reflects what the agent actually sees.
Persistence, projections and the transcript remain consistent because the plugin
only appends valid, typed session events.

⚠️ Requirements & limitations

• Only user messages can be edited; recall works on user and assistant
  messages. Tool results are shadowed along with the recalled range but are not
  themselves recall targets.
• The agent must be idle: while a reply is streaming you must stop it
  (⏹) before recalling or editing. The Host rejects with agent-busy
  otherwise.
• Recall/edit operate on the active model surface: a message that was
  already compacted away or previously recalled is rejected
  (target-shadowed).
• Regenerate re-sends only the text of the original prompt; prompts that
  carried images fall back to the text-only content.

🗺️ Roadmap

• [ ] Version timeline / reroll — browse and jump between past rewinds of a message
• [ ] Forked-session edit — edit a past message and continue in a branched session
• [ ] More locales beyond 简体中文 / English

🛠️ Development

# structure
lib/host-core.js       # transport-neutral host logic (no imports)
lib/index.js           # published Host: harness RPC + HTTP route
lib/client.js          # client SOURCE (React via import; pluggable transport)
lib/client.bundle.js   # BUILT client bundle — the self-registering loader entry
                       # (`window.__ModuleLoader__.load`) served by client-modules
lib/dynamic-host.js    # GENERATED dynamic Host half (from lib/host-core.js)
lib/dynamic-client.js  # GENERATED dynamic Client half (from lib/client.js)
scripts/build-client.mjs      # bundle lib/client.js → lib/client.bundle.js
scripts/generate-dynamic.mjs  # generate both dynamic entries from the canonical sources
scripts/check-dynamic.mjs     # syntax-check the dynamic entries (function bodies)
test/                 # vitest suite: host-core ops + generated-entry smoke tests
.github/workflows/    # CI (syntax + build-sync + tests) and npm publish (v* tags)
cordis.patch.yml      # dsh.bundle profile patch layer

pnpm install          # install dev dependencies (vitest, esbuild)
pnpm check            # syntax-check sources AND the generated dynamic entries
pnpm build            # regenerate lib/dynamic-*.js + lib/client.bundle.js
pnpm test             # run the host-core unit tests
npm pack --dry-run    # verify the published file list

⚠️ Generated files. lib/dynamic-host.js, lib/dynamic-client.js and
lib/client.bundle.js are built artifacts generated from lib/host-core.js
and lib/client.js — never edit them by hand. CI fails when a committed
artifact is stale (git diff --exit-code), so run pnpm build before
committing. The dynamic client reuses the same client source as the published
one and only swaps the transport (host.call vs the HTTP route) via
__setMessageEditorWire.

PRs and issues are welcome — see CONTRIBUTING (coming soon)
and the issue tracker.

📚 Ecosystem

Listed on the dsh-plugin topic and
installable from dsh-market. For a
curated overview of the DeepSeek Harness plugin ecosystem, see
awesome-dsh-plugin.

👥 Team

Built by the OfferKuai team — an AI job application
assistant on a mission that "users need results, not repeated conversations".
Founder: Zhaofeng (Yaming). This plugin is released as open source for the
DeepSeek Harness community.

📄 License

MIT