dsh-agent-mode-switcher

DeepSeek Harness plugin: switch the current conversation's agent preset (mode) after the model finishes answering, and keep chatting in the same session.

What it does

The session header already shows which mode (agent preset) a conversation runs — 标准模式 / PTC 模式 / 极简模式 or a custom preset. In a stock harness that label is read-only once the conversation started, because agentPreset.select refuses every non-blank session. This plugin:

1. Replaces the shipped header label with a live switcher chip (the same conversation.session.header.actions cell, id agent-preset).
2. Picking another preset calls the standard agentPreset.select RPC, which recomposes the session's agent from that preset. The conversation, its history, and its workspace stay exactly where they are — only the toolset/prompt composition changes, recorded in the session log so a resume or fork rebuilds the same mode. The select echo updates the session summary immediately (the shipped seat's own pattern), so the header chip reflects the new mode without a page refresh.
3. The chip is disabled while the model is answering and the host refuses the swap during a running turn regardless.

Requirements

The harness does not ship the idle-session preset switch yet (no upstream release channel), so a source checkout of deepseek-harness with the bundled patch applied is required today. On a harness without the patch the host still answers agent-preset-locked for started sessions and the switcher cannot change modes.

Install

The plugin itself never needs to be built. The repository ships the prebuilt host entry and browser bundle in lib/ (committed), so installing is a clone/pull plus one CLI command — no pnpm install in this repo, no prepare scripts, no allowBuilds approvals. The only build in the whole flow is the harness's own, for the core patch in step 1.

1. Patch the harness core

From this repository, against your deepseek-harness checkout:

node scripts/apply-patch.mjs /path/to/deepseek-harness
cd /path/to/deepseek-harness
npm run build:lib:host

The patch (patches/dsh-agent-preset-mid-conversation-switch.patch) relaxes the agentPreset.select guard from "blank only" to "not running", letting an idle session that already started switch modes. It is additive only — no existing behavior changes for blank sessions.

2. Install the plugin (no build needed)

From a local clone (recommended for iterating) — installs as a link: the committed lib/ is served as-is, and git pull in the clone updates the plugin without any build:

git clone https://github.com/my-dsh-plugin/agent-mode-switcher.git
pnpm dsh plugin add --profile web /path/to/agent-mode-switcher

(dsh CLI from your harness checkout; set DSH_HOME to your harness home if it is not the default ~/.dsh.)

Straight from git — pnpm fetches the repository and uses the committed lib/; no build script runs:

pnpm dsh plugin add --profile web github:my-dsh-plugin/agent-mode-switcher

dsh plugin add adds the dependency and reconciles the dsh.profile.bundles layer list. The manual equivalent is editing the profile's package.json:

"dependencies": {
  "dsh-agent-mode-switcher": "link:/path/to/agent-mode-switcher"
}

"dsh": {
  "profile": {
    "bundles": ["@deepseek-ai/dsh-base", "@deepseek-ai/dsh-web-app", "dsh-agent-mode-switcher"]
  }
}

then pnpm install inside the profile directory.

3. Restart and verify

Restart the harness (npx @deepseek-ai/dsh web or however you launch it). Open any started conversation: the header's agent-preset label is now a switcher chip. After the model finishes answering, pick another mode — the chip updates immediately and the conversation continues in the same session.

• The chip is there but selecting a mode does nothing → the core patch is not active in the running build (check the rebuild step, or that the old process was fully stopped).
• The chip is missing entirely → the plugin is not in the running profile's bundle layer (re-run dsh plugin add, verify the bundles list).

Maintenance

The patch is pinned to the current api-proxy select guard, so it drifts as the harness upstream moves. When your checkout updates, regenerate and re-verify the patch before committing:

node scripts/regenerate-patch.mjs /path/to/deepseek-harness
git -C /path/to/deepseek-harness stash
git -C /path/to/deepseek-harness apply --check /path/to/agent-mode-switcher/patches/dsh-agent-preset-mid-conversation-switch.patch
git -C /path/to/deepseek-harness stash pop

The regeneration covers only this plugin's core extension (the three api-proxy files). Set DSH_PATCH_BASE=<commit> when the extension lives on a committed branch instead of the working tree.

Development

Building is only for changing the plugin itself — consumers never build. It requires the sibling deepseek-harness checkout (../deepseek-harness) because the client bundle is produced by the shared harness preset:

pnpm install
pnpm test       # vitest: controller and select-echo suites
pnpm typecheck  # tsc -b over src + tests against the harness checkout
pnpm build      # tsc declarations + tsdown host + client bundle into lib/

After a build, commit lib/ so consumers keep getting the prebuilt artifacts (a git pull is all a link-installed profile needs to pick up a change).

Trade-offs

Switching a started conversation's mode leaves any history produced under the previous composition — logged tool calls the new toolset cannot make are the caller's accepted cost. The switch is refused while a turn runs, so the running request's tool schemas never change mid-answer.

Known Limitations and Deferred Work

• The switch is offered wherever the header renders; a session with pending interactions (an approval dialog) may still switch, and the host guard (running only) is the only enforcement.
• Roster refresh rides the settings/document-updated event and the component's own load; a preset authored on another machine requires a session reload to appear, matching the shipped agent-preset surfaces.
• Other open tabs see the new mode only after a refresh (same as the shipped agent-preset surfaces).

License

Apache-2.0