dsh-completion-playground

A fullscreen raw-prompt playground for DeepSeek Harness, packaged as an agent mode. Every "Run" is an ordinary conversation turn whose model call happens to be DeepSeek's raw /beta/completions endpoint — so the harness keeps doing the work it already knows: Chat, the built-in Trajectory, the session list, token accounting and cancellation.

What you get

• New Session — the mode picker offers Playground Mode. The hero screen stays exactly as shipped (mode chip, workspace picker, normal composer).
• After the first turn — a Completion tab appears with a fullscreen prompt editor. The shipped Chat and Trajectory tabs are untouched, and the native composer is vacated, so no second input box floats under the transcript.
• Logprob — a sibling tab that runs the prompt as an ordinary turn while asking the endpoint for logprobs, then draws every position's candidate distribution as probability bars. It has its own max_tokens and candidate count (1–20). The run is a real turn, so it appears in Trajectory; the distribution rides the assistant message, so the bars come back after a tab switch or a reload.
• Run — the whole document is sent as one user message, the completion streams back and is appended to the document. Stop cancels the turn (and the upstream request).
• max_tokens (1–4096; the endpoint documents a 4K cap) and a stop list with escape sequences, sent as the request's max_tokens / stop.
• Trajectory — each Run is a real turn: one user record (the document at that moment) plus one assistant record (what the model wrote).

Install

From GitHub (pin a tag for reproducibility):

dsh plugin --profile <your-profile> add github:Zhichii/dsh-completion-playground#v0.3.0

Or open Settings → Plugins → Install, paste the same spec, and confirm. A git install tracks the default branch unless you pin #<ref>.

The package ships plain ESM (index.js) and a plain browser bundle (client.js) — there is no build step and no install script, so the install needs no build approval.

Configure the API key

The Host half resolves the reference DEEPSEEK_API_KEY through the harness credential seam (ctx.credentials) and falls back to the process environment. The value is sent only to the configured upstream. Optional environment overrides:

Variable	Meaning
DSH_COMPLETION_BETA_BASE_URL	Upstream base URL, default https://api.deepseek.com/beta
DSH_COMPLETION_ATTRIBUTION	User-Agent sent upstream (white-labelling)

Use

1. New Session → pick Playground Mode.
2. Write the first prompt in the normal composer. The mode has already claimed the session, so that first request is already pinned to the completion route.
3. Click the Completion tab → the fullscreen editor. Write the document, then Run (or Ctrl/Cmd + Enter).
4. Inspect the Trajectory tab: the Run is recorded as a turn.

Stop sequences

The stop field is single-line, so line breaks are typed as escapes (decoded in the browser, sent upstream as real characters):

Typed	Sent as the stop sequence
\n	one line feed
\n\n	two consecutive line feeds (stop at a blank line)
\nUser:	line feed followed by User:
\t	one tab
\\	one literal backslash

Unknown escapes (\q) keep their backslash. At most 16 sequences; empty rows are dropped.

Gemini mode

The bundle also ships a second mode, Gemini 模式 — a chat wrapper over the same endpoint that speaks a Gemini-style transcript instead of raw documents:

User
<your message>

Gemini
**Analyzing

• No system prompt. The adapter never reads system or developer messages, and the mode's preset declares no prompt sections, so nothing the deployment would normally inject reaches the model. Persona and style live in the few-shot rounds instead.

• No tool calls. The preset composes no tools and the adapter ignores tool declarations and tool results; the request carries no tools field.

• Few-shot prefix: gemini-prefix.md ships beside the module — a real Gemini-style transcript (four rounds, <think> blocks and all) that leads every request. It carries both the format and the persona, because this mode sends no system prompt.

• Postfix: the prompt closes with Gemini\n<think>\n**Analyzing (the heading is left open), so the model completes it, writes the analysis, closes </think> and then answers. Setting geminiPostfix: "Gemini\n" instead makes the model emit <think> itself, so the stored answer begins with the tag.

• Both are overridable through the bundle row's config:

  - id: completion-plan-b-host
    name: 'dsh-completion-playground'
    config:
      geminiPrefix: |          # replaces the shipped transcript entirely
        User
        你是谁
  
        Gemini
        …
      geminiPostfix: "Gemini\n<think>\n**Analyzing"   # or "Gemini\n" to let the model write <think>
      geminiStops: ['\nUser\n']    # default: keep the model from speaking as the user
      geminiEcho: true               # echo the primed opener into the answer (default)
      geminiMaxTokens: 2048          # default output cap
  

• The shipped prefix is a stable ~4K-token prefix, so the provider's prompt cache should absorb most of its cost from the second request on.

• Echo: the primed opener is re-emitted by the adapter as the answer's first text delta, so the saved assistant message starts with <think>\n**Analyzing… even though that text was primed into the prompt. The echo is derived from the postfix (the Gemini\n role tag is stripped), so it follows whatever opener you configure; geminiEcho: false turns it off and leaves the prefill out of the stored message.

• Routing is decided on the Host from the session's own preset: a completion-gemini provider route is pinned per request, so the mode needs no browser half and uses the shipped Chat UI (no extra tab, no custom editor).

• The composer's model selector may still show the deployment default: the pin is applied per request rather than written into the session's model selection, because that write would persist the route as the whole deployment's default model.

Preset id completion-gemini (recorded per session, therefore stable); display name Gemini 模式.

How it works

Run ──▶ ctx.remote.session.prompt(document)
     ──▶ Agent loop assembles a real turn
     ──▶ provider route `completion-beta` (host half): prompt = the last human user message
     ──▶ POST {baseURL}/completions   stream: true
     ──▶ assistant text streams back; the editor appends it to the document

• Host half (index.js): ctx.llm.registerAdapter(['completion-beta'], adapter); a /completion-plan-b/max-tokens route receives each session's max_tokens + stop + logprobs; an agent/request waterfall listener (prepended) pins the route and the request controls for claimed sessions, so the deployment's default model is never rewritten.
• Browser half (client.js): a watcher in conversation.input.dock notices the session's agentPreset projection, claims it, and — once the session has history — registers this mode's two conversation.view entries plus a conversation.composer chain takeover that vacates the native composer.
• How the Logprob tab gets its distribution: logprobs is not a field of LlmCallConfig, so it travels on the same claim channel as max_tokens and the adapter reads it back through the request's own sessionId. The per-position distribution leaves on the finish chunk's replayState ({response}, with blocks omitted), which the harness stores on the assistant message's source. The run is therefore an ordinary turn — in Trajectory, cancellable, metered — and the browser half only reads the log back. At most 512 positions are stored; beyond that only the count is kept.
• Why a provider route instead of a private HTTP call: the harness drives the request, so nothing about Trajectory, paging, accounting or cancellation has to be reimplemented. The trade-off is that Trajectory stores the document snapshot per Run rather than a diff.

Compatibility

Developed and tested against DSH 0.1.7-rc.2, and re-verified against 0.2.0-rc.2 (engines.dsh). It binds to these extension points, which are public but may move between harness versions:

conversation.view / conversation.composer / conversation.input.dock slots · agentPreset session projection · agentPreset declarations · agent/request waterfall and LlmCallConfig · ctx.llm.registerAdapter · ctx.webServer.register · ctx.credentials · the session.prompt / session.cancel remotes · session.blank.

A session's mode is locked once its first turn commits (a harness invariant — even a failed turn opens one), so the Completion tab's session keeps that mode; start a new session to pick another mode.

Development

npm test        # activation smoke tests for both halves + stop-escape decoding

Both halves are dependency-free by design: the Host half uses only services reachable from ctx, and the browser half never imports a harness client package.

Installing without pnpm (development)

If the deployment has no pnpm on PATH, mount the plugin by path in the profile's own patch layer instead — dsh-client-modules resolves the nearest package.json from the row's module location, so both halves load with no package install:

# <profile>/cordis.patch.yml
- insert:
    - id: completion-plan-b-host
      name: /absolute/path/to/dsh-completion-playground/index.js
    - id: preset-completion-plan-b
      name: '@deepseek-ai/dsh-agent-preset'
      config:
        id: completion-plan-b
        name: Playground Mode
        description: Fullscreen raw-prompt editor plus a Logprob probe; every Run is a real turn (/beta/completions).
        order: 30
        plugins: []

Host-side changes need a process restart (the loader does not re-import a changed module, and the profile's HMR watches patch files, not module code). The browser bundle is stat-polled and republished by content hash, so an open page picks a client.js edit up by itself. The diagnostic GET /completion-plan-b/rev reports which revision the running Host loaded.

The connection trust gate is looked up per request rather than captured in apply(): a cold boot can mount this row before the connection service exists, and a captured undefined would leave the private routes answering unauthenticated callers for the rest of the process.

Publishing to npm (optional)

The manifest is npm-ready (publishConfig.access: public, files whitelist, no build step):

npm publish --access public

repository / homepage / bugs already point at this repo. Tag a release in git so the GitHub install spec above can pin it (v0.3.0 matches this version).

License

MIT — see LICENSE.