Humanizer tools for DeepSeek Harness

A host-only plugin that gives an agent two tools backed by a dedicated Humanizer llama-server:

Tool	Purpose
dsh_humanize_text	Rewrite a prose draft and return preservation warnings.
dsh_humanizer_status	Check server readiness and per-slot context without inference.

The agent supplies a draft, reviews the returned text, and decides how to use it. This plugin does not intercept every answer, download weights, start a server, edit files, or send messages. It has no browser bundle or configuration card.

Installation

This initial implementation is prepared on the DSH v0.2.0 development line. A source branch or PR is not an npm release. Until a registry release exists, build a local tarball from that checkout:

pnpm install --frozen-lockfile
pnpm --filter @yadsh/dsh-plugin-log build
pnpm --filter @yadsh/dsh-humanizer build
pnpm --dir plugins/dsh-humanizer pack --pack-destination ./dist
dsh plugin --profile web add <absolute-path-to-package.tgz>

The profile must supply the Host's tools service. If your deployment restricts tools with an allow-list, grant the two names in the table explicitly.

Start the model service

Follow the upstream installation and runtime guide. Use a recent llama.cpp build compatible with the model. For example:

hf download jialinyyzz/humanizer humanizer-12b-Q8_0.gguf prompt_format.json --local-dir ./humanizer-model
llama-server -m ./humanizer-model/humanizer-12b-Q8_0.gguf -c 8192 -np 1 -ngl 99 --host 127.0.0.1 --port 8080

The Q8 file is about 12.7 GB; upstream recommends at least 32 GB of system memory for it and offers smaller GGUF files. The actual GPU/CPU performance and available memory must be checked on the deployment machine. The model runs offline after its weights have been downloaded.

Use a dedicated server loaded with Humanizer. Readiness does not verify weight identity or SHA. The plugin requires /health, /props, /tokenize and /completion; a generic chat API alone is insufficient. Use an authenticated HTTPS reverse proxy when the service is outside the Host machine.

Configuration

Configuration belongs to the dsh-humanizer composition row in the selected Host profile. For example, the row inserted by the plugin can carry:

- id: dsh-humanizer
  name: "@yadsh/dsh-humanizer"
  config:
    endpoint: http://127.0.0.1:8080
    timeoutMs: 120000
    maxChunkBytes: 6000
    maxOutputTokens: 2048

Field	Default	Meaning
enabled	true	When false, no tools are registered.
endpoint	http://127.0.0.1:8080	Bare HTTP(S) origin; no path, URL credentials, query or fragment.
apiKey	empty	Bearer key; empty falls back to DSH_HUMANIZER_API_KEY.
timeoutMs	120000	Overall call budget, including every group; clamped to 1000–900000 ms. Status is capped at 5000 ms.
maxInputBytes	65536	Total draft UTF-8 bytes; clamped to 1024–262144.
maxChunkBytes	6000	Maximum paragraph group bytes; clamped to 256–16384 and the input cap.
maxOutputTokens	2048	Generation token reservation per group; clamped to 128–4096.
maxOutputBytes	131072	Total returned text bytes; clamped to 1024–524288.

Use the environment variable for deployment secrets; never put a real key in a committed profile. Config changes take effect when the plugin is reapplied.

Tool contract

Example arguments to dsh_humanize_text:

{
  "text": "Demo Product shipped 12 units on 2026-10-12.",
  "protectedPhrases": ["Demo Product"]
}

protectedPhrases is optional. It accepts up to 32 exact strings, each present in the draft and at most 512 characters long. These strings are checked after generation; the model can still change them.

The tool returns:

• rewrittenText: generated text for review.
• chunkCount and totalChunks: groups processed and planned.
• truncated: true if generation stopped early, completion status was unconfirmed, or the returned text hit its byte cap.
• warnings: number/phrase mismatches, unexpected language, incomplete text, and a reminder to review meaning and facts.

Prose is grouped at blank-line boundaries without slicing a paragraph. Separators between groups survive; the model may change formatting within a group. Oversized drafts or individual paragraphs fail explicitly. Fenced code blocks are rejected. For Markdown tables, links, code, DOCX or PDF, extract the prose you intend to rewrite before calling this tool.

Before any inference, every prompt is tokenized by the loaded model. The reserved output plus prompt must fit the smaller of 8192 and the server's reported per-slot context. A smaller server context or a dense paragraph can therefore require a smaller group or output reservation. Context is never silently shifted to make a draft fit.

dsh_humanizer_status returns status (ready, loading or unavailable), contextTokens (0 when unavailable/loading), and a sanitized errorCode. Loading is a 503 health response; ready requires health and context checks.

Quality limits

Upstream documents and evaluates English and Chinese. Russian and other languages are experimental; this plugin reports that limitation for detected Cyrillic and text without Latin/Han characters. Mixed languages and other Latin-script languages are not reliably detected.

Number and exact phrase checks are heuristics. They do not establish factual correctness or preserve units, names, dates, quotations or meaning. A clean warning list cannot replace review, and a detector outcome is not guaranteed. An incomplete result must not be presented as a full document.

Security model

The endpoint is operator configuration, never an agent argument. Redirects are refused, including redirects from completion requests. The entire draft and returned text are sent to this service; choose its placement accordingly.

All HTTP responses are byte-bounded, including error responses. Transport failures expose stable codes such as UNAVAILABLE, TIMEOUT, HTTP_401, INVALID_RESPONSE or CONTEXT_TOO_LARGE, without echoing server bodies, URLs or secrets. Caller cancellation and plugin disposal abort active requests. Removing the plugin unregisters its tools.

Plugin logs contain only completion counts, sizes and warning counts, not drafts, rewrites, phrases or keys. The Host may still journal tool arguments and results under its normal session policy. Generated text is untrusted data and must not be interpreted as instructions to execute.

Compatibility

This implementation is developed and checked against DSH 0.2.0-rc.2 and Cordis 4.0.4, on the v0.2.0 branch. It requires tools/register and Node ^22.19.0 || >=24.0.0. There is no client module or Remote service.

Automated HTTP fixtures verify the integration protocol and lifecycle. They do not measure real-model quality or deployed Host behavior.

Development

pnpm --filter @yadsh/dsh-humanizer check
pnpm tarball:verify:packages plugins/dsh-humanizer

License

Plugin code is MIT. The pinned Humanizer prompt is adapted from Apache-2.0 upstream material; see THIRD_PARTY_NOTICES.md and LICENSE-APACHE-2.0. Weights are downloaded separately and are not included in this package.