DSH SuperAdvisor 0.1.7

DSH SuperAdvisor 0.1.7

An on-demand second opinion for your main model, approved by you. This release renames DSH Advisor to DSH SuperAdvisor, with repository and package name dsh-super-advisor.

The main model still calls ask_advisor; you review, edit, approve, or reject each request. The advisor returns text advice for the main model to verify and use. Model services and credentials come from DSH Models. Output defaults continue to follow the model service, with no additional received-text size cap; service limits and the 120-second request timeout still apply.

Typical setups

• Local main model + cloud advisor: use a tool-capable local model for routine work and approve focused consultations with a stronger cloud model when needed.
• Flash main model + Pro advisor: use Flash for execution and ask Pro for difficult questions or focused reviews. The aim is to approach stronger-model quality with lower usage and cost; no comparative quality or cost benchmark has been performed.

Install

Requires Node 24+, pnpm, and DSH. Verified on Linux with a local POSIX filesystem and DSH 0.1.3-alpha.2.

curl -fL https://github.com/kvmem/dsh-super-advisor/releases/latest/download/dsh-super-advisor.tgz -o dsh-super-advisor.tgz
dsh plugin --profile web add ./dsh-super-advisor.tgz
dsh web

In Settings → Models, configure the service, saved Base URL, credentials, and model. In Settings → Plugins → DSH SuperAdvisor, choose the advisor and save. Settings and result cards now show the SuperAdvisor name. The main model remains independently selectable; saving settings does not run inference.

Upgrade from DSH Advisor

Finish or stop active tasks and exit DSH first. After downloading the new package, replace the old package in the same profile:

dsh plugin --profile web remove dsh-tool-advisor
dsh plugin --profile web add ./dsh-super-advisor.tgz
dsh web

Keep the same DSH_HOME, settings, audit directory, and any custom storageDir. The advisor settings namespace and patch row, DSH_ADVISOR_* initial-value variables, and ask_advisor tool are retained for compatibility. Keep one version installed per profile. See the complete upgrade instructions.

The versioned dsh-super-advisor-0.1.7.tgz and fixed dsh-super-advisor.tgz assets are identical. The legacy dsh-tool-advisor.tgz asset preserves the original 0.1.6 bytes so old download commands do not silently change package identity. Use the new filename for 0.1.7. SHA256SUMS covers all three assets. Existing 0.1.6 release assets and Git history remain intact.

Validation

51 automated tests, source/test type checks, client syntax checks, and real DSH Loader checks pass. Isolated installation and Chromium checks cover the renamed package, settings card, and upgrade from 0.1.6 with retained settings and audit state. Testing uses local fixtures and controlled adapters; no new cloud-model calls were made. Historical cloud acceptance is recorded separately in ACCEPTANCE.md.

This release is distributed through GitHub; it has not been published to npm.

English README · 中文说明 · Community discussion

DSH SuperAdvisor 0.1.6 (legacy DSH Advisor)

DSH SuperAdvisor 0.1.6

Historical release, originally published as DSH Advisor (dsh-tool-advisor). The original 0.1.6 assets are preserved. For the renamed package and upgrade instructions, see DSH SuperAdvisor 0.1.7.

Unofficial project, independently developed and maintained by community members.

DSH SuperAdvisor lets the main model ask another model for focused advice. You review the complete request, edit or remove evidence, and approve each send. The advisor returns text; the main model continues to verify and execute the task.

Typical setups

• Local main model + cloud advisor: a local model connected to DSH and capable of calling tools handles routine work, then asks a stronger cloud model for focused advice after user approval.
• Flash main model + Pro advisor: deepseek-v4-flash handles the main task, while deepseek-v4-pro provides advice on difficult questions or focused reviews. Flash verifies the advice and continues execution.

The aim is to approach the quality of using a stronger model throughout with fewer stronger-model calls and lower cost. Quality and savings depend on the task, models, context size, and consultation frequency; comparative quality and cost have not been benchmarked.

Included in this release

• Configure the advisor in Settings → Plugins → 顾问模型, using services already configured in DSH Models.
• Inherit the model service's token budget by default, with no additional plugin response-size cap. Existing custom caps can be disabled in Output settings.
• Review the exact destination and request before each send. Editing the request or changing configuration requires a fresh approval.
• Read a short result preview, expand the advice, and inspect the original text and request ID.
• Persist send markers and results to prevent automatic duplicate sends after interruption or restart.

Install

Use Node 24+, pnpm, and DSH 0.1.3-alpha.2. This is the tested DSH version; compatibility with other versions has not been verified.

Download dsh-tool-advisor.tgz from this release, then run from the download directory:

dsh plugin --profile web add ./dsh-tool-advisor.tgz
dsh web

The .tgz contains the built plugin. The versioned dsh-tool-advisor-0.1.6.tgz attachment has identical contents. The README also provides a fixed latest-release download URL for installations without version numbers. GitHub's automatically generated source archives require a build before installation. If DSH is not installed yet:

npm install -g pnpm @deepseek-ai/dsh@0.1.3-alpha.2

In Settings → Models, configure the model service, including its saved Base URL, credentials, and model. In Settings → Plugins → 顾问模型, select the service and advisor model and save. The main task's model stays independently selectable. Plugin settings currently use Chinese labels.

Verification and limits

51 automated tests pass, together with type checks, build checks, and actual DSH Loader tests for the built and unpacked module. Chromium checks cover settings persistence, restart recovery, custom limits, and concurrent editing. This release's tests use controlled adapters and local fixtures; earlier Flash → Pro → Flash cloud acceptance is documented separately in ACCEPTANCE.md.

Verified on Linux with a local POSIX filesystem. Model capabilities, service budgets, and the default 120-second request timeout still apply. Each advisor call uses your configured model service and may incur its normal charges. No npm package has been published as part of this GitHub release.

English README · 中文说明 · Report an issue