DSH Second Opinion
简体中文
dsh-second-opinion adds an asynchronous second-model review to DeepSeek Harness. The primary model's answer is committed immediately; a reviewer model then checks the completed assistant message, records quote-anchored concerns, and optionally asks the primary model to reconsider.
Quote-anchored second opinion
Features
- Reviews each committed assistant message without delaying or failing the primary turn.
- Runs reviews serially within one Session and concurrently across Sessions.
- Supports advisory review and bounded automatic reconsideration.
- Anchors comments to exact quoted text and keeps general comments when no quote applies.
- Persists pending, completed, and failed reviews in a Storage Domain sidecar.
- Exposes retry, per-comment reconsideration, and whole-review reconsideration in the Web UI.
- Gives the reviewer no primary-agent tools, so a review cannot perform tool side effects.
Requirements
- DeepSeek Harness
0.1.2-rc.1 (the tested release target)
- Node.js
^22.19.0 || >=24.0.0
pnpm available to the dsh plugin command
Installation
Install the published npm package into the Web profile:
dsh plugin --profile web add dsh-second-opinion
Or install the current GitHub source:
dsh plugin --profile web add github:scwlkq/dsh-second-opinion
Restart the DSH Web process after installation. The bundle activates both the Host plugin and its browser contribution.
Usage
The default advisory mode needs no extra configuration. Start DSH, send a prompt, and open the speech-bubble action under a completed assistant message. The dialog first shows Review pending, then the reviewer verdict and comments. You can send one comment or the complete review back to the primary agent for reconsideration.
dsh --profile web
The reviewer route defaults to the provider and model that produced the target message. Set both reviewer route variables to use a dedicated B model:
export DSH_REVIEW_PROVIDER='<provider-id>'
export DSH_REVIEW_MODEL='<model-id>'
dsh --profile web
Modes
| Mode | Behavior |
|---|
advisory | Stores and displays the review. A person chooses whether to reconsider one comment or the complete review. This is the default. |
auto | When the verdict is concern or disagree, automatically steers the primary agent to reconsider. The original answer remains in the transcript. |
Enable automatic reconsideration for the bundle layer:
export DSH_REVIEW_MODE=auto
dsh --profile web
Automatic reconsideration is limited by autoRevisionsPerTurn and does not review the plugin's own advice chain, preventing an unbounded review loop.
If the Session advances before a slow review completes, the result remains available as advice but is not injected into the newer task.
Configuration
The bundle reads these environment variables at profile startup:
| Environment variable | Default | Meaning |
|---|
DSH_REVIEW_MODE | advisory | Selects advisory or auto. |
DSH_REVIEW_PROVIDER | unset | Reviewer provider ID. Must be set together with DSH_REVIEW_MODEL. |
DSH_REVIEW_MODEL | unset | Reviewer model ID. Must be set together with DSH_REVIEW_PROVIDER. |
For every option, add an ID-targeted override to the profile's cordis.patch.yml:
- id: second-opinion
config:
mode: advisory
provider: '<provider-id>'
model: '<model-id>'
reasoningEffort: high
maxTokens: 900
maxContextMessages: 20
autoRevisionsPerTurn: 1
reviewToolCalls: true
| Option | Default | Meaning |
|---|
mode | advisory | Review-only or automatic reconsideration behavior. |
provider | target message route | Dedicated reviewer provider ID. Configure with model. |
model | target message route | Dedicated reviewer model ID. Configure with provider. |
reasoningEffort | unset | Optional reasoning effort forwarded to the reviewer route. |
maxTokens | 900 | Maximum reviewer output tokens. |
maxContextMessages | 20 | Maximum recent transcript messages sent with the target answer. |
autoRevisionsPerTurn | 1 | Maximum automatic reconsiderations for one primary turn. |
reviewToolCalls | true | Whether to review assistant messages that contain no non-empty text, including tool-call-only messages. |
provider and model must both be present or both be omitted. Invalid self-contained configuration fails when the plugin loads.
Privacy, Cost, and Failure Behavior
- The reviewer provider receives the target answer and up to
maxContextMessages recent visible transcript messages. This content can include prompts, tool-call arguments, tool-result markers, and image markers; hidden reasoning is not included.
- Each reviewed assistant message adds one model request. In
auto mode, a non-agreeing review can add another primary-model request.
- A reviewer failure never blocks, delays, or replaces the primary answer. Invalid reviewer output becomes a failed review that the user can retry.
- The reviewer receives only the structured
submit_review response tool, not the primary agent's executable tools.
Current Limitations
- Exact quote anchors are displayed inside the review dialog; version
0.1.0 does not decorate or select the original transcript DOM.
- Reviews start after an assistant message is committed, not while tokens are streaming.
- Automatic reconsideration appends a new decision instead of rewriting the original answer.
- The first release targets the DSH Web profile for interactive review actions.
Uninstallation
dsh plugin --profile web remove dsh-second-opinion
Restart the DSH Web process after removal.
Development
pnpm install
pnpm test
pnpm run typecheck
pnpm run lint
pnpm run build
To verify a packed artifact against a local DeepSeek Harness checkout:
DSH_SOURCE_ROOT=/path/to/deepseek-harness pnpm run smoke:install
The detailed Chinese walkthrough is in docs/使用指南.md.
License
MIT
