@deepseek-ai/dsh-research-profiles-default

@deepseek-ai/dsh-research-profiles-default

English | 中文

The default model profiles for the DSH research service, registered on ctx.research under the exact ids the risk router selects.

Profiles

id	provider/model	cost tier	data classes
screening-cheap	deepseek / deepseek-chat	cheap	public
extraction-standard	deepseek / deepseek-chat	standard	public, unpublished
verification-independent	deepseek / deepseek-reasoner	standard	public, unpublished
methods-strong	deepseek / deepseek-reasoner (effort high)	strong	public, unpublished, sensitive
adjudication-strong-local	deepseek / deepseek-reasoner	strong	public, unpublished, sensitive

The profiles are frozen constant data, so every route reads the same model routes — deterministic and replayable from the session log. DEFAULT_PROFILE_BY_TASK mirrors the risk router's task-kind map for deployments that want to read the intended profile without routing.

Integration

The plugin registers a ResearchProfileProvider on ctx.research (name defaults by default) and unregisters it on unload for HMR safety. A cross-package test in dsh-research-router-risk pins this package's profile ids against the router's selection list so a rename cannot silently degrade high-impact routes.

Model Experience

Model profile selection

What the model sees

The package registers no prompt, tool, or schema of its own. Its frozen profiles are the data the risk router selects: each routed child is started on the profile's provider/model, with the profile's maxTokens ceiling and (for methods-strong) a high reasoning effort — so a delegation to verification-independent runs the child on a different model than the lead's route, deterministically.

Token effect

The package itself adds no tokens; the selected profile changes which model and effort tier the child's requests use.

KV Cache effect

Delegated children start fresh conversations per assignment, so there is no package-level request prefix; the profile choice does not alter any prompt text.

Known Limitations and Deferred Work

• Profiles are frozen static data — there is no dynamic or learned profile selection; every route reads the same five constants, so provider/model availability changes require a code update.
• Profiles assume the DeepSeek model family — the five profiles name DeepSeek models; a deployment using other providers must register its own profile provider instead.

Source: packages/research/dsh-research-profiles-default/src/index.ts