Senparc.Xncf.PromptRange.Abstractions 0.3.2-preview6

Senparc.Xncf.PromptRange.Abstractions

Senparc.Xncf.PromptRange.Abstractions contains the event contracts shared by PromptRange and optional prompt-optimization or evaluation modules.

Features

• Prompt initialization request/response events.
• Prompt optimization request/response events with model-parameter context.
• Prompt-test completion events carrying output, score, and evaluation metadata.
• Immutable records that can cross NCF EventBus boundaries without a PromptRange implementation reference.

Installation

<PackageReference Include="Senparc.Xncf.PromptRange.Abstractions" Version="0.2.2-preview2" />

Key API

• PromptInitRequestEvent and PromptInitResponseEvent correlate prompt initialization.
• PromptOptimizationRequestEvent and PromptOptimizationResponseEvent carry optimization input, output, score, and error state.
• OptimizationContext and OptimizedParameters describe model-generation parameters.
• PromptTestFinishedEvent records a completed prompt test and evaluation result.
• Skills/SkillContracts.cs defines the portable Skill catalog, execution, and package contracts.
• Skills/SkillExceptions.cs defines explicit not-found and access-denied failures.

Skill portability

PromptRange exposes its versioned prompt tree as Skills through ISkillCatalog and ISkillExecutor. A Skill uses a stable SkillCode and a concrete PromptCode version (RangeName-Tactic-Aiming). Consumers such as AdminChat and AgentsManager should reference this abstractions package rather than PromptRange database entities.

Portable Skill packages are ZIP streams with a skill.json manifest and the Semantic Kernel-compatible config.json/skprompt.txt files. Secrets, provider credentials, and executable binaries must not be included in a package.

The PromptRange module registers the ISkillCatalog, ISkillExecutor, and ISkillPackageService implementations. Consumers may resolve these interfaces from dependency injection without referencing PromptRange entities.

External installation accepts a local ZIP upload or an HTTPS package URL. The package is validated for its manifest, size, archive paths, and executable content, then installed as Draft. An administrator must explicitly publish the imported PromptCode through the Skill management API before AdminChat or AgentsManager can execute it.

AdminChat exposes three Skill operations: list available Skills, load a selected Skill for the current turn, and execute it. After an administrator explicitly requests installation, AdminChat may call the HTTPS installation function; it only creates a Draft and never publishes or executes the newly installed Skill automatically.

Event payloads may contain prompts and model output. Apply authorization, tenant ownership, redaction, and retention controls before publishing or persisting them.

PromptCode reference integration

References/PromptReference.cs supplies optional, implementation-independent integration:

• PromptReference: PromptItemId, exact full PromptCode, UTF-16 Start and Length. Store metadata separately from the original content; never send labels or scores to the LLM. ValidateRanges checks ordered, nonoverlapping spans, with a maximum of 32 references.
• IPromptReferenceService.ValidateAsync additionally verifies the accessible version and exact text.
• GetInfoAsync, ScoreAsync and ScoreWithAIAsync expose version-level information and scoring. Only AI and manual source scores exist; whichever was saved last becomes CurrentScore. Scores are 0–10, at most one decimal; -1 in info means unrated.
• AI scoring requires a benchmark and a nonempty result; context/result each allow 64,000 characters. The caller must authorize the stored reference and the corresponding result, not accept arbitrary message IDs or another user's conversation. A missing provider must reject annotated submissions explicitly while leaving plain-text behavior unchanged.

The Admin shared components provide draft tracking, history labels, manual scoring and quick view. Hosts can supply their own authorized AI callback without a dependency on AdminChat entities. See the Admin integration guide.