- CHANGELOG.md:v2.8.0 条目(中英双语); - README.md:版本行 2.7.1→2.8.0 + 愿景第 6 条(知识可证明); - doc/releases/v2.8.0.md:双语发版说明(亮点/用法/已知限制); - skills/marqdo-dev/index.mq.md:新增 prove 任务(路径/目录模式/ require_lean/json;参数用 + 行声明升参,check 与 narrative-ratio 通过); - 设计文档 §12.5 标记已合入;发版链接统一指向 doc/releases/v2.8.0.md。
2.9 KiB
Executable File
markup, title, import
| markup | title | import | ||||
|---|---|---|---|---|---|---|
| occlusion | Provider adapters — requests, configuration and usage |
|
This directory owns provider adaptation. OpenAI-compatible requests encode chat messages, decode replies and usage, and retain streaming/error conventions. Ollama selects its documented local defaults while reusing compatible transport. The parent model API creates author model resources; the AI root explains agent and trial responsibilities.
This offline usage example inspects response data only. It does not read credentials, load a plugin or contact the provider:
marqdo call ext/ai/llm/index.mq.md usage 'data={"usage":{"prompt_tokens":2,"completion_tokens":3,"total_tokens":5}}' --json
For local configuration, resolve_local reads the documented environment fallbacks and therefore needs environment.read. Sending chat_completions is a real HTTP effect and requires an explicitly configured endpoint, timeout and credential handling. Do not print a live configuration record containing secrets merely to prove that an adapter exists.
When adding a provider, preserve semantic message/event/result shapes at the public model boundary. Keep provider-specific headers and endpoint paths in the adapter, document supported response variants and test them against local fixtures before live traffic. A usage fixture does not verify provider availability, streaming latency or authentication.
usage
| 参数 | 类型 | 默认 | 说明 |
|---|---|---|---|
| data | map | required | Already received OpenAI-compatible response record |
| 返回 | 类型 | 说明 |
|---|---|---|
| result | any | Public prompt_tokens/completion_tokens/total_tokens usage projection |
The data record is an observation rather than a request body. Missing counters follow the
public adapter convention; the example above contains two input tokens, three output tokens and
five total tokens. We return Read actual usage counters from the supplied response
without claiming to have measured a new model invocation.
resolve_local
| 参数 | 类型 | 默认 | 说明 |
|---|---|---|---|
| model | any | None | Explicit local model override; None uses the adapter fallback |
| base_url | any | None | Explicit local API base URL; None uses the documented environment/default |
| 返回 | 类型 | 说明 |
|---|---|---|
| result | any | Actual Ollama configuration selected by the existing resolver |
The caller may supply model=None and base_url=None to use the shared local configuration
rules. Resolving the configuration does not launch a service or ask the model for a response.
We return [Resolve local model fields with the existing adapter](ollama.resolve(model=model, base_url=base_url)).
Treat the resulting configuration as runtime data and only execute a request after granting its effects.