opensource/dify
2
0
Fork 0
mirror of https://github.com/langgenius/dify.git synced 2026-05-09 12:59:18 +08:00
Code Issues Packages Projects Releases Wiki Activity

remove the agent notes

Browse Source
This commit is contained in:
Yanli 盐粒 2026-01-27 01:09:33 +08:00
parent 2b787bff72
commit 688b2cfc48
4 changed files with 0 additions and 57 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

0
agent-notes/.gitkeep
View File

25
api/agent-notes/core/model_runtime/entities/message_entities.py.md
View File

@ -1,25 +0,0 @@
## Purpose
`core/model_runtime/entities/message_entities.py` defines the canonical in-memory Pydantic entities for model runtime
prompt messages and multi-modal message content. These entities are used across providers (built-in and plugin-backed)
and are serialized/deserialized when exchanging prompt/response payloads between layers.
## Key invariants
- `PromptMessage.content` is either a `str`, a list of typed content items (discriminated by `type`), or `None`.
- `PromptMessage.validate_content` normalizes dict/content-model inputs into the correct concrete content classes using
`CONTENT_TYPE_MAPPING`.
- `PromptMessage.serialize_content` ensures a list of content items is emitted as a list of plain dicts.
- `AssistantPromptMessage.tool_calls` may coexist with text/multi-modal content and is considered part of "non-empty".
## Opaque pass-through fields
- `opaque_body` is an optional JSON value on `PromptMessageContent` and `AssistantPromptMessage`.
- It is treated as an uninterpreted provider-specific payload and must be passed through unchanged between Dify and
plugin LLM providers (no validation/transformation beyond JSON serialization).
## Safety / compatibility notes
- Do not make `opaque_body` required; existing providers/plugins may not send it.
- Keep `type` discrimination stable; content subclasses must continue to be selectable via `Field(discriminator="type")`.

20
api/agent-notes/core/model_runtime/model_providers/__base/large_language_model.py.md
View File

@ -1,20 +0,0 @@
## Purpose
`core/model_runtime/model_providers/__base/large_language_model.py` defines the base `LargeLanguageModel` interface used
by model providers, including plugin-backed providers via `PluginModelClient`.
## Plugin invocation flow
- For plugin-based providers, `invoke()` delegates to `PluginModelClient.invoke_llm(...)`, which streams
`LLMResultChunk` objects from the plugin daemon.
- Dify yields chunks to callers and also aggregates chunks to fire `after_invoke` callbacks (and to construct a
blocking `LLMResult` when `stream=False`).
## Key invariants / edge cases
- When aggregating chunks into an `LLMResult`, preserve provider-specific fields on the assistant message:
- `AssistantPromptMessage.opaque_body` (pass-through, uninterpreted JSON).
- Incremental `tool_calls` (merge deltas via `_increase_tool_call`).
- Chunk `.prompt_messages` may be empty for plugin responses (compat layer for the plugin daemon); Dify re-attaches the
original request `prompt_messages` for downstream consumers.

12
api/agent-notes/tests/unit_tests/core/model_runtime/__base/test_llm_invoke_opaque_body.py.md
View File

@ -1,12 +0,0 @@
## Purpose
Unit tests for plugin-backed `LargeLanguageModel.invoke()` behavior around preserving provider pass-through data.
## What it covers
- `AssistantPromptMessage.opaque_body` from plugin `LLMResultChunk` deltas is preserved:
- On the returned `LLMResult` in blocking (`stream=False`) mode.
- On the aggregated `LLMResult` passed to `on_after_invoke` callbacks in streaming mode.
- Streaming mode also verifies that `chunk.prompt_messages` is re-attached to the original request prompt messages.
- Streaming aggregation merges incremental `tool_calls` across chunks.