Persist the agent's `AgentContext` so that it works in rehydrated agent instances. This makes context usable in the `AgentProtocolServer`, where the agent instance is loaded and destroyed for every step.
- Make `AgentContext` a Pydantic model
- Add `context` parameter to `ContextComponent.__init__` so we can pass in an existing instance
- Add `context: AgentContext` to `AgentSettings` so it is persisted
- Add `type` attribute to `ContextItem` implementations as a discriminator
- Rename `ContextItem` base class to `BaseContextItem` and make new `ContextItem` type alias (union of the implementation types)
Documentation files were in docs/content/AutoGPT/components, symlinks in autogpts/autogpt/autogpt/(agents|commands).
Chef doesn't allow symlinks that point to locations outside of package dir.
Replacing the documentation files with symlinks, and symlinks with the actual documentation files, should fix this.
- feat(agent/core): Add `AnthropicProvider`
- Add `ANTHROPIC_API_KEY` to .env.template and docs
Notable differences in logic compared to `OpenAIProvider`:
- Merges subsequent user messages in `AnthropicProvider._get_chat_completion_args`
- Merges and extracts all system messages into `system` parameter in `AnthropicProvider._get_chat_completion_args`
- Supports prefill; merges prefill content (if any) into generated response
- Prompt changes to improve compatibility with `AnthropicProvider`
Anthropic has a slightly different API compared to OpenAI, and has much stricter input validation. E.g. Anthropic only supports a single `system` prompt, where OpenAI allows multiple `system` messages. Anthropic also forbids sequences of multiple `user` or `assistant` messages and requires that messages alternate between roles.
- Move response format instruction from separate message into main system prompt
- Fix clock message format
- Add pre-fill to `OneShot` generated prompt
- refactor(agent/core): Tweak `model_providers.schema`
- Simplify `ModelProviderUsage`
- Remove attribute `total_tokens` as it is always equal to `prompt_tokens + completion_tokens`
- Modify signature of `update_usage(..)`; no longer requires a full `ModelResponse` object as input
- Improve `ModelProviderBudget`
- Change type of attribute `usage` to `defaultdict[str, ModelProviderUsage]` -> allow per-model usage tracking
- Modify signature of `update_usage_and_cost(..)`; no longer requires a full `ModelResponse` object as input
- Allow `ModelProviderBudget` zero-argument instantiation
- Fix type of `AssistantChatMessage.role` to match `ChatMessage.role` (str -> `ChatMessage.Role`)
- Add shared attributes and constructor to `ModelProvider` base class
- Add `max_output_tokens` parameter to `create_chat_completion` interface
- Add pre-filling as a global feature
- Add `prefill_response` field to `ChatPrompt` model
- Add `prefill_response` parameter to `create_chat_completion` interface
- Add `ChatModelProvider.get_available_models()` and remove `ApiManager`
- Remove unused `OpenAIChatParser` typedef in openai.py
- Remove redundant `budget` attribute definition on `OpenAISettings`
- Remove unnecessary `usage` in `OpenAIProvider` > `default_settings` > `budget`
- feat(agent): Allow use of any available LLM provider through `MultiProvider`
- Add `MultiProvider` (`model_providers.multi`)
- Replace all references to / uses of `OpenAIProvider` with `MultiProvider`
- Change type of `Config.smart_llm` and `Config.fast_llm` from `str` to `ModelName`
- feat(agent/core): Validate function call arguments in `create_chat_completion`
- Add `validate_call` method to `CompletionModelFunction` in `model_providers.schema`
- Add `validate_tool_calls` utility function in `model_providers.utils`
- Add tool call validation step to `create_chat_completion` in `OpenAIProvider` and `AnthropicProvider`
- Remove (now redundant) command argument validation logic in agent.py and models/command.py
- refactor(agent): Rename `get_openai_command_specs` to `function_specs_from_commands`
* Introduce `BaseAgentActionProposal`, `OneShotAgentActionProposal`, and `AssistantThoughts` models to replace `ThoughtProcessResponse`, `DEFAULT_RESPONSE_SCHEMA`
* Refactor and clean up code because now we don't need to do as much type checking everywhere
* Tweak `OneShot` response format instruction
Granular:
* `autogpt.agents.prompt_strategies.one_shot`
* Replace ThoughtProcessResponse`, `DEFAULT_RESPONSE_SCHEMA` and parsing logic by `AssistantThoughts` and `OneShotAgentActionProposal`
* (TANGENTIAL) Move response format instruction into main system prompt message
* (TANGENTIAL) Adjust response format instruction
* `autogpt.agents.base`
* Add `BaseAgentActionProposal` base model -> replace `ThoughtProcessOutput`
* Change signature of `execute` method to accept `BaseAgentActionProposal` instead of separate `command_name` and `command_args`
* Add `do_not_execute(proposal, feedback)` abstract method, replacing `execute("human_feedback", ..., feedback)`
* Move `history` definition from `BaseAgentSettings` to `AgentSettings` (the only place where it's used anyway)
* `autogpt.models`
* Add `.utils` > `ModelWithSummary` base model
* Make the models in `.action_history` (`Episode`, `EpisodicActionHistory`) generic with a type parameter for the type of `Episode.action`
* `autogpt.core.resource.model_providers.schema`
* Add `__str__` to `AssistantFunctionCall` which pretty-prints the function call
All other changes are a direct result of the changes above.
## BREAKING CHANGE:
* Due to the change in `autogpt.models.action_history`, the application after this change will be unable to load/resume agents from before this change and vice versa.
* The `additional_output` field in the response of `execute_step` has changed slightly:
* Type of `.thoughts.plan` has changed from `str` to `list[str]`
* `.command` -> `.use_tool`
* `.command.args` -> `.use_tool.arguments`
* Fix ImportError for validate_yaml_file function
* Fix ImportError for validate_yaml_file function
---------
Co-authored-by: suyun <sylearn@foxmail.com>
Co-authored-by: Nicholas Tindle <nicholas.tindle@agpt.co>
### Background
Follow up after merging https://github.com/Significant-Gravitas/AutoGPT/pull/7054, old plugins will no longer be used.
### Changes 🏗️
- Removed all dead code needed to load and use plugins.
- Removed `auto-gpt-plugin-template` dependency
- Removed `rev=` from `autogpt-forge` dependency (the set `rev` had incompatible `duckduckgo-search` versions)
- Kept `--install-plugin-deps` CLI option and dead code associated (may be needed for new plugins)
* feat: add all the new component docs to the site
* fix(docs): relative links and markdown warnings
* feat(docs): How to contribute to the docs as a docs section
* fix(docs): missed docs page for developer setup
* fix(docs): re-add configurations options
* fix(docs): bad link to components fixed
* fix(docs): bad link to components fixed
* ref(docs): reorder some items to make more sense
* fix(docs): bad indentation and duplicate block
* fix(docs): warning about out of date markdown extension
* fix(docs): broken links fixed
* fix(docs): markdown formatter complaints
This incremental re-architecture unifies Agent code and plugins, so everything is component-based.
## Breaking changes
- Removed command categories and `DISABLED_COMMAND_CATEGORIES` environment variable. Use `DISABLED_COMMANDS` environment variable to disable individual commands.
- Changed `command` decorator; old-style commands are no longer supported. Implement `CommandProvider` on components instead.
- Removed `CommandRegistry`, now all commands are provided by components implementing `CommandProvider`.
- Removed `prompt_config` from `AgentSettings`.
- Removed plugin support: old plugins will no longer be loaded and executed.
- Removed `PromptScratchpad`, it was used by plugins and is no longer needed.
- Changed `ThoughtProcessOutput` from tuple to pydantic `BaseModel`.
## Other changes
- Created `AgentComponent`, protocols and logic to execute them.
- `BaseAgent` and `Agent` is now composed of components.
- Moved some logic from `BaseAgent` to `Agent`.
- Moved agent features and commands to components.
- Removed check if the same operation is about to be executed twice in a row.
- Removed file logging from `FileManagerComponent` (formerly `AgentFileManagerMixin`)
- Updated tests
- Added docs
See [Introduction](https://github.com/kcze/AutoGPT/blob/kpczerwinski/open-440-modular-agents/docs/content/AutoGPT/component%20agent/introduction.md) for more information.
- Set default values for `ProviderBudget` / `ModelProviderBudget` fields
- Remove redundant field redefinitions on `ModelProviderBudget` class
- Change `ModelProviderUsage.update_usage(..)` and `ModelProviderBudget.update_usage_and_cost(..)` signatures for easier use
- Change `ModelProviderBudget.usage` from `ModelProviderUsage` to `defaultdict[str, ModelProviderUsage]` for per-model usage tracking
- Fix `ChatModelInfo`/`EmbeddingModelInfo` `service` attribute: rename from `llm_service` to match base class and fix types.
This makes it unnecessary to specify the `service` field when creating a `ChatModelInfo` or `EmbeddingModelInfo` object.
- Use `defaultdict(ModelProviderBudget)` for task budget tracking in agent_protocol_server.py
* refactor(agent/core): Rearrange and split up `OpenAIProvider.create_chat_completion`
- Rearrange to reduce complexity, improve separation/abstraction of concerns, and allow multiple points of failure during parsing
- Move conversion from `ChatMessage` to `openai.types.ChatCompletionMessageParam` to `_get_chat_completion_args`
- Move token usage and cost tracking boilerplate code to `_create_chat_completion`
- Move tool call conversion/parsing to `_parse_assistant_tool_calls` (new)
* fix(agent/core): Handle decoding of function call arguments in `create_chat_completion`
- Amend `model_providers.schema`: change type of `arguments` from `str` to `dict[str, Any]` on `AssistantFunctionCall` and `AssistantFunctionCallDict`
- Implement robust and transparent parsing in `OpenAIProvider._parse_assistant_tool_calls`
- Remove now unnecessary `json_loads` calls throughout codebase
* feat(agent/utils): Improve conditions and errors in `json_loads`
- Include all decoding errors when raising a ValueError on decode failure
- Use errors returned by `return_errors` instead of an error buffer
- Fix check for decode failure
SwiftyOS
Reinier van der Leer
Nicholas Tindle
王苏云
Krzysztof Czerwinski
Fullstop000