Files
talemate/docs/user-guide/agents/director/settings.md
veguAI 89d16ae513 0.33.0 (#229)
* linting

* Add cleanup function for recent scenes in config to remove non-existent paths

* remove leghacy world state manager buttons

* move world state scene tools into sub component

* linting

* move module properties to navigation drawer

* Update icons in NodeEditorLibrary and NodeEditorModuleProperties for improved UI clarity

* prompt tweaks

* director chat prompt simplifications

* more prompt fixes

* Enhance type hints for duration conversion functions in time.py

* narrate time action now has access to response length instructions

* Add IsoDateDuration node for ISO 8601 interval string construction

* Update advance_time method to include return type annotation and return message

* Add AdvanceTime node to world state for time advancement with duration and narration instructions

* linting

* Add agent state exclusions to changelog with a TODO for module migration

* Add message emission for actor, narrator, and scene analysis guidance in respective components. Enhance AgentMessages and SceneTools for better message handling and visual feedback.

* Remove agent messages from state when opening agent message view in SceneTools component.

* linting

* openroute fetch models on key set

* Add input history functionality to message input in TalemateApp component. Implement keyboard shortcuts for navigating history (Ctrl+Up/Down) and limit history to the last 10 messages. Update message sending logic to store messages in history.

* Update message input hint in TalemateApp component to include keyboard shortcuts for navigating input history (Ctrl+Up/Down).

* node updates

* unified data extraction function

* prompt tweaks

* Add gamestate context support in BuildPrompt and corresponding template. Introduced new property for gamestate context and updated rendering logic to include gamestate information in prompts.

* Refactor Prompt class by removing LoopedPrompt and cleaning up related methods. Update data response parsing to streamline functionality and improve clarity. Adjust imports accordingly.

* Add 'data_multiple' property to GenerateResponse class to allow multiple data structures in responses. Update output socket type for 'data_obj' to support both dict and list formats.

* Add DictUpdate node

* Add UnpackGameState node to retrieve and unpack game state variables

* gamestate nodes

* linting

* Enhance scene view toggle functionality to support shift-click behavior for closing all drawers when hiding the scene view.

* immutable scenes should reset context db on load

* linting

* node updates

* prompt tweaks

* Add context type output and filtering for creative context ID meta entries in PathToContextID and ContextIDMetaEntries nodes

* Add string replacement functionality and Jinja2 formatting support in nodes. Introduced 'old' and 'new' properties for substring replacement in the Replace node, and added a new Jinja2Format node for template rendering using jinja2.

* Add additional outputs for context validation in ValidateContextIDItem node, including context type, context value, and name.

* prompt tweaks

* node adjustments

* linting

* Add data_expected attribute to Focal and Prompt classes for enhanced response handling

* node updates

* node updates

* node updates

* prompt tweaks

* director summary return appropriately on no action taken

* Enhance action handling in DirectorChatMixin by skipping actions when a question is present in the parsed response, ensuring better response accuracy.

* Enhance ConfirmActionPrompt component by adding anchorTop prop for dynamic alignment and adjusting icon size and color for improved UI consistency.

* anchor clear chat confirm to top

* responsive layout fixes in template editors

* linting

* relock

* Add scene progression guidance to chat-common-tasks template

* Refactor push_history method to be asynchronous across multiple agents and scenes, ensuring consistent handling of message history updates.

* Update chat instructions to clarify user intent considerations and enhance decisiveness in responses. Added guidance on distinguishing between scene progression and background changes, and refined analysis requirements for user interactions.

* Enhance DirectorConsoleChatsToolbar by adding a usage cheat sheet tooltip for user guidance and refining the Clear Chat button's UI for better accessibility.

* store character data at unified point

* fix button

* fix world editor auto sync

* Shared context 2 (#19)

Shared context

* Refactor NodeEditorLibrary to improve search functionality and debounce input handling. Updated v-text-field model and added a watcher for search input to enhance performance.

* Refactor NodeEditor and TalemateApp components to enhance UI interactions. Removed the exit creative mode button from NodeEditor and updated tooltips for clarity. Adjusted app bar navigation icons for better accessibility and added functionality to switch between node editor and creative mode.

* comment

* Character.update deserialize voice value correctly

* Enhance SharedContext.update_to_scene method to properly add or update character data in the scene based on existence checks. This improves the synchronization of character states between shared context and scene.

* shared context static history support
fix context memory db imports to always import

* Update WorldStateManagerSceneSharedContext.vue to clarify sharing of character, world entries, and history across connected scenes.

* linting

* Enhance chat modes by adding 'nospoilers' option to DirectorChat and related payloads. Update chat instructions to reflect new mode behavior and improve UI to support mode-specific icons and colors in the DirectorConsoleChatsToolbar.

* Comment out 'repetition_penalty_range' in TabbyAPIClient to prevent unexpected "<unk><unk> .." responses. Further investigation needed.

* linting

* Add active_characters and intro_instructions to Inheritance model; implement intro generation in load_scene_from_data. Update WorldStateManagerSceneSharedContext.vue to enhance new scene creation dialog with character selection and premise instructions.

* rename inheritance to scene initialization

* linting

* Update WorldStateManagerSceneSharedContext.vue to conditionally display alert based on scene saving status and new scene creation state.

* Refine messages for shared context checkboxes in WorldStateManagerCharacter and WorldStateManagerWorldEntries components for clarity.

* Add scene title generation to load process and update contextual generation template. Introduced a new method in AssistantMixin for generating scene titles, ensuring titles are concise and free of special characters. Updated load_scene_from_data to assign generated titles to scenes.

* linting

* Refactor GameState component to integrate Codemirror for JSON editing, replacing the previous treeview structure. Implement validation for JSON input and enhance error handling. Remove unused methods and streamline state management.

* Add lastLoadedJSON property to GameState component for change detection. Update validation logic to prevent unnecessary updates when game state has not changed.

* Remove status emission for gameplay switch in CmdSetEnvironmentToScene class.

* allow individual sharing of attributes and details

* linting

* Remove redundant question handling logic in DirectorChatMixin to streamline action selection process.

* Update EXTERNAL_DESCRIPTION in TabbyAPI client to include notes on EXL3 model sensitivity to inference parameters. Adjust handling of 'repetition_penalty_range' in parameter list for clarity.

* director chat support remove message and regenerate message

* Refactor ConfirmActionInline component to improve button rendering logic. Introduced 'size' prop for button customization and added 'comfortable' density option. Simplified icon handling with computed property for better clarity.

* linting

* node updates

* Add appBusy prop to DirectorConsoleChats and DirectorConsoleChatsToolbar components to manage button states during busy operations.

* Refactor DirectorChatMixin to utilize standalone utility functions for parsing response sections and extracting action blocks. This improves code clarity and maintainability. Added tests for new utility functions in test_utils_prompt.py to ensure correct functionality.

* Update clear chat button logic to consider appBusy state in DirectorConsoleChatsToolbar component, enhancing user experience during busy operations.

* linting

* Remove plan.md

* Add chat template identifier support and error handling in ModelPrompt class

- Implemented logic to check for 'chat_template.jinja2' in Hugging Face repository.
- Added new template identifiers: GraniteIdentifier and GLMIdentifier.
- Enhanced error handling to avoid logging 404 errors for missing templates.
- Introduced Granite.jinja2 template file for prompt structure.

* node fixes

* remove debug msg

* Enhance error handling in DynamicInstruction class by enforcing header requirement and ensuring content defaults to an empty string if not provided.

* recet scene message visibility on scene load

* prompt tweaks

* Enhance data extraction in Focal class by adding a fallback mechanism. Implemented additional error handling to attempt data extraction from a fenced block if the initial extraction fails, improving robustness in handling responses.

* linting

* node fixes

* Add relative_to_root function for path resolution and update node export logic

- Introduced a new function `relative_to_root` in path.py to resolve paths relative to the TALEMATE_ROOT.
- Updated the `export_node_definitions` function in registry.py to use `relative_to_root` for module path resolution.
- Added a check to skip non-selectable node definitions in litegraphUtils.js during registration.

* show icons

* Improve error handling in export_node_definitions by adding a try-except block for module path resolution. Log a warning if the relative path conversion fails.

* typo

* Refactor base_attributes type in Character model to a more generic dict type for improved flexibility

* relock

* ensure character gets added to character_data

* prompt tweaks

* linting

* properly activate characters

* activate needs to happen explicitly now and deactivated is the default

* missing arg

* avoid changed size error

* Refactor character removal logic in shared context to prevent deletion; characters are now only marked as non-shared.

* Add update_from_scene method calls in SharedContextMixin for scene synchronization

* Add ensure_changelogs_for_all_scenes function to manage changelog files for all scenes; integrate it into the server run process.

* Enhance backup restore functionality by adding base and latest snapshot options; improve UI with clearer labels and alerts for restore actions.

* Update _apply_delta function to enhance delta application handling by adding parameters for error logging and force application of changes on non-existent paths.

* Skip processing of changelog files in _list_files_and_directories function to prevent unnecessary inclusion in file listings.

* Update IntroRecentScenes.vue to use optional chaining for selectedScene properties and enhance backup timestamp display with revision info.

* linting

* Refactor source entry attribute access in collect_source_entries function to use getattr for optional attributes, improving robustness.

* Implement logic to always show scene view in scene mode within TalemateApp.vue, enhancing user experience during scene interactions.

* prompt tweaks

* prompt tweaks

* Update TalemateApp.vue to set the active tab to 'main' when switching to the node editor, improving navigation consistency.

* Add active frontend websocket handler management in websocket_endpoint

* agent websocket handler node support

* Refactor init_nodes method in DirectorAgent to call superclass method and rename chat initialization method in DirectorChatMixin for clarity.

* Add characters output to ContextHistory node to track active participants in the scene

* Add Agent Websocket Handler option to Node Editor Library with corresponding icons and labels

* Add check for node selectability in NodeEditorNodeSearch component to filter search results accordingly.

* Add SummarizeWebsocketHandler to handle summarize actions and integrate it into SummarizeAgent

* nodes

* Add data property to QueueResponse class for websocket communication and update run method to include action and data in output values.

* Update manual context handling in WorldStateManager to include shared property from existing context

* Enhance GetWorldEntry node to include 'shared' property in output values from world entry context

* Update scene loading to allow setting scene ID from data and include ID in scene serialization

* Update icon for AgentWebsocketHandler in NodeEditorLibrary component to mdi-web-box

* Refactor WorldStateManager components to enhance history management and sharing capabilities. Added summarized history titles, improved UI for sharing static history, and integrated scene summarization functionality. Removed deprecated methods related to shared context settings.

* linting

* Change log level from warning to debug for migrate_narrator_source_to_meta error handling in NarratorMessage class.

* Update GLM-no-reasoning template to include <think></think> tag before coercion message for improved prompt structure.

* allow prompt templates to specify reasoning pattern

* Add Seed.jinja2 template for LLM prompts with reasoning patterns and user interaction handling

* Enhance NarratorAgent to support dynamic response length configuration. Updated max generation length from 192 to 256 tokens and introduced a new method to calculate response length. Modified narration methods to accept and utilize response length parameter. Added response length property in GenerateNarrationBase class and updated templates to include response length handling.

* Update response length calculation in RevisionMixin to include token count for improved text processing.

* Refactor response identifier in RevisionMixin to dynamically use calculated response length for improved prompt handling.

* linting

* allow contextual generation of static history entries

* Add is_static property to HistoryEntry for static history entry identification

* Add "static history" option to ContextualGenerate node for enhanced contextual generation capabilities.

* Add CreateStaticArchiveEntry and RemoveStaticArchiveEntry nodes for managing static history entries. Implement input/output properties and error handling for entry creation and deletion.

* nodes updated

* linting

* Add assets field to SceneInitialization model and update load_scene_from_data function to handle scene assets. Update WorldStateManagerSceneSharedContext.vue to include assets in scene initialization parameters.

* Refactor CoverImage component to enhance drag-and-drop functionality and improve styling for empty portrait state.

* Add intent_state to SceneInitialization model and update load_scene_from_data function to handle intent state. Introduce story_intent property in Scene class and reset method in SceneIntent class. Update WorldStateManagerSceneSharedContext.vue to include intent state in scene initialization parameters.

* Refactor WorldStateManagerSceneSharedContext.vue to improve cancel functionality by introducing a dedicated cancelCreate method and removing the direct dialog toggle from the Cancel button. This enhances code clarity and maintainability.

* Update SharedContext to use await for set_shared method, ensuring proper asynchronous handling when modifying character sharing status.

* Add MAX_CONTENT_WIDTH constant and update components to use it for consistent max width styling

* fix issue with data structure parsing

* linting

* fix tests

* nodes

* fix update_introduction

* Add building blocks template for story configuration and scene management

* Refactor toggleNavigation method to accept an 'open' parameter for direct control over drawer visibility in TalemateApp.vue

* Update usageCheatSheet text in DirectorConsoleChatsToolbar.vue for clarity and add pre-wrap styling to tooltip

* Add cover image and writing style sections to story and character templates; update chat common tasks with new scene restrictions and user guide reference.

* linting

* relock

* Add EmitWorldEditorSync node to handle world editor synchronization; update WorldStateManager to refresh active tab on sync action.

* Update Anthropic client with new models and adjust default settings; introduce limited parameter models for specific configurations.

* director action  module updates

* direct context update fn

* director action updates

* Update usageCheatSheet in DirectorConsoleChatsToolbar.vue to include recommendation for 100B+ models.

* Remove debug diagnostics from DirectorConsoleChats.vue to clean up console output.

* Update card styles in IntroRecentScenes.vue for improved visual consistency; change card color to grey-darken-3 and adjust text classes for titles and subtitles.

* Update EmitWorldEditorSync node to include websocket passthrough in sync action for improved event handling.

* Increase maximum changelog file size limit from 500KB to 1MB to accommodate larger change logs.

* linting

* director action module updates

* 0.33 added

* Add Nexus agent persona to talemate template and initialize phrases array

* Add support for project-specific grouping in NodeEditorLibrary for templates/modules, enhancing organization of node groups.

* docs

* Enhance NodeEditorLibrary by adding primary color to tree component for improved visibility and user experience.

* docs

* Enhance NewSceneSetupModal to include subtitles for writing styles and director personas, improving context and usability.

* Update agent persona description in WorldStateManagerTemplates to specify current support for director only, enhancing clarity for users.

* Refine agent persona description in WorldStateManagerTemplates to clarify assignment per agent in Scene Settings, maintaining focus on current director-only support.

* fix crash when attempting to delete some clients

* Add TODO comments in finalize_llama3 and finalize_YI methods to indicate removable cruft

* Add lock_template feature to Client configuration and update related components for template management

* linting

* persist client template lock through model changes

* There is no longer a point to enforcing creative mode when there are no characters

* fix direct_narrator character argument

* Update CharacterContextItem to allow 'value' to accept dict type in addition to existing types

* docs

* Update lock_template field in Client model to allow None type in addition to bool

* Remove unused template_file field from Defaults model in Client configuration

* Refactor lock_template field in Client model and ClientModal component to ensure consistent boolean handling

* Add field validator for lock_template in Client model to ensure boolean value is returned

* fix issue where valid data processed in extract_data_with_ai_fallback was not returned

* Update default_player_character assignment in ConfigPlugin to use GamePlayerCharacter schema for improved data validation

* linting

* add heiku 4.5 model and make default

* opse 4.5 isnt a thing

* fix issue where  fork / restore would restore duplicate messages

* improve autocomplete handling when prefill isn't available

* prompt tweaks

* linting

* gracefully handle removed attributes

* Refactor scene reference handling in delete_changelog_files to prevent incorrect deletions. Added a test to verify proper scene reference construction and ensure changelog files are deleted correctly.

* forked scenes reset memory id and are not immutable

* emit_status export rev

* Update RequestInput.vue to handle extra_params more robustly, ensuring defaults are set correctly for input.

* only allow forking on saved messages

* linting

* tweak defaults

* summarizer fire off of push_history.after

* docs

* : in world entry titles will now load correctly

* linting

* docs

* removing base attrib ute or detail also clears it from shared list

* fix issue where cancelling some generations would cause errors

* increase font size

* formatting fixes

* unhandled errors at the loop level should not crash the entire scene

* separate message processing from main loop

* linting

* remove debug cruft

* enhance error logging in background processing to include traceback information

* linting

* nothing to detemrine of no model is sent

* fix some errors during kcpp client deletion

* improve configuration issue alert visibility

* restore input focus after autocomplete

* linting
2025-10-25 14:06:55 +03:00

8.3 KiB

Settings

General

Director agent settings

Direct

If enabled the director will attempt to direct the scene.

This is currently only relevant if the scene loaded comes with a script file. For the scenes that are currently included that is only true for the Simulation Suite.

Turns

How many turns to wait before the director makes a decision.

Direct Scene

If enabled the director will attempt to direct the scene through narration.

Direct Actors

Enables the director to direct the actors in the scene.

Right now this is only triggered manually by the player when the players uses the Direct actor toolset from the Scenario tools.

Actor Direction Mode

When an actor is given a direction, how is it to be injected into the context

  • Direction
  • Inner Monologue

If Direction is selected, the actor will be given the direction as a direct instruction, by the director.

If Inner Monologue is selected, the actor will be given the direction as a thought.

Long Term Memory

--8<-- "docs/snippets/tips.md:agent_long_term_memory_settings"

Dynamic Actions

Dynamic actions are introduced in 0.28.0 and allow the director to generate a set of clickable choices for the player to choose from.

Director agent dynamic actions settings

Enable Dynamic Actions

If enabled the director will generate a set of clickable choices for the player to choose from.

Chance

The chance that the director will generate a set of dynamic actions when its the players turn.

This ranges from 0 to 1. 0 means the director will never generate dynamic actions, 1 means the director will always generate dynamic actions.

Number of Actions

The number of actions to generate.

Never auto progress on action selection

If this is checked and you pick an action, the scene will NOT automatically pass the turn to the next actor.

Instructions

Allows you to provide extra specific instructions to director on how to generate the dynamic actions.

For example you could provide a list of actions to choose from, or a list of actions to avoid. Or specify that you always want a certain action to be included.

Guide Scene

Director agent guide scene settings

The director can use the summarizer agent's scene analysis to guide characters and the narrator for the next generation, hopefully improving the quality of the generated content.

!!! danger "This may break dumber models" The guidance generated is inserted after the message history and right before the next generation. Some older models may struggle with this and generate incoherent responses.

Guide Actors

If enabled the director will guide the actors in the scene.

Guide Narrator

If enabled the director will guide the narrator in the scene.

Max. Guidance Length

The maximum number of tokens for the guidance. (e.g., how long should the guidance be).

Auto Direction

A very experimental first attempt at giving the reigns to the director to direct the scene automatically.

Currently it can only instruct actors and the narrator, but different actions will be exposed in the future. This is very early in the development cycle and will likely go through substantial changes.

!!! note "Both overall and current intent need to be set for auto-direction to be available" If either the overall or current scene intention is not set, the auto-direction feature will not be available.

![Auto Direction Unavailable](/talemate/img/0.30.0/auto-direction-unavailable.png)

Story and scene intentions are set in the [Scene Direction](/talemate/user-guide/world-editor/scene/direction) section of the World Editor.

Director agent auto direction settings

Enable Auto Direction

Turn auto direction on and off.

!!! note "Auto progress needs to also be enabled" If auto direction is enabled, auto progress needs to be enabled as well.

![Auto Progress On](/talemate/img/0.30.0/auto-progress-on.png)

Natural flow

Will place strict limits on actor turns based on the provided constraints. That means regardless of what the director would like to do, the actor availability will always take precedence.

Max. Auto turns

Maximum turns the AI gets in succession (spread accross characters). When this limit is reached, the player will get a turn no matter what.

Max. Idle turns

The maximum number of turns a character can go without speaking before they are automatically given a turn by the director. (per character)

Max. Repeat Turns

The maximum number of times a character can go in succession without speaking before the director will force them to speak. (per character)

Instructions

Instruct Actors

Allow the director to instruct actors.

Instruct Narrator

Allow the director to instruct the narrator.

Instruct Frequency

Only pass on instructions to the actors or the narrator every N turns.

!!! note "Evaluation of the scene happens regardless" The director will evaluate the scene after each round regardless of the frequency. This setting merely controls how often the instructions are actually passed on.

Evaluate Scene Intention

Allows the director to evaluate the current scene phase and switch to a different scene type or set a new intention.

The number of turns between evaluations. (0 = NEVER)

!!! note "Recommended to leave at 0 (never)" This isn't really working well at this point, so recommended to leave at 0 (never)

Director Chat

!!! example "Experimental" Currently experimental and may change substantially in the future.

The Director Chat feature allows you to interact with the director through a conversational interface where you can ask questions, make changes to your scene, and direct story progression.

Director Chat Settings

Enable Analysis Step

When enabled, the director performs an internal analysis step before responding. This helps the director think through complex requests and plan actions more carefully.

!!! tip "Recommended for complex tasks" Enable this when working on complex scene modifications or when you want more thoughtful responses. Disable it for simple queries to get faster responses.

Response token budget

Controls the maximum number of tokens the director can use for generating responses. Higher values allow for more detailed responses but use more tokens. Default is 2048.

Auto-iteration limit

The maximum number of action-response cycles the director can perform in a single interaction. For example, if set to 10, the director can execute actions and generate follow-up responses up to 10 times before requiring your input again. Default is 10.

Retries

The number of times the director will retry if it encounters an error during response generation. Default is 1.

Scene context ratio

Controls the fraction of the remaining token budget (after fixed context and instructions) that is reserved for scene context. The rest is allocated to chat history.

  • Lower values (e.g., 0.30): 30% for scene context, 70% for chat history
  • Higher values (e.g., 0.70): 70% for scene context, 30% for chat history

Default is 0.30.

Stale history share

When the chat history needs to be compacted (summarized), this controls what fraction of the chat history budget is treated as "stale" and should be summarized. The remaining portion is kept verbatim as recent messages.

  • Lower values (e.g., 0.50): Summarize less (50%), keep more recent messages verbatim
  • Higher values (e.g., 0.90): Summarize more (90%), keep fewer recent messages verbatim

Default is 0.70 (70% will be summarized when compaction is triggered).

Custom instructions

Add custom instructions that will be included in all director chat prompts. Use this to customize the director's behavior for your specific scene or storytelling style.

For example, you might add instructions to maintain a particular tone, follow specific genre conventions, or handle certain types of requests in a particular way.