Conversation Config controls global chat behavior, including title and label generation, the global default system prompt, context compression, user option pass-through, and built-in Skills/Prompts.
Entry Point#
Open Conversation in the admin console. The page is organized into Conversation settings, Context compression, Option pass-through, and the built-in prompt library.
These settings affect later conversations for all users. Existing generated titles, labels, and messages are not rewritten automatically.
For how pass-through works with capabilities, dangerous-field blocking, and native tools, see Capabilities & Native Tools. For how context compression works with RAG and file evidence, see Files and Retrieval Advanced Guide.
Task Model#
The task model is used for conversation title and label generation. It can follow the current model or use a stable, inexpensive, fast chat model.
If users often choose expensive models, do not let title and label tasks follow the current model. A dedicated internal task model lowers cost and keeps title style consistent.
Global Default System Prompt#
The global default system prompt is injected as the platform-wide behavior instruction for all chat models. Empty means no global prompt. Model-level prompts are appended after it.
Use it for general platform requirements such as accuracy, clarity, organizational safety, and sensitive-information handling. Do not put one project, one user, or one temporary task requirement into the global prompt.
Title and Label Prompts#
The title prompt controls automatic title generation. The label prompt controls automatic label generation. Both can insert conversation message content and should return stable JSON structures.
Keep output structure stable when editing these prompts, or titles and labels may fail to parse. Validate in a small scope before production use.
Context Compression#
Context compression helps long conversations and large evidence sets stay within model context limits. Administrators can allow compression globally, while users can still choose whether to use it in their chat settings.
| Setting | Purpose |
|---|---|
| Token budget truncation | Trims sendable history based on the current model context window. |
| Turn and token thresholds | Trigger compression by conversation length or token size. |
| Recent turns preserved | Keeps recent raw turns and summarizes older content. |
| Highlights and snippet limits | Control template-summary density and individual snippet length. |
| Evidence retention days | Controls retention for RAG results, historical evidence, and tool evidence. |
| Background compression | Runs compression after the current response to reduce blocking. |
| LLM compression / compression model | Uses a model for semantic summaries; choose a stable, cost-controlled chat model. |
| Failure threshold | Falls back to template summaries after repeated LLM compression failures. |
| Summary prompts | Customize full and light summary instructions to preserve facts, numbers, constraints, and open tasks. |
Disabling compression can reduce background cost, but long conversations are more likely to lose early context or hit model context limits.
Option Pass-Through Policy#
Option pass-through controls whether user-defined model options in conversation settings can reach upstream calls. It balances flexibility and governance.
| Policy | Behavior | Best For |
|---|---|---|
| Allowlist | Only listed paths pass; everything else is dropped. | Recommended default for production. |
| Denylist | Most fields pass except listed or system-blocked paths. | Trusted users or internal debugging. |
| Disabled | User-defined options do not pass through. | Strict control, cost-sensitive, or compliance-heavy environments. |
System-dangerous fields are always blocked and cannot be allowed. This prevents users from bypassing platform model, tool, safety, or route policies.
Path Rules and Preview#
Allowlist and denylist use path notation. Top-level fields use field names; nested fields use dots. Rules can be default for all protocols or added for specific protocols.
The pass-through section provides a live preview showing which fields pass, which are filtered, and why. The guide explains path notation, strategy differences, protocol keys, and system-blocked fields.
Before saving, use the preview to confirm the outcome, especially when switching denylist or disabled modes.
Built-In Skills and Prompts#
The prompt library at the bottom of Conversation Config maintains platform built-in Skills and Prompts. Built-in items are visible to users, but users can only view and use them, not edit or delete them.
| Type | Admin manages | User-side effect |
|---|---|---|
| Prompts | Name, description, content, enabled state | Users select them from / to insert text into the composer. |
| Skills | Name, description, SKILL.md, enabled state | Users select them from / to load context for the current turn; no commands are run. |
Deleting a built-in item does not affect user-created Skills or Prompts. Disable items for temporary removal; delete only when the template is no longer maintained.
Practical Tips#
Use allowlist policy in production and expose only necessary parameters. Keep the global system prompt short and stable; put business-specific requirements into project prompts, model prompts, or built-in Skills. After saving conversation settings, test once from a normal user perspective.
Setting Key Reference#
Conversation Config maps to platform-level chat.* settings and affects later conversations for all users.
| Group | Setting Keys |
|---|---|
| Task model and prompts | chat.conversation_task_model, chat.conversation_title_prompt, chat.conversation_labels_prompt, chat.default_system_prompt |
| Context compaction switches | chat.context_compact_enabled, chat.context_token_budget_enabled |
| Compaction trigger and retention | chat.context_max_turns, chat.context_compact_trigger_tokens, chat.context_compact_preserve_recent_turns, chat.context_compact_highlights_per_role, chat.context_compact_snippet_chars, chat.context_artifact_retention_days |
| Compaction execution | chat.compact_async_enabled, chat.compact_llm_enabled, chat.compact_task_model, chat.compact_max_failures, chat.compact_system_prompt, chat.compact_light_prompt |
| Option pass-through | chat.model_option_policy_mode, chat.model_option_allowed_paths, chat.model_option_denied_paths |
The latest frontend recognizes these pass-through protocol keys: default, openai_chat_completions, openai_responses, openrouter_responses, openai_image_generations, openai_image_edits, anthropic_messages, gemini_generate_content, google_image_generation, xai_responses, xai_image, and xai_image_edits.