02.7b: Verbatim Tool Definitions

Miễn trừ trách nhiệm

Trang này sử dụng bản nháp tạm thời do AI tạo sinh viết và duy trì. Dù đã được xác minh tính chính xác, bạn vui lòng đối chiếu lại với mã nguồn.

The verbatim tool definitions contributor emits an in-band JSON dump of active tool schemas for models operating under verbatim tool-calling mode.

Flow and ownership

Certain local or OpenAI-compatible custom endpoints accept the API tools parameter without exposing function schemas to the underlying model. To support tool use on these endpoints, TomoriBot uses verbatim tool-calling:

1. Gating predicate: Evaluates shouldInjectVerbatimToolCallingNudge(tomoriState) in src/utils/tools/verbatimToolCalling.ts. The contributor activates only when llm.verbatim_tool_calling === true, the provider is custom, and llm.has_tools === true.
2. Schema assembly: buildVerbatimToolDefinitionsJson assembles ToolStateForContext, queries getAvailableToolsWithMCP for built-in and MCP tools matching feature flags and Deliberate Tool Mode scopes, and formats them into OpenAI-compatible function definitions using OpenAICompatibleToolAdapter.
3. Context emission: buildVerbatimToolDefinitionsContextItem in src/utils/text/context/toolDefinitions.ts formats the schemas into a fenced JSON block preceded by calling instructions, and emits a user-role item tagged KNOWLEDGE_VERBATIM_TOOL_DEFINITIONS.

Coordination with dialogue nudge

Verbatim tool calling is split into two coordinated halves:

• Schema definitions (stage 07b): Serializes tool signatures in the prompt’s reference zone (before RAG documents), placing static schema definitions in the prompt-cache-friendly prefix.
• Behavioral nudge (stage 11): Injects VERBATIM_TOOL_CALLING_NUDGE at depth 3 in Dialogue History, instructing the model how to format tool invocations in its output text.

Both halves share the identical gating predicate, so they always activate and deactivate together.

When generation fallbacks switch between native and verbatim models across retry attempts, prepareProviderContextItems in src/utils/chat/generationTurn.ts dynamically adds or removes this block to match the candidate model.

Constraints and rationale

• Prompt cache stability: Placing tool schemas in the prefix rather than at the dialogue tail prevents tool definitions from invalidating prompt cache prefixes across successive turns.
• Fail-safe fallback: If schema resolution fails, the error is logged and returns null, allowing context assembly to proceed without crashing the turn.

Source pointers

• src/utils/text/context/toolDefinitions.ts: buildVerbatimToolDefinitionsContextItem.
• src/utils/tools/verbatimToolCalling.ts: shouldInjectVerbatimToolCallingNudge and VERBATIM_TOOL_CALLING_NUDGE.
• src/providers/openaiCompatible/openaiCompatibleToolAdapter.ts: OpenAICompatibleToolAdapter.
• 11: Dialogue History: behavioral verbatim nudge injection.
• Chunk Normalization: stream-time text parsing of verbatim tool calls.