Skip to main content
Chronologue’s function calling interface allows agents, language models, and interactive tools to invoke structured actions through a well-defined schema. This layer bridges natural language intent with precise, executable instructions—powering planning, memory logging, calendar scheduling, and feedback loops. Function calls can be invoked via:
  • OpenAI’s functions or tool_choice API
  • Internal agent runtime (/functions/call)
  • Manual interface triggers (e.g., UI buttons or editable tables)

​
2. Use Cases

  • Translate user prompts into agent_plan actions
  • Reflect on memory with structured create_reflection calls
  • Log new observations or task completions
  • Schedule calendar events based on structured parameters
  • Edit traces or plans via interactive table or UI interface

​
3. Function Calling Schema

Each function includes:
FieldTypeDescription
namestringUnique identifier for the function
descriptionstringNatural language summary of the function
parametersobjectJSON Schema defining input parameters
Example: 
{
  "name": "schedule_task",
  "description": "Create an agent plan to schedule a task with a specific time and duration.",
  "parameters": {
    "type": "object",
    "properties": {
      "content": { "type": "string" },
      "scheduled_for": { "type": "string", "format": "date-time" },
      "duration_minutes": { "type": "integer" }
    },
    "required": ["content", "scheduled_for"]
  }
}

​
4. Defined Functions in Chronologue

Function NamePurpose
schedule_taskCreate a new agent plan or calendar event
create_reflectionLog a reflection trace
log_observationRecord real-time outcome or observation
query_memoryFetch traces within a given filter context
submit_feedbackRate or comment on an existing memory trace
propose_revisionSuggest edits to an agent plan or schedule

​
5. API Integration with FastAPI

Function calls are processed via: POST /functions/call Request body: 
{
  "name": "schedule_task",
  "arguments": {
    "content": "Write CUDA blog post",
    "scheduled_for": "2025-05-13T16:00:00Z",
    "duration_minutes": 90
  }
}
The API dispatches the call, validates parameters with Pydantic, and returns a result:
  • A new memory trace
  • A calendar_event block
  • A confirmation or summary message

​
6. Function Calling with LLMs

Chronologue registers these functions with tool-enabled LLM runtimes (e.g. OpenAI, Claude). Prompt:
“Schedule a writing session for tomorrow afternoon.”
Resulting call: 
{
  "function_call": {
    "name": "schedule_task",
    "arguments": {
      "content": "Writing session",
      "scheduled_for": "2025-05-13T14:00:00Z",
      "duration_minutes": 120
    }
  }
}
The returned trace is optionally previewed before saving.

​
7. Design Principles

  • Schema-driven: Validated with JSON Schema and Pydantic
  • Composable: Output can be stored, revised, or chained
  • Reversible: Every function call result is a trace
  • Minimal: Only expose essential agent operations

​
8. Error Handling and Feedback

When a function call fails:
  • Returns a 400 or 422 response with detail
  • trace_id is included for log correlation
  • Suggestions (hint) may be returned for retry
Example error response: 
{
  "error": "Missing required field 'scheduled_for'",
  "status": 422,
  "hint": "Include a valid ISO timestamp for the schedule."
}

​
9. Manual Function Calls from UI (Editable Calendar/Table)

Chronologue supports invoking functions directly from:
  • Calendar editors (e.g., dragging a block)
  • Markdown tables in chat interface (editable fields)

​
Example Table (Editable in UI):

ActionTimeDurationFeedbackEdit
Reflect2025-05-10 22:0015 min4/5[Edit]
Call Mom2025-05-11 10:0030 min-[Reschedule]
Plan Week2025-05-12 08:0045 min5/5[Adjust + Repeat]
On interaction, each Edit button triggers a corresponding function call:
  • propose_revision
  • reschedule_task
  • update_trace_field
This interface allows users to manually invoke functions, not just through LLMs.
Chronologue’s function calling interface enables structured, transparent, and user-aligned interaction with agents. Whether invoked by an LLM, clicked in a UI, or scheduled via planner, each function brings natural language closer to structured, traceable execution.